Development and Applications of Localised Numerical Weather Prediction Models in Building Energy Management

Development and applications of localised Numerical Weather Prediction models in building energy management

Dimitris Lazos

A thesis in fulfilment of the requirements for the degree of Doctor of Philosophy

School of Photovoltaics and Renewable Energy Engineering

Faculty of Engineering

August 2016

‘I hereby grant the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole or part in the University libraries in all forms of media, now or here after known, subject to the provisions of the Copyright Act 1968. I retain all proprietary rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertation. I also authorise University Microfilms to use the 350 word abstract of my thesis in Dissertation Abstract International (this is applicable to doctoral theses only). I have either used no substantial portions of copyright material in my thesis or I have obtained permission to use copyright material; where permission has not been granted I have applied/will apply for a partial restriction of the digital copy of my thesis or dissertation.'

Signed ……………………………………………......

Date ……………………………………………......

AUTHENTICITY STATEMENT

‘I certify that the Library deposit digital copy is a direct equivalent of the final officially approved version of my thesis. No emendation of content has occurred and if there are any minor variations in formatting, they are the result of the conversion to digital format.’

Signed ……………………………………………......

Date ……………………………………………...... Acknowledgments This thesis is dedicated to Nick “The Machine”. You may not with us now, but you are not forgotten. Victory or Death, my friend.

During the writing of the thesis, Mei’s help was much appreciated and I would like to thank her for being earnest and eager to assist. I would also like to thank Yelena, Maria and Stavros for the inspiration.

Last, but not least I would like to express my gratitude to Merlinde and Alistair for their valuable guidance and being extremely understanding and flexible throughout the project.

Contents Acknowledgments ______i

List of Figures ______vi

List of Tables ______x

List of Abbreviations ______xii

1 Introduction ______1

1.1 Research background ______1

1.2 Thesis outline ______3

2 Literature review ______5

2.1 Chapter outline ______5

2.2 Forecasting ______6 2.2.1 Background______6 2.2.2 Forecasting techniques ______7 2.2.2.1 Time-series and regression model forecasting ______7 2.2.2.2 Machine learning forecasting ______8 2.2.2.3 Physical model forecasting ______9 2.2.2.4 Technique comparison ______11 2.2.3 Weather variable forecasting ______13 2.2.3.1 Time series and regression weather forecasting ______14 2.2.3.2 Machine learning weather forecasting ______15 2.2.3.3 Numerical weather predictions (NWP) ______16 2.2.3.4 Weather forecasting design considerations ______24 2.2.4 Load forecasting ______27 2.2.4.1 Commercial building loads______27 2.2.4.2 Time series and regression load forecasting ______30 2.2.4.3 Machine learning load forecasting ______32 2.2.4.4 Physical model load forecasting ______34 2.2.4.5 Hybrid models ______36 2.2.4.6 Peak load predictions ______37 2.2.4.7 Forecasting with Degree Hours & Days ______40 2.2.4.8 Load forecasting summary ______43 2.2.5 Generation forecasting ______44

2.3 Building energy management systems ______46 2.3.1 Model Predictive Control systems ______48 ii

2.3.2 Weather forecasting in BEM ______50 2.3.3 Building conditioning ______53 2.3.3.1 Dynamic conditioning ______53 2.3.3.2 Preconditioning ______55 2.3.3.3 Effects and modelling of thermal mass ______59 2.3.3.4 Forecasting and effects of occupancy ______61 2.3.4 Energy generation and storage management ______63

2.4 Summary and research opportunities ______65 2.4.1 Inclusion of weather predictions in building energy systems ______65 2.4.2 Research gaps ______67

3 Short term numerical weather forecasting ______70

3.1 Model outline ______70

3.2 Weather prediction model design ______72 3.2.1 Data acquisition ______72 3.2.2 Prediction model architecture ______73 3.2.3 Base prediction models______76 3.2.3.1 Persistence prediction model ______76 3.2.3.2 Numerical predictions in TAPM ______77 3.2.4 Hybrid prediction models ______79 3.2.4.1 Linear regression weighted forecasting (WF) model ______79 3.2.4.2 Historical data weighted forecasting (WFS) model ______81 3.2.4.3 ARX prediction model ______85 3.2.5 Forecast sensitivity to output update intervals______86 3.2.6 Correction algorithm for extreme heat events ______87

3.3 Results ______89 3.3.1 Temperature predictions ______89 3.3.2 Relative humidity predictions ______93 3.3.3 Wind speed predictions ______95 3.3.4 Abrupt change predictions ______96 3.3.5 Extreme heat event predictions ______99 3.3.6 Peak load predictions ______100 3.3.7 Value of localisation ______101

3.4 Discussion of the model applicability ______104

4 Peak load forecasting with an ensemble of weather forecasts ______110

4.1 Ensemble forecasting model design ______110 4.1.1 Model outline and rationale ______110 iii

4.1.2 Ensemble branch parameters ______111 4.1.3 Model outputs and validation ______114

4.2 Peak load prediction model ______116 4.2.1 Potential peak periods ______116 4.2.2 Detection of significant peak loads ______118

4.3 Case study & results ______124 4.3.1 Potential peak period detection ______124 4.3.2 Significant peak detection and relative magnitude predictions ______126

4.4 Discussion of the peak load prediction model ______132

4.5 Conclusions ______135

5 Weather analysis tool for evaluation of the precooling potential ______137

5.1 Tool design outline ______137

5.2 Model design ______139 5.2.1 Simulations outline ______139 5.2.2 Site selection ______139 5.2.3 Precooling conditions ______141 5.2.4 Precooling ratios ______142 5.2.4.1 Precooling potential ______144 5.2.4.2 Precooling utilisation ______145 5.2.4.3 Precooling frequency and annualised ratios ______146 5.2.4.4 Theoretical precooling value ______146

5.3 Simulation results and discussion ______148 5.3.1 Results summary______148 5.3.2 Precooling frequency ______150 5.3.3 Precooling potential ______153 5.3.4 Precooling utilisation ______155 5.3.5 Precooling value ______157 5.3.6 Precooling dependence on local climate factors ______159 5.3.7 Simulations for different climates ______160

5.4 Conclusions ______166

6 Development of a building load predictive control algorithm based on enthalpy forecasts ______168

6.1 Model outline ______168

6.2 Predictive control algorithms ______169 iv

6.2.1 Weather forecasting inputs and horizons ______169 6.2.2 Enthalpy of air ______171 6.2.3 Correlation of enthalpy differences of ambient and interior air with cooling load _____ 172 6.2.4 Additional inputs______175

6.3 Weather dependent response control ______176 6.3.1 Dynamic internal temperature control algorithm ______176 6.3.2 Peak load control algorithm ______181 6.3.3 Preconditioning control (PC) algorithm______183

6.4 Case study ______184 6.4.1 Building characteristics ______184 6.4.2 Application of the NARX in the case study building ______186 6.4.3 Estimation of savings with predictive control algorithms ______188 6.4.4 Precooling control algorithm effects ______194 6.4.5 Effects of thermal comfort zone boundaries ______198

6.5 Conclusions ______199

7 Conclusion ______201

Appendix______205

Bibliography ______211

List of Figures

Figure 1: Outline of the research streams in the thesis ...... 3 Figure 2: Classification of physical models according to the information known about the internal system relationships and data ...... 10 Figure 3: Comparison of the forecasting skill of a statistical AR prediction model (top), versus a NWP model (bottom) in 5-day horizon temperature forecasts (Zavala et al., 2010) ...... 17 Figure 4: TAPM graphical user interface. The grid parameterisation is visible in regards to spacing, resolution, location and temporal options ...... 19 Figure 5: Parameterisation menu for the lower order domains (up to 4) ...... 20 Figure 6: A sample grid for the greater Sydney area ...... 21 Figure 7: Summary of key research considerations in the field of weather forecasting for building energy management ...... 26 Figure 8: Typical commercial building load profile on two consecutive weekdays (Mathieu et al., 2011) ...... 28 Figure 9: Daily load profiles for a university building in Spain in 2009 (Penya et al., 2011) ...... 28 Figure 10: Steam load prediction in a building using an ANN architecture with weather inputs for both training and forecasting (Kusiak et al., 2010)...... 33 Figure 11: A 3R2C building thermal network model used for analysing the thermal response and load forecasting for a case study building in the US (Lee and Braun, 2007) ...... 35 Figure 12: Comparison of an AR, machine learning and hybrid model in terms of errors in load forecasting for a commercial building in China (Xuemei et al., 2010a) ...... 37 Figure 13: Correlation of daily peak load and average temperature for gas and electrical heating in various commercial buildings in Sydney (Steinfeld et al., 2011) ...... 39 Figure 15: Summary of features and applications of generation forecasting models in existing literature, with a clear distinction between renewable energy (RE) and non-renewable energy sources ...... 46 Figure 16: Comparison of features of commercial and residential energy management ...... 47 Figure 17: Example of a typical MPC system with weather forecasting inputs (Oldewurtel et al., 2010) ...... 50 Figure 18: Correlation of discomfort cost (Jd) and energy cost (Je) for a range of controllers (Kummert et al., 2000)...... 51 Figure 19: Comparison of 5-day ahead operating HVAC strategies, with predicted interior temperatures (grey), actual observed temperatures (blue) and the comfort zone (between the black lines) (Zavala et al., 2010) ...... 52 Figure 20: Comparison of the effects of a non-predictive approach with preconditioning strategies on interior temperature and daily cooling loads (Lee and Braun, 2007) ...... 57 Figure 21: Effects of thermal mass on internal temperature (Lockerbie, 2016) ...... 59

Figure 22: Architecture of a BEMS that integrates weather, load and generation forecasts for developing appropriate responses. This thesis will investigate several pathways of this modular approach...... 68 Figure 23: Architecture of the proposed model, able to generate wind speed, temperature and relative humidity prediction outputs ...... 74 Figure 24: Percentage hourly variation of temperature and humidity ...... 75 Figure 25: Comparison of weighting contribution of each base model to the WF linear prediction model ...... 81

Figure 26: Weights (wx) of the persistence model for temperature for each hybrid model at each location and time series point. The NWP weighting is equal to 1-wx ...... 82 Figure 27: Weights (wx) of the persistence model for relative humidity for each hybrid model at each location and time series point. The NWP weighting is equal to 1-wx ...... 83

Figure 28: Weights (wx) of the persistence model for wind speed for each hybrid model at each location and time series point. The NWP weighting is equal to 1-wx ...... 83 Figure 29: Correlation of temperature and humidity for a 9 year period for the airport site ...... 88 Figure 30: Flowchart of adjusting TAPM predictions of high temperatures according to predicted humidity with a correction algorithm ...... 89 Figure 31: Comparison of effects of update frequency of the persistence model on the accuracy of temperature prediction ...... 91 Figure 32: Comparison of hourly performance of the four prediction models (airport site) ...... 92 Figure 33: Comparison of monthly performance of the four prediction models (airport site) ...... 93 Figure 34: Comparison of effects of update frequency of the persistence model on the accuracy of relative humidity prediction ...... 95 Figure 35: Performance of WFS model with 6 hourly and 3 hourly updates, during a typical hot day with a sudden change in temperature ...... 97 Figure 36: Performance of ARX model with 6 hourly and 3 hourly updates during a typical hot day with a sudden change in temperature ...... 97 Figure 37: Algorithm for the detection of potential peak periods in the summer ...... 117 Figure 38: Significant peak detection algorithm...... 121 Figure 39: Prediction of potential peak periods according to temperature and relative humidity over 3 day period ...... 125 Figure 40: Building load over three day sample period, including the actual peak load and the potential peak load periods as predicted by TAPM ...... 125 Figure 41: Distribution of branch predictions for temperature over a three day sample period, including the potential peak period for each day ...... 126 Figure 42: Distribution of daily peak loads in relation to mean peak period temperature (recorded onsite) ...... 127 Figure 43: Distribution of daily peak loads in relation to mean peak period rel. humidity (recorded onsite) ...... 128

vii

Figure 44: Distribution of daily peak loads in relation to daily CDH (recorded onsite) ...... 128 Figure 45: Comparison of predicted and actual daily peak loads for the 2012-13 summer period in the TETB ...... 130 Figure 46: Comparison of predicted and actual daily peak loads for the 2013-14 summer period in TETB ...... 131 Figure 47: Map of the five Sydney sites ...... 140 Figure 48: Daily temperature profile (as simulated) for Penrith between 27th January 20:00 and 28th January 20:00 in 2014. The upper and lower boundaries can be seen at 25 and 20 degrees respectively. The DH for each time step are calculated as the difference from the base temperature at each period...... 143 Figure 49: Number of days that precooling conditions were met per year for each site ...... 151 Figure 50: Average number of days that precooling conditions are met across the five Sydney sites ..152 Figure 51: Summary of precooling potential across sites in Sydney ...... 154 Figure 52: Annualised potential ratio R1 for the Sydney sites for each simulation year ...... 155 Figure 53: Precooling utilisation summary for sites across Sydney...... 156 Figure 54: Annualised utilisation ratio (R2) for Sydney sites for each simulation year ...... 156 Figure 55: Comparison of precooling days per year and the magnitude of the mean annual diurnal temperature difference ...... 160 Figure 56: Timeline of operation of the modules in daily horizon control. The timeline shows when does each forecasting component is scheduled to run the simulations and develop predictions throughout a day ...... 170 Figure 57: Time series nonlinear input-output neural network architecture for the correlation of enthalpy change and cooling load ...... 174 Figure 58: Dynamic internal temperature control algorithm for cooling based on the outputs of the ST model ...... 180 Figure 59: Peak load control algorithm with gradual IT rises during the potential peak period ...... 183 Figure 60: Precooling control algorithm at every time step before the occupied period ...... 184 Figure 61: Western facade of the L5 building (Gollings, 2006) ...... 185 Figure 62: Summary of the time series neural network performance for the training, validation and testing data ...... 186 Figure 63: NARX performance in correlating cooling load to enthalpy changes ...... 187 Figure 64: Prediction error distribution. The majority (90%) of the errors are within 44 kW of the actual observed load ...... 188 Figure 65: Simulation result distribution for each NN iteration - predictions of average annual cooling load and comparison with standard control ...... 190 Figure 66: Simulation result distribution for each NN iteration - predictions of maximum peak load and comparison with standard control ...... 191 Figure 67: Comparison of total building and cooling loads (standard control) with ambient temperature in a period of 2 days (7-8th January 2013) ...... 192

viii

Figure 68: Evolution of internal temperature with standard control and predictive control (simulated according to algorithms in section 6.3) for the period of 7-8th January 2013. The dashed lines represent the thermal comfort boundaries (22-25°C) ...... 193 Figure 69: Cooling load comparison for a two day period (7th and 8th January 2012) between standard control and predictive control. The daily peak loads are indicated for both models...... 194 Figure 70: Comparison of internal temperature control with the standard or predictive setup for a sample day (1st November 2012) - precooling control is active in order to shift some of the load overnight in expectation for the following hot day. The dashed lines represent the thermal comfort zone boundaries (22-25°C) ...... 196 Figure 71: Load profile with standard and predictive control for 1st November 2012. The effects of PC are visible after the reference hour (01:00)...... 197 Figure 72: Comparison of predictions for temperature using the overall ensemble mean vs the ensemble mean from branches within 1 standard deviation for a sample 11 day period ...... 205 Figure 73: Comparison of predictions for relative humidity using the overall ensemble mean vs the ensemble mean from branches within 1 standard deviation for a sample 11 day period ...... 206 Figure 74: Time series ensemble forecasting for temperature for a sample 11 period, showing the actual temperature (BOM in red), the ensemble z1 predicted temperature (in green) and the predictions of individual branches ...... 207 Figure 75: Time series ensemble forecasting for temperature for a sample 11 period, showing the actual rel. humidity (BOM in red), the ensemble z1 predicted rel. humidity (in green) and the predictions of individual branches ...... 208 Figure 76: Quartiles of ensemble branch predictions comparison to observed temperature (in purple) for a sample 11 day period ...... 209 Figure 77: Quartiles of ensemble branch predictions comparison to observed rel. humidity (in purple) for a sample 11 day period ...... 210

List of Tables

Table 1: Comparison of common forecasting techniques ...... 11 Table 2: Terminology used in reference to TAPM simulations ...... 18 Table 3: TAPM accuracy comparison of different domain orders ...... 78 Table 4: Comparison of each domain’s ability to predict bigger changes in temperature and relative humidity ...... 79

Table 5: Values of wx weightings of the persistence model for temperature prediction post-processing for each hour after the reference hour (x) and each location ...... 84

Table 6: Values of wx weightings of the persistence model for relative humidity prediction post- processing for each hour after the reference hour (x) and each location ...... 84

Table 7: Values of wx weightings of the persistence model for wind speed prediction post-processing for each hour after the reference hour (x) and each location ...... 84

Table 8: ARX parameters for each site, calculated from the historical data analysis - ax represents the coefficient of the observation x time steps before the prediction, and b1 represents the external input (TAPM) ...... 86 Table 9: MAE of each prediction model for temperature ...... 90 Table 10: RMSE of each prediction model for temperature ...... 90 Table 11: Comparison of persistence update frequency for temperature ...... 90 Table 12: MAE of each prediction model for relative humidity ...... 93 Table 13: RMSE of each prediction model for relative humidity ...... 94 Table 14: Comparison of persistence update frequency for relative humidity ...... 94 Table 15: MAE of each prediction model for wind speed ...... 95 Table 16: RMSE of each prediction model for wind speed ...... 96 Table 17: Comparison of performance of prediction models in forecasting sudden hourly changes ...... 98 Table 18: Comparison of accuracy of predictions of each model for extreme heat events (temperatures over 32°C) ...... 99 Table 19: Comparison of accuracy of predictions of each model for extreme heat events (temperatures over 32°C) with and without the correction algorithm ...... 100 Table 20: Comparison of absolute peak temporal difference for all models ...... 101 Table 21: Summary of locations chosen to test the value of localisation ...... 102 Table 22: Comparison of MAE for temperature and humidity predictions between BOM and WFS for the three sites for simulation year 2015 ...... 103 Table 23: Summary of the two tailed t-test for Bexley (temperature only) ...... 103 Table 24: Summary of the two tailed t-test for Bondi Beach ...... 103 Table 25: Summary of the two tailed t-test for Macquarie University ...... 104 Table 26: Comparison of the skill of each model according to the results obtained from the simulations ...... 107 Table 27: Individual ensemble branch characteristics ...... 112 Table 28: Comparison of computational time for each ensemble branch ...... 113 x

Table 29: Comparison of performance from WF model and ensemble mean from 12 branches for predictions of temperature and relative humidity ...... 114 Table 30: Comparison of prediction performance using ensemble mean or the mean of the branches within one standard deviation ...... 115 Table 31: Comparison of forecasting skill depending on weather input source ...... 132 Table 32: Characteristics of the five sites ...... 140 Table 33: Calculation of the DH values for each time step for the sample day. The summation of the values in the end represents the values of DHL, DHN, and DHU used in the precooling ratios...... 144 Table 34: Summary of results across five sites in Sydney for each simulation year ...... 148 Table 35: Summary of precooling potential and utilisation across the five sites in Sydney...... 149 Sydney...... 153 Table 36: Comparison of distribution of precooling days ...... 153 Table 37: Summary of precooling value ratios for the five sites in Sydney ...... 157 Table 38: Sensitivity of the value ratio (v) to the selection of boundary temperatures (Site 2 – Airport) ...... 158 Table 39: Sensitivity of the value ratio (v) to the selection of boundary temperatures (Site 5 – Penrith) ...... 158 Table 40: Comparison of mean monthly diurnal temperature differences for each simulation year and 20 year average for Penrith ...... 159 Table 41: Comparison of precooling ratios from the simulations in different climates in Australia ....162 Table 42: Comparison of precooling ratios from the simulations in different climates globally ...... 163 Table 43: Comparison of precooling ratios from the simulations in different climates globally ...... 164 Table 44: Dynamic temperature control responses for each decision branch ...... 179 Table 45: Summary of cooling load and peak reduction with predictive weather control ...... 189 Table 46: Comparison of demand and peak loads for the sample two day period for each control method ...... 193 Table 47: Comparison of daily load distribution for days that the precooling conditions are met over the simulation period in the simulation time series (2012-2014) ...... 198 Table 48: Comparison of savings for different control strategies according to the thermal comfort range boundaries ...... 199

List of Abbreviations ANN: Artificial Neural Network AR: Autoregressive ARX: Autoregressive with External Input ASHRAE: American Society of Heating, Refrigeration and Air-Conditioning Engineers BEMS: Building Energy Management System BOM: Bureau of Meteorology BP: Building peak load CDH: Cooling Degree Hours CSIRO: Commonwealth Scientific and Research Organisation DD: Degree Days DG: Distributed Generation DH: Degree Hours DL: Building dynamic load EW: Exponential Weighting HDH: Heating Degree Hours HVAC: Heating, Ventilation and Air-conditioning IT: Interior Temperature LT: Lower temperature boundary of the thermal comfort zone MA: Moving Average MAE: Mean Absolute Error MPC: Model Predictive Control MT: Mean temperature of the thermal comfort zone NCEP: National Centers for Environmental Prediction NWP: Numerical Weather Prediction PC: Preconditioning Control PV: Photovoltaic RE: Renewable Energy RH: Relative Humidity RMSE: Root Mean Square Error ST: Short Term Numerical Weather Forecasting

xii

STLF: Short Term Load Forecasting SVM: Support Vector Machine TAPM: The Air Pollution Model TETB: Tyree Energy Technologies Building TMY: Typical Meteorological Year UNSW: University of New South Wales UT: Upper temperature boundary of the thermal comfort zone WF: Weighted numerical Forecast - linear WFS: Weighted numerical Forecast - optimised

xiii

1 Introduction 1.1 Research background In several countries around the world, including Australia, there is increased interest in the energy performance of commercial buildings and their responsiveness to changes in external and internal conditions. The integration of sustainability concepts such as energy efficiency, onsite distributed generation and passive energy designs is becoming more common in both research, construction and building operations, as they enable significant savings to the facility bills of the building. In addition to the financial savings that may be realised with such measures, the potential demand and carbon emission reductions are significant drivers towards the optimisation of the energy performance of commercial buildings. In the 21st century world, with increasing awareness of the negative impacts of fossil fuel generation and wasteful energy consumption, there are plenty of regulations aiming to reduce the footprint of commercial buildings.

Historically and as activities in commercial building become more energy intensive, the complexity of energy management and control systems increases as well. Furthermore, advancements such as electricity, the telegraph, the telephone and the internet necessitated the development of centralised control nodes and the management of interactions between buildings with grids. Initially, the grids were localised and connected buildings within towns or neighbourhoods, but by the middle of the 20th century grids expanded dramatically to cover regions at a city, state and national levels (Burn, 2012). Recently, telecommunications and often electricity grids span across borders as well.

The main characteristic of grid-connected commercial buildings in the previous century was that control of energy was rather inflexible. Before the advent of computers and telecommunications, there were limited options for managing the energy use, and hence the energy was managed in an inefficient way. Modern building systems however, have the capacity to actively manage energy and minimise waste according to their tailored needs (Lazos et al., 2014). While such processes increase efficiency, they do require a variety of input data to function properly.

The primary motivation for this thesis is aligned with the trend of producing tailored inputs for dynamic energy management systems in modern day commercial buildings. Specifically, the majority of the work in this project is related to developing and utilising tailored weather predictions in a range of building energy management applications. Weather conditions have a range of direct and indirect effects on building energy demand, generation and management. Among others, the weather conditions:

1. Directly affect the HVAC system loads 2. Directly affect the generation from solar modules or wind converters installed in the building 3. May directly affect the natural lighting availability 4. Indirectly affect the operation and efficiency of certain devices

However, the integration of weather inputs in building energy predictors and demand response decision making is limited. Inclusion of weather inputs in the design and modelling stage of new building construction is in general higher, although there is room for improved integration in this field as well. One of the main reasons that weather inputs are not broadly utilised is the difficulty in obtaining accurate information for the present and future states of weather at any location and also the complexity in integrating dynamic and historical weather trends in control strategies.

As such, the primary research question that this thesis attempts to address is whether numerical weather predictions can be integrated in building energy management system applications. A number of solutions based on a numerical weather prediction model and statistical processing that are easy to implement at any location will be proposed and their applicability will be analysed.

Towards that goal, there are two main streams of research. Firstly, the development of short-term numerical predictions of weather variables (few hours to days ahead) allows for appropriate demand response measures to be realised. The predictions are statistically post-processed in a fashion that is meaningful for building energy management control. This constitutes the primary research stream in this thesis and the proposed outputs are expected to be useful in dynamic building energy

2 management and control systems. The second stream involves long-term analysis of the local climate effects, the outputs of which are expected to be useful for both energy management decisions, as well as new building design. Figure 1 shows a schematic of the research streams of this work along with the chapters that refer to the respective topics.

Short term Demand responses predictions (chapters 3,4) (chapter 6) Numerical weather predictions Long term climate Building design analysis(chapter 5)

Figure 1: Outline of the research streams in the thesis 1.2 Thesis outline Apart from the introduction, there are 6 additional chapters in this thesis.

Chapter 2 (literature review) offers detailed background information related to the research fields addressed in the thesis: forecasting, building energy systems and predictive control. The review of existing literature will be selectively focused on information that help understand the usefulness and challenges of integrating weather prediction applications in building energy systems. By the end of the chapter it will be possible to highlight and justify a number of research gaps associated with the main objectives of this thesis.

Chapter 3 discusses the methodology of developing numerical weather forecasts and a range of statistical post-processing techniques. The predictions are designed in such ways that maintain a balance between accuracy, computational time and ease of implementation. An extensive assessment of the predictions’ accuracy shows the improved value of the numerical predictions over existing reference models and predictions obtained from third parties.

Chapter 4 narrows down the predictions in order to detect peak loads. A novel ensemble approach is proposed and two robust, yet easy to implement peak 3 detection algorithms are described. The peak detection, in line with the concept of ensemble forecasting, is probabilistic and may be integrated in building energy management systems in order to develop appropriate weather dependent responses. There is a case study for a university campus building, which validates the applicability of these predictions in peak load detection.

Chapter 5 belongs to the long-term analysis of weather trends, unlike chapters 3 and 4 that focus on short-term predictions of hours or days ahead. In this research stream, a method is proposed that allows the estimation of preconditioning potential for the cooling season in different locations within the same metropolitan area, as well as different climates within Australia. The value of localisation is demonstrated with a range of results and useful conclusions are made about the non-building related factors that affect precooling potential.

Finally, chapter 6 proposes three algorithms for the integration of the numerical weather predictions in predictive building control. The implementation process is explained in detail and there is a case study for a university campus building that simulates the energy savings and peak load reduction and compares them to the standard non-weather responsive control.

2 Literature review

 Detailed background information for relevant research fields of forecasting and building energy management  Review and discussion of existing literature  Addressing the research question regarding the integration of weather predictions in building energy management  Contextualising the PhD research project by defining the gaps in existing literature

2.1 Chapter outline The Literature Review chapter primarily aims to provide the necessary background information for the concepts related directly or indirectly with the thesis. Through the critical review of existing literature, it will be possible to obtain an insight to the question regarding the value of NWP in building energy management. Chapter 2 is divided in thematic sections according to the respective research fields.

Section 2.2 is related to forecasting. After outlining the types of forecasting in the relevant literature, examples of studies will be discussed in an attempt to contextualise the choices and assist the reader in understanding the proposed thesis’ forecasting methodology. After a brief introduction to forecasting (section 2.2.1), section 2.2.2 will review and compare common groups of forecasting models in the field. Sections 2.2.3, 2.2.4 and 2.2.5 will narrow down, discuss and comment on the applications of these forecasting groups in weather, load and generation forecasts respectively.

Section 2.3 is concerned with building energy management and especially the inclusion of forecasting in it. Part of the thesis involves the applications of forecasts in novel modules of energy management for various horizons. This section will review the operating principles of typical management systems and explain the importance of forecasting and weather integration in them. The section begins with a discussion of the advantages and weaknesses of Model Predictive Control (section 2.3.1), and continues with evaluating the inclusion of weather forecasts in building operations (section 2.3.2). Section 2.3.3 reviews and comments on building

5 conditioning principles and finally section 2.3.4 outlines the main findings from the literature on generation and storage roles in building energy management.

At the end of the chapter (section 2.4), and after summarising the findings from the literature review it will be possible to highlight the research gaps that this thesis is aiming to fill and set up the basis for discussing the methodology of each step carried out through the PhD project in chapters 3-6.

The majority of the work presented in this chapter has been published in two research papers (Lazos et al., 2014, Lazos et al., 2015).

2.2 Forecasting 2.2.1 Background Forecasting is among the primary principles involved in every step of this PhD thesis. There are two types of forecasting that the thesis is concerned about: weather and energy forecasting. Throughout the thesis, weather forecasts function as inputs for energy forecasting. In this section, the principles of forecasting techniques for both types and their effectiveness will be reviewed. It will also be possible to discern the gaps in the existing literature that justify the implementation of the forecasting methodology that is described in the following chapters of the thesis.

The main purpose of forecasting is constructing predictions about the value of a number of variables and/or the state of a system at a future time. These predictions are typically constructed based on a set of past and/or present data processed under a number of assumptions forming the prediction model (Box and Jenkins, 1970). In the context of energy management, forecasting models serve a variety of purposes and may be applied to predict a broad range of variables.

In the following section, an outline of the main groups of forecasting techniques applied in the broader research field of weather and energy management will be given. This outline will highlight the principles and features of forecasting model groups in general. However the applications and the evaluation of specific model’s usefulness to the thesis development will be discussed in section 2.3.

2.2.2 Forecasting techniques Forecasting may be carried out via time-series, machine learning and physical models. Hybrid models are not uncommon, as groups contain features that can complement each other and alleviate any shortcomings. The algorithm groups described below, have evolved in the last two decades to a point where they can predict the future state of the variables in question with high degrees of accuracy for a range of applications in weather and energy management modelling. However, the groups have their distinct strengths and weaknesses. This thesis’ methodology involves components of physical forecasting and time-series statistical forecasting, often in a hybrid form.

2.2.2.1 Time-series and regression model forecasting A major stream of forecasting models is based on processing a set of data points in a time series of a single or multiple variables. The time series regression can be used to forecast the future values of the same variable(s), according to specific rules. For example, by considering the energy generation from a solar array over the course of time, predictions about generation in the future can be obtained. Another stream of regression models is concerned with the correlations of number of input variables to a number of output variables, for instance examining how the building load is affected by ambient temperature, the occupancy patterns and the day of the week. Both streams were useful for the development of the methodology of the PhD thesis and hence discussed in this section.

There are several techniques in the field of weather and energy forecasting, which are applied in countless variations in the literature. Linear regression (LR) is among the simplest statistical approaches. LR models predict the value of the output variable(s) through multiple linear relationships of appropriately weighed coefficients and a number of input variables. The weights are typically derived from historical observations. Another group of common statistical time-series techniques are based on the Box-Jenkins model (Box and Jenkins, 1970, Hagan and Behr, 1987, Loveday and Craggs, 1993) and combine an autoregressive (AR) and a moving average (MA) part. The AR and MA parts are responsible for linking the present value of the time series to its past values, as well as some past random error. When there is an integrated part added to ARMA methods, it is possible to remove any 7 non-stationarity from the data (Wan Ahmad and Ahmad, 2013); these models are referred to as ARIMA. The AR group of forecasting techniques is capable of simulating processes that are governed by variables that are subject to randomness (stochastic forecasts) and they are naturally applied to processes, such as the weather or the energy consumption.

Stochastic forecasts may be combined with deterministic models that can predict the evolution of variables independently of random errors. Exponential weighting (EW) is an example of such a technique. The principle of EW approaches is assigning appropriate weights to observations in a time series in order to develop forecasts for future points. The Holt-Winters (HW) approach (Holt, 2004, Taylor, 2003) is of particular interest to energy consumption forecasting as it accounts for seasonality patterns in a time series. These patterns may be linked in an additive or a multiplicative manner. The presence of seasonality in variables, such as ambient temperature or energy consumption allows the use of Fourier series as an alternative method to approximate the wavelike behaviour of the variable and make predictions in the frequency domain.

The groups of statistical techniques outlined above relate directly to the methodology of the proposed thesis. The main reason for this is that both weather and energy consumption variables demonstrate trends that are easy to individually identify, analyse and predict. Furthermore, there are several correlations between variables that may be modelled and understood based on historical archives. A key element towards both goals is weight estimation from past data.

2.2.2.2 Machine learning forecasting A different approach in forecasting, is realised in the form of algorithms simulating learning processes, mainly Artificial Neural Networks (ANN). While this thesis does not incorporate any significant machine learning components, a review of the main techniques and applications is necessary to obtain a better understanding of the forecasting field.

Rather than using input data to decompose the time series and develop a fit using a number of parameters, ANN attempt to simulate the non-linear and non-stationary univariate or multivariate dependencies through networks similar to those found in

8 the central nervous systems of mammals. The network consists of layers of nodes that form a number of staged connections of different weights between inputs and outputs. Training data are fed in the input layer and change as they propagate towards the final layer, in an attempt to match the observed data as closely as possible. The most common algorithm governing this process is known as back propagation; the weights of each node are being continuously modified according to a stream of feedback that flows backwards and describes how accurate the configuration is compared to the target results (Rumelhart et al., 1986). Support Vector Machines (SVM) are another machine learning group utilised in forecasting and can model non-linear relationships based on a structure risk minimisation principle that aims to minimise the upper bounds of error of the object function. Their distinct advantage over ANN is the ability to locate global minima rather than local minima in the solution space, as well as the ability to solve non-linear systems with a smaller training dataset (Xuemei et al., 2010a).

2.2.2.3 Physical model forecasting Physical models for forecasting are based on attempting to mathematically model the physical processes that characterise a system in order to predict its future state. Complex systems are commonly governed by a range of processes. For example, in the case of building energy systems, the physical processes include the structural and thermodynamical aspects of the building, as well as the interactions with its internal and external environment. While historical data are typically not required for most physical models, the amount of inputs differs from case to case resulting in a variety of complexities.

According to the amount of physical relationships and internal system information known and accounted for, physical models can be further classified in the literature as white box (for a high number of relationships/internal system data), grey box (for a lesser amount of relationships/ internal system data) or black box (for a minimal number of relationships/ internal system data). A summary of the classification explained earlier can be seen in Figure 2.

Black Box Very low number of physical relationships between inputs and outputs known Low availability of internal system information Predominantly based on input and output data correlation

Grey Box Large number of physical relationships between inputs and outputs known High availability of internal system information Both input/output data and intermediate data are used

White Box Large number of physical relationships between inputs and outputs known High availability of internal system information High reliance on accurate estimation of intermediate system parameters

Figure 2: Classification of physical models according to the information known about the internal system relationships and data In the case of building energy systems, an example of a physical grey box model is the development of a thermal network analogous to an electrical circuit. Such a model is able to simulate and predict the building’s thermal behaviour by taking into account and thermally connecting the heat sources (internal appliances, occupants and solar radiation), thermal resistors (walls, floors, ceilings, external building envelope) and capacitors (thermal mass elements) (Zhou et al., 2008).

In the case of weather forecasting, a common type of physical forecasting models are known as Numerical Weather Prediction (NWP) models. Numerical methods refer to the analysis of the evolution of a set of variables in the atmosphere in a multidimensional calculation space, governed by differential equations of the thermodynamics, fluid dynamics and chemical reactions of the constituents of air. Obtaining consistently accurate weather forecasts of high temporal and spatial resolution via numerical methods is relatively impractical and computationally complex for small scale applications. As such, NWP applications for building energy management have not been extensively researched and will form the basis of the majority of the methodological steps in the current PhD thesis.

2.2.2.4 Technique comparison In this section, an outline of common forecasting techniques will be provided and a brief comparison will be conducted. The comparison is considering each technique’s applicability in any stage of building energy management processes. The applications of several of these models are of relevance to the development of the thesis, and will be further discussed in the following chapters.

A summary of the features of the different groups of forecasting approaches and a number of examples from the literature related to building energy management can be seen in Table 1.

Table 1: Comparison of common forecasting techniques

Forecasting Method Features Major Example studies family limitations Time series & Autoregressive Simple, fast, Historical data are (MacArthur et al., 1989, Ren and regression moving average relatively high needed, weak in Wright, 2002, Yoshida and Terai, models (ARMA accuracy, ability modelling non- 1991, Yoshida and Terai, 1992, & ARIMA) to account for linear patterns Mustafaraj et al., 2010, Chowdhury seasonalities to and Rahman, 1987, Kimbara et al., some extent, 1995, Borges et al., 2011, Fernandez short forecasting et al., 2011, Xuemei et al., 2010a, Yao horizons et al., 2004, Kawashima et al., 1995, Penya et al., 2011) Autoregressive As ARMA, with As ARMA, plus (Rios-Moreno et al., 2007, Soleimani- models with enhanced ability require Mohseni et al., 2006, Bacher et al., exogenous to account for availability of 2009, Bacher et al., 2011, Cai et al., inputs (ARX) recent exogenous 2010, Kyungtae et al., 2012) exogenous variable changes monitoring Linear Simple, fast, fair Weighing of (Zhang and Hanby, 2007, Huang et regression (LR) accuracy coefficients is al., 2011, Pedersen et al., 2008, challenging, weak Kyungtae et al., 2012, Kawashima et in modelling non- al., 1995) linear patterns and seasonalities Machine Artificial neural Accurate, no Reliance on (Ruano et al., 2006, Soleimani- Learning networks need for historical data, Mohseni et al., 2006, Gonzalez and (ANN) supervision, computationally Zamarreno, 2005, Ferrano and able to model complex Wong, 1990, Lanza and Cosme, 2001, non-linear Mustafaraj et al., 2011, Argiriou et patterns, high al., 2000, Argiriou et al., 2004, running speed Ferreira et al., 2012a, Ferreira et al., 2012b, Ahmed, 1999, Gouda et al., 2002, Yang and Kim, 2004, Huang et al., 2013, Chen et al., 2011, Chow et al., 2012, Al-Messabi et al., 2012, Yokoyama et al., 2009, Kwok, 2011, Kwok et al., 2011, Shi and Wang, 2009, Kusiak et al., 2010, Kreider and Wang, 1992, Ben-Nakhi and Mahmoud, 2004, Hou et al., 2006, Rivard et al., 2005, Yao et al., 2004,

Kawashima et al., 1995, Kawashima et al., 1996, Penya et al., 2011)

Support vector Accurate, more Computationally (Xuemei et al., 2010a, Li et al., 2010b, machines (SVM) solid complex, low Li et al., 2009b, Li et al., 2009a, Li et architecture running speed al., 2010a, Cherkassky et al., 2011, than NN, able to Xuemei et al., 2010b) model non- linear patterns, needs less training data than ANN Physical & Engineering Highly accurate, Multiple inputs (Keller and Costa, 2011, Braun, 1990, Numerical white and grey do not rely on needed, can be Braun and Chaturvedi, 2002, Braun box methods historical data, complex and slow et al., 2001, Lee and Braun, 2007, Lee physical running speed and Braun, 2004, Xu, 2004, Al- interpretation Rabghi and Al-Johani, 1997, Fraisse et al., 2002, Chen and Yu, 2009, Rabl and Norford, 1991, Westphal and Lamberts, 2004, Luo and Ariyur, 2010, Wang and Xu, 2006, Pakanen and Karjalainen, 2009, Oldewurtel et al., 2012, Lehmann et al., 2013, Wen and Smith, 2007, Sun et al., 2010b, Sun et al., 2013a, Sun et al., 2013b, Escriva-Escriva et al., 2010) Numerical Highly accurate, Uncertainty (Zavala et al., 2010, Zavala et al., weather do not rely on present, resource 2009, Kwak et al., 2013, Zhou et al., prediction historical data, intensive and time 2008, Nagai, 2007) (NWP) & physical consuming, low physical interpretation, temporal weather long forecasting resolutions if forecasts using horizons external data are external inputs used

According to the review of the literature in sections 2.2.2.1-2.2.2.3 and the examples in Table 1, a number of findings may be extracted. Firstly, the advantages of statistical regression are that it is simple in terms of implementation, demonstrates high simulation speeds and requires low computational power. However, regression techniques are challenged when trying to predict non-linear relationships that may not follow a clear pattern. They also rely heavily on consistent past data, for both regression processes and weight calculation. Machine learning techniques are considered as an alternative able to capture non-linear relationships without manual estimation of the parameters. Nonetheless, the complexity of such algorithms is higher than statistical methods and the reliance on reliable and large amounts of archived data is still an issue. In addition, problems, such as overfitting to the training datasets may arise. Finally, physical forecasting methods attempt to analyse the underlying principles that govern the system, rather than trying to “guess” the input-output relationships. Parameterisation of physical 12 models poses a challenge and in many cases estimations have to be made according to the availability of internal system parameters and design complexity limitations. It is worth noting that grouping forecasting models according to their operational principles, is not the only classification system in the literature. Very often, forecasting models are distinguished according to their forecasting horizon, or the period in the future for which they can make predictions. There is no clearly agreed definition of the boundaries for different horizons in forecasting, but in general short term forecasting refers to predictions of up to hours or a few days ahead, medium term forecasting refers to predictions of up to several days ahead, and long term forecasting may refer to horizons of weeks, months or even years.

2.2.3 Weather variable forecasting A major focus of this thesis is to develop tailored weather forecasts in order to use their outputs for energy management of commercial buildings. The reason for this is that many of the operational traits of the buildings and behavioural aspects of occupants are inherently affected by weather variables. Onsite solar and wind power generation, HVAC load patterns and to a certain extent lighting load and occupancy habits are predominantly dependant on the ambient temperature, humidity, incident solar radiation, cloud formation and wind (Pedersen, 2007). This section will discuss existing applications of weather forecasting in the context of building energy management and highlight the importance of including them for enhanced system performance. Through the discussion, primarily in sections 2.2.3.3 and 2.2.3.4, it will be possible to identify certain research gaps and justify the design of the weather prediction component of the thesis.

Weather is perceived as the macroscopic result of numerous interactions between particles of in the atmosphere and the surface of the Earth. Naturally, it is nigh impossible to obtain information about the state of every particle participating in those interactions. Hence, weather forecasting relies on data assimilation conducted via spatially and temporally scattered observations (Zavala et al., 2010). Among the groups of forecasting techniques described in section 2.2.2, modern weather predictions in the context of energy management, favour physical models. However, there are earlier examples in the literature based on time series and regression weather forecasting instead. 13

2.2.3.1 Time series and regression weather forecasting Among the simplest statistical techniques, useful for short term forecasting and significant for the development of this thesis, is the persistence model. The persistence model assumes that the state of the atmosphere is stationary and does not experience significant changes. In simple terms, it assumes that the weather remains relatively constant in short periods of time and as such, weather variables demonstrate only minor variations for short-term horizons up to 3 hours ahead (Nielsen et al., 1998). The persistence model is often used as a reference model in studies attempting to validate other more complex weather forecasting models and evaluate their skill (Mittermaier, 2008). Furthermore, persistence is useful for the estimation of individual component model weighting in predictions produced by hybrid or the weighting of individual branches in ensemble models (Greybush et al., 2008). Wind speed predictions are commonly utilising the persistence model and several applications have been developed for the assessment of the power output of a wind farm or micro siting of the turbines (Agüera-Pérez et al., 2013). However, predictions of other weather variables, such as temperature (Abdel-Aal, 2004) and rainfall (Landman et al., 2012), referred to the persistence as well, in order to evaluate the skill of the model of each study. In addition to its usefulness as a short- term horizon forecasting tool, the use of persistence has been justified for even larger temporal scales in the domain of climatology (Bunde and Havlin, 2002). As with the examples discussed above, the persistence assumptions will be used in this thesis for both forecasting purposes and as a reference model. According to the literature the persistence model is under-utilised in regulating numerical weather predictions in short-term horizon and high resolution forecasts, which constitutes a notable research opportunity.

An early attempt to predict ambient temperature in a more complex and realistic way than the persistence model was carried out with the help of an ARIMA model (MacArthur et al., 1989); its outputs were used in a cooling load prediction model. Decomposing such AR models in a stochastic part and a deterministic EW model to predict ambient temperature reportedly improved the accuracy by almost 10% compared to the base AR case (Ren and Wright, 2002, Yoshida and Terai, 1991). This

14 technique was modified to predict humidity and solar radiation, by using a Fourier time series for the deterministic part (Yoshida and Terai, 1992).

Besides ambient temperature, solar radiation is often considered as an important weather variable in the literature. Solar radiation was forecasted using historical data and validated for modelled buildings (Chen and Athienitis, 1996). Furthermore, local solar radiation was predicted via a deterministic EW technique (Ren and Wright, 2002). A different approach through simulation of the atmospheric state in MATLAB was proposed (Keller and Costa, 2011). A multiple LR technique for short term solar radiation forecasting for buildings was implemented by Zhang & Hanby (Zhang and Hanby, 2007) using a combination of onsite observations and third party weather forecasts. ARIMA and LR based models are mainly challenged in predictions of non-linear patterns, especially when the availability of historical data is limited (Jetcheva et al., 2014).

Understanding the correlation of changes in weather variables in a set rather than individually is another example of statistical forecasting model. For instance, high temperatures in the summer are very often observed in parallel with low relative humidity and taking into account their correlation can enhance the accuracy of weather forecasts. This forms another research gap that the thesis aims to fill.

2.2.3.2 Machine learning weather forecasting Approaches based on ANN of different architectures have been also proposed in the field of weather variable forecasting, such as predictions of ambient temperature patterns for short term horizons (Lanza and Cosme, 2001) or general integrated multivariate weather forecasting modules (Argiriou et al., 2000, Argiriou et al., 2004). Another type of ANN included a sensor that directly obtains and feeds in data of temperature, solar radiation and cloud coverage to the network (Ferreira et al., 2012a). Common elements of these algorithms include the necessity for onsite weather observations and an indexing method associating these observations to particular times, days and seasons. The use of ANN over statistical methods in weather variable forecasting is justified due to their ability to capture non-linear patterns in the evolution of weather variables (Florita and Henze, 2009). Improved forecasting skill has been observed when ANN were hybridised with statistical approaches, for instance adding an ARMA component to forecast solar radiation 15 after removing non-stationarity from the time series (Ji and Chee, 2011). It should be noted however, that the added complexity and costs of developing and training ANN compared to simpler statistical approaches are often not offset by the boosts in forecasting performance (Florita and Henze, 2009).

2.2.3.3 Numerical weather predictions (NWP) As discussed in section 2.2.3.1, this thesis aims to utilise statistical forecasting in the context of building energy management. The statistical forecasts will be combined with numerical weather predictions to develop predictions of higher accuracy in various horizons.

The uncertainty related to weather variables imposes a challenge to predictions of longer horizons. Furthermore, certain events characterised by sudden changes in the atmosphere, such as a sudden drop in temperature, or the occurrence of gusts of wind are not easily predicted via statistical or ANN approaches. Both issues arise from the chaotic nature of weather and its sensitivity to initial conditions. To address such challenges, weather variable evolution can be modelled and predicted via NWP models. As explained in section 2.2.2.3, NWP models are able to solve fundamental equations governing the evolution of atmospheric variables in three spatial dimensions and through time and predict its future state. Most modern NWP models are able to downscale weather data from the synoptic scale (typically areas with dimensions of at least a few hundred square kilometres) to smaller size grids (typically areas of dimensions of as low as 1 square kilometre) nested within each other.

Outputs from a NWP model (Weather Research & Forecast Model) have been used in order to quantify the sensitivity of weather to the initial conditions and their inherent uncertainty (Zavala et al., 2010). It was confirmed that the NWP model was more accurate than the statistical AR model it was compared to, especially for horizons longer than 8 hours. It can be seen in Figure 3 that even though the regression approach was able to predict the temperature trends, the fact that it deviates notably from the actual observed values undermines its applicability to energy system management algorithms (Zavala et al., 2010).

Figure 3: Comparison of the forecasting skill of a statistical AR prediction model (top), versus a NWP model (bottom) in 5-day horizon temperature forecasts (Zavala et al., 2010) In the development of this thesis, NWP are derived exclusively from the software TAPM, or The Air Pollution Model. Since TAPM is critical to each of the following chapters of the thesis, a separate section describing its operation modes is necessary.

2.2.3.3.1 The Air Pollution Model (TAPM) TAPM was developed by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) mainly for modelling atmospheric changes for use within the fields of environmental science and pollution research. However, it also contains a meteorology component that can be utilised to obtain weather variable outputs for any location in the world. The base meteorological variables are determined in TAPM as follows (Hurley et al., 2005):

 The orthogonal horizontal plane components of wind speed (u and v) are determined from the momentum and terrain following vertical wind speed equations  The terrain following vertical velocity vector () is determined from the continuity equation  The potential virtual temperature (θv) is determined from the conservation of heat and vapour  The Exner Pressure () is determined as the sum of hydrostatic (H) and non- hydrostatic pressure components (N); the non-hydrostatic component of the equation is optional

The derivation of the above equations can be found in CSIRO’s documentation (Hurley, 2008a). Apart from the base variable equations and grid resolution, parameterisation affecting the local meteorology includes soil and vegetation types, precipitation micro-physics, radiative and turbulence fluctuations and degree of urbanisation (Hurley et al., 2005). The derivation of equations for each of the above parameterisations can be also found in CSIRO’s documentation (Hurley, 2008a). TAPM is able to generate hourly numerical predictions for a variety of weather variables. Table 2 shows a list of the terminology used in this section, as well as throughout the following chapters in the methodology of the thesis when TAPM is involved.

Table 2: Terminology used in reference to TAPM simulations

Term Description Domain The domain is a 3-D space of custom dimensions in the shape of a rectangular prism. TAPM simulations and predictions occur within a particular domain. Domain order Higher order domains refer to domains of greater size, while lower order domains refer to domains of smaller size. Domains in TAPM exist within each other in successively lower orders. Grid points Each domain is divided in discrete grid points in all dimensions (x,y,z). Physical features, such as latitude, longitude, and elevation or vegetation index are stored in TAPM’s database for every grid point. Simulation results can be obtained at any grid point of any domain (but not at any other point of the domain). Grid resolution Grid points within the same domain are separated by a custom constant distance, which can be defined as the resolution. The resolution can be the same or different for each dimension. Nesting ratio The nesting ratio is defined as the ratio of the total areas on the ground level (x and y dimensions) of two consecutive order domains.

As with most modern NWP platforms, TAPM is able to run simulations in up to five 3-dimensional domains nested within each other. The user can define the dimensions of each domain, as well as the number of grid points in each domain that the simulations will take place. Synoptic data is used to initialise the simulations at every grid point and set the boundary conditions of the highest order (largest) domain. Synoptic weather observations can be taken directly from meteorological databases (such as the National Centre for Environmental Prediction - NCEP). Figure 4 shows the user interface of TAPM, with the primary options for parameterising the domain.

Figure 4: TAPM graphical user interface. The grid parameterisation is visible in regards to spacing, resolution, location and temporal options The simulations are then run in succession at the lower order domains with progressively higher spatial resolution. The lower order domains use the outputs from the previous higher order domain to determine the boundary conditions. The simulations are computationally non-intensive and they can be run on a normal computer. Thus, it is possible to use the synoptic data to generate very highly localised meteorological prognostics in a reasonable amount of time. Figure 5 shows the parameterisation menu for the domains of lower order (2-5).

Figure 5: Parameterisation menu for the lower order domains (up to 4) The analysis of the results can be obtained for any domain and for any grid point within the domain, as well as different altitudes. For the work presented in this thesis, the outputs were obtained from the grid point with coordinates (0,0) (at the centre of the map) and at ground level. Figure 6 shows a sample domain grid as depicted by the TAPM user interface.

Figure 6: A sample grid for the greater Sydney area There have been a few studies utilising the meteorological component of TAPM in the existing forecasting literature. Solar global horizontal irradiance was predicted with a model based on TAPM and using domains with resolution of 45km. It was found that the predictions overestimated the irradiance, especially during periods of rapid cloud coverage changes (Dehghan et al., 2014). More relevant to this thesis, accurate temperature and wind speed predictions were produced at domains with 3km resolution (Thatcher and Hurley, 2010) in certain Melbourne regions. TAPM was also used to predict hourly values of radiation and temperature in a site in Western Australia, even though post-processing was required for certain cases, such as cloudy days and high irradiance periods (Hibberd, 2011).

However, there is so far a lack of studies examining the potential of using TAPM to generate useful outputs for the energy management of commercial buildings at very high resolution domains. This thesis attempts to fill in this research gap. Furthermore, besides assessing TAPM’s forecasting skill in general predictions, new applications of TAPM outputs that have not been discussed in the literature will be generated. These include among others, predictions of abrupt changes and extreme heat events, as well as the correlations of weather variables, such as temperature and relative humidity. Such findings are of use to energy management systems, as 21 with appropriate infrastructure it may be possible to achieve significant energy savings. Another novelty that the methodology of this thesis is proposing is ensemble forecasting, which is explained in detail in the following section.

2.2.3.3.2 Ensemble forecasting Most often, NWP models and especially those referring to synoptic scale forecasting consist of multiple iterations of predictions for the same time frame and same geographical area. The output predictions from each individual iteration are combined to form a bundle, or an ensemble, of forecasts.

Typically, ensembles of NWP can be formulated in three ways (World Meteorlogical Organization, 2012):

 Running parallel simulations of the same prediction model with slightly different initial conditions for one or more variables.  Running parallel simulations of the same prediction model with constant initial conditions, but different parameterisation of the grid’s simulation geometry (number of grid points, resolution, and dimensions).  Running parallel simulations of different prediction models.

As it is impossible to measure the state and model the behaviour of each particle in the atmosphere individually, weather predictions are always associated with some uncertainty and randomness; hence they are defined as non-deterministic. Ensemble forecasting is common in modern meteorology as it encompasses this non-deterministic nature of weather and instead provides a probabilistic perspective on the evolution of the future state of the atmosphere. This means that the prediction outputs convey the likelihood that the weather variables will be within a certain range according to their distribution in the ensemble bundle, rather than a certainty that they will have a specific value. While ensembles have been a staple in synoptic scale forecasting and medium-term predictions up to a fortnight ahead, they can also been implemented in regional and convective scales (grids with dimensions of a few kilometres) and short-term horizons (World Meteorlogical Organization, 2012).

The obvious applications for ensemble NWP are directly related to weather dependent systems and meteorological prognostics, such as daily weather 22 predictions or predicting extreme phenomena (tornadoes, hurricanes etc). However, short or medium term horizon ensemble predictions at regional or local scales have also been successfully utilised in various engineering applications outside of the field of meteorology. Some examples include: predictions and management of air traffic and airport capacity (Kicinger et al., 2012, Yousefi et al., 2013, Stobie and Avjian, 2009), management of winter sport events (Mailhot et al., 2010), irrigation scheduling (Cai et al., 2011) and shipping lane routing optimisation (Hinnenthal and Clauss, 2010). Renewable energy generation forecasting has also utilised ensembles to minimise uncertainty in energy predictions for wind (Al- Yahyai et al., 2012, Ancell et al., 2015), solar (Mathiesen, 2013, Zavala, 2013, Alessandrini et al., 2015) and hydroelectricity (Wang et al., 2012).

Ensemble weather forecasting has also been applied to electricity demand predictions for large scale systems. Investigations found that outputs from such models can account for the uncertainty in the evolution of future demand better than deterministic inputs (Ranaweera et al., 1996, Yang et al., 2007). Forecast errors in demand predictions with NWP ensembles was found to be lower than in deterministic approaches based on archived data in short and medium term horizons (Taylor and Buizza, 2002) for regional level electrical loads. In regards to extreme weather events that may lead to very high peak cooling loads, such as high temperatures, deterministic predictions from NWP models may produce mean absolute errors of up to 15% (Lazos et al., 2015). This may be reduced via the implementation of a probabilistic approach (ensemble). Ensemble NWP outputs have also been used to confirm and quantify the correlation of peak loads at a regional scale with ambient temperature (De Felice et al., 2015). These applications involved load predictions and modelling at a local or regional scale and horizons of up to a few days ahead.

For energy management purposes at the building level, ensembles of NWP are more challenging to implement, due to the underlying complexity of NWP models, their intense computational requirements (may be up to 5-6 hours) and the often long lead-in time necessary to establish initial conditions (Mathiesen, 2013). Instead, data-driven grey or black box approaches, such as ANN ensembles for load prediction can be used and have proven effective in short-term forecasts of the 23 behaviour of the building energy system with reported mean percentage errors of up to 5% (Kusiak et al., 2010, Fan et al., 2014, De Felice and Yao, 2011, Jetcheva et al., 2014, Burger and Moura, 2015).

NWP ensembles are also known to add value to forecasts by minimising the error consistently for longer horizons of a few days ahead. The superiority of NWP ensembles compared to statistical time series models for these horizons has been confirmed, as the standard error was reduced notably for temperature, radiation and wind predictions (Zavala et al., 2009).

Regardless of the application or the horizon, ensemble NWP models are treated as effective solutions to the challenge of minimising the uncertainty error in predictions of variables related to weather and climate. While most such models are applied in medium-term horizons and regional or national levels, a few short-term horizon applications at lower scale spatial levels have also been proposed. NWP ensembles however, have never before been used in the context of building load predictions, which constitutes a research gap that this thesis aims to address. Since the resolutions of certain ensemble branches is as high as 100m, they may be used for predictions at the individual building level. However, sharing between buildings may also be possible after adjusting the grid resolutions accordingly.

2.2.3.4 Weather forecasting design considerations In the context of energy management, weather forecasting functions as the first layer of prediction inputs. Naturally, a question about the selection of an appropriate weather forecasting framework arises, when designing such a management system. There are numerous factors that need to be considered to answer this question. Firstly, forecasting may refer to specific weather variables. In most cases ambient temperature is the most useful variable for energy management systems. However, other variables, such as relative humidity and cloud coverage may be of interest depending on the energy system’s design.

The second factor refers to the forecasting horizon. As discussed earlier, NWP models provide the highest forecasting skill, especially for horizons of some hours up to days ahead. For shorter horizons, statistical or ANN are able to produce excellent and accurate predictions instead.

Another important consideration is related to the availability of historical and recent input data. Both statistical and ANN prediction models are heavily data driven, as they rely on a specific amount of accurately calculated weights to develop predictions (Bacher et al., 2009). On the other hand, numerical models typically rely less on historical data and can function with only recent observations. Frameworks that combine data driven predictions from historical data with external recent inputs naturally complement each other and may result into even better forecasts compared to individual models. The external inputs in such models necessitate that some onsite observation and logging equipment is present (weather monitoring station). In these forecasting frameworks, the weights and general trends of weather variable changes may be identified via data driven approaches and corrected by regularly updated data inputs from a recent observations. Hence, the availability of historical and/or recently observed weather data dictates the major elements of a forecasting framework.

Of course, onsite weather monitoring may not always be available for every application – especially for small scale predictions, such as forecasting at the level of a building or a neighbourhood. Due to this constraint, weather forecasting is commonly obtained from third parties (weather stations, airports, meteorology bureaus), instead of being developed onsite. The advantage of this approach is that predictions can be obtained without committing significant resources. The predictions may be modified to match the need of the respective application. For instance, Kwak et al. used readily available weather forecasts from the Korean Meteorological Administration and developed their own numerical model for solar irradiation on a case study office building (Kwak et al., 2013). This allowed for more flexible predictions, as the outputs were tailored to the building energy system in terms of resolution, horizons and precision. In a similar manner, raw real time weather observations from the Hong Kong Observatory were obtained and fed into modules able to process and generate more relevant information for the building, such as solar heat gains, relative humidity and localised temperature evolutions (Zhou et al., 2008). External weather forecasts have been also combined with onsite observations to produce temperature and humidity forecasts for building control purposes (Nagai, 2007).

NWP models offer an alternative for the provision of weather predictions at any location, as they can easily downscale synoptic data to any domain and allow localised forecasts to be made. The provision of localised weather data inputs for developing forecasts is rather significant: any forecasting error at this level will propagate and may be magnified towards higher level predictions that depend on weather variable inputs (such as energy demand). Regardless of whether the forecasting is based on statistical, machine learning or numerical methods, localisation of input data helps avoid spatial and temporal errors when predicting the values of weather variables. Forecasts generated onsite with localised data (whether from onsite observations or downscaled numerical data) offer a range of benefits: the ability to better capture the microclimate of the location, the ability to tailor the prediction outputs depending on the application and the capacity to develop custom spatiotemporal resolutions (Zavala et al., 2010). A summary of the basic design considerations as discussed in the existing literature for a weather forecasting model in the context of energy management can be seen in Figure 7.

Figure 7: Summary of key research considerations in the field of weather forecasting for building energy management

2.2.4 Load forecasting In the context of energy management, the main application of weather forecasting is in the field of load and generation predictions. Specifically, the management of energy systems of all levels (national, regional, urban, building) is associated with the prediction of the load and, where available, onsite generation from renewable sources, which in turn receive weather inputs. Accurate forecasting information may assist in minimising the energy costs and adding value to the generated energy, especially during peak load periods.

This section will attempt to answer the question about whether the accuracy of load forecasts can be enhanced when using weather variable inputs. Changes in weather conditions and especially ambient temperature and relative humidity, affect the cooling and heating demand and as a result pass on some of the inherent uncertainty in weather forecasting towards the load forecasting level (Lu et al., 2009, Perez- Lombard et al., 2008, Steinfeld et al., 2011). The scope of this thesis is investigating load forecasting predominantly at the building level and specifically for commercial buildings and hence a discussion of the main traits of such loads is necessary.

2.2.4.1 Commercial building loads Commercial building loads are distinctive as they generally vary in regular diurnal, weekly and seasonal patterns, which are to a large extent related to the weather and provide grounds for optimisation as long as the patterns are well understood (Gould et al., 2008). On an intra-day basis, notable similarities appear between typical commercial building loads: the presence of a base load, a morning ramp-up, an afternoon peak followed by a “shoulder” and finally an evening recession towards the base load (Mathieu et al., 2011). These load profile features can be seen clearly repeating for a two-day period in Figure 8.

Figure 8: Typical commercial building load profile on two consecutive weekdays (Mathieu et al., 2011) The main reason for this inter-day regularity of load profiles is that occupancy patterns follow the working hours and are consistent in most commercial buildings. Additionally, the occupants tend to carry out activities of the same energy intensity on average.

Figure 9 shows the load profiles for each day in a year for a university building.

Figure 9: Daily load profiles for a university building in Spain in 2009 (Penya et al., 2011) The similarity in shape between weekdays is notable. Weekend load profiles are usually of the same shape as weekday profiles only reduced in magnitude. However, for certain buildings Sunday or public holiday loads may be negligible (for example education buildings, as seen in Figure 9.

Typically, major efforts have been made in predicting and optimising the thermal aspects of building energy systems, because heating, ventilation and air

28 conditioning (HVAC) accounts for a major (often the largest) part of the overall energy demand in commercial buildings (Perez-Lombard et al., 2008, Lu et al., 2009). Moreover, evidence shows that HVAC is majorly responsible for the magnitude of peak loads (Steinfeld et al., 2011). Ambient dry bulb temperature is recognised as the primary contributing variable to commercial building loads (Reddy and Claridge, 1994), a now commonly accepted fact in load prediction related research.

For cooling loads in particular, there are two components related to the heat gains in a building. The heat originating from external sources, due to temperature, as well as the heat originating from internal sources, such as lighting, electronics or even occupants is characterised as sensible heat and acts to directly increase the indoor temperature of a building (Engineering Toolbox, 2016a). Additionally, there is latent heat, due to the moisture originating from sources like occupant respiration or the humidity in the atmosphere. Unlike sensible heat, latent heat changes do not change the air temperature (Engineering Toolbox, 2016a). Most forecasting techniques account for both types of heat in the building to calculate the HVAC load.

Commercial building loads are significant contributors to the energy consumption mix globally and are expected to expand their contribution in the future (Day et al., 2009); as such, the development of a multitude of approaches to load forecasting has been necessitated (Perez-Lombard et al., 2008). There are several studies that offer extensive reviews on this field of load forecasting: Zhao and Magoules (Zhao and Magoules, 2012) reviewed the recent trends in forecasting building energy consumption with particular emphasis on machine learning techniques. Foucquier et al. (Foucquier et al., 2013) complemented their effort by expanding the list of algorithms that appear in the literature, comparing and classifying them. Another recent effort by Sun et al. (Sun et al., 2013b) focused on examining methods of shifting peak load in buildings.

There is clearly significant interest in the literature in load forecasting for commercial buildings, due to the regularity of trends, potential for savings and ease of implementing active energy management policies. The following sections will

29 review existing approaches for building load forecasting and comment on their usefulness and performance, as well as their relevance to the current thesis.

For most commercial buildings, historical load data is readily available. As a result, data driven forecasting for both short and long horizons is favoured. The simplest models treat total or heating/cooling load as a time series, without any correlation to weather outputs. However, it is common to use externally observed or locally generated weather inputs in order to enhance the accuracy of load predictions. This section will focus on the latter group, as they are more relevant to the aims of the thesis, starting with statistical models.

2.2.4.2 Time series and regression load forecasting Time series and regression methods have been broadly used in the literature for load forecasts. There is a multitude of statistical techniques depending on the application, required accuracy and resolution, however the typical approach is to use archived data to generate a time series that fits the actual load data as closely as possible. Using historical data to develop time series algorithms is a common practice in order to produce load forecasts for horizons up to a few hours ahead (Taylor, 2012), commonly referred to as Short Term Load Forecasting (STLF). The reason for that is that demand response (DR) measures implicate significant potential for savings if applied timely (Mathieu et al., 2011) and as section 2.3 will outline, most DR measures are implemented within short horizons. This section will outline several models, many of which utilise weather data to develop load forecasts. The advantage of using weather inputs will be clear in the analysis of the forecasting skill and application of these models in section 2.3.

Polynomial regression methods are commonplace in load forecasting as a time series. Electrical and heat loads have been predicted using LR and generalised long- term profiles for different types of buildings (Pedersen et al., 2008). Fernandez et al. (Fernandez et al., 2011) tested a prediction algorithm with a polynomial of varying degrees with hourly data from a university building in Spain, but found it can be improved if AR models were applied. Either way, these algorithms are able to run without the need for any weather inputs.

Another common method for modelling the load time series for STLF is based on EW techniques. The original Holt-Winters model is based on three components: level, trend and seasonality (Wan Ahmad and Ahmad, 2013) and several modern models utilise and build on these features. The exponential weighting can be complemented by a moving average (MA) part for the load, as well as ambient temperature inputs (Seem and Braun, 1991). Taylor applied double and triple exponential smoothing in order to capture additional seasonal patterns in electrical loads (Taylor, 2012, Taylor, 2010, Taylor and Snyder, 2012), however the algorithms have not been thoroughly tested at the building load level. Towards the opposite direction of simplifying the model by modifying the level and removing the trend component, He and Zhang (He and Zhang, 2005) validated their approach by forecasting the AC load of an office building.

An alternative statistical method based on historical data in building STLF is the AR family, especially with the integrated part included (Kimbara et al., 1995). Fernandez et al. tested an ARIMA model (Fernandez et al., 2011) and then improved it by introducing a variable learning window mechanism that accounts for days of the same type (weekdays, weekends) and using weight factors to magnify the importance of the most recent observations in the time series (Borges et al., 2011). ARIMA models may utilise weather inputs, especially temperature and/or relative humidity.

Another interesting approach was that of Frank & Sen (Frank and Sen, 2011), in which peak and overall loads of certain buildings were calculated based on an algorithm that manipulates general climate data from Energy Databases (such as Typical Meteorological Year data), rather than utilising specific on site data. Finally, a more complex, yet effective method that is able to capture the non-linear correlations between temperature and commercial building load based on Fourier series has been discussed in the literature (Dhar et al., 1999a, Dhar et al., 1999b).

In addition to STLF, there have also been studies that attempt to predict the long term effects of climate changes in the energy consumption of commercial buildings (Yu et al., 2012). Van Paassen and Luo (Van Paassen and Luo, 2002) developed a statistical weather generator model that can be applied towards that end as well.

Another approach was based on generating weather simulations from existing data and utilising them for long term load predictions (David et al., 2005).

The usefulness of time series load forecasts is then obvious, as they can be applied in time-series regressions of load for multiple horizons, as well as load forecasts developed as a function of external input variables, such as ambient or internal temperature. In this thesis, load forecasting will be include various statistical components.

2.2.4.3 Machine learning load forecasting The notion of using weather data to assist with predictions of load is not uncommon in machine learning based methods. These models are data-driven as well, but instead of well-defined regressions, they attempt to predict load using a set of layers interconnected in multiple dimensions. As with most statistical methods, the energy demand is expressed as a time series and the changes in its value depend on a variety of inputs. Similarly, and like statistical methods, the most common inputs that are considered are temperature and relative humidity (Yokoyama et al., 2009, Ferrano and Wong, 1990, Gonzalez and Zamarreno, 2005).

An increasingly popular group of load forecasting methods involve the use of ANN. ANN parameterisation is highly variable and depends on the architecture proposed in each study. The load is expressed as a time series and its value is affected by a number of inputs in these models, with temperature and relative humidity being the most common weather inputs used in the literature (Yokoyama et al., 2009, Ferrano and Wong, 1990, Gonzalez and Zamarreno, 2005). Historical weather data can be used for model training purposes, while weather forecasts can be used for future load predictions once training is complete; a typical ANN system of this architecture can be seen in Figure 10.

Figure 10: Steam load prediction in a building using an ANN architecture with weather inputs for both training and forecasting (Kusiak et al., 2010).

Regarding commercial building STLF, there is no consensus towards a single architecture that stands out. A multitude of algorithms have been tried and demonstrated superior performance in correlating weather and temporal inputs with load outputs compared to statistical regression models. Examples of networks include simple back-propagation ANN (Shi and Wang, 2009, Gonzalez and Zamarreno, 2005), multiple perceptron architecture (Kwok et al., 2011, Kusiak et al., 2010, Kreider and Wang, 1992), general regression architecture (Ben-Nakhi and Mahmoud, 2004), hybridisation with Rough Sets (Hou et al., 2006), real-time adaptive ANN with dynamic structure (Rivard et al., 2005), recurrent ANN emphasising on load dependencies on time (Ahmed, 1999) and ANN with global solutions (Yokoyama et al., 2009). The methods display a range of differences in acquiring, pre-processing, training, weighting and post processing the data. However, most ANN need access to an extensive historical data archive. Another common trait is the specificity of such models to the training data.

Techniques based on SVM are sometimes preferred to ANN in load forecasting, due to their advantages discussed in section 2.2.2. The parameterisation of the SVM model impacts its accuracy and may be done via particle swarm optimisation algorithms (Xuemei et al., 2010b, Li et al., 2010a) or data clustering (Cherkassky et al., 2011). In their papers, Fernandez et al. (Fernandez et al., 2011, Penya et al., 2011) concluded that a multidimensional SVM model clearly outperformed an ANN model with a hidden layer of 10 neurons in terms of forecasting accuracy. In addition the ANN design and parameterisation poses a greater challenge. The improved 33 performance of SVM over ANN in predicting commercial loads has been also confirmed in other studies (Li et al., 2009a, Li et al., 2009b).

For STLF, many of the papers reviewed in sections 2.2.4.2 and 2.2.4.3 report low errors and superior prediction accuracy of machine learning methods compared to statistical regression methods. Nevertheless, it shall be noted that for both statistical and machine learning forecasting, the training is implemented using historical data from the case study building. This gives rise to two issues: the ability to adapt to dynamic changes in the building is limited, and it is not guaranteed that the algorithm can provide consistent results if applied to a different building.

2.2.4.4 Physical model load forecasting The shortcomings mentioned above (specificity of the algorithms to particular sets of data and inability to model dynamic and abrupt changes), can be overcome with the application of physical models in load forecasting. Physical models are commonly part of a broader energy management system in commercial buildings. Powerful simulation software suites that are based on physical methods include Energy Plus, DOE-2 and TRNSYS (Sun et al., 2013a, Chou and Chang, 1993), but lightweight computer modelling algorithms have been proposed as well, such as by Karmacharya et al. (Karmacharya et al., 2012).

Thanks to their flexibility and modularity, such simulation models are used to examine commercial buildings’ energy profile from various perspectives and to different degrees of detail and develop load forecasts. The models are of interest to researchers, building designers and energy managers and are able to handle archived and real-time load and weather data in characterising a building and/or developing load predictions (Cloudt et al., 2013). Prediction accuracies using energy simulations have been reported to be as high as 99% for the prediction of HVAC loads (Li et al., 2013, Raftery et al., 2011); however, there are also limitations for instance when predicting solar heat gains effects on load (Loutzenhiser et al., 2007). Nevertheless, in a study by Chua and Chou the simulations managed to accurately evaluate the dependence of HVAC load to certain weather parameters (Chua and Chou, 2011). Energy simulation models have been also successful in generating accurate predictions of load in mixed purpose commercial buildings (Gao et al., 2013). Another successful application that is useful for implementing forecasts for 34 optimised controllers was the prediction of the effects of using blinds for shading via building energy models on building load (Kotey et al., 2009, Daum and Morel, 2010).

The parameterisation of buildings or zones within buildings in these models relies on the equations governing energy transfers via conduction, convection and radiation as well as the rates of energy transfer and the geometry of the building (Fraisse et al., 2002). Once the model is setup, it can receive ambient weather inputs and calculate the energy response of the building for a given horizon based on the aforementioned equations.

A common modelling technique in the literature for load forecasting is based on using an electrical circuitry analogy to analyse the thermal behaviour of different zones within a building. In the example of Figure 11, temperature differences are equivalent to voltage differences, resistors and capacitors represent the thermal resistance and thermal storage capacity of the materials of the building and heat sources are modelled as current sources.

Figure 11: A 3R2C building thermal network model used for analysing the thermal response and load forecasting for a case study building in the US (Lee and Braun, 2007) Typically, each wall (external and internal), as well as the ceiling and floor are represented by a set of three resistors and two capacitors, while glazed surfaces (eg. windows) are represented by a single resistor (Lee and Braun, 2007, Luo and Ariyur, 2010). The American Society of Heating, Refrigerating and Air Conditioning (ASHRAE) method uses a set of transfer functions parameterised appropriately and

35 based on such thermal networks to convert energy gains into building loads and has been since considered as the starting point in various other research papers for the prediction of commercial building load (Al-Rabghi and Al-Johani, 1997, Chen and Yu, 2009). Rabl and Norford (Rabl and Norford, 1991) developed an algorithm using similar physical state equations to describe the thermal behaviour and load of office buildings, while more recent research (Braun et al., 2001, Braun and Chaturvedi, 2002, Lee and Braun, 2007) achieved the same goal through training the model with onsite weather data or external weather data from third parties (Westphal and Lamberts, 2004). However, non-parametric state models are also viable (Pakanen and Karjalainen, 2009).

Variations in heat transfers, potentially caused by changes in occupant numbers or stochastic patterns, like leaving doors or windows open, have a significant effect in the HVAC load. An advantage of physical load forecasting models is that they can easily simulate energy transfers between zones within the same building. This allows the dynamic parameterisation of the connectivity between different zones and more accurate assessment of the load (Luo and Ariyur, 2010). The thermal state at different zones within a building can be modelled as a function of weather inputs, internal heat gains and structural parameters, which in turn allows the forecasting of its future values and thus the energy demand (Oldewurtel et al., 2012). Similar techniques, where the changes in the state and demand of a zone within a commercial building are estimated via thermal networks (Lehmann et al., 2013) or thermodynamical and fluid mechanics equations (Wen and Smith, 2007) have been proposed. A different approach is energy auditing and performance modelling of individual air conditioning units in different locations of the same building, the aggregation of which allows for the prediction of the total load (Escriva-Escriva et al., 2010).

2.2.4.5 Hybrid models Hybridisation of forecasting methods is not uncommon in the literature. By combining principles from various techniques, more accurate load forecasts may be produced. The reason for this is twofold. Firstly, the ideal type and amount of data is not always available, and hence many modelling designs may have to receive inputs from other models instead. Furthermore, each technique performs better 36 when forecasting loads under different circumstances, and hence by combining they may complement each other.

For instance, Xuemei et al. report that the errors of an ARIMA algorithm were reduced by roughly 50% by post-processing the outputs with a SVM model in predicting the cooling loads of a commercial building (Xuemei et al., 2010a). The effects on errors and the improvement with the hybrid forecasting model can be seen in Figure 12 for the case study building in that study.

Figure 12: Comparison of an AR, machine learning and hybrid model in terms of errors in load forecasting for a commercial building in China (Xuemei et al., 2010a)

Similar results were achieved in an effort to hybridise an SVM and a genetic algorithm model as the load forecasting error was reduced by almost 15% (Li et al., 2010b). A hybrid model with LR and ARX has also been implemented using indices and succeeded in minimising the size and complexity of the dataset, with similar results in improving accuracy compared to the individual models (Kyungtae et al., 2012). Gould et al. (Gould et al., 2008) expanded the established approaches of forecasting via ARIMA or EW by combining them with a multiple seasonality model with the extra capacity to capture seasonal cycles in the time series. The algorithm was tested on utility data, however it could potentially apply to commercial loads as well. ANN models have been used for HVAC load forecasting in combination with ARIMA and LR models after being weighed accordingly using an hierarchical process (Yao et al., 2004). In a number of comparative studies the load prediction accuracy of these ANN hybrid models is notably higher than statistical methods (Kawashima et al., 1995, Fernandez et al., 2011, Penya et al., 2011).

2.2.4.6 Peak load predictions So far, the methods described in this section were predominantly focused on demand predictions. However, predicting the peak load occurrence in buildings is often equally (or even more) significant for various reasons. Peaks in the energy

37 demand of commercial buildings, even when short in duration, impose significant costs to the building energy bills (Sun et al., 2010a). In a typical commercial or services building, peak loads tend to occur as a result of increased HVAC system loads (Steinfeld et al., 2011, Lazos et al., 2014, Biyik et al., 2014). Predictions of demand and detection of peak loads are inherently dynamic processes, as there are multiple variables that affect the evolution of the building load in an intra-day basis (Wang et al., 2014, Taylor, 2012). While certain factors of the HVAC load, such as the occupancy levels in the building, are relatively straightforward to predict, others – for instance, the weather conditions – are not.

It has been established in the literature that there is a direct correlation between predictions of peak demand in air-conditioned buildings and ambient temperature inputs (Lam, 1999, Pedersen et al., 2008, Fernandez et al., 2011, Lazos et al., 2015, Massana et al., 2015). This is an easily understood relation, as the hotter (or colder) the temperature, the higher the demand for cooling (or heating respectively). Other weather variables that are considered in peak load predictions are relative humidity, as it affects the latent heat transfers, and solar radiation, as it affects the sensible heat transfers between the building and its environment (Zhou et al., 2008, Li and Lam, 2001). The dependence of peak building loads for various cases in Sydney can be seen in Figure 13.

Figure 13: Correlation of daily peak load and average temperature for gas and electrical heating in various commercial buildings in Sydney (Steinfeld et al., 2011)

Peak load predictions are not only significant for the energy management in individual buildings, but for the electrical utility providers and grid operators as well. The capacity of generation and hence the commitment of generation units at any given time is closely tied to the aggregate peak loads of the end-users (Grant et al., 2014). Regional and to an extent national economies can benefit from accurate peak demand forecasting, as energy, carbon emissions and monetary waste may be minimised and technical failures may be averted (Mtembo et al., 2014, Hoffman, 1998). Machine learning models are particularly useful in short term peak load predictions at the utility level, as data is readily available and most often there is an abundance of historical data that can be used for learning and validation purposes. In addition, peak load patterns tend to be more obvious and easily recognisable at the utility level compared to individual building peak loads, hence machine learning models often are reported to outperform other prediction models (Grant et al., 2014). As with individual building demand predictions, weather inputs are a major part of forecasting models at this level too, as electricity demand and spot pricing

39 depends on the present and future values of weather variables (Zavala et al., 2010, Zavala et al., 2009, Huurman et al., 2012).

At both the building and utility levels, peak load predictions are undoubtedly important and associated with multiple benefits in energy management and sustainability. As discussed in this section, peak prediction models benefit from the inclusion of weather-related inputs, as peaks and response measures are linked to variables such as the ambient temperature, relative humidity and radiation. Hence, weather predictions are regarded as especially important for building systems capable of dynamic control, in order to develop short-term peak predictions of less than 24-hour horizons. While there are meteorological monitoring installations in several commercial buildings, most often they serve only archiving purposes; weather forecasts are either obtained from external meteorological entities or developed via non-numerical models. This thesis aims to approach the challenge of peak load forecasting via the generation of numerical forecasts in ensemble.

2.2.4.7 Forecasting with Degree Hours & Days Regardless of the method, predicting the cooling and heating requirements of a building is a remarkably complex task, as the load depends on a range of factors in addition to the weather, such as the internal heating gains, the building envelope and insulation types, and the acceptable range of thermal comfort of the occupants. There is a variety of models and assumptions in the literature about the estimation and forecasting of these factors, however the focus of this thesis is predominantly on the effects and value of weather forecasting in managing the building’s energy demand. As such, this section will introduce and discuss the applicability of the statistical concepts of degree-hours (DH) and degree-days (DD) in load forecasting. In addition to reviewing the relevant literature, detailed explanation of the terminology and calculations will be provided.

The DH are calculated the amount of hours that the ambient temperature (Ti) exceeds or is below a certain threshold multiplied by the magnitude of the difference (Day, 2006). The calculation of DH can be conducted at different horizons, such as daily, monthly or annually. For determining cooling loads, cooling degree hours (CDH) are used, while for determining heating loads, heating degree hours (HDH) are used instead (Letherman and Al-Azawi, 1986). 40

Cooling in buildings is necessary when the temperature is too high, and hence the CDH can be calculated as the number of degree hours above a base temperature

(TBC). The period in consideration is treated as a time series, and each time interval (i) is hourly. Assuming a period of n time steps, the CDH would be calculated as:

푛

퐶퐷퐻 = ∑(푇푖 − 푇퐵퐶) (1) 푖

Accordingly, heating is necessary when the temperature is too low, and hence the HDH are calculated as the number of degree hours below a heating base temperature (TBH) as follows:

푛

퐻퐷퐻 = ∑(푇퐵퐻 − 푇푖) (2) 푖

The base temperature is set according to the requirements of the model, but generally speaking it refers to the upper and lower temperatures of the comfort range zone. The sums include only time steps where the difference is positive in both cases. In order to estimate the actual energy requirements from DH, the heat transfer in the building needs to be considered. Letherman & Al-Azawi suggested that the heat transfer factor (L) to and from the building is dependent on the ventilation rate, the glazing percentage and the construction materials of the envelope (Letherman and Al-Azawi, 1986). Figure 14 shows an indicative list of heat transfer factors (L) for energy calculations from DH, which are derived as the product of the heat transfer factor and the CDH/HDH multiplied by 3,600 (to convert to seconds).

Figure 14: Loss or gain of heat in W/K as a function of ventilation rate (in AC/h) and glazing percentage (Letherman and Al-Azawi, 1986)

The number of occupants in a building is an additional factor that may be considered in the estimation of the heat loss factor. The higher the number of occupants, the lower the L for a constant ventilation rate within a building (Durmayaz et al., 2000)

If the sums of equations 1 and 2 are divided by 24, then the cooling degree-days (CDD) and heating degree-days (HDD) are obtained respectively. Typically, these metrics are useful for modelling of the building’s annual energy profile. When hourly data are not available, ASHRAE suggests that the DD are calculated as the differences of base temperature and the average of daily minimum and maximum temperatures instead (ASHRAE, 2009). However, the advantage of using hourly data for the calculation of DD has been accentuated in the literature in order to improve accuracy (De Rosa et al., 2014, Day and Karayiannis, 1998, Waide and Norton, 1995). The algorithms for predictions in this thesis use hourly weather data for any module involving DD calculations.

The advantage of modelling loads with DH or DD compared to other regression or physical models is the relative simplicity and direct correlation of the building’s load to the weather conditions (Krese et al., 2011, Layberry, 2008, Krüger et al., 2010). However, there are noteworthy limitations to DH or DD models, as in the base forms they only take into account the ambient temperature into consideration. The result of that is that the calculations are useful to estimate the sensible heat loads, but not the latent heat loads, which can be significant for cooling demand in the summer (see section 2.2.4.1). Relative humidity, wind speed and solar radiation all affect the demand and hence they need to be accounted for to improve the accuracy of these models (Krese et al., 2011).

To address the issue of lack of inclusion of the effects of humidity, an alternative to DH and DD was proposed: the concept of enthalpy latent hours (ELH) and enthalpy latent days (ELD) (Huang et al., 1987). The calculations for ELH are modified to the summation of differences of outdoor enthalpy (hi) at a time step (i) and the base enthalpy (hB) at a specific temperature and humidity:

푛

퐸퐿퐻 = ∑(ℎ푖 − ℎ퐵) (3) 푖

While for heating loads, the HDD/HDH calculations are fairly accurate, cooling load estimation improved significantly with the use of ELD when tested with real data in case study buildings (Huang et al., 1987). The ELH method for calculating cooling loads has been applied and showed improvements over CDH calculations in two case study buildings in Slovenia (Krese et al., 2011).

When long-term analysis of the energy needs of a building is required, DD modelling is preferred to DH modelling, as the values are smaller and easier to understand. Furthermore DD modelling is popular in studies analysing the climatic effects on building energy profile across countries or globally. A variety of studies has used this technique to categorise building performance across regions (De Rosa et al., 2014, Pusat and Ekmekci, 2016). Finally, such models may be also used to help with design and optimisation of building features, like insulation (Bolatturk, 2008) or preconditioning (Zhou et al., 2011). In this thesis, DH and DD modelling will be a central theme for the algorithms described in chapter 5.

2.2.4.8 Load forecasting summary Regardless of the approach, the scope or the horizon, load forecasting are valuable inputs for manual and automated decision making in any energy management system. Physical models are based on established physical laws and equations, which are universal; thus the main hurdle for forecasting accuracy is posed by resource limitations when modelling the system. Hybridisation or statistical post- processing could further enhance their accuracy, but this raises the complexity and computational costs.

As discussed earlier, locally generated weather inputs for load forecasts when available, such as temperature or solar heat gains, result in improved performance. Consequently, it is suggested that since all approaches for STLF demonstrate comparable errors, typically ranging between 1 and 10% (Li et al., 2009b, Zhao and Magoules, 2012, Foucquier et al., 2013) and possess unique advantages and disadvantages, the defining factor for utilisation by a commercial building should be the ability to integrate weather and load forecasts within an optimisation framework. For studies concerned with the integration of weather in building energy management, like the current thesis, the concepts of DH and DD are of particular interest and utilised broadly. 43

2.2.5 Generation forecasting The penetration of DG, such as solar panels, wind turbines, cogeneration, fuel cells or other types of batteries in commercial buildings introduces a new challenge in forecasting and energy management. In the case of cogeneration and batteries, the generation can be adjusted at will and therefore optimisation algorithms are mainly concentrating on the forecasts of energy costs. However, with intermittent energy sources such as solar radiation or wind power, forecasting the availability of energy and matching it with the demand constitutes a major aspect of energy management systems.

Solar based generation (PV or solar thermal) is by far the main topic in the field of energy generation forecasting for buildings. In terms of solar power generation, which is the renewable energy form with the highest penetration in commercial buildings, there is a variety of forecasting software applications that can provide accurate estimates for generation over certain periods of time. Based on models that take into account solar radiation data, system specs and efficiencies, these methods are very effective in predictions of the system’s overall performance in long-term horizons. However, they fail to capture the solar generation in real time or short horizons, thus they have limited uses in dynamic energy management systems (Tamizh-Mani et al., 2008).

Forecasts of real-time and short term power output of photovoltaic (PV) systems are much more valuable than long-term performance predictions for commercial building energy management. Recurrent cloud formation patterns have been recognised as being notoriously difficult to predict and since the power output of PV system is directly proportional to the incident sunlight, this poses a challenge to forecasting attempts. Such a stochastic ARIMA model for cloud coverage was partially successful (Chowdhury and Rahman, 1987).

For generation forecasts, time series and physical models are often combined to increase prediction accuracy. Typically, weather inputs from a NWP model are known to improve the accuracy of generation forecasts. For instance, an ARX model with NWP temperature inputs for online solar PV generation prediction was proposed by Bacher et al. (Bacher et al., 2009) and demonstrated significantly

44 improved accuracies by 35% compared to the persistence model. The model was expanded to incorporate NWP inputs for online solar thermal generation predictions with similarly high accuracy (Bacher et al., 2011). For both cases, it was reported that NWP were most useful in horizons longer than 4 hours up to several days ahead, while the AR part was sufficient for short term generation forecasting. This is in alignment with the findings from the weather variable prediction section (2.2.3). Another statistical LR method for long horizon PV power generation has been carried out and utilised ambient temperature and solar radiation as exogenous inputs. The mean monthly errors reported were as low as 5% (Huang et al., 2011). Compared to a simple time series regression forecasting model, ARX can reportedly improve the prediction accuracy by 13% (Yang and Xie, 2012).

A similar trend can be observed with hybrid generation forecasting models between ANN and NWP. For instance, a range of NWP outputs were used as inputs to an ANN for the real-time prediction of PV solar power and produced average forecast accuracies of 90% (Chen et al., 2011) Another ANN approach to real-time power generation forecasting has been implemented with weather data as the inputs and it was shown that it could generate forecasts of up to 30 minutes ahead with accuracy within the 95% confidence intervals (Chow et al., 2012). However, it should be noted that ANN of different architectures with no weather inputs at all have been designed and used for power generation forecasting as well (Al-Messabi et al., 2012).

Instead of using NWP outputs, third party weather observations may be obtained and used for the prediction of solar generation via ARX frameworks. In a comparative study, it was shown that this method produced 15% higher accuracy than a non-adaptive ANN (Cai et al., 2010).

Additionally, solar radiation predictions may be utilised in solar energy generation forecasts in an intra-day basis; solar energy plays a significant role in DR measures that aim to reduce peak loads, as typically high solar energy generation coincides with high peaks in the middle of the day (Chen et al., 2011).

Information about the future availability of energy from DG sources augments the value of the energy generated, since it can assist with peak shaving, demand

45 matching and DR implementations. Hence, the savings potential from generation in buildings increases with accurate generation forecasts. The main variables associated with solar generation (solar radiation) and wind generation (wind speed) are not primarily favoured in load forecasts as a consequence of the limited computational resources and complexity of such predictions. Additionally, since not all buildings are designed with DG capacity, the integration of such forecasts in existing building management systems may be too mostly irrelevant. Figure 15 summarises the main considerations for a model able to deliver onsite generation forecasts for a building and their research/practical applications.

Figure 15: Summary of features and applications of generation forecasting models in existing literature, with a clear distinction between renewable energy (RE) and non-renewable energy sources 2.3 Building energy management systems The forecasting section discussed the various dimensions involved in predictions related to the energy management of buildings. The research in this thesis is focusing on applications of highly accurate, localised numerical forecasts in commercial energy management systems, which compared to residential buildings, have certain features that are arguably favoured in the field of forecasting and optimisation. A summary of these features can be seen in Figure 16.

Figure 16: Comparison of features of commercial and residential energy management Energy management in a commercial building involves generating energy onsite, managing the demand via energy efficient designs, upgrades and response (DR) policies; often a combination of several of these modules is present. These systems are subject to high degrees of optimisation in terms of energy savings as the value of energy is variable and affected by a range of factors, such as the time of day, season and the energy source (Foucquier et al., 2013). Ideally, during periods of high energy cost, consumption should be minimised and generation, where available, should be maximised. Thermal or electrical storage are frequently being used as energy buffers and regulators of demand during the day. The key to optimisation algorithms in building energy management is analysing the system’s behaviour and predicting its future state on a rolling horizon of some minutes up to several days ahead, so that generation and demand can be matched and produce the maximum potential savings (Sun et al., 2013b).

Towards that goal and based on forecasting information and data about the past and current states of weather variables, load or generation, Building Energy Management Systems (BEMS) can make decisions, such as adjusting the set points 47 for HVAC, regulating the flow and consumption of energy within the building or managing the exchange of energy with the grid. The decisions of a BEMS can be either short term, which usually aim to reduce imminent demand peaks, or long term, which usually aim to minimise overall energy costs (Perez-Lombard et al., 2008).

A major role of a BEMS is controlling certain internal conditions (temperature, relative humidity, ventilation rate) and ensuring they are kept within a predefined range of values. Depending on the ambient weather and environmental conditions, significant amounts of energy may be spent towards that goal via the HVAC system. The purpose of keeping the values within a predefined range is to ensure that the thermal comfort of the occupants is within an acceptable range. The thermal comfort of an individual depends on various factors, predominantly the air temperature, relative humidity and radiative heat exchange, however there are other non- weather related factors such as the ventilation rate, metabolism, clothing, and type of activity (Havenith et al., 2002).

This thesis will develop a number novel modules for BEMS involved in forecasting and decision making for both short and long term horizons. In this section of the literature review, examples of BEMS applications will be discussed and it will be possible to understand the connection between forecasting and decision making processes as well as optimisation of energy management systems in terms of savings and thermal comfort. Emphasis will be placed on the ability of BEMS to manage HVAC loads, which are the types of building loads predominantly associated with the content of this thesis. However, it should be noted that most BEMS control and optimise various other energy systems within a building, such as lighting, telecommunications and safety systems.

2.3.1 Model Predictive Control systems Sets of rules in the form of controllers regulate the parameters of operation of individual BEMS and can integrate inputs from various sources in decision making. These sources may be controlled by the BEMS itself (sensors within the building, onsite weather monitoring equipment, software models) or just provide data externally (weather stations, electricity grid, urban authorities). Unlike Rule Based

Control, where a set of rules govern the behaviour of the BEMS deterministically, Model Predictive Control (MPC) is a framework that seeks an optimal solution to an objective function of load bound to certain constraints (tenant comfort, system utility, energy costs and building characteristics). Most systems with MPC, are able to utilise forecasts related to the energy system, the capacity of thermal or electrical storage and the interdependencies on weather or occupancy patterns (Oldewurtel et al., 2012). This thesis assumes that MPC capacity is present for the application of the proposed techniques.

Decision making for MPC systems is dynamic and capable of developing short-term responses to changes in the internal or external conditions. While often computationally and operationally intensive, such strategies can minimise the energy costs without compromising the comfort of the occupants as a response to external stimuli. This is mainly achieved by exploiting the whole range of the thermal comfort zone throughout the day. Typical predictive control involves optimisation based on energy pricing, energy generation and exchange with the grid, as well as building load (Kang et al., 2014). External stimuli, such as temperature or solar heat gains are realised in the form of disturbances and integrated in the algorithm in a real-time manner. MPC can be adjusted to optimise the system flows on varying horizons depending on the energy system capacity and structure of the building (Krarti et al., 1999). Sensors may in certain cases improve savings further, when placed at critical points within the building (Kang and Park, 2013).

One way to deal with prediction inaccuracies is to use auto-regressive models and the assumption that noise follows a normal (Gaussian) distribution to estimate errors at a future point in the time series (Oldewurtel et al., 2008, Oldewurtel et al., 2012). Predictors using an unbiased Gaussian noise assumption with time dependent variance outperformed conventional strategies for estimating the thermal state of a building by up to 18% and were unaffected by forecasting errors (Henze et al., 2005, Henze and Krarti, 1999). Regardless of the technique, MPC demonstrates tolerance to the inherent inaccuracies of the weather and demand forecasts that provide inputs (Henze and Krarti, 1999). A flowchart demonstrating a typical MPC architecture as described in (Oldewurtel et al., 2012) can be seen in 49

Figure 17. In this architecture, weather predictions complement the feedback loop of demand monitoring in the attempts to optimise the energy system.

Figure 17: Example of a typical MPC system with weather forecasting inputs (Oldewurtel et al., 2010) 2.3.2 Weather forecasting in BEM Weather forecasting for MPC mainly incorporates predictions of temperature and sometimes humidity and solar radiation (Cooperman et al., 2010). As discussed previously, the literature indicates the value of using weather forecasts for load predictions (especially locally generated). This section will outline the importance of weather forecasting in MPC and evaluate its contribution towards energy system optimisation in both energy (and cost) savings as well as occupancy comfort. Occupancy comfort is actually regarded as an equally important objective to energy savings in the optimisation literature. As seen in Figure 18 there is often higher margin for improvement and optimisation in terms of occupancy comfort compared to energy savings. An optimal MPC system with perfect weather forecasting would achieve the lowest energy costs at the lowest possible discomfort cost. Most conventional controllers though are operating without weather inputs and can only achieve low occupant discomfort at relatively high energy costs (Kummert et al., 2000).

Figure 18: Correlation of discomfort cost (Jd) and energy cost (Je) for a range of controllers (Kummert et al., 2000). Superior optimisation results in terms of both savings and occupancy comfort have been observed when weather forecasts have been used as inputs for MPC, compared to the systems without weather integration in several studies (Chen and Athienitis, 1996, Ferrano and Wong, 1990, Candanedo et al., 2013b, Zhou et al., 2008, Oldewurtel et al., 2012, Aswani et al., 2012, Cigler and Privara, 2010). Zavala et al. (Zavala et al., 2010, Zavala et al., 2009) proposed a dynamic real-time optimisation framework that incorporates weather forecasts of different horizons and concluded that savings up to 30% can be achieved for one day ahead horizons in a large case study building compared to the base scenario of reactive energy management (without weather forecasting). In addition, the weather component resulted into alleviating one of the main limitations of optimisation routines, namely the inability to utilise existing trends in the time series of demand. It can be seen in Figure 19 that the statistical forecasts without any weather inputs resulted in higher error and as a result the BEMS decided to set the HVAC temperatures outside the comfort zone.

Figure 19: Comparison of 5-day ahead operating HVAC strategies, with predicted interior temperatures (grey), actual observed temperatures (blue) and the comfort zone (between the black lines) (Zavala et al., 2010) On the other hand, weather predictions helped maximise the savings by matching predicted and actual temperatures much more closely and correctly adjust HVAC set temperatures within the comfort zone. Similarly, dynamic optimisation with weather inputs are considered and have been validated via mixed integer linear programing (Collazos et al., 2009) or ANN controllers (Argiriou et al., 2004, Argiriou et al., 2000, Ruano et al., 2006) with reported energy savings in the vicinity of 25- 30%.

Weather inputs are often post-processed with occupancy data to improve the potential for optimisation of a BEMS. Significant savings can be realised in systems where weather and occupancy patterns are correlated, (Dong et al., 2011). Using a numerical analysis it has been concluded that savings of up to 50% may be achieved with real-time optimisation systems in place receiving weather sensitive and occupancy inputs (Zavala, 2013).

Undeniably, many sources in the literature indicate the importance of weather forecasting in optimisation and management of a BEMS. This thesis is proposing a 52 number of modules related to the optimisation of building conditioning based on numerical weather outputs that the literature has not yet investigated for both the forecasting and management components of a BEMS. These modules include the prediction of abrupt changes, ensemble forecasting and predictions of extreme weather events.

2.3.3 Building conditioning A considerably large part of BEMS control systems is designed for analysing outputs from physical models in order to manage internal conditions (conditioning the building). These models are able to account for a range of factors, such as the building structural composition, the building geometry, and the interactions with its surroundings. Often, conditioning occurs proactively: the building internal conditions are set at certain levels at specific periods in advance, in order to achieve desirable conditions a few hours later (typically during the occupied peak consumption period). However, dynamic conditioning is significant in any BEMS and is constantly modified in response to external or internal changes. Conditioning can also be active or passive: active refers to responses that consume energy to change the internal conditions (such as using the HVAC system to affect internal temperature), while passive refers to conditioning that occurs due to the interactions of the building envelope and the environment and the energy stored in its thermal mass. In both situations, weather conditions play an important role in optimising the response and developing an appropriate strategy for maximising savings and occupancy comfort.

This section will start by reviewing the research related to dynamic conditioning in buildings and the importance of weather inputs in it.

2.3.3.1 Dynamic conditioning It has been demonstrated in several studies about control and optimisation of BEMS that dynamically conditioning a building is an effective means of generating energy savings and ensuring occupant comfort stays within acceptable levels throughout the day (Henze et al., 2005, Oldewurtel et al., 2008, Braun, 1990, Nagai et al., 2002, Hajiah and Krarti, 2012, Ruud et al., 1990, Xu, 2004, Lee and Braun, 2004, Morgan and Moncef, 2010).

While the amount of inputs, BEMS architecture and complexity of layers used to model the building and its zones vary, weather related variables such as the temperature and relative humidity are vital to ensure efficient management of the load and occupancy comfort (Gunay et al., 2014, Oldewurtel et al., 2012). MPC has the capacity to utilise the past values and forecasts of weather variables and assist with dynamic conditioning.

Dynamic conditioning mainly involves adjusting of HVAC set points according to the external and internal conditions and building zone dynamics (Cooperman et al., 2010, Lehmann et al., 2013, Li et al., 2013, Oldewurtel et al., 2012, Ferreira et al., 2012b). Most of these methods attempt to operate the HVAC system at the point of maximum efficiency (lowest energy costs) for a certain horizon and within well- defined boundaries of thermal comfort. This allows the system to develop response measures to any external changes and reduce peak energy consumption and hence minimise energy costs (Luo et al., 2010). Dynamic conditioning via the adjustment of HVAC set points has been a known technique for decades (Ferguson and Winn, 1989).

Apart from adjusting the HVAC set temperatures, mixed-mode strategies can make use of accurate weather information in an attempt to dynamically condition a building. In certain cases with the appropriate infrastructure, such as automatic control of shading devices, simulations of a building in a warm temperate climate during periods of 6 summer days have suggested that the energy consumption may be reduced by up to 80% with a MPC implementation compared to the manual control scenario (Hu and Karava, 2014). This strategy included a mix of both active and passive cooling, ventilation and shading plan. Additionally, the comfort of occupants was not violated at any time during the simulation period. Moreover, a single zone MPC has been proposed using the outputs from a NN (Ferreira et al., 2012b). A design able to generate a real-time model of the case study building’s thermal performance and then automate the BMS optimisation processes has been developed (Hagras et al., 2008). In the controller discussed in another study solar heat gains were modelled thoroughly for the energy management of a passive solar commercial building (Kummert et al., 2000).

Certain studies are also concerned with the optimisation of individual components of the HVAC load, rather than developing a set point tracking plan. In a consolidated study (Vakiloroaya et al., 2011) a variety of strategies of operation were tested for each component of the cooling system in order to realise the optimal mode. Sun et al. (Sun et al., 2010b) proposed an optimisation algorithm to determine the most effective schedule in starting up chillers in case study office buildings.

2.3.3.2 Preconditioning Preconditioning refers to the process of adjusting the internal building conditions in advance, so that during the following time periods there is reduced energy consumption by the HVAC system and hence lower costs (Rabl and Norford, 1991). Temperature predictions of appropriate horizons are always required to assess the preconditioning potential savings and plan the control strategies. Depending on the building model, location and characteristics, savings up to 40% have been realised in many case study buildings around the world with preconditioning (May- Ostendorp et al., 2011, Karava et al., 2012). Preconditioning processes may be realised in various ways, but generally can be classified as either passive or active.

Passive preconditioning (ventilation via open windows, ducts, and utilisation of energy of the building’s thermal mass) takes advantage of the naturally occurring changes in the external conditions and the building design to condition the interior. The notion of taking advantage of diurnal temperature differences without significant energy expenditure, can result in “low cost” savings depending on the infrastructure in place. The opportunity for applying such measures has been demonstrated with studies that indicate night ventilation results in annual electrical cost reductions of up to 17% (Braun and Zhong, 2005, Peterson and Hunn, 1985). In addition to energy consumption reductions, in some instances passive preconditioning allows the replacement of the building air mass with fresher air from outside, improving the occupancy comfort (Chenvidyakarn and Woods, 2005). Another advantage is that there is a net energy demand decrease, compared to active methods where the total energy required to “shift” some of the load to prior time zones is higher than the energy saved (Cole et al., 2014).

Active preconditioning with the HVAC system operating fully or partially, uses cheaper off-peak energy in advance to minimise peak energy demand during the 55 day. Generally, these strategies are more reliable and consistent than passive ones as they can be applied regardless of the external temperature differences and can generate more significant peak load reductions. The peak load shaving is achieved by using the HVAC system overnight to precondition the building and hence reduce the energy needed for conditioning the day after (Roth et al., 2009). In some cases the peak shaving is so effective (up to 75% reduction) that these strategies may be a viable alternative to upgrading components of the HVAC (Keeney and Braun, 1997). Nevertheless, the external weather conditions impact the cost of the implementation and hence weather forecasting is necessary for optimisation (Keeney and Braun, 1997).

Lee & Braun (Lee and Braun, 2007) studied two active preconditioning strategies and compared them to the typical deterministic night setup. It can be seen in Figure 20 that while the daily load overall is not significantly affected, the peak load is reduced thus leading to cost savings without compromising the comfort of the occupants. The effectiveness of such strategies assumes that electricity price information is available and integrated in the MPC (Greensfeldera et al., 2011).

Figure 20: Comparison of the effects of a non-predictive approach with preconditioning strategies on interior temperature and daily cooling loads (Lee and Braun, 2007) Preconditioning planning involves the prediction of the internal temperature of a building zone as a function of the ambient temperature and the building features. The internal temperature must remain within certain comfort boundaries throughout the occupied period. Depending on the construction, the thermal mass of the building may significantly alter the effects of ambient weather on the indoor conditions. Envelopes constructed from heavy materials not only attenuate the changes in ambient temperature, but also add a notable temporal delay to the changes that follow in internal conditions of temperature, which is taken into account during preconditioning planning. On the contrary, lightweight constructions result in internal temperature ranges that are comparable to the ambient temperature and occur with very little lag (Lockerbie, 2016). Regardless, both construction types may benefit from appropriately designed preconditioning strategies that match the local climate (Peng Xu, 2009).

It has been shown that due to the effects of the thermal mass of the building, small forecasting errors in ambient temperature (below 2-3°C) tend to have relatively

57 minor effects in optimisation of preconditioning and the thermal comfort (reductions of up to 20% reported) (Hu and Karava, 2014). In essence, the indoor temperature display a lag –the magnitude of which depends on the building envelope – compared to the ambient temperature (Kruger and Fernandes, 2006, Spindler and Norford, 2009b).

The literature also suggests that the long-term climate at the location of the building under question is a major factor for decision making in preconditioning strategies (Braun and Zhong, 2005, Rasouli et al., 2013). The variations in preconditioning effectiveness are easy to pinpoint when comparing sites across completely different climates (Kintner-Meyer and Emery, 1995). Apart from the climate, geographical factors may affect the development of preconditioning policies. For instance, proximity to large masses of water typically diminishes diurnal differences in temperature due to the thermal mass of the ocean, and sea breezes. Hence passive preconditioning to cool down a building in the summer may not be possible. Other factors affecting local micro-climates and hence the preconditioning potential include vegetation coverage, landscape elevation, built environment density, type of urban zone and pollution (Todhunter and Terjung, 1988, Santamouris et al., 2001). Numerical modelling is particularly useful in such high spatial resolution analyses (Eichhorn et al., 1988). Of course the building characteristics are also vital when managing preconditioning strategies. Generally, heavy building envelopes result in superior peak load reductions than lighter construction buildings (Khaled and Krarti, 2007, Peng et al., 2009).

Most existing literature is concerned with the design of energy controllers and modelling the thermal response of the building to preconditioning rather than assessing the potential due to the local climate (Briller, 2012). This thesis aims to address this research gap by developing a weather forecasting and a climate assessment tool to assist with development of preconditioning strategies. In essence, the methodology of chapters 3, 4 and 5 aim to offer highly valuable localised numerical forecasts for use within models for the energy operations of a building.

2.3.3.3 Effects and modelling of thermal mass Building conditioning decisions require inputs of weather data, which are used to understand the effects of thermal mass. Depending on the construction, the thermal mass of the building may alter the effects of external weather on the conditions inside the building and hence the HVAC loads. Specifically, envelopes constructed from heavy materials (high thermal mass) attenuate the changes in ambient weather conditions, and introduce a temporal delay to the changes that follow in internal conditions. On the contrary, lightweight constructions (low thermal mass) result in internal conditions ranges that are comparable to the ambient ones and occur with very little lag. This phenomenon can be clearly seen in Figure 21.

Figure 21: Effects of thermal mass on internal temperature (Lockerbie, 2016) Models of thermal mass control are typically “white” or “grey” box: they attempt to emulate the building’s thermal response with a set of parameters that depend on the construction materials, zone orientation and geometry of the building. These parameters all affect the heat exchanges between the building and its environment, as well as flows of heat within the building itself. Sometimes referred to as “inverse models”, such techniques use weather and demand data to estimate the parameters of the building and utilise them for future predictions and optimisation (Braun et al., 2001). Given the same ambient weather forecasts, inclusion of the thermal mass in physical load forecasting always produces more accurate results compared to models that do not take it into account (Kossak and Stadler, 2015).

Considerable research is being conducted on developing thermal mass models for the building operations. Typically, these models are necessary for developing conditioning strategies: knowing the dynamic response of the building, the discharge rates and energy storage in the thermal mass as a function of the current 59 and past weather conditions allows one to modify the HVAC schedule and develop relevant responses to minimise energy costs (Henze et al., 2007). The most recent white box models function within multi-objective optimisation frameworks that take into account not only weather inputs, but also energy pricing and occupancy patterns (Li and Malkawi, 2016). Simulations for modelling the effects of thermal mass are typically conducted in specialised building software, such as EnergyPlus, Simulink, Matlab or TRNSYS (Li and Malkawi, 2016, Rackes and Waring, 2014, Ma and Li, 2010, Rempel et al., 2016). Dedicated control systems that respond to external changes after modelling their attenuation and lag due to the thermal mass have been proposed (Armstrong et al., 2006). This is not a surprise, as savings of up to 30% in peak costs and up to 50% in energy demand may be realised with dynamic thermal mass control via building conditioning strategies (Liu and Henze, 2007).

Another field of thermal mass research is concerned with the building design and construction optimisation. A high thermal mass construction is considered as an effective design in hot and arid climates, as diurnal temperature differences may be smoothened and cooling loads may be reduced. In a case study high thermal mass building in Israel it was found that the cooling requirements were dropped from 328 Cooling Degree-Days (CDD) to 28 CDD (Krüger et al., 2010). Further discussion of the CDD concept and its correlation to cooling loads may be found in section 2.2.4.8. The advantages of high thermal mass have also been recognised in alleviating overheating due to climate changes (Kendrick et al., 2012). Furthermore, studies of thermal mass effects in heating applications have been studied: in general heavy constructions that are well insulated are able to perform more efficiently in terms of energy demand during the winter (Rempel et al., 2013). However, thorough understanding of the local climate trends is necessary for thermal mass design and control (Rempel et al., 2016), which enhances the notion of the need for accurate localised weather forecasts and historical data.

It has been shown that due to the effects of the thermal mass of the building, forecasting errors in ambient weather conditions affect the optimisation of the building energy management and the thermal comfort (Hu and Karava, 2014). In essence, the interior temperature changes display attenuation and temporal lag compared to the ambient weather changes, and the magnitudes depend on the 60 thermal mass parameters (Kruger and Fernandes, 2006). Hence, the errors in interior zone conditions are tied to any errors in ambient weather predictions both in terms of magnitude and time (Spindler and Norford, 2009b). This observation necessitates the existence of accurate weather forecasts for optimising the building operation in thermal mass control strategies, as well as good quality long-term weather data for modelling the thermal mass parameters and building responses at different frequencies.

Modelling a building’s thermal mass and estimating the building’s thermal factors is a very well covered research field, and outside the scope of this thesis. However, the importance of accurate localised weather forecasts in control strategies cannot be overstated. Additionally, thermal mass design simulations require an understanding of the local climate, which again may be analysed using numerical methods. This is a clear research opportunity, that the methodology of this thesis (specifically the forecasting methodology in chapters 3, 4 and 5) aims to address. The advantage of the proposed methodology compared to existing studies is the development of localised weather forecasts and historical weather data analyses that can be easily tailored to any location.

2.3.3.4 Forecasting and effects of occupancy This section will outline and discuss the effects of occupancy patterns in commercial building energy demand and conditioning. Occupancy is not directly related to the weather conditions and hence is not in the primary scope of this thesis; regardless, it may affect demand and forecasting attempts significantly in certain cases. The reason for that is if future occupancy is known, the HVAC setpoints and other demand response strategies may be triggered to match the expected demand. For instance, if a very hot day is expected that happens to coincide with a number of significant conferences/meetings in a building (hence, extra occupants will be present and likely activity will increase), additional peak shaving measures may be necessary. Occupancy levels are generally associated with internal heat gains. Firstly, humans radiate energy at varying rates depending on their age, sex and physical characteristics. The radiation rates are modified by the metabolism and activity that occupants carry out. Additionally, the occupancy levels affect the extent of use of lighting, electronic and electrical devices, which in turn contribute to 61 internal heat gains. In fact occupancy affects “plug-loads” (the load from electrical/electronic devices plugged in a wall socket) much more significantly than total building load (including HVAC and supplementary systems) (Kim and Srebric, 2015)

In many ways, occupancy forecasting is similar to the weather forecasting: AR, ANN, SVM may be used in short-term horizons (Chen and Soh, 2016, Kwok et al., 2011) with the same advantages and limitations as the ones described in section 2.2.2.4. It is not uncommon in the literature for occupancy predictions to be considered in parallel with weather predictions in order to forecast future demand and develop responses (Oldewurtel et al., 2012, Siroky et al., 2011). However, unlike weather forecasts, occupancy is generally considered more predictable in medium and long term horizons: the number of occupants does not fluctuate wildly in many commercial buildings throughout the year.

Occupancy modelling and forecasts for use in HVAC control often require monitoring with sensors (commonly acoustic or infrared) in certain zones in the building. Depending on the building function and layout, the locations for the sensors can be strategically chosen so that the data can be generalised. Using sensor historical data it is possible to train machine learning algorithms and then use them in time-series occupancy models (usually Markov Chains) to predict future occupancy levels (Dobbs and Hencey, 2014, Adamopoulou et al., 2016, Howard and Hoff, 2013). Data from real observations are very significant compared to simulation data for occupancy predictions, as trends may change in specific contexts: for instance activities like meetings, workshops or celebration events may not be completely understood by simulations alone. Hence, context awareness in algorithms is proposed for usage of occupancy predictions in HVAC control

(Adamopoulou et al., 2016). Environmental data, such as interior CO2 levels have also been proposed as inputs for predicting occupancy levels (Ryu and Moon, 2016). When direct data for occupancy are not available, models have been proposed that take into account indirect observations, such as the amount of vehicles in the building’s parking with reasonable success in correlating to occupancy levels (correlation coefficient of 84%) (Oliveira-Lima et al., 2016). It is also suggested that zonal investigation of occupancy schedules is sufficient in demand control, as 62 variations in individual numbers of actual occupants present at their offices has little effect to building load (Shinkawa and Nobe, 2006).

The actual contribution of occupancy and its relative weight compared to weather- related factors depends on various factors, such as the climate, the type of activities and the type of building. While an estimate of the relative contribution of occupancy in generating internal heats and affecting HVAC demand is difficult to generalise outside of specific case studies, it is shown that inclusion of occupancy forecasts of any methodology in control are generally able to reduce HVAC demand by 10%-20% compared to conventional non-occupancy strategies (Dobbs and Hencey, 2014, Yang et al., 2016). The savings are mainly realised in avoiding over-conditioning, or cooling down/heating up spaces that are not used or will not be used in the near future (Erickson and Cerpa, 2010). This may be achieved via reducing or shutting down the HVAC for empty zones, or dynamically modifying the set points accordingly prior to and during periods that occupancy fluctuates.

This section outlined the role of occupancy levels and schedules in building conditioning: predictions of both the number of occupants and the timing they are active in certain zones is important for HVAC control. There are various time-series approaches in existing literature that produce highly useful forecasts of occupancy. Their implementation in HVAC control is always considered in parallel to weather predictions, which do not have a prominent effect in the occupancy levels.

2.3.4 Energy generation and storage management Management and optimisation of DG schedules and energy storage, where available, is equally important to demand management in buildings. This section will outline existing research trends and practices in this field, with emphasis on the importance of weather forecasting integration.

The management of generation is much easier when energy can be generated on demand, rather than intermittently, for instance via a cogeneration unit. Cogeneration systems able to produce both heat and electricity in commercial buildings, and optimisation processes regarding their scheduling have been proposed using mixed integer linear programming (Asano et al., 1994). A network flow model based on the physical operation modes of a cooling, heating and power

63 system (CHP), as well as dynamic states of the grid has been described and utilised for savings optimisation (Cho et al., 2009), while the same goal was reached with an even more simplified flow approach (Okamoto, 2010). Weather conditions responsible for transient trends in a CHP system have been considered and allowed the development of a predictive control algorithm (Yun et al., 2011).

When renewable energy sources are part of the DG mix, optimisation problems become more complex due to the erratic nature of power generation. The controller in addition to optimising load variables according to the predicted states of the system, has to be able to forecast the generation. PV generation, which is the most common source of renewable energy in buildings is strongly affected by weather parameters and especially solar radiation and ambient temperature. In buildings with PV capacity, MPC systems are able to forecast how much energy the PV system will generate, taking into account weather observations. In addition to weather data, real-time pricing of energy is typically used as input to optimise the energy flows of buildings with capacity to export electricity (Zong et al., 2012). Monitoring and optimisation of solar PV generation is dynamic and online and hence the importance of frequent weather forecast updates is highlighted (Ferhatbegovic et al., 2012). Simple controllers that disregard weather inputs can still improve the utilisation of solar PV and thermal systems as seen in a small building case study (Wahab et al., 2011).

For certain buildings, BEMS include modules for optimising electrical energy storage (batteries). Time-series statistical approaches have been developed, in which energy storage charge and discharge rates are governed by the feedback generated from previous responses of the system in terms of costs (Liu and Henze, 2007). Batteries, apart from being energy storage media, can be regarded as additional nodes of DG in a building, especially mobile ones that are not permanently connected to the building. Electric vehicle batteries connectivity to a commercial building has been modelled using a mixed integer linear program in order to predict and optimise the behaviour of the energy system (Stadler et al., 2013). In addition, the effects of MPC on buildings with both active and passive energy storage have been demonstrated (Henze et al., 2004). Typically, electrical storage capacity is not directly affected by the weather conditions. 64

Ice storage is an alternative energy storage method. The operating principle of ice storage system is cooling water down to form ice during low energy cost times and releasing its latent heat during peak times to provide cooling. As with battery storage, it is possible to minimise the costs with accurate load predictions. MPC algorithms with distinct emphasis on ice storage have been described (Ferrano and Wong, 1990, Candanedo et al., 2013a, Hajiah and Krarti, 2012, Kawashima et al., 1996). Typically, ice storage optimisation is conducted on a daily horizon, however it was shown that control based on calculations of the load and weather evolution over a weekly period are viable (Henze et al., 1998). More so than battery storage, ice storage is affected by the existing weather conditions.

2.4 Summary and research opportunities 2.4.1 Inclusion of weather predictions in building energy systems As discussed in sections 2.2 and 2.3, weather predictions are regarded as vital input data in BEMS, especially the ones with predictive control capacity. Air temperature and relative humidity directly affect the heat exchange between the building and the environment as well as the HVAC load and occupant comfort. Temperature, solar radiation and wind speed also affect the capacity to generate energy onsite from distributed generation systems, where available. Finally, wind speed may affect ventilation effectiveness depending on the building design.

According to the review findings, weather forecasting may assist in:

 Adjusting the dynamic optimisation of HVAC set points according to the weather conditions and building zone dynamics (Cooperman et al., 2010, Lehmann et al., 2013, Li et al., 2013, Oldewurtel et al., 2012, Ferreira et al., 2012b). Most of these methods attempt to operate the HVAC system at the point of maximum efficiency (lowest energy costs) for a certain horizon and within well-defined boundaries of thermal comfort.  Predictions of onsite energy generation (where available): solar generation is directly affected by incident radiation and is negatively affected by higher temperatures and cloud formation (Suzuki et al., 2012, Zong et al., 2012, Chow et al., 2012, Yang and Xie, 2012, Al-Messabi et al., 2012, Huang et al., 2011, Bacher et al., 2011, Chen et al., 2011). Furthermore the operation

schedule of cogeneration sources may be optimised with predictions of future weather conditions (Luo et al., 2010, Collazos et al., 2009, Cho et al., 2009).  Enabling preconditioning (night-control) processes to shave mid-day peaks by shifting some of the load to low tariff off-peak times (Spindler and Norford, 2009a, Greensfeldera et al., 2011, Xu and Haves, 2006, Rabl and Norford, 1991, Ruud et al., 1990). Preconditioning strategies depend on the external weather conditions, as the thermal mass models require accurate weather data to both estimate the building’s parameters and forecast future loads.  Enabling various other demand response measures: MPC systems are able to respond to hourly energy prices more efficiently with accurate weather inputs (Luo et al., 2010, Candanedo et al., 2013b, Mathieu et al., 2011, Hagras et al., 2008, Lee and Braun, 2004)

Direct comparison of the reported accuracy in predictions of weather variables and load, as well as the energy savings via optimisation algorithms based on these predictions is not particularly meaningful, as most of the techniques are relevant and tailored to specific buildings and depend on the availability of data and required outputs. Hence, it is not simple to establish an absolute conclusion regarding which weather prediction methods result to highest savings in a BEMS.

After reviewing the literature, it is evident that weather and specifically the ambient temperature, relative humidity and solar radiation are among the main factors of uncertainty in building load and generation predictions and in turn for energy management optimisation. As such, the major finding from sections 2.2-2.3 is that the inclusion of weather variable forecasting produces improved results in the prediction and optimisation of the performance of a building energy system compared to a deterministic approach without any weather data inputs. Specifically, in studies where optimisation and control frameworks utilised weather inputs, savings in energy costs of up to 40% are not uncommon compared to the systems without weather inputs.

2.4.2 Research gaps A common BEMS design is to incorporate predictions of temperature, which are in turn applied to load forecasts and management of the building conditioning or other DR measures. For larger commercial buildings with a range of generation and management options, predictions of additional weather variables may be necessary for the optimal operation of the energy system. However, the implementation of an energy system able to utilise several weather inputs appears to be problematic. The introduction of additional input variables increases the degree of complexity and in turn the required computational resources. Furthermore, predictions of multiple weather variables in parallel require non-linear models that cannot be accurately designed using statistical methods alone. The incorporation of localised numerical weather forecasts is a solution to this challenge as it allows development of non- linear models and while still there is an increase in complexity and cost of the energy system, the savings potential is promising.

So far, the literature has not yet considered the integration of localised numerical weather forecasting at high resolutions in commercial BEMS, which constitutes a clear research opportunity. The gap refers to both deterministic and ensemble numerical forecasting. Furthermore, there is a distinct lack of decision algorithms for the inclusion of numerical predictions of peak loads, sudden weather changes, extreme heat events and preconditioning assessment in building demand management. Finally, there is little evidence in current forecasting research of utilising the persistence model to improve numerical weather predictions.

Using a set of lightweight numerical prediction tools and statistical processing, this thesis is proposing a modular approach based on weather predictions that is able to provide valuable inputs for optimising a variety of aspects of energy management in commercial buildings and addressing the research gaps described above. The framework of the proposed integrated architecture consisting of a three layer forecasting and optimisation system can be seen in Fig. 17. The interdependencies between many individual modules of successive layers have been established in the literature and reviewed in the sections above.

A framework such as the one shown in Figure 22 is meaningful for commercial buildings with enhanced energy generation and management capacity, so that the potential added savings can justify the increased model cost.

Figure 22: Architecture of a BEMS that integrates weather, load and generation forecasts for developing appropriate responses. This thesis will investigate several pathways of this modular approach. The main advantage of this modular approach is the added value in a range of building energy management components of the third tier (responses). The uninterrupted flow of inputs and outputs from each layer to the next, the compatibility of data formatting and tailored control over all aspects of energy management are also noteworthy benefits. As the architecture is modular, certain elements may be excluded from the BEMS, or additional layers may be added. In the context of this thesis, chapter 3 will focus on the first level predictions of temperature, humidity and wind speed, while chapters 4 and 5 will extend the applications to certain second level applications (load forecasting, passive storage). Finally chapter 6 will discuss response applications (level 3), based on the outputs of the previous layers.

Since the uncertainty associated with weather forecasting in level 1 of the architecture propagates towards the higher levels, it is important to be able to have access to accurate forecasts for the range of variables required for different horizons, preferably localised. Weather variable predictions in level 1 feed into higher level modules and assist with the respective predictions of load and generation. Temperature is the most important variable and can be predicted using a time series or a ANN approach especially for short horizons; however, NWP models are favourable for such an integrated framework as they can predict factors like humidity, wind speed and direction and cloud formation that contribute to the system’s performance. In this thesis a combination of NWP and statistical forecasting will form the foundation of the architecture in weather forecasting.

Level 2 consists of modules related to the load, the availability of energy and the building internal and external interactions. Generation forecasts from RE are useful for managing peak and base loads, as well as energy flows to and from the grid. Passive energy storage in the thermal mass of the building can also be modelled towards the same end. Other important parameters depending on weather forecast outputs include solar heat gains and occupant comfort. In addition, long-term effects of the climate may be possible to predict, which will be useful for future planning and upgrading. At this level, electricity pricing is considered as a major non-weather dependent component subject to forecasting.

Finally, level 3 consists of the decision making regarding the BEMS’s responses, with MPC capability. Besides HVAC set point optimisation, active storage where available, is relying on inputs from the previous layer. Electricity flows within the building (from DG sources to load and/or storage), as well as exchange with the grid can be planned accordingly. External factors that need to be considered include the building type, size, equipment and occupancy patterns. The cost of the complexity of the system and higher spatial and temporal resolutions should always be weighed against the potential savings in each layer.

3 Short term numerical weather forecasting

 Development of a set of weather forecasting models based on numerical predictions  Statistical post-processing was implemented to increase accuracy in short-term horizons  Forecasting skill can be further improved with more frequent onsite weather inputs  Control modules may select outputs from different models depending on BEMS operation mode 3.1 Model outline Based on the findings of chapter 2, there is clearly significant potential for adding value to the energy management systems of commercial buildings with weather prediction generated locally. The work in this chapter attempts to develop a variety of useful localised weather forecasts, specifically predictions of air temperature, relative humidity and wind speed of up to 6 hours ahead, tailored for building energy management purposes.

Specifically, the aim of this chapter is to describe the methodology of the design and development of a hybrid localised numerical prediction model with statistical regression post processing, which may be used in a control and optimisation framework in a BEMS. The justification of this hybridisation is to combine the advantages of individual numerical and regression prediction models in a high spatial resolution. Specifically, numerical models perform better in instances of rapid weather changes, while regression models enable the utilisation of existing patterns in weather. Furthermore, localisation adds value to the forecasts through increased accuracy compared to forecasts obtained from third parties.

The chapter begins with discussing the details of the development and parameterisation of the proposed models, as well as their constituent base components in section 3.2. Section 3.2 also describes the acquisition process and characteristics of data used in this study. The following part (section 3.3) includes a range of results and useful findings from the simulations, which are discussed and evaluated extensively in section 3.4. Additionally, there is a discussion of the limitations and next steps in the context of the development of a complete prediction and optimisation model in this thesis.

The advantage of short-term horizon forecasting is that the use of recent weather observations can be applied to improve accuracy. Furthermore, it is a horizon that has been used in existing control strategies as it allows for both a reasonable time frame to develop optimal energy management plans, and a margin for adjustments if needed (Candanedo et al., 2013a, Hazyuk et al., 2012). Finally, a 6-hourly horizon allows to align relatively closely to the persistence assumption, which is a major constituent of the model. Longer horizon forecasts based on the same methodology may be considered, however the weighting of the numerical component should be dominant. Chapters 4 and 5 demonstrate the value of these longer numerical forecasts in specific peak load and preconditioning applications.

The hybridisation was based on the outputs of two base prediction models: a reference model, which generates forecasts from onsite observations and assumes that the state of the atmosphere remains unchanged during the next hours and a numerical prediction model, which is able to downscale synoptic data to generate localised forecasts. The design requires onsite weather observations for the statistical post processing steps, however since many buildings do not have access to weather monitoring it could run with inputs from external entities. This is considered as a significant benefit as the numerical prediction component can generate simulations from synoptic scale data at any location and hence provide the basis for the final outputs, even if the statistical post processing is based on data taken from a different location.

There are two different approaches in post processing the numerical weather predictions and the development of the hybrid model: a weighted regression and an autoregressive model with external outputs. Furthermore, the effects of updating the reference model inputs (onsite observations) more frequently were investigated. All versions of the base and hybrid models were assessed in terms of their accuracy, requirements and limitations.

The majority of the work presented in this chapter has been published as an individual research paper (Lazos et al., 2015).

3.2 Weather prediction model design This section describes the details of each step during the development of the weather prediction models. The first step includes a description of the data used in the model, as well as the locations chosen for the validation simulations (section 3.2.1). Section 3.2.2 describes the model architecture and explains a variety of design decisions, while section 3.2.3 provides details about the two base prediction components: persistence and numerical. The three different methods (linear weighting, optimised weighting and ARX) for generating the hybrid predictions from the two base models are discussed in section 3.2.4, followed by the overview of an algorithm that enhances accuracy with more frequent persistence updates in section 3.2.5. Finally, section 3.2.6 describes a correction algorithms that also improves accuracy specifically for days with extreme heat events.

Before discussing the design, it is deemed necessary to elaborate on the acquisition and processing of the data used in this study. The data not only allowed to apply and assess the accuracy of the model, but was an integral part of it, since it is required for calculating weights of certain parameters.

3.2.1 Data acquisition The primary data for the development and comparison of the persistence model, as well as the validation of the hybrid models was obtained from weather observations at three locations around Sydney, NSW:

 the Sydney Airport Weather Station (33°57’ S, 151°10’ E, elevation 6m above sea level)  the Bankstown Airport Weather Station (33°55’ S, 150°59’ E, elevation 6.5m above sea level)  the Canterbury Racecourse Weather Station (33°55’ S, 151°7’ E, elevation 3m above sea level) All of the above sites are located in urban or suburban environments, purposefully selected in order to observe the model capability in predicting weather variables in a setting that resembles a typical commercial building locale. The data was taken directly from weather stations operated by the Bureau of Meteorology (BOM) at each location for three years (2005, 2006 and 2010) and were measured at 10m above ground level. The temporal resolution is half-hourly. It should be noted that 72 the data provided by the BOM database are averaged over the half-hourly period they correspond to.

In addition to the weather data from the weather stations, synoptic scale data for use with the NWP software was required. Synoptic data for this thesis were taken from the National Centre for Environmental Prediction (NCEP) databases. It should be noted that data may also be obtained from the European Centre for Medium- Range Weather Forecasts (ECMWF), however formatting them in a way that is compatible for use within TAPM is more challenging (ECMWF, 2015) - unlike NCEP data that can be easily imported into the software.

3.2.2 Prediction model architecture The proposed forecasting model is designed to receive inputs from two sources: a persistence model and a numerical weather prediction model. The persistence and NWP are referred to as the base models and both were designed to generate hourly forecast outputs for a 6 hour horizon. The outputs from the base models are post- processed using three different statistical techniques in order to generate the final prediction models, referred to as the hybrid models. The observed data not only serve as a basis for comparison of the accuracy of the models, but also assist with estimating certain weights of the regression.

Figure 23 outlines the architecture of the prediction framework and the utilisation of data, for designing the base and hybrid models. As explained earlier, observation and synoptic data can be used as both inputs for the persistence model, as well as for the estimation of statistical regression parameters of the hybrid models. The prediction model architecture outlined in Figure 23 is unique and has not been implemented yet in any other numerical prediction studies.

Figure 23: Architecture of the proposed model, able to generate wind speed, temperature and relative humidity prediction outputs Each prediction day is segmented into four 6-hour long parts, through which the weather variables are assumed to stay constant according to the persistence assumption. Accordingly, there are four reference times during a day, at 01:00, 07:00, 13:00 and 19:00. The prediction model consists of four iterations daily generated at the reference times and predicting the hourly values of the weather variables for the following six hours. For instance the first iteration at 01:00 will produce forecasts for 02:00, 03:00, 04:00, 05:00, 06:00 and 07:00.

The rationale behind the selection of these specific reference times is to align the forecasts with the progressively weaker validity of the persistence hypothesis. The further away from the time point when a variable was measured, the less likely it will remain constant, due to weather systems changing over the course of a few hours. As such, the persistence model will be less capable of accurate predictions for longer horizons than a few hours. For a commercial BEMS, the most important time frames are usually in the middle of the day, as commercial loads have a peak in the afternoon and are very low and mostly independent of weather variables outside working hours (Mathieu et al., 2011). The greatest variations of temperature occur during the morning after the sun rises and in the evening as the sun sets. As such, if the reference hour is set at 7:00 it is hypothesised that by the next iteration which occurs at 13:00 the persistence model will have minimum contribution, because the temperature at the middle of the day is expected to be higher than the early morning temperature. Similarly, the temperature at 19:00 is expected to be notably lower than the midday temperature. This is a necessary assumption in order to maximise the accuracy of the hybrid models. 74

The trend of higher variability of weather variables, and especially ambient temperature at certain hours in the middle of the day can be justified by the results shown in Figure 24. The figure shows the average percentage variation of both observed temperature and observed relative humidity at each hour over the course of a day for 3 years, as derived from the weather data from 3 weather stations.

Hourly variability of weather during the day 9.00%

8.00%

7.00%

6.00%

5.00%

4.00%

3.00%

Relative hourly variation 2.00%

1.00%

0.00% 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Hour of day

TEMP HUM

Figure 24: Percentage hourly variation of temperature and humidity It can be seen in Figure 24 that average hourly temperature varies more during the period starting at 8:00 and keeps varying significantly until 13:00 compared to the rest of the day.

Practical reasons were considered in addition to modelling requirements for the selection of the reference hours. Since commercial building occupancy is predominantly between 8:00 and 18:00 it would make sense to attempt to produce forecasts as accurately as possible for this period in order to ensure occupant comfort and optimise energy savings. As such the reference hour choice of 7:00 would allow to produce forecasts for the morning block with the most recent observed values taken into consideration. Similarly, the most recent observed values will be used at reference hour 13:00 to produce forecasts for the afternoon block, which is typically associated with high HVAC loads especially in the summer. 75

Furthermore, choosing a reference hour at 1:00 allows us to predict the evolution of weather variables during the night and help with decision making for preconditioning.

The effectiveness of the hybrid model depends on the chosen horizon: in general, time steps closer to the reference hour will generate more accurate forecasts, since it is unlikely that the weather has changed significantly from the observed state. It follows that more frequent occurrence of a reference hour will improve the overall accuracy. The details for the sensitivity of the model to interval length are discussed in section 3.2.5. Section 3.3 explains the simulation results with different update intervals and chapter 6 adopts an hourly update approach to demonstrate the proposed model’s advantages.

3.2.3 Base prediction models 3.2.3.1 Persistence prediction model The tool proposed in this chapter includes a persistence component, which uses the observations from weather stations as inputs. The weather variable inputs include observations for air temperature, relative humidity and wind speed. Assuming the reference time point is occurring at hour n, then Aobs,n is the observed reference value of the weather variable A averaged between hours n-1 and n. For example, the reference temperature at 01:00 is the average temperature observed at the station’s location between 00:00 and 01:00. For the following six hours (n+1 to n+6) the persistence model assumes that the temperature stays unchanged and is equal to the temperature at reference hour n.

The predicted average values of the variable A for each of the following hour intervals (An+x) according to the persistence component are then:

퐴푛+1 = 퐴표푏푠,푛

퐴푛+2 = 퐴표푏푠,푛

퐴푛+3 = 퐴표푏푠,푛

퐴푛+4 = 퐴표푏푠,푛

퐴푛+5 = 퐴표푏푠,푛

퐴푛+6 = 퐴표푏푠,푛

퐴푛+푥 = 퐴표푏푠,푛 (3)

3.2.3.2 Numerical predictions in TAPM The second and main base model component is based on numerical predictions simulated in TAPM. The simulations may be initialised by the synoptic data and run for the same 6-hourly intervals as the persistence model. During the design and validation stages of this chapter, the simulations in TAPM were run in hind cast mode with the synoptic data from NCEP.

The spatial resolution of the model was configured with the following constraints:

 5 grids nested within each other  Ratio of grid dimensions 81:27:9:3:1 (within the recommended limits for optimal simulations from the developer (Hurley, 2008b))  Outer grid domain resolution of at least 40 km x 40 km to remove boundary conditions as far away from the central grid point – this is necessary as accuracy of TAPM suffers at grid points close to the domain boundaries (Zoras et al., 2007)  Inner grid domain resolution of at most 500 m x 500 m to capture localised effects  25x25x25 grid (x,y,z) (recommended for optimal simulations from the developer (Hurley, 2008b))

The higher the spatial resolution the longer it takes for the simulations to run. However, the forecasting skill was found to be similar even for very high resolutions (down to 100m x 100m inner grids). Thus the domain configuration was selected to be able to produce high forecasting accuracy at a reasonable running time (around 6 minutes per forecasting day on a middle-end laptop computer):

40.5푘푚 ∶ 13.5푘푚 ∶ 4.5푘푚 ∶ 1.5푘푚 ∶ 0.5푘푚

TAPM simulations can produce predictions of weather variables averaged over one hour periods. Table 3 shows the mean absolute forecasting error and root mean square error of the predictions obtained from different domain orders, for each weather variable averaged across the three sites for a yearly set of data. It was

77 observed that the lowest order domain produced the highest accuracy and this trend was true for all variables and all weather stations. Hence, the selected domain for weather predictions in this work was the lowest order domain #5. However it should be pointed that the error variation between different orders is relatively small – especially for temperature (difference in MAE of 0.1°C and RMSE of 0.3°C) and relative humidity (difference in MAE of 0.75% and RMSE of 1.4%). Wind speeds were found to produce the highest errors, which may be attributed to the fact that TAPM simulations calculate the wind speed at 10m, while the comparison data are taken from stations at different altitudes (see section 3.2.1). While the altitude difference has little effect on temperature and relative humidity, wind speeds are highly dependent on elevation (generally speaking higher elevations are associated with higher wind speeds due to reduced friction).

Table 3: TAPM accuracy comparison of different domain orders

Domain 1 Domain 2 Domain 3 Domain 4 Domain 5 Domain resolution 40.5 km 13.5 km 4.5 km 1.5 km 0.5 km Temperature 1.98 1.95 1.93 1.89 1.86 MAE(°C) Temperature RMSE 2.77 2.69 2.62 2.60 2.44 (°C) Humidity MAE (%) 12.50 12.41 12.25 11.98 11.75 Humidity RMSE (%) 17.33 17.02 16.72 16.55 15.93 Wind MAE (m/s) 3.51 2.99 2.76 2.50 2.39 Wind RMSE (m/s) 3.99 3.86 3.85 3.72 3.31

Furthermore, the performance of TAPM predictions was evaluated according to their ability to correctly predict substantial changes in the weather variables. Such changes and especially ones that result in rises or drops of temperature are significant for the management of energy in buildings. Domain #5 demonstrated slightly higher accuracy for most predictions of substantial changes, thus justifying its selection for the analysis of results. Table 4 summarises the findings of each domain’s ability to predict substantial changes in temperature and relative humidity. For this evaluation of TAPM prediction skill, all changes resulting in hourly variation of a weather variable higher than the 90 percentile or lower than the 10 percentile mark were treated as substantial changes.

Table 4: Comparison of each domain’s ability to predict bigger changes in temperature and relative humidity

Domain Dom. Dom. Dom. Dom. Dom. #1 #2 #3 #4 #5 Temperature changes 90- 2.01 2.01 2.04 1.99 1.92 percentile MAE (°C) Temperature changes 10- 1.60 1.61 1.52 1.53 1.53 percentile MAE (°C) Relative humidity changes 11.42 11.40 11.29 11.34 11.53 90-percentile MAE (%) Relative humidity changes 11.60 11.15 11.11 10.96 10.91 10-percentile MAE (%)

3.2.4 Hybrid prediction models Statistical post-processing is able to reduce the magnitude of the forecasting error of TAPM substantially. As discussed in chapter 2, statistical post-processing based on trends of the time series may be able to minimise the errors in forecasting weather variables that are used as input in controlling the operations of a BEMS. In most cases, while the comfort zone is relatively wide for temperature (between 5- 7°C) (Zavala et al., 2010, Mishra and Ramgopal, 2014), errors within the range of 1- 2°C in ambient temperature forecasts can still enable optimal operation of the energy system. There is higher tolerance for relative humidity errors, as the comfort zone lies typically within a wider range (20-80%).

The statistical post processing of the numerical predictions of TAPM in this thesis are carried out in two ways: a weighted regression and an auto-regressive model with external input (ARX). This section will describe the methodology and design of the hybrid forecasts.

3.2.4.1 Linear regression weighted forecasting (WF) model The linear regression WF model is the basic hybrid model and was developed to utilise outputs from both base models with varying weights. The main assumption is that the numerical forecast outputs become progressively more significant for predictions of temperature further away from the reference hour (Bacher et al., 2009). As the day progresses the persistence hypothesis is becoming less valid, as weather conditions change. With the effects of daily weather cycles becoming more prevalent, the numerical prediction framework is more capable in predicting the

79 values of weather variables. On the other hand, for predictions closer in time to the reference hours, the persistence model is more significant since the state of the atmosphere changes only marginally.

A linear weighting regression was implemented to express the varying weights of the two base models for the six hourly blocks after each reference hour. The forecasts of variable A for each hour (An+x) were obtained by weighing the observed temperature at reference hour (Aobs,n) by a factor of wn+x and the numerical prediction (ANWP,n+x) at hour n+x by a factor of (1-wn+x). n+x is the point in the time series with distance of x hours from the reference hour n. Thus the post-processed predictions for weather variables (An+x) are derived at each time step using the following equations:

퐴푛+1 = 퐴표푏푠,푛푤푛+1 + 퐴푁푊푃,푛+1(1 − 푤푛+1)

퐴푛+2 = 퐴표푏푠,푛푤푛+2 + 퐴푁푊푃,푛+2(1 − 푤푛+2)

퐴푛+3 = 퐴표푏푠,푛푤푛+3 + 퐴푁푊푃,푛+3(1 − 푤푛+3)

퐴푛+4 = 퐴표푏푠,푛푤푛+4 + 퐴푁푊푃,푛+4(1 − 푤푛+4)

퐴푛+5 = 퐴표푏푠,푛푤푛+5 + 퐴푁푊푃,푛+5(1 − 푤푛+5)

퐴푛+6 = 퐴표푏푠,푛푤푛+6 + 퐴푁푊푃,푛+6(1 − 푤푛+6)

퐴푛+푥 = 퐴표푏푠,푛푤푛+푥 + 퐴푁푊푃,푛+푥(1 − 푤푛+푥) (4)

For An+1 the weighting factor wn+1 was set to 100%, representing 100% contribution of the persistence model and then linearly decreased by 1/6 for each subsequent point of the time series. Accordingly, the weighting factor of the numerical predictions, 1-wn+x started from 0% and increased in a linear fashion by 1/6 for each hourly interval. The 1/6 decrease represents the negative gradient of the linear weighting, as each prediction iteration occurs in a 6 hourly horizon. The weighting factors of both persistence and NWP models over the duration of each six-hourly prediction interval can be seen in Figure 25. The WF model resulted in notable improvement of the forecasting accuracy over both the persistence model and TAPM as will be later discussed in the results section.

Weight factors for the WF model

100% 90% 100.00% 83.33% 80% 83.33% 66.67% 70% 60% 66.67% 50.00% 50% 40% 33.33% 50.00% 30% 16.67% 33.33%

20% Weighting Weighting contribution 10% 0.00% 16.67% 0% 1 2 3 4 5 6 Hour after reference hour (n+x)

Persistence contribution NWP contribution

Figure 25: Comparison of weighting contribution of each base model to the WF linear prediction model 3.2.4.2 Historical data weighted forecasting (WFS) model

Rather than assuming a linear decrease of the weighting factor wn+x the further away the time series is from the reference hour, the WFS model uses heuristics to find the optimal weightings for each point in the 6 hourly blocks for a specific location. This process is carried out by utilising long-term archived observed weather data for each specific location. One training year was used for each site to estimate the weighting factors, and these factors were applied to assess the skill of the WFS model for all data. Otherwise, the hybridisation is carried out using the equation (II).

The primary objective of the WFS algorithm was to minimise the average mean absolute forecast error. To that end, the optimal weights from historical data were found with the help of the generalised reduced gradient algorithm, since it provides a fast and efficient approach to solve linear problems. The algorithm searches for minima where the function gradient is equal to zero, by continuously modifying each of the six weighting factors and examining the effects on the partial derivatives. The central differencing technique was implemented, which modifies the weighting factors in both directions in an attempt to find an accurate solution.

The optimisation algorithm was run multiple times from different starting points in order to avoid local minima and the results converged to the values shown in Figure

26, Figure 27 and Figure 28 using one year long training data for each site. The results confirm that the initial approach of linear weighting was reasonable for all variables in all sites, but the linear regression can be improved further with post processing.

Weighting of the persistence model for temperature in hybrid models 100% 90% 80% 70% 60% 50% 40% 30% 20%

10% Persistence Persistence model contribution 0% 0 1 2 3 4 5 6 Hour after the reference hour (x)

WF (all sites) WFS (Airport) WFS (Bankstown) WFS (Canterbury)

Figure 26: Weights (wx) of the persistence model for temperature for each hybrid model at each location and time series point. The NWP weighting is equal to 1-wx

Weighting of the persistence model for humidity in hybrid models 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% Persistence Persistence model contribution 0% 0 1 2 3 4 5 6 Hour after the reference hour (x)

WF (all sites) WFS (Airport) WFS (Bankstown) WFS (Canterbury)

Figure 27: Weights (wx) of the persistence model for relative humidity for each hybrid model at each location and time series point. The NWP weighting is equal to 1-wx

Weighting of the persistence model for wind speed in hybrid models 100% 90% 80% 70% 60% 50% 40% 30% 20%

10% Persistence Persistence model contribution 0% 0 1 2 3 4 5 6 Hour after the reference hour (x)

WF WFS (Airport) WFS (Bankstown) WFS (Canterbury)

Figure 28: Weights (wx) of the persistence model for wind speed for each hybrid model at each location and time series point. The NWP weighting is equal to 1-wx Table 5, Table 6 and Table 7 show the calculated weightings for each location and each weather variable from the historical data analysis.

Table 5: Values of wx weightings of the persistence model for temperature prediction post-processing for each hour after the reference hour (x) and each location

Hour (x) Airport Bankstown Canterbury 1 92.86% 93.33% 88.34% 2 71.43% 44.34% 60.00% 3 50.00% 34.51% 37.86% 4 23.08% 22.55% 21.95% 5 15.00% 6.65% 11.09% 6 11.36% 8.17% 6.01%

Table 6: Values of wx weightings of the persistence model for relative humidity prediction post- processing for each hour after the reference hour (x) and each location

Hour (x) Airport Bankstown Canterbury 1 89.30% 88.21% 88.13% 2 76.47% 75.78% 72.10% 3 66.98% 69.55% 55.64% 4 59.60% 47.71% 41.71% 5 51.43% 32.21% 33.57% 6 45.57% 16.01% 25.87%

Table 7: Values of wx weightings of the persistence model for wind speed prediction post-processing for each hour after the reference hour (x) and each location

Hour (x) Airport Bankstown Canterbury 1 93.42% 79.79% 89.65% 2 85.83% 65.17% 77.93% 3 75.13% 48.94% 70.24% 4 70.73% 42.16% 54.43% 5 65.31% 34.54% 40.37% 6 63.04% 30.97% 32.92%

It is obvious that localisation is important, as there are substantial differences in weightings across sites. Another observation is that for wind speed predictions, the persistence component tends to be relatively more significant, even for 5 or 6 hours after the reference hour. This implies that the numerical predictions do not perform

84 as well for wind predictions after hours, compared to temperature and relative humidity predictions.

3.2.4.3 ARX prediction model Using the TAPM forecasts as inputs (B(t)) and the past observations (A(t)) it was possible to approach the hybridisation forecasting problem from a different perspective than a time series weighted regression. An auto-regressive model with external output (ARX) was implemented that is able to correlate the value of the variable at a time point t (A(t)) to a finite number of observed values from the past A(t-k) and external inputs B(t-k). The order of the model is described as a set of three integer parameters that dictate the architecture:

 ma is the number of time steps of past output observations used

 mb-1 is the number of time steps of past inputs used

 mk is the dead-time of the system – in this analysis it is kept at 1 as observed temperature depends on the immediately previous observation

The ARX model of order ma:mb:1 can be then written as:

퐴(푡) + 푎1퐴(푡 − 1) + 푎2퐴(푡 − 2) + ⋯ + 푎푚푎퐴(푡 − 푚푎)

= 푏1퐵(푡 − 1) + 푏2퐵(푡 − 2) + ⋯ + 푏푚푏퐵(푡 − 푚푏)

The equation can be rewritten in a simpler form using a time-shift operator q-m representing the difference between the current output and the mth time step:

−1 −2 −푚푎 퐴(푡) + 푎1푞 퐴(푡) + 푎2푞 퐴(푡) + ⋯ + 푎푚푎푞 퐴(푡)

−1 −2 −푚푏 = 푏1푞 퐵(푡) + 푏2푞 퐵(푡) + ⋯ + 푏푚푏푞 퐵(푡)

The ARX model can be then expressed in a compact form with two weighting polynomials and the input/output time series:

퐴(푡)푎(푞) = 퐵(푡)푏(푞)

Finally, a noise component e(t) with constant variance can be added to the model to simulate the random factors contributing to the evolution of the system:

퐴(푡)푎(푞) = 퐵(푡)푏(푞) + 푒(푡) (5)

For this project and in order to compare models on the same basis a 6:1:1 ARX model was implemented. The weights a(q) = (a1, a2, … , a6) and b(q) = (b1) in equation III 85 were estimated for each order using the MATLAB identification system using the observation data for each site. Table 8 summarises the calculated the values of the coefficients in the polynomials a(q) and b(q), as derived from the analysis of 3 years of historical data for each location.

Weight Airport Bankstown Canterbury

a1 1.2640 1.2730 1.2221

a2 -0.1509 -0.2541 -0.3067

a3 -0.1255 -0.1367 -0.0402

a4 -0.0674 -0.0622 -0.0663

a5 -0.0859 -0.0554 -0.0103

a6 -0.0864 0.0649 0.0145

b1 0.0787 0.1659 0.1808

3.2.5 Forecast sensitivity to output update intervals The post-processed forecast model performance can be enhanced further if the input values of the variables from the persistence model are updated in more regular intervals of 3 hours (3U), 2 hours (2U) and one hour (1U) respectively. The variable values from the numerical model predictions are still obtained from 6 hourly simulations. Of course this may be more difficult to implement from a planning point of view, since it means that the BEMS will have to be updated more often with weather prediction inputs. The research results demonstrated that the forecasting performance is increased with more frequent updates and more importantly, the forecasts become more responsive and able to detect sudden changes and peaks; however, decisions for demand response measures, scheduling HVAC operations and committing energy sources may have to be changed more frequently. In buildings with limited infrastructure and inflexible decision making policies, this may pose constraints to an implementation with more frequent updates.

3.2.6 Correction algorithm for extreme heat events Information about the occurrence of extreme events is of interest for the optimal operation of BEMS. In this context, particular emphasis is placed on the prediction of periods of high temperatures as they are responsible for major peaks in HVAC loads and related to the potential activation of demand response measures. Furthermore, during hot days generation plants, such as cogeneration or solar panels may need to shut down or operate at reduced output mode, in order to avoid potential breakdowns from extreme loads and high temperature build-up.

In order to enhance the accuracy of the predictions for extreme heat events, the raw weather data from the stations can be used to develop a correction algorithm. The correction algorithm is based on the correlation of the respective distributions of ambient temperature and relative humidity during these events. To obtain these distributions and develop the correction algorithm, long term historical data is necessary for each specific location. Figure 29 shows the distribution of temperature-humidity observations as obtained from historical data over a period of 9 years at the Airport site.

1200

1000

800

Counts 600

400 41 35 200 29 23 0 17 Temperature 100 90 11 80 70 60 50 40 30 5 20 10 Humidity 0

0 10 20 30 40 50 60 70 80 90 100

Figure 29: Correlation of temperature and humidity for a 9 year period for the airport site

For this location, it was found that the majority of the extreme heat events occur when both of the following conditions are met at the same time:

1. Ambient temperature is above 32°C 2. Relative humidity is below 50%

With the temperature-humidity distribution in consideration it is possible to improve the skills of the hybrid models. Specifically, for the intervals that the hybrid model predicted both a drop in humidity below 50% and a rise of temperature above 32°C, the peak temperature prediction can be adjusted according to the average forecasting error in extreme heat predictions (as derived from simulations and compared to historical data). The correction algorithm for predictions of extreme heat occurrences can be developed according to the flowchart shown in Figure 30.

Figure 30: Flowchart of adjusting TAPM predictions of high temperatures according to predicted humidity with a correction algorithm The thresholds for temperature and rel. humidity can be of course adjusted accordingly for each specific site, as the distributions of the two variables and their correlation during extreme heat events may vary. While the correction algorithm is relatively simple (and effective, as will be shown in section 3.3.5) its disadvantage is that relies on the presence of a long-term historical record of observations of temperature and humidity in order to understand their correlation.

3.3 Results The main metric used for assessing and comparing forecast accuracy is the mean absolute error (MAE). This is preferred over percentage errors (MAPE), as MAE is able to compare variable prediction on a normalised basis across the year. For instance, a temperature error of 1° would produce the same MAE regardless of season; however if MAPE was used the winter value would be significantly higher as the temperatures are generally lower in the winter. RMSE is also a useful metric, since it allows us to compare the magnitude of variation in errors across models.

3.3.1 Temperature predictions Table 9 and Table 10 show the average MAE and RMSE respectively for each model (including the base models) and site.

Table 9: MAE of each prediction model for temperature Prediction model MAE Persistence TAPM WF WFS ARX Average Airport (°C) 1.93 1.69 1.32 1.30 1.51

Average Canterbury (°C) 2.59 1.99 1.74 1.61 1.89

Average Bankstown (°C) 2.81 1.92 1.82 1.64 1.76

Average Total (°C) 2.44 1.86 1.62 1.52 1.72 Accuracy increase from - 23.69% 33.52% 37.88% 29.60% persistence

Table 10: RMSE of each prediction model for temperature Prediction model RMSE Persistence TAPM WF WFS ARX Average Airport (°C) 2.88 2.22 1.86 1.79 2.17

Average Canterbury (°C) 3.56 2.63 2.16 2.16 2.51

Average Bankstown (°C) 4.04 2.47 2.41 2.14 2.40

Average Total (°C) 3.49 2.44 2.14 2.03 2.36 Accuracy increase from - 30.27% 38.73% 41.81% 32.40% persistence

For temperature predictions the results demonstrated a superiority of the WFS model for 6 hourly updates of the observed values. As seen in Tables 2 and 3, the WFS hybrid model improved the persistence predictions by 37.88% (MAE) and 41.81% (RMSE). Regarding the absolute prediction errors, the WFS model improved the persistence predictions by 1.52°C (MAE) and 2.03°C (RMSE). The effects of more frequent updates of the persistence component inputs (instead of every 6 hours, having update intervals of 3, 2 or 1 hours) as discussed in section 3.2.5 can be seen in Table 11.

Table 11: Comparison of persistence update frequency for temperature 6 hourly 3 hourly 2 hourly 1 hourly Prediction model updates updates updates updates (6U) (3U) (2U) (1U) Persistence MAE (°C) 2.44 1.53 1.20 0.86 Relative accuracy increase from 6U - 37.21% 50.89% 64.57% WFS MAE (°C) 1.60 1.25 1.09 0.84 Relative accuracy increase from 6U - 22.08% 31.88% 47.50% ARX MAE (°C) 1.72 1.10 0.90 0.66 Relative accuracy increase from 6U - 36.24% 47.67% 61.43% 90

It is worth noting that the greatest improvement compared to the 6U models occurred for the ARX model. In fact, for the 3U, 2U and 1U cases the ARX outperformed the WFS in terms of prediction accuracy (MAE) across all sites. While the difference in performance appears to be modest, selecting the ARX model over the WFS is preferred under the assumptions discussed in section 3.2.5 (more frequent updates of the persistence input every 3, 2 or 1 hours instead of every 6 hours in the original design). Further discussion about the advantages and disadvantages of each model can be found in section 3.4. Figure 31 displays a comparison of the effects of update frequency of the persistence inputs to the WFS and ARX models.

Comparison of MAE for prediction of temperature

3.00

2.50 2.44 2.00

1.72 1.50 1.60 MAE(C) 1.53

1.00 1.25 1.20 1.10 1.09 0.90 0.86 0.84 0.50 0.66

0.00 6U 3U 2U 1U Update frequency

PERS WFS ARX

Figure 31: Comparison of effects of update frequency of the persistence model on the accuracy of temperature prediction

As ambient temperature is the variable of highest significance in the context of energy management, further results were generated to illustrate the performance of the base and hybrid models in regards to it. Figure 32 shows the average MAE for each hour of the day as obtained for the Airport site for each of the base and hybrid models over the three-year period of the simulations.

Comparison of temperature prediction MAE per hour

6.00

5.00

4.00

3.00 MAE(C) 2.00

1.00

0.00 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Hour of day PERS TAPM WF WFS ARX

Figure 32: Comparison of hourly performance of the four prediction models (airport site)

It is evident that TAPM generates consistent predictions and in fact is superior to other models for the morning hours (9AM to 1PM). This can be explained due to the presence of relatively high gradients at these times (since the temperature typically rises faster as seen in Figure 24). This has implications for load forecasting, as raw TAPM forecasts could be used in place of the post-processed models during those times for any optimisation routines. Archived data of onsite observations need to be analysed to assess the frequency of these instances, as well as their magnitude and time of occurrence. These factors are site specific and subject to both the landscape (proximity to the coast, shading, terrain roughness) as well as the microclimate of the region in question. The jagged shape of the curve is a result of the hybrid prediction models receiving inputs from the persistence model every 6 hours. The trend clearly shows that the persistence assumption is fairly accurate for the first few hours (hence higher weighting is required), but becomes erroneous for the later hours of the block (hence lower weighting is required).

In regards to seasonal performance, the numerical predictions appear to be slightly less accurate in temperature forecasts during the summer months, while the persistence model is less accurate during the rest of the year. No clear trends could be recognised for the hybrid models. The comparison per month can be seen in Figure 33.

Comparison of temperature prediction MAE per month

2.5

1.5

MAE(C) 1

0.5

0 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month

PERS TAPM WF WFS ARX

Figure 33: Comparison of monthly performance of the four prediction models (airport site) 3.3.2 Relative humidity predictions Table 12 and 13 show the MAE and RMSE for the predictions of relative humidity across all sites.

Table 12: MAE of each prediction model for relative humidity

Prediction model MAE Persistence TAPM WF WFS ARX Average Airport (%) 9.48 12.74 8.03 7.66 8.55

Average Canterbury (%) 12.29 12.37 8.97 8.91 9.35

Average Bankstown (%) 11.93 10.82 8.16 8.12 8.95

Average Total (%) 11.17 11.75 8.26 8.09 8.95 Accuracy increase from - -5.18% 26.03% 27.54% 20.32% persistence

Table 13: RMSE of each prediction model for relative humidity

Prediction model RMSE Persistence TAPM WF WFS ARX Average Airport (°C) 13.45 15.93 10.96 10.62 12.09

Average Canterbury (°C) 17.81 15.71 12.23 11.99 13.18

Average Bankstown (°C) 17.14 16.14 11.99 11.95 14.89

Average Total (°C) 16.13 15.93 11.73 11.52 13.39 Accuracy increase from - 1.27% 27.31% 28.59% 17.00% persistence

The numerical model did not perform as well for predictions of humidity as it did for predictions of temperature, however the WFS was still the superior model with an overall improvement compared to persistence of 27.54% (MAE) and 28.59% (RMSE). The effects of more frequent updates of persistence to the accuracy of the WFS and ARX models can be seen in Table 14.

Table 14: Comparison of persistence update frequency for relative humidity

6 hourly 3 hourly 2 hourly 1 hourly Prediction model updates updates updates updates (6U) (3U) (2U) (1U) Persistence MAE (%) 11.24 7.24 5.84 4.33

Relative accuracy increase from 6U - 35.59% 48.07% 61.51%

WFS MAE (%) 8.23 6.37 5.41 4.22

Relative accuracy increase from 6U - 22.60% 34.26% 48.76%

ARX MAE (%) 8.95 6.24 5.14 3.89

Relative accuracy increase from 6U - 30.32% 42.53% 56.57%

A similar trend as the one observed for temperature predictions was seen for relative humidity predictions. The ARX model demonstrated the most significant improvement with more frequent persistence updates and outperformed the WFS for the 3U, 2U and 1U models. Figure 34 summarises the above results.

Comparison of MAE for prediction of relative humidity

12.00

11.24 10.00

8.95 8.00 8.23 7.24 6.00 6.37 6.24

5.84 MAE(%) 5.41 5.14 4.00 4.33 4.22 3.89

2.00

0.00 6U 3U 2U 1U Update frequency

PERS WFS ARX

Figure 34: Comparison of effects of update frequency of the persistence model on the accuracy of relative humidity prediction 3.3.3 Wind speed predictions Finally the summary accuracy results for wind speed can be seen in Table 15 and 16

Table 15: MAE of each prediction model for wind speed

Prediction model MAE Persistence TAPM WF WFS ARX Average Airport 1.72 3.52 1.99 1.68 2.71 (m/s) Average Canterbury 1.53 1.77 1.36 1.34 1.54 (m/s) Average Bankstown 1.71 1.89 1.50 1.49 1.68 (m/s) Average Total (m/s) 1.65 2.39 1.62 1.50 1.98 Accuracy increase - -44.61% 2.20% 9.20% -19.41% from persistence

Table 16: RMSE of each prediction model for wind speed

Prediction model Persistence TAPM WF WFS ARX RMSE Average Airport 2.96 5.23 3.04 2.94 3.54 (m/s) Average Canterbury 1.98 2.28 1.80 1.75 1.92 (m/s) Average Bankstown 1.94 2.43 1.80 1.75 1.89 (m/s) Average Total (m/s) 2.29 3.31 2.22 2.15 2.45 Accuracy increase - -44.46% 3.44% 6.39% -6.78% from persistence

TAPM appeared to show less skill in wind speed predictions and in fact was significantly less accurate than the persistence model in all instances. Among the hybrid models, WF and WFS only resulted in minor accuracy increases compared to persistence, and in fact ARX generated less skilful forecasts compared to persistence.

3.3.4 Abrupt change predictions Sections 3.3.1-3.3.3 described the results obtained from predictions of weather throughout three years of simulations. Often, the ability to focus on predictions of abrupt changes in weather variables is equally or more significant for the energy management of commercial buildings, as such changes may contribute to unforeseen load spikes and lead to deviations from the optimal planning. In the context of this thesis, the following changes are considered abrupt:

 Rise or drop of hourly average temperature of more than 10°C  Rise or drop of hourly average relative humidity of more than 10%  Rise or drop of hourly average wind speed of more than 6 m/s (Xinmei Huang, 2008)

As both WFS and ARX models receive inputs from recent observations with a delay of up to 6 hours for the 6U versions and up to 3 hours for the 3U versions (depending on the time compared to reference hour), it follows that if a sudden change occurs in the prediction block it will be difficult to detect. This can be seen in Figure 35 and Figure 36, where WFS and ARX predictions respectively lag behind the actual sudden change that occurred at 15:00.

Abrupt temperature prediction - WFS

45 40 35 30 25 20

15 Temperatue Temperatue (C) 10 5 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour

BOM TAPM WFS WFS 3U

Figure 35: Performance of WFS model with 6 hourly and 3 hourly updates, during a typical hot day with a sudden change in temperature

Abrupt temperature prediction - ARX

15 Temperature (C) 10

0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour

BOM TAPM ARX 6U ARX 3U

Figure 36: Performance of ARX model with 6 hourly and 3 hourly updates during a typical hot day with a sudden change in temperature

In both figures the BOM curve represents the actual observed hourly average temperature, while the rest of the values are the predicted values of temperature as generated by the prediction models.

These figures are an extreme example showing a 17 degrees drop in a single day, which happens rarely. Considering all abrupt changes as defined earlier, Table 17 summarises the results for the forecasting error of predictions in such instances, whenever they occur. The results are related to the predictions obtained from the simulations across the three sites for three years.

Table 17: Comparison of performance of prediction models in forecasting sudden hourly changes Type of Persistence TAPM WF WFS ARX WFS ARX abrupt MAE MAE MAE 6U 6U 3U 3U change MAE MAE MAE MAE Temp. rises 6.09 2.30 4.89 3.94 4.98 3.72 4.00 (°C) Temp. 6.86 4.29 5.82 5.40 5.60 4.35 4.47 drops (°C) Humidity 16.8 11.9 12.5 12.5 15.5 14.1 12.4 rises (%) Humidity 19.7 11.5 14.7 15.2 18.2 14.0 14.0 drops (%) Wind speed 5.91 5.41 5.71 5.67 5.88 5.98 5.91 rises (m/s) Wind speed 3.82 2.08 2.74 2.25 3.11 3.77 2.57 drops (%)

It can be seen in Table 17 that TAPM performs better than the rest of the models in predicting abrupt weather changes. Another notable observation is that using more frequent inputs in the 3U model improves the performance of the WFS and ARX for temperature and relative humidity skill, but not for wind speed. It should however be noted, that albeit their importance, this type of abrupt changes is relatively rare. Specifically, in 9 years’ worth of hourly historical observations across the three sites, there were in total:

 225 cases of abrupt changes in temperature, spread almost evenly across the three sites (representing 0.3% of a year’s worth of observations)

 179 cases of abrupt changes in relative humidity, spread almost evenly across the three sites (representing 0.23% of a year’s worth of observations)

 304 cases of abrupt changes in wind speed, the majority of which were observed at the Bankstown site (representing 0.23% of a year’s worth of observations (representing 0.39% of a year’s worth of observations)

Partly, this relates to the relative mild climate of Sydney and it is expected that in climates with a bigger range of weather variations, these abrupt changes may be observed more often. 3.3.5 Extreme heat event predictions For the three sites, it was found that periods of extreme heat (temperature of more than 32°C) were just above the 97 percentile mark of the dataset and represent a meaningful value for defining an extremely hot day in Sydney’s climate. This value of course may be adjusted according to the climate of the location the model is applied, as well as the specific demand response requirements and automation processes of the energy system. Table 18 summarises the results of the forecasting skill of each model in predicting such extreme events and compares it to the prediction skill throughout the year.

Table 18: Comparison of accuracy of predictions of each model for extreme heat events (temperatures over 32°C) Persistence TAPM WF WFS ARX WFS ARX 6U 6U 3U 3U MAE of extreme 5.03 3.64 3.16 2.94 3.18 2.40 1.70 events (°C) Average yearly 2.44 1.86 1.62 1.52 1.72 1.25 1.10 MAE (°C)

As seen in Table 18, both the persistence and TAPM models experience a significant decrease in skill when predicting the magnitude of extreme events. Accordingly, and since they receive inputs from the base models, the hybrid models demonstrate a similar decrease in forecasting accuracy. The ARX 3U appears to be the best performing model for this type of forecasts. The main reason for this trend is that both base models tend to underestimate the magnitude of such high temperatures. Specifically, TAPM underestimates the temperatures of extreme heat periods around 91% of the time and persistence underestimates the temperature of extreme heat around 77% of the time.

The correction algorithm described in section 3.2.6 proved useful for improving the prediction skill of an extreme heat event. Table 19 shows the MAE of each hybrid

99 model (the persistence and TAPM predictions remain unchanged) after adjusting for low humidity with the correction algorithm.

Table 19: Comparison of accuracy of predictions of each model for extreme heat events (temperatures over 32°C) with and without the correction algorithm WFS ARX WFS ARX 6U 6U 3U 3U MAE of extreme events (°C) – normal 2.94 3.18 2.40 1.70 MAE of extreme events (°C) – correction 1.76 2.13 1.35 1.18 algorithm

It can be seen that the correction algorithm enables the error to be limited significantly in predictions of extreme heat events. While there are still higher errors than the average (Table 18), the difference can be minimised and hence more consistent (and valuable) predictions of extreme heat occurrences may be obtained.

3.3.6 Peak load predictions Another interesting metric is related to peak load predictions. As commercial building energy management is greatly concerned with peak load predictions, forecast models can be validated according to their ability to accurately forecast when a peak in temperature will occur, which is the main contributing factor of peak loads.

The differences between the actual and predicted peak temperatures were compared for each model and then averaged over a simulation period of two years across all sites. Table 20 shows the results, which demonstrate a superiority of the WFS model, as it was able to predict peaks slightly more timely than the rest of the models, on average 1.72 hours away from the actual peak time. The numerical model (TAPM) had an average difference of 1.77 hours and the ARX model an average difference of 2.23 hours. Both the hybrid models were improved by changing the update interval to 3 hours (3U).

100

Table 20: Comparison of absolute peak temporal difference for all models

Prediction model TAPM WF WFS WFS 3U ARX 6U ARX 3U Average peak 1.77 1.73 1.72 1.60 2.23 1.80 difference 2010 (hours) Average peak 2.00 1.65 1.56 1.46 2.01 1.56 difference 2006 (hours) Average peak 1.89 1.69 1.64 1.53 2.12 1.68 difference (hours)

Since peak load predictions necessitate a very prompt response, this approach may not be as valuable. Section 4.1 will discuss an improved algorithm for peak load predictions based on an ensemble of forecasts instead. Furthermore the duration of these peak temperature events are important for HVAC control. Appropriate strategies to account for these peak events are discussed in section 4.2 and implemented in a case study in chapter 6.

3.3.7 Value of localisation An important research opportunity as explained in Chapter 2, is investigating the value of localisation of high resolution numerical forecasts. It is hypothesised that localised forecasts would be more accurate and in turn allow for optimal building operations management compared to forecasts taken from external sources that may be at a distance. The simple reasoning behind this hypothesis is that even within a few kilometres of distance, weather conditions tend to vary, especially in regions with distinct geographical boundaries (for instance land-sea boundaries or plains-hills boundaries) or segregated landscape features (for instance urban- forested areas).

In order to test this assumption, the methodology described in section 3.2 was applied to three locations for 1 year-long simulations (for year 2015). The application utilised localised numerical predictions that were used to develop forecasts according to the WFS 6U model. The results were then compared to external forecasts taken from the closest BOM weather stations to each location in

101

6 hourly blocks for the simulation year. The reason that three new locations were tested was that the three original sites as explained in section 3.2.1, were at the same location as existing weather stations. Table 21 shows a summary of three new locations in increasing distance from their closest weather stations.

Table 21: Summary of locations chosen to test the value of localisation

Location Closest weather Distance to station weather station Bexley Sydney Airport 5 km Bondi Beach Sydney Harbour 8 km Macquarie University Parramatta 11 km

The WFS and BOM forecasts were then compared against onsite observations for the simulation year 2015, kindly provided by weather monitoring logs at Bexley Public School, Bondi Surf Life Saving Club and Macquarie University. It should be noted that wind speed measurements were not available at all locations and hence the comparison was only carried out for temperature and relative humidity. Furthermore, the data from Bexley were missing relative humidity observations and several hourly time steps for the temperature data series (approximately 500 out of 8760) so the comparison omitted the missing time steps. The comparison was based on the MAE in degrees for temperature and MAE in % for relative humidity.

The Null Hypothesis (H0) for the simulations states that there is no improvement in accuracy (MAE) from using localised forecasts compared to the distant weather station forecasts. The Alternative Hypothesis (HA) instead states that there is a significant improvement in accuracy (MAE) from using localised forecasts compared to the distant weather station forecasts. After the WFS simulations were run and the BOM forecasts were processed the MAE results were summarised in Table 22.

102

Table 22: Comparison of MAE for temperature and humidity predictions between BOM and WFS for the three sites for simulation year 2015

Temperature MAE (°C) Relative Humidity MAE (%) Location WFS BOM WFS BOM Bexley 1.753 1.761 N/A N/A

Bondi Beach 1.598 1.617 8.001 9.231 Macquarie U 1.994 2.052 9.508 11.277 Average 1.782 1.810 8.755 10.254

Two tailed t-tests with α=0.05 (confidence interval 95%) were applied to the data in order to decide on the rejection of the H0. The summary of the t-tests for each site can be seen in Table 23, 24 and 25

Table 23: Summary of the two tailed t-test for Bexley (temperature only) Model WFS BOM Mean 1.753 1.761 Variance 0.082 0.094 Observations 8220 8220 Hypothesized Mean Difference 0.000 t Stat -1.598 t Critical two-tail 1.960

Table 24: Summary of the two tailed t-test for Bondi Beach Variable Temperature Rel. Humidity Model WFS BOM WFS BOM Mean 1.598 1.617 8.001 9.231 Variance 0.083 0.089 2.978 3.459 Observations 8760 8760 8760 8760 Hypothesized Mean Difference 0.000 0.000 t Stat -4.340 -45.379 t Critical two-tail 1.960 1.960

103

Table 25: Summary of the two tailed t-test for Macquarie University Variable Temperature Rel. Humidity Model WFS BOM WFS BOM Mean 1.994 2.052 9.508 11.277 Variance 0.090 0.101 2.061 5.857 Observations 8760 8760 8760 8760 Hypothesized Mean Difference 0.000 0.000 t Stat -12.508 -58.826 t Critical two-tail 1.960 1.960

The results show that since the t-stat at the Bexley site (-1.598) is less than the absolute value of the critical t-value (1.960), the alternative hypothesis can be rejected and hence BOM generated forecasts and data may be used. It may be claimed that since the Bexley site is the closest to the weather station (approximately 5 km) and there are no significant geographical boundaries between the station and the site, the insignificant difference in the results makes sense.

However, for both temperature and relative humidity for the other two sites the null hypothesis can be rejected, as the t-stats are greater than the absolute value of the critical t-values. Hence, it can be claimed that localised high resolution numerical forecasts improve the accuracy of predictions compared to BOM forecasts taken from a distant station.

3.4 Discussion of the model applicability The evaluation of the short term numerical weather forecasting model and its associated modification algorithms described in the previous section has to be carried out not just in terms of their skill, but taking into consideration their usefulness and limitations of their applicability in a BEMS. In this context, it is concluded that both versions of the hybrid models (WFS and ARX) provide significant advantages in relation to both forecasting skill and applicability over the base models in most of the cases.

Specifically, for all variables considered, WFS’s forecasting skill was found to be superior to both the base models (reference and TAPM), as well as the linear WF and ARX models under the initial 6-hourly update assumption. Since the weighting factors are calculated from archived observations on site, it is expected that using longer databases may improve the hybrid model’s skill further. 104

The ARX model outperformed the WFS in any simulation where the reference input interval was shorter than 3 hours. More frequent updates of the reference component resulted into higher skill not just for the ARX, but across all hybrid models. While three different variations were tested (3, 2 and 1 hourly update intervals), it was mainly the 3-hourly model that was considered for practical reasons. Updating the reference component requires inputs from onsite observations and may lead to adjustments in demand responses or HVAC set points that take some time to implement. Certain dynamic optimisation systems may have the ability to utilise shorter update horizons, improving the potential benefits even further. For instance, modules responsible for adjusting the HVAC set points could foresee a peak in temperature larger than initially estimated and attempt to bring the set point closer to the upper bound of the comfort zone before the peak occurs (resulting in peak load reduction). While the computation time of the simulations by the proposed models is short, dynamic energy system optimisation may be limited by various factors in a building, such as decision making policies and lack of infrastructure allowing to adjust the energy flows in a dynamic manner. Thus, while shorter update intervals may produce more accurate predictions, they may be inapplicable if the energy system cannot respond in a timely manner to the new outputs.

Regarding the prediction of abrupt changes, TAPM appeared to have the highest skill overall even compared to the hybrid models. This is justified by the fact that usually these changes are caused by larger scale weather phenomena, such as cold fronts, which TAPM can model more accurately from the synoptic data. On the other hand the hybrid models show inferior skill in predictions of sharp gradients, as they include a statistical component that assumes a steady state of the atmosphere. To address this issue, and in line with the rationale of hybridisation, a simple algorithm may be incorporated in the control system that uses 100% weighting of the numerical prediction component if it indicates that such an abrupt change will occur within the next period. As it was seen in Table 17, this results in a substantial error drop, especially for abrupt rises of temperature, which are primarily responsible for unforeseen load peaks (from 4° and 5°C for the hybrid models to 2.3 °C with 100% numerical weighting). Furthermore, as discussed in section 3.3.4, while abrupt

105 changes in weather conditions are of great importance to building energy management, the analysis showed that they do not occur that often in the simulation locations. As such, the advantage of a hybrid model becomes clear, since it is able to outperform in accuracy the numerical model for the majority of the days within a year (when gradients are low), but can also handle the abrupt change predictions by using the simple algorithm described above.

Another algorithm that can be used and shows the value of hybridisation was described in Figure 30. The correction algorithm may be used to significantly improve the accuracy of the models in predicting extreme heat events and the presence of peaks in temperature as it reduced the error by 1°C in most cases and aligned it more closely with the annual average error.

Overall, while both WFS and ARX perform better than the base models in most metrics, it was found that the WFS is more consistent and able to generate more accurate predictions. However, instead of suggesting the use of a single model for all purposes, different modules of a control system may utilise the outputs from the various components, which is after all a benefit of hybridisation since each model complements the other. Depending on what outputs are needed, certain components may be called by the BEMS controller for input. For instance, the 3U ARX generated better predictions of extreme heat events especially after considering the humidity correlation, while WFS was better in predicting the timing of such peak events. The methodology detailing an implementation algorithm for the selection and utilisation of prediction modules is described in chapter 6 and applied to a case study university building.

It should be noted that while TAPM performed relatively well individually in predictions of temperature, it was actually not as skilful in predictions of relative humidity and wind speed compared to the reference model. The hybridisation value was realised in both cases, where application of the regression algorithms improved the forecasting accuracy by up to 28%.

The domain selection in TAPM was carried out in order to provide both high spatial resolution and minimise running times. As such, the computational resources necessary to run the simulations and generate the 6 hourly forecasts are very low.

106

On a typical office computer, it took no longer than 5 minutes to run each iteration, which is significant for the timely adjustments and optimisation of energy management in a building.

Table 26 shows a qualitative comparison of the performance of each individual component, as well as the hybrid models as obtained from the results in the previous section.

Table 26: Comparison of the skill of each model according to the results obtained from the simulations

Model Persistence TAPM WF/WFS ARX Forecasting skill Low Moderate High High (temperature) Forecasting skill Low Low High Moderate (humidity) Forecasting skill (wind Low Very low Moderate Low speed) Effects of persistence Very positive N/A Positive Very input frequency positive Abrupt changes Very low High Low Low forecasting skill Extreme heat events Very low Low Low Low forecasting skill (default) Extreme heat events N/A N/A High High skill (correction algorithm) Peak time prediction Very low Moderate High Moderate skill

The main advantage of the models analysed in this chapter is that they are able to provide a comprehensive set of hourly weather predictions for short-term horizons with relatively low computational needs. Savings when using weather forecasts may be realised in various ways: studies discussed in chapter 2 indicate energy and peak load savings in the range of 10-40% from weather-related optimisation of load management, compared to a deterministic strategy without any weather inputs. If the building system has the capacity to generate energy onsite, conduct 107 preconditioning and other demand response measures, the savings are cumulative. Savings will be furthered discussed and modelled in chapter 6.

Another advantage of the proposed model is that the error of the onsite generated predictions may be substantially smaller than the error of predictions taken from a weather station, albeit the magnitude depends on the distance of the station as well as the environment around the building. Section 3.3.7 showed that there is significant difference in accuracy between the localised forecasts and external forecasts taken from 8 and 11km away.

There are of course certain limitations to the model proposed in this chapter. Firstly, the proposed forecasts include ambient temperature, relative humidity and wind speed predictions, but not solar radiation ones. As discussed in section 2.3, solar radiation may be a significant factor of the thermal behaviour of buildings. However, there are certain challenges in localised radiation forecasts. To begin with, unlike temperature and humidity, high resolution solar radiation predictions are much more sensitive to local conditions that cannot be modelled well in TAPM. These include, but are not limited to, shading from trees or other buildings and cloud formation. Furthermore, the building envelope and orientation greatly affect solar radiation and hence predictions of an ambient solar radiation value do not make as much sense as ambient temperature or humidity. For example, the solar radiation incident on a flat roof will differ significantly from the radiation incident on a southern vertical façade of the same building at any time step. The modelling challenge also involves study of the envelop materials, as glazing, brick walls, concrete or other types of construction allow varying amounts of solar radiation to permeate and hence directly affect the conditions of the building. Finally, the indirect effects of radiation stored in the thermal mass of the building are also not simple and have to be modelled specifically for each building. As a result, several studies for weather related building conditioning do not consider solar radiation, but still manage to produce accurate predictions based on temperature and/or humidity inputs alone (Tyagi et al., 2011, Penya et al., 2011, Mathieu et al., 2011, Zavala et al., 2009). Hence, as the scope of this chapter is to develop localised forecasts that may be applied to any site (regardless of the building type), solar radiation predictions were omitted. 108

Another significant issue is the necessity of archived observations that are required to evaluate the weightings of the hybrid models and develop the correction algorithm. This issue is not as troublesome, since using a linear regression for hybridising the reference and TAPM outputs as explained in the WF model still produces sufficient results. In fact, for most metrics the improvement with WFS over WF is in the range of 5-10%. More importantly, the application of the hybrid models implies the presence of onsite weather monitoring equipment, which may not be available in certain buildings. An additional issue is that TAPM uses synoptic data inputs, which have to be obtained from entities like NCEP. While the hybridisation can be carried out in a variety of common commercial database platforms, TAPM licenses have to be obtained further increasing the total cost of the model. Finally, a more rigorous significance test process may be conducted with increased number of simulation years, sites and a variety of distances to better evaluate the localised forecasts.

Future work may also assess the ability to predict solar radiation in a more generic way that may be applied with minor modifications to any building rather than modelling for a particular construction only. Different forecasting horizons that may be useful for building energy management may also be examined. Both are expected to increase the usefulness and practical applications of the proposed models. Also the value of localisation may be examined, by comparing the performance of the onsite models against external station predictions for buildings with onsite weather monitoring at varying distances from the stations. Finally, as TAPM was designed originally for modelling of atmospheric pollution, there is room for research that may improve its meteorological prediction capacity and especially accounting for urban microclimate effects. TAPM offers a reasonable land-use classification system for different types of urban areas (depending on the density and average building height) (Adams et al., 2015). However, the categorisation is rather basic and potentially more complex and sophisticated models may further improve simulation precision an account for effects such as urban heat islanding. A tailored numerical prediction model for energy management could be integrated into a control framework and improve the potential value of the energy system.

109

4 Peak load forecasting with an ensemble of weather forecasts

 Design of an ensemble forecasting model for peak load prediction  Ensemble branches based on numerical predictions in 15 different domains  Peak load prediction algorithm includes two components to predict the timing of the peak and the relative magnitude (compared to typical peaks in the building)  Case study results showed promising accuracy of the probabilistic forecasts in two-yearly simulations 4.1 Ensemble forecasting model design 4.1.1 Model outline and rationale In the context of this PhD thesis, TAPM is proposed as the tool for developing an ensemble of weather forecasts by running parallel simulations with different parameterisation. The software interface does not allow direct manipulation of the synoptic data, thus changing the initial conditions is not practical. The proposed methodology is explained in detail in this chapter (section 4.2). Following this, the model is applied for predictions of peak loads on a case study building. The results of the application are described in section 4.3.

The discussion in section 2.2.3.3.2 highlighted that ensemble forecasting is common in modern meteorology as it accounts for the non-deterministic nature of the weather. However, according to the review, there is a clear research opportunity, as the literature has not examined yet any implementations of ensemble forecasting at high resolution for building energy management.

Furthermore, the discussion in section 2.2.4.5, highlighted the importance of peak load forecasting in commercial building energy systems: significant rolling annual costs are due to high consumption peaks, which in turn are in many cases dependent on intensive use of the HVAC, for instance during hot summer days. According to the literature, substantial efforts are being made to develop prediction models for these peak HVAC loads, so that appropriate demand responses may be scheduled. This chapter attempts to approach the issue from a novel perspective: using ensembles of forecasts to predict the likelihood and relative magnitude of expected peaks during the summer.

110

4.1.2 Ensemble branch parameters As discussed earlier, there are in general three ways of developing an ensemble of weather forecasts (World Meteorlogical Organization, 2012):

 Running parallel simulations of the prediction model with slightly different initial conditions of the weather variables  Running parallel simulations of the prediction model with different parameterisation (however the same initial conditions)  Running parallel simulations of different prediction models

While TAPM does not offer features to adjust the model physics and the governing equations, there are parameters that can be changed related to the geometry of the forecasting domains and the nesting ratios. Specifically, the user can modify the nesting ratio (the ratio of the lengths of two consecutive square domains), the size of each domain, as well as the number of grid points that the simulations take place in three dimensions. Synoptic data is used to initialise the simulations at every grid point and set the boundary conditions for the highest order domain. The simulations are then run in succession at each point of each lower order domain. The lower order domains receive their boundary conditions as inputs from the immediately higher order domain. Hence, it is possible to generate a number of different predictions for the same location and horizon (ensemble branches) from the same initial weather conditions (synoptic data) by:

 Changing the number of domains  Changing the number of grid points (x,y,z) of each domain  Changing the nesting ratio between each domain  Changing the resolution of the lowest order (smallest area) domain (and consequently the higher order domains)

The forecasting horizon of each ensemble branch is 24h ahead. The values of weather variables at any grid point of a single domain are calculated through a set of equations describing the atmospheric conditions, as well as the spatial and temporal interactions between neighbouring grid points, each of which has distinct characteristics (elevation, vegetation coverage, landscape type). (Hurley, 2009). Therefore the weather predictions obtained at each grid point are affected not only 111 by the domain characteristics that were mentioned earlier, but also the interactions with its adjacent grid points and of course the characteristics of terrain at the point.

In the proposed ensemble forecasting model, 15 different ensemble branches are implemented. The characteristics of each branch are given in Table 27.

Table 27: Individual ensemble branch characteristics

Ensemble Number Grid Nesting Lowest order Highest order Branch of points ratio domain domain area domains (x:y:z) resolution (m) (km2) E1 5 25:25:25 3 500 1,025,156.25 E2 5 25:25:25 3 300 369,056.25 E3 5 25:25:25 3 100 41,006.25 E4 5 10:10:20 3 300 59,049 E5 5 10:10:20 3 100 6,561 E6 3 35:35:25 4 1000 313,600 E7 3 35:35:25 4 500 78,400 E8 3 35:35:25 4 250 19,600 E9 3 50:50:25 4 1000 640,000 E10 3 50:50:25 4 500 160,000 E11 4 75:75:20 2 1000 360,000 E12 4 75:75:20 2 100 3,600 E13 4 25:25:25 2 2500 250,000 E14 4 25:25:25 2 1500 90,000 E15 4 25:25:25 2 500 10,000

As an example, the ensemble branch E1 consists of 5 domains. Each domain consists of 25:25:25 grid points in the x:y:z directions respectively. The resolution of the lowest order domain is 500m, which means that each grid point has a distance of 500m from its neighbouring points on the x, y and z directions. The nesting ratio for E1 is 3, which means that second lowest order domain has a resolution of 1,500m (ratio 1:3). Since the E1 branch consists of 5 domains, the highest order domain has a resolution of 40,500m.

Branches with highest order domains whose areas are above 250,000 km2 were selected to simulate large-scale synoptic weather phenomena, while branches with 112 highest order domain areas below 20,000 km2 were selected to simulate convective local-scale weather phenomena.

The geometrical parameterisation was implemented according to the recommended range of values (Hurley, 2009) as well as with the total computational time into consideration. High resolutions, high number of grid points and high number of domains all contribute to a multiplicative increase in computational time. All branches are able to produce 24-hour ahead forecasts with hourly resolution in under 35 minutes of simulation time. The forecasts include predictions at each time step for ambient temperature and relative humidity at the selected site. The average computational time for each branch on a middle-spec laptop computer can be seen in Table 28.

Table 28: Comparison of computational time for each ensemble branch

Ensemble Computational time (simulation minutes for a single 24-hourly prediction) E1 20 E2 28 E3 35 E4 13 E5 19 E6 7 E7 15 E8 22 E9 10 E10 18 E11 10 E12 22 E13 5 E14 9 E15 14

A day-ahead forecasting horizon was selected as it is both practical and provides accurate numerical predictions. A variety of responses, such as preconditioning the building may occur several hours before the actual peak occurs and hence shorter horizons may not allow sufficient time to prepare. Chapter 6 describes the application of this methodology to predict the daily peak loads of a case study university building. In this application the ensemble forecasting runs once per day

113

(at the beginning of the day) and may receive minor corrections throughout the course of day at hourly intervals.

The ensemble mean is the most common output metric from ensemble models (World Meteorlogical Organization, 2012). The ensemble mean for each weather variable can be thought of as the expected value of the probability density function of that variable at any point in the time series (Taylor and Buizza, 2002).

4.1.3 Model outputs and validation In order to steer away from a deterministic approach, a range of values can be considered instead, based on the distribution of predictions of the branches in applications of BEMS related forecasts. The value of considering the branch distribution was illustrated via a validation algorithm of the predictions versus observed values for 2 years’ worth of data (2011-2012) at the University of New South Wales campus. Firstly, weather predictions were derived using TAPM following the forecasting setup described in section 3.2 (without any post- processing, i.e. only the base NWP component). Then 6-hourly horizon ensemble predictions were generated and the ensemble mean at each point of the time series was calculated. Only 12 ensemble branches were used for this analysis.

Table 29 compares the forecasting skill of the ensemble branches versus the numerical prediction from TAPM (without post-processing).

Table 29: Comparison of performance from WF model and ensemble mean from 12 branches for predictions of temperature and relative humidity

Temperature (°C) Relative humidity (%) Ensemble mean Single TAPM Ensemble mean Single TAPM Average MAE 2.13 2.05 13.42 15.57 RMSE 2.74 2.87 16.31 18.55

While temperature predictions were relatively unaffected by the implementation of an ensemble approach, the forecasting skill improved clearly for relative humidity for the simulation period.

Instead of relying on the ensemble mean, predictions may be made based on the distribution of branches. The notion behind this approach is that certain TAPM

114 branch configurations tend to over or underestimate the predicted weather variables. Since the branches initialise from the same synoptic data, it is hypothesised that if these outliers are discarded, the accuracy would improve. In order to obtain and compare the results of this methodology, the forecast at each point of the time series was calculated as the unweighted mean of the ensemble branches that were within one standard deviation of the average prediction. The results are summarised in Table 30 with ensemble z1 column representing the performance of the ensembles within 1 standard deviation:

Table 30: Comparison of prediction performance using ensemble mean or the mean of the branches within one standard deviation

Temperature (°C) Relative humidity (%) Ensemble Ensemble z1 Ensemble Ensemble z1 mean mean Average MAE 2.13 1.99 13.42 9.50 RMSE 2.74 2.68 16.31 12.14

The results showed that the ensemble z1 predictions were on average more accurate than the ensemble mean, albeit the difference was more notable for relative humidity predictions.

The forecasting skill of the ensemble mean model, ensemble z1 model and deterministic TAPM predictions can be visualised in Figure 72 and Figure 73, over an 11-day sample period of the time series from the simulations (as the figures are wide, they are attached in landscape orientation in the Appendix). The sample period constitutes of “interesting” weather phenomena, such as extremely hot days, or abrupt changes in temperature over a short period. It can be seen that error is comparable for all prediction approaches for temperatures; however, predictions of relative humidity are more accurate when disregarding the outliers (ensemble z1). Interestingly, the periods of high error for all models coincide, indicating the inability of the numerical model to capture certain weather phenomena. Further research needs to be conducted to attempt to recognise the occurrence of such patterns.

115

Figure 74 and Figure 75 show the ensemble branch prediction spread and the ensemble z1 predictions compared to the observed values for the same 11-day period of simulations in the UNSW campus (as the figures are wide, they are attached in landscape orientation in the Appendix). As discussed earlier, the improvement in accuracy over individual numerical predictions (and taking the ensemble mean) is more obvious for relative humidity, especially for peak humidity. When most individual branches tend to overestimate the predicted values of relative humidity, the ensemble z1 falls much closer to the observed values, as it disregards the extreme values of individual branch predictions.

While the forecasting skill of the ensemble z1 predictions improved over the ensemble mean and individual branch predictions, there are still times when the actual value of the variable falls outside the ensemble distribution of predictions. As stated earlier, further research needs to be conducted to attempt to recognise the occurrence of such patterns in order to improve the forecasting skill of the model. Figure 76 and Figure 77 highlight the issue by comparing the observed values to the quartile distribution of ensemble predictions (as the figures are wide, they are attached in landscape orientation in the Appendix). The comparison shows the minimum, maximum, median as well as 1st and 3rd quartiles of the ensemble prediction at each time step. This implies that a probabilistic prediction model may be more relevant to a deterministic one.

4.2 Peak load prediction model Based on the analysis of the results from 4.1, a novel approach for peak cooling load prediction based on weather forecasts is proposed and described in section 4.2. The predictions are based on a probabilistic approach, derived from the distribution of the ensemble branch predictions as simulated according to section 4.1.

4.2.1 Potential peak periods In the summer, peak loads associated with cooling in mild climates like Sydney, typically occur as a result of high temperatures and low relative humidity, as discussed in section 3.2.6. The first component of the peak load prediction model is to predict the period of a day that a peak load may occur, as a result of high

116 temperature and low relative humidity. The algorithm for the detection of potential peak periods within a day is outlined in Figure 37.

Figure 37: Algorithm for the detection of potential peak periods in the summer It should be noted that the proposed algorithm may be modified according to the local climate – for instance in tropical regions like Singapore, peak loads are more often associated to high temperatures and high relative humidity instead and hence the threshold values for comparison at the two decision nodes may be different.

The two decision steps are designed to integrate the probabilistic element of ensemble forecasting in the peak load predictions. In each of these decision steps the distribution of branch forecasts is considered, since specific branches perform differently in predicting different scale weather phenomena. Firstly, if the upper quartile of branch predictions for temperature is close to the maximum daily temperature a potential peak is identified. In simple words this occurs if at least 4 branches predict a temperature within 10% of the maximum daily temperature at any time point. The 10% threshold is set to account for the mean percentage error of TAPM forecasts when predicting extreme events (such as high temperatures). For the points in the predictions that the temperature clause is met, a second comparison is made: if the lower quartile of branch predictions for relative humidity

117 is within 10% of the minimum daily relative humidity the potential peak is confirmed.

4.2.2 Detection of significant peak loads While the detection of potential peak periods within the day indicates the necessity for demand response measures during these times, it is not by itself sufficient to provide information about the magnitude of the potential peak. This type of information is useful, as typically only the highest peaks over a long period (which may vary between a month and a year in most cases) in commercial buildings are associated with significant energy costs and occasional energy system shortcomings. The practical implications of this is that even if a peak is predicted to occur within a certain period, it may not be significant enough in terms of cost to trigger any response measures. Hence, the algorithm in Section 4.2.1 can be complemented by an additional algorithm that provides an insight into the magnitude of daily peak loads according to a variety of weather related factors.

The first factor affecting peak loads is seasonality – commercial building loads demonstrate distinct patterns between weekdays and weekends, with the latter being associated with much lower overall demand and peak loads and hence lower optimisation potential. Hence, the significant peak load algorithm is used only for weekdays. Furthermore, in countries like Australia in the Southern hemisphere there are significant holiday periods over the summer (Christmas and New Year Day), where typically commercial buildings experience a major drop in demand and hence peak loads. Such holidays are excluded from the algorithm as well.

The second factor is the mean temperature within the potential peak period (as predicted in 4.2.1). As discussed in section 2.2.4, high peak cooling loads are most often associated with relatively high temperatures. This depends on the local climate, as a summer day in Sydney with average temperature of 30°C during the day would be considered relatively very hot compared to a day in Singapore with the same average temperature. The reason for that is that over a year in Sydney’s climate, such days occur less frequently than in Singapore’s tropical climate. Accordingly, HVAC demand is expected to peak in days with relatively high average daytime temperature for a specific climate. The temperature node of the algorithm

118 compares the predicted mean temperature within the potential peak period to the mean summer temperature at the specific location as derived from historical data. The comparison is made for the same time frame (i.e. the “hottest” zone of the day as predicted by the ensemble, which is typically close to the afternoon).

The third factor affecting the magnitude of peak loads is the mean relative humidity within the potential peak period (as predicted in section 4.2.1). In a similar manner as above, the humidity decision node compares the predicted mean relative humidity within the potential peak period to the mean summer relative humidity at the specific location as derived from historical time series data. In the case of Sydney, Australia, lower relative humidity indicates a higher likelihood of a significant peak load (as discussed in 2.2.4).

Finally, the fourth factor is linked to the Cooling Degree Hours (CDH). The concept of CDH is used broadly in the research field of building demand and is a useful metric to describe the total cooling needs over a period of time. In the context of peak load predictions, it is expected that high values of CDH over a day, would be associated with increased cooling loads. The CDH can be calculated in respect to an upper bound comfort temperature, which depends on the climate, the building’s characteristics and occupant activity levels. For example in Sydney, the upper comfort temperature for a typical commercial building with light occupant activity (walking or working in an office) is 25°C. However, as the cooling system must remove heat generated from internal sources and radiative gains, the CDH may be calculated based on a lower ambient temperature of 21°C. Hence the CDH for a case study building in the Sydney climate can be calculated as:

퐶퐷퐻 = ∑(푇푖 − 21) (6) 푖=1

Ti represents the prediction for temperature at the hour i. Any negative values of CDH are disregarded, as they would imply no cooling is necessary during that time. As opposed to the first two nodes, the CDH is calculated for the whole 24-h period of the forecast, instead of only within the potential peak periods. This is because the CDH factor is related to the components of peak loads with higher temporal periods – for instance a few hot hours in the morning would increase the overall demand 119 and hence raise the baseline for the afternoon peak period of the same day. Similarly to the previous nodes, the CDH is compared to mean summer historical figures for the location.

The outputs from the three nodes are used to estimate the likelihood of a significant peak occurring in the building for each day, under a simple assumption. The assumption states that significant peaks would occur during the times that weather conditions favour higher demand (high mid-day temperatures, low mid-day relative humidity and high daily CDH) compared to the historical averages. Hence when all three weather factors are favourable, there is a greater probability of a significant peak occurring. When fewer favourable factors are present in a single day, the probability of a significant peak would be accordingly lower. It should be noted that this hypothesis assumes that non-weather related factors are controlled: occupancy levels, activity type and HVAC control policies remain constant in order to identify the effects of weather on peak loads alone.

The classification of each summer day according to the proposed model assigns “high”, “moderate”, “low” or “negligible” chances of significant peak occurrence. These qualitative characterisations correspond to three, two, one or no factors among the mean peak period temperature, mean peak period relative humidity and daily CDH favouring the presence of a significant peak. As shown earlier, high temperature, low relative humidity and high CDH can all be regarded as favourable weather factors for the occurrence of high peak cooling loads.

The outline of the qualitative peak magnitude estimation algorithm can be seen in Figure 38.

120

Figure 38: Significant peak detection algorithm 121

The algorithm of Figure 38 offers the foundation for assessing the likelihood of a significant peak occurring without the need for any historical load data. Its advantage is that it offers a quick insight of the relative magnitude of a peak with no need of any historical weather or load data (as weather simulations may be run in hindcast mode in the numerical prediction software).

Where historical load data are available, a more precise quantitative estimation of the relative magnitude of daily peak loads in the building is possible. The prediction is based on studying the correlation of daily cooling peak loads with each one of the weather-related factors independently and assuming that the non-weather related factors are constant and their statistical distribution. The consideration of two time series is required: the first time series contains the historical daily peak loads of summer days (weekdays and non-holidays only) of the building in question. The second time series contains the historical values of mean peak period temperature, mean peak period relative humidity and daily CDH for the same period and the same location as the building, calculated according to the methodology described in section 4.2.1.

After the two time series are processed, it is possible to obtain the correlation coefficient (r) as an indication of any effects of weather factors on building loads. The correlation coefficient may be obtained from the time series easily using a statistical software such as Excel or R. The correlation coefficients between each weather variable and the building load can serve as weights for the predictions of any future significant peaks. This means that each weather variable may affect the load to a different extent for a specific building. Hence, the “high” likelihood of a significant peak occurring may be a result of two important weather factors (with high r) being very favourable, instead of all three.

To quantitatively predict the level of peak load for a specific day, firstly the z-scores of predicted mean peak zone temperature, mean peak zone relative humidity and daily CDH are determined. The mean and standard deviation required to determine the z-scores are obtained from the historical time series data of each factor. The z- scores for each of the three factors are then multiplied by their respective weights (correlation factors) in order to account for their respective significance in affecting the load. 122

For example, let the correlation factor for temperature be 0.6 and for CDH 0.4, and that on for a single day predicted temperature and CDH are both having a z-score of 1. The weighted z-score should then be 0.6 for temperature and 0.4 for CDH.

Finally, analysis of the historical time series data can reveal the statistical distribution of daily peak loads according to the weighted z-scores for each factor at each time step (day). A factor may be considered “favourable” when its weighted z- score is higher than 1.

In the above example, temperature conditions would be considered favourable if the z-score of the predicted mean peak zone temperature was 1.67 (hence the weighted z-score would be1.67 × 0.6 = 1). On the other hand CDH conditions would be considered favourable if the z-score of the predicted daily CDH was 2.5 (hence the weighted z-score would be2.5 × 0.4 = 1). This practically means that because higher temperatures were found to have a more profound effect (r=0.6) in peak loads than CDH (r=0.4), a potential significant peak is identified for relatively lower values of temperatures than CDH. It should also be noted that for relative humidity, the absolute value of the weighted z-score may be used, as it may demonstrate a negative correlation with peak loads in the summer.

The final step is determining four distinct levels of statistically expected daily peak, according to the number of favourable factors (weighted z-score higher than 1). For this step the historical time series are required once more. Each day in the time series is classified according to the number of favourable factors present and as such four groups are generated:

1. High chance of significant peak load (3 favourable factors) 2. Moderate chance of significant peak load (2 favourable factors) 3. Low chance of significant peak load (1 favourable factors) 4. Negligible chance of significant peak load (0 favourable factors)

The average daily peak loads are calculated for each of the groups and function as the prediction (expected) peak load levels.

An application of this methodology in a case study building is presented in section 4.3.2.

123

4.3 Case study & results The case study building for the peak load forecasting model is a 6 Star Green Star Design rating (the highest possible rating in Australia) university building in Sydney, Australia. The Tyree Energy Technologies Building (TETB) in the University of New South Wales (UNSW) campus is a five-level, 15,000 m2 floor area building opened in 2012 (University of New South Wales Facilities Management, 2015). The building has 4 levels and includes offices, a few labs and several amphitheatres and classrooms. There is also a coffee shop operating during the day.

The numerical simulations covered a period of 2 years (2012-2014) and the outputs from the algorithms described in 4.2.1 and 4.2.2 were used to detect potential peak loads in the building and their significance. The validation of the model was carried out by comparison of the predictions from simulations with the actual peak loads of TETB.

4.3.1 Potential peak period detection For predictions of potential peak load periods the algorithm described in 4.2.1 was used. For summer peak loads (arbitrarily chosen between 15th November and 15th March for the case study region) the algorithm predicted average potential peak periods of 4.5 hours in each day. In these periods, predictions indicated high temperatures and low relative humidity and hence a potential peak in building load. The results showed that the actual peak load in the building occurred within these periods 90.6% of the days. For the days that the peak load was not in the predicted period, the average error was 1.25 hours. This means that the actual peak occurred within 1.25 hours on average before or after the predicted peak period.

Figure 39 and Figure 40 illustrate the results for peak load predictions for a sample 3-day summer period with consecutive days of temperatures greater than 30 degrees. During these days in the Sydney climate, cooling peak loads are highly likely. The temperature and relative humidity values in Figure 39 are the predictions calculated from the algorithm described in Section 4.2.1. Finding the potential peak periods takes the distribution of ensembles into consideration. The green zones in both figures represent the potential peak period for each day as predicted by the algorithm of Section 4.2.1. In Figure 40 it is possible to see the actual building load

124 at each hourly time step for the same 3-day period. The peak daily load falls within the predicted peak period for all three days.

Figure 39: Prediction of potential peak periods according to temperature and relative humidity over 3 day period

Figure 40: Building load over three day sample period, including the actual peak load and the potential peak load periods as predicted by TAPM

125

Figure 41 shows the distribution of branch predictions of temperature for the same time period. It can be seen that only the time steps with a high upper quartile of branch predictions are identified as potential peak periods (yellow zones) in the first decision making step (temperature clause).

Figure 41: Distribution of branch predictions for temperature over a three day sample period, including the potential peak period for each day 4.3.2 Significant peak detection and relative magnitude predictions In order to determine whether the daily peak is significant and whether several demand responses should be implemented, access to historical data and statistical analysis was required. For this case study, the daily TETB building peak loads over the summer period (excluding holidays) were correlated with the observed mean peak period temperature, mean peak period relative humidity and CDH from the UNSW weather monitoring logs (taken from a weather station on top of a building 500 meters from the TETB) over the period 2012-2015 (according to the proposed methodology in section 4.2.2).

The distribution of daily peak loads for both summers can be seen in the following figures, plotted against the predicted mean peak period temperature (Figure 42), against the mean peak period relative humidity (Figure 43) and against the daily CDH (Figure 44). It should be noted that the daily CDH – load correlation had 2 126 outliers of 120+ CDH in a single day, with relatively moderate peak loads. These were removed from the figure, as during hot days a bit part of the HVAC load is covered from the UNSW campus network rather than the case study building.

Peak load correlation with temperature 700

600

500

400

300

Dailypeak load(kW) 200

100

0 15 20 25 30 35 40 Mean peak period temperature

Figure 42: Distribution of daily peak loads in relation to mean peak period temperature (recorded onsite)

127

Peak load correlation with relative humidity 700

600

500

400

300

200 Dailypeak load(kW)

100

0 0 10 20 30 40 50 60 70 80 90 100 Mean peak period relative humidity

Figure 43: Distribution of daily peak loads in relation to mean peak period rel. humidity (recorded onsite)

Peak load correlation with CDH 700

600

500

400

300

DailyPeak Load(kW) 200

100

0 0 10 20 30 40 50 60 70 80 CDH in a single day

Figure 44: Distribution of daily peak loads in relation to daily CDH (recorded onsite)

128

The results as seen in Figure 42, Figure 43 and Figure 44 showed that the correlation coefficient of daily peak load with mean peak period temperature was 0.58, the correlation coefficient of daily peak load with mean peak period relative humidity was -0.53 and the correlation coefficient of daily peak load with daily CDH was 0.36 respectively. The negative correlation coefficient between daily peak load and humidity shows that peak loads are actually higher when relative humidity is at low level. This validates the assumptions made in 4.2.2 about the effects of the three factors on the magnitude of the peak loads.

The main source of error in peak load predictions can be attributed to the fact that the TETB HVAC system is not handling the cooling load of the building exclusively; part of the cooling load is managed externally by the UNSW campus network and hence is not logged in the TETB building load. Another notable characteristic of the TETB is that being a new building, the occupancy levels in its first year (2012) were lower, which was reflected at the energy demand and peak loads. Hence the correlation coefficients were derived separately for each summer period (2012 and 2013) and averaged for the purposes of this study. Another possible reason for the presence of outliers and errors may be attributed to significant variations in any of the non-weather factors that may have in turn affected the peak load. For instance, there may have been an exhibition or exam in the building that may have raised the occupant number to much higher levels than usual – in turn this may have had a significant impact on cooling load.

Using the significant peak detection algorithm and the ensemble predictions for each day, it was finally possible to classify each summer day as showing “high”, “moderate”, “low” or “negligible” chances of significant peak occurrence as described in section 4.2.2. For the TETB summer period 2012-13, it was found that the magnitude of the daily peak loads for each level were:

 Three favourable factors were associated with daily peak loads that were at least 44% higher (452 kW) compared to the average baseline daily peak (no favourable factors – 314 kW)

129

 Two favourable factors were associated with daily peak loads that were at least 33% higher (419 kW) compared to the average baseline daily peak (no favourable factors– 314 kW)  One favourable factor was associated with daily peak loads that were at least 22% higher (383 kW) compared to the average baseline daily peak (no favourable factors – 314 kW)  No favourable factors were associated with the baseline daily peak loads of 314 kW.

Based on historical data, the detection algorithms and the ensemble outputs were able to predict the presence of all significant peaks for the periods that the model was run. The comparison between the predicted daily peak loads from the algorithm explained in 4.2.2 and the actual daily peak loads for the two summer periods separately can be seen in Figure 45 and Figure 46. The difference in levels between summers of 12-13 and 13-14 is due to the building being rather new the first period and hence had a lower number of occupants and activities going on.

Peak load predictions summer 2012-13

500

450

400

350

300

250

200

Dailypeak load(kW) 150

100

0 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 Day

Actual load Predicted load -314

Figure 45: Comparison of predicted and actual daily peak loads for the 2012-13 summer period in the TETB

130

Peak load predictions summer 2013-14 700

600

500

400

300

200 Daily peak load (kW) load peak Daily

100

0 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 Day

Actual load Predicted load

Figure 46: Comparison of predicted and actual daily peak loads for the 2013-14 summer period in TETB In these figures, the distinct levels of relative peak load magnitude can be seen. These levels correspond to the number of “favourable factors” predicted for each time step (day). There were certain periods, where a significant peak was predicted but did not occur. Specifically for 18 out of 148 days of the simulation period, the model indicated a false positive peak cooling load. The magnitude of the daily peak loads was also fairly accurately predicted and the mean absolute percentage error was 11.4%. Notably the error in almost all cases was overestimating the magnitude of the peak load.

To illustrate the advantage of the proposed generation of onsite ensemble predictions with the numerical model, both algorithms (peak timing and peak relative magnitude) were run with weather forecast inputs from the Bureau of Meteorology, as taken from the closest weather station to the TETB case study building (about 6km away from the building, in Sydney Airport). It was shown that for the same period, the numerical onsite predictions resulted in higher accuracy compared to weather station data. Interestingly, both iterations resulted in the same number of false positive predictions of significant cooling load peaks at the same

131 periods (hot, dry days but with relatively low actual load). The comparison can be seen in Table 31.

Table 31: Comparison of forecasting skill depending on weather input source

Onsite ensemble Weather numerical station data predictions Accurate predictions of peak load timing 90.6% 40.4% Peak load magnitude MAPE 11.4% 20.6% False positive significant peaks 18 18

4.4 Discussion of the peak load prediction model The proposed model in this chapter aims to provide a platform for the prediction of peak cooling loads in buildings, by utilising an ensemble of localised weather forecasts and two statistical algorithms. The predictions convey important information for building energy management regarding the time of occurrence of a peak load, as well as the likelihood that this peak is high, and hence associated with significant costs.

Regarding the timing of daily peak loads, the results showed that model was able to predict 4.5 hour long periods on average, within which the peak load was expected to occur (potential peak periods). In simple terms this indicates that the daily cooling peak would occur within a period of 4.5 hours on average, as predicted by the first algorithm and according to the case study simulations. The prediction was successful for 134 out of 148 days in the simulation period; for the rest of the days when the peak load actually occurred outside the prediction zone, the average timing error was slightly over 1 hour. The ensemble was able to predict shorter periods of potential daily peaks during periods with sudden changes in weather conditions. The reason for this, is that under these circumstances, fewer time steps (hours) would trigger a potential peak over a single day. Accordingly, in days with relatively low rates of change of temperature and relative humidity, the ensemble would indicate longer intraday potential peak periods.

The main implication of these results is associated with the probabilistic nature of ensemble forecasting. Specifically, a deterministic prediction about the exact time 132 that a peak cooling load may occur within a day would not be possible solely based on a weather forecast. This was confirmed by the results, as individual ensemble branches were unable to consistently produce a deterministic prediction of the peak load time.

Furthermore, demand response measures in buildings are proactive and typically span over a period of time within a day, rather than being switched on and off instantaneously; examples of such measures include shutting down certain units, increasing the AC set points, committing onsite energy generation sources or rescheduling energy flows within the building. Hence the proposed ensemble model’s temporal predictions about the occurrence of peak loads align closely with the typical time frames that demand responses are triggered and implemented.

An advantage of the ensemble prediction model is its flexibility, as prediction accuracy can be traded off with reduced length of the potential peak period. According to the proposed algorithm in 4.2.1 the peak period detection was based on comparing each branch to the extreme daily values within a range of 10%. This figure was in turn derived from the forecasting skill of the numerical model when used for this kind of predictions from chapter 3. Reducing the range to lower values than 10%, would exclude some ensemble branches from triggering a potential peak and hence prediction accuracy would decrease. However, the average potential peak period would be shorter, as less branches would be considered.

As discussed above, not all summer cooling peak loads are equally significant in magnitude. The reason for that is that in certain climates, like Sydney, several summer days are associated with temperatures in the low 20s and hence the daily cooling demand would be low. Hence, while the first algorithm would detect a potential peak for such “colder” summer days, it would not be significant to the energy management system, and would not contribute to high energy costs. The second proposed algorithm, as described in section 4.2.2 aims to classify each day’s predicted peak load according to the likelihood of being high and hence causing high energy costs. Some information about the historical peak loads of the building are necessary for the statistical analysis upon which the second algorithm is based. Specifically, historical data of daily summer peak loads were classified according to

133 the number of favourable factors present (high temperature, low relative humidity and high number of CDH). It was found that for the case study building there were distinct bands of peak load magnitudes according to the number of favourable factors. The results showed that all significant peaks could be successfully predicted for the two summer periods of the simulation, however there were also a few false positive predictions. This would mean that the predicted high cooling peak load that the model predicted did not actually occur. The impacts of false positives and the implementation cost of demand responses will be investigated in chapter 6.

The results indicate the necessity for inclusion of additional factors for more accurate prediction of cooling peak loads that are non-weather related, such as past load values as time-series data, occupancy patterns, HVAC operation policies and onsite generation capacity. The inclusion of such factors in the prediction algorithms would not only reduce the probability of false positives, but would also improve the predictions regarding the actual peak load magnitude. The proposed model utilises historical building load data to produce predictions of four distinct levels of peak loads. While this is sufficiently accurate for most days, as seen in Figure 46, there are several instances, where the model correctly predicted a significant peak load, but underestimated its magnitude. The reason for this type of error is that the algorithm predicting the relative magnitude of a daily peak based on statistical averages of past peak loads and their correlation with weather factors. The effects of the non-weather related factors add extra layers of complexity in capturing the exact magnitude in this weather-based model, and as such there are instances that the magnitude of the actual peak load is higher than predicted.

As seen in the results the weather related factors forming the basis of the prediction algorithms may be complemented by non-weather related factors to alleviate the shortcomings explained in this section and improve the peak load magnitude prediction accuracy; and this will be addressed in future work. It should be noted, that inclusion of additional factors is of course expected to increase the complexity of the algorithms.

134

4.5 Conclusions This chapter investigated a novel approach for the prediction of peak cooling loads in buildings. The predictions are based on the generation of an ensemble of day ahead weather forecasts for a case study building in Sydney, Australia. The outputs of the ensemble are then used in two statistical algorithms.

The first algorithm pinpoints the potential time period within the next day that the peak building load would occur. The prediction is based on the distribution of the results from the ensemble branches and consideration of the predicted temperature, relative humidity and cooling degree hours. In the case study building and for a period of two years’ worth of simulations it was shown that the actual peak occurred within the predicted period for 90.6% of the period under consideration. The average potential peak period for each day was roughly 4.5 hours long.

However, not all peak loads are important to energy management – typically, building energy costs are increased only by the highest peak loads over a period of time. Hence, the second algorithm aims to classify each predicted daily peak according the chances it has of being significantly high. This probabilistic approach is inherently tied to the generation of an ensemble of forecasts. Once more, the distribution of the ensemble outputs, the predicted temperature, relative humidity in and cooling degree hours were considered. In addition the second algorithm utilises statistically processed historical data from the building peak loads to generate forecasts of peak load magnitudes. It was found that all significant peak loads were detected successfully for the case study building over the simulation period; however there were 18 false positives over the period of 148 simulation days. Regarding the peak load magnitude, the mean absolute percentage error was around 12%.

While the algorithm provides a novel and robust approach to peak load prediction for any building based on weather predictions, there is certainly room for improvement in terms of prediction accuracy and minimising the amount of erroneous peak load detections (magnitude and time-wise). Future research may attempt to incorporate non-weather related factors, linked to building

135 characteristics and statistical analysis over longer periods in order to improve the robustness of the predictions.

136

5 Weather analysis tool for evaluation of the precooling potential

 Design of a tool to assess precooling potential for a given location  Assessment based on numerical analysis of past weather trends, independent of building characteristics  Comparison of different sites with three dimensionless ratios  Best precooling potential in sites with high diurnal temperatures confirmed  Additional favourable factors include higher distance from the coast and lower urban density 5.1 Tool design outline Passive energy management is of particular interest for building energy systems, as clever design or operation policies can result in significant financial or energy savings at little or no cost. Examples of passive measures include the optimisation of orientation of the building related to the sun’s path, the utilisation of natural lighting or natural ventilation, using the thermal mass of the building to discharge energy with a phase lag and the preconditioning of buildings. Preconditioning of buildings describes a set of measures that make it possible to shift parts of the HVAC load to different time periods in order to save on energy costs (and if done without any energy expenditure to reduce total consumption as well). Precooling a building during the summer the night before using natural ventilation is an example of a common preconditioning practice (Rabl and Norford, 1991).

In this chapter, a tool is proposed that aims to help with assessing the potential of precooling at a specific site, based on historical weather data. It should be emphasised that in this thesis, “potential” refers to the theoretical precooling that may be utilised at a given location. The calculation of the potential ratio is explained clearly in section 5.2.4.1.

Knowing this information can assist both designers of new buildings to add features that take advantage of high precooling potentials if available, as well as building managers seeking to upgrade existing BEMS in order to improve the energy performance. As with the models in previous chapters, the use of TAPM is proposed. In addition to forecasting, TAPM can be used to analyse past synoptic weather data for any location and allow the extraction of information about any patterns. The 137 proposed model is validated with the analysis of 7 years’ worth of data across 5 sites in the Sydney metropolitan area, in New South Wales, Australia.

Section 5.2 will provide the details of the model design. This includes information about the simulation parameterisation in TAPM (section 5.2.1), the site selection (section 5.2.2) and the development of a range of ratios that would help assess the precooling value (section 5.2.4) based on a set of predefined conditions (section 5.2.3). Following that, the results will be thoroughly presented and discussed in section 5.3, including a section dedicated to analysing the effectiveness of the model in different climates in Australia and globally (section 5.3.7).

It is hypothesised that the theoretical precooling potential varies significantly for different areas. In addition to its applications in BEMS, the tool attempts to extract some findings about the effects of a range of factors in regards to precooling potential. Specifically, the primary factor is expected to be the magnitude of diurnal temperature differences in the summer. Additionally, factors such as the proximity to large bodies of water, urban density and vegetation are considered in the analysis.

The assessment of the theoretical precooling potential, as defined in this chapter at a given location is derived from the local climatic conditions and is independent of the building characteristics. The performance ratios involved in the analysis are dimensionless and allow for easy comparison across various locations. The proposed model is tailored for precooling during the summer, which is more meaningful than preheating during the winter. The reason for that is in almost any climate, the temperatures during the non-occupied period (overnight) are lower than the occupied period (daytime) throughout all seasons. Hence any space may be cooled down during the night, which makes sense in the summer, but not “warmed up” during the winter nights.

It should also be noted that large portions of the work presented in this chapter have been accepted for publication in a paper for the Energies journal. At the time of writing of the thesis the publication process of the paper was at the final stage, but not yet complete.

138

5.2 Model design 5.2.1 Simulations outline The proposed tool architecture is based on analysing downscaled synoptic data from past years at the location in question in order to develop a statistical analysis of the local climatic conditions. The database is then used to assess the precooling potential in the area.

Three nested domains were used in the simulations. Domain #1 has dimensions 40.5km × 40.5km, domain #2 has dimensions 13.5km × 13.5km and domain #3 has dimensions 4.5km × 4.5km. The resolution and number of domains is lower compared to the forecasting algorithms of the previous chapters in order to minimise computational time without compromising forecasting accuracy. Specifically, selecting three domains results in an average decrease of computational time of roughly 25% compared to a five domain simulation of similar lowest order resolution. At the same time, the accuracy of predictions is comparable: for ambient temperature the MAPE increased from 2.5% to approximately 3.3% when three domains were used instead of five. The selected configuration generates forecasts at low simulation times (average time of 15 minutes per simulation day in a mid- end desktop computer). The number of grid points in each domain for the model (x,y,z) was 25 × 25 × 25. The simulations were run at each site in hind-cast mode for seven years at five different locations. The primary output from the simulations consists of the ambient (dry bulb) temperature averaged for each hourly period of each year. The data processing to obtain meaningful findings for use within building energy optimisation was implemented via algorithms developed in Matlab.

5.2.2 Site selection The simulation sites belong to the Sydney metropolitan area and were selected to represent different weather locales. Figure 47 shows the locations of the sites, with the main factor during site selection being the proximity to the coast. Built environment density and vegetation were also considered.

139

Figure 47: Map of the five Sydney sites The traits of each site are summarised in Table 32.

Table 32: Characteristics of the five sites

Location Distance to coast Urban density Vegetation Site 1 Bondi Coastal High Low Site 2 Airport 1 km Low Low Site 3 Canterbury 5 km High Low Site 4 Bankstown 12 km Moderate Moderate Site 5 Penrith 40 km Low Moderate

It should be noted that the urban density (Adams et al., 2015) and vegetation characterisations have been derived from TAPM’s database for each respective site. Specifically, TAPM’s database contains information about the vegetation type and leaf area index (area of green leaf coverage per unit ground area). Vegetation may affect the local climate in various complex ways, but generally speaking, it may reduce the warming impacts of the greenhouse effect (due to the increased absorption of CO2) and also increase radiation reflection due to high canopy albedo (Brovkin, 2002). Furthermore, the presence of trees provides shading in their immediate vicinity, which may reduce the ambient temperature compared to surfaces that are exposed to direct sunlight. Finally, humidity exchange between the plants and the atmosphere (due to respiration from leaves) is found to further affect 140 the local microclimate and slightly reduce daytime temperatures (Armstrong et al., 2016).

While the five sites are within a radius of about 35 km, the diurnal temperature differences vary and it is hypothesised that this will result in differences in evaluation of the precooling potential.

5.2.3 Precooling conditions In order for the tool to assess the potential for precooling, it needs to be configured to detect favourable conditions for the process to occur. For Sydney, during the summer months (15th November to 15th March) these conditions are met when:

 During the occupancy times (08.00 – 20.00) of a weekday, the ambient

temperature rises above the upper boundary of acceptable indoor

temperature (25°C)

 During the previous night (20.00-08.00), the ambient temperature falls

below the lower boundary of acceptable indoor temperature (20°C)

 Relative humidity is maintained between 25 and 85%

These boundaries reflect typical Australian office indoor comfort conditions (WorkSafe Victoria, 2008), adjusted for the effects of the thermal mass (attenuation, temporal lag) and internal heat gains. Internal heat gains are highly variable and depend on the number of occupants, the occupant activity, operation of electronic and electrical devices, as well as the space layout (Oldewurtel et al., 2012). It should be noted that the thermal comfort of an individual depends on various factors as well, such as the air temperature and relative humidity and are not constant throughout the year or even the course of a day. Other non-weather related factors include the ventilation rate, metabolism, clothing, and type of activity (Havenith et al., 2002).

When the precooling conditions are met, energy may be saved by precooling the building the night before so that the cooling load of the daytime is reduced as discussed earlier. The precooling conditions were derived in relation to the comfort 141 zone for each climate and the dampening of the changes in weather variables from outdoor to indoor environment (as discussed in section 3.2.4). Additionally, the precooling conditions were established assuming light activity of occupants (representing typical commercial building occupant activity).

5.2.4 Precooling ratios The proposed methodology in this chapter, which may be used to assess the precooling potential and its theoretical value according to the ambient weather conditions, is based on a range of statistical ratios. The ratios are in turn based on the concept of degree-hours (DH). The base temperatures are set according to the precooling conditions described in section 5.2.3. The advantage of modelling with DH compared to other regression or physical models is the relative simplicity and direct correlation of the building’s load to the weather conditions (Krese et al., 2011, Layberry, 2008).

Three DH metrics were used in the development of the precooling ratios:

 DHL: Represents the amount of degree-hours that the ambient temperature falls below the lower boundary (20°C) during the night non-occupied period

 DHU: Represents the amount of degree-hours that the ambient temperature rises above the upper boundary (25°C) during the daytime occupied period

 DHN: Represents the amount of degree-hours that the ambient temperature is below the upper boundary (25°C) during the daytime occupied period Additionally, the number of days that the precooling conditions are met in a single year (n) is used in the evaluation of the precooling ratios.

An illustration of a daily profile showing the calculation of the DH values for a sample day can be seen in Figure 48.

142

Figure 48: Daily temperature profile (as simulated) for Penrith between 27th January 20:00 and 28th January 20:00 in 2014. The upper and lower boundaries can be seen at 25 and 20 degrees respectively. The DH for each time step are calculated as the difference from the base temperature at each period.

For the sample day displayed in Figure 48, the DHL is calculated as the sum of the differences of the ambient temperatures below the lower threshold (20°C) at each time step during the vacant period (time steps 0 to 12). The DHN and DHU are calculated as the sums of differences of the ambient temperatures below and above the upper threshold (25°C) respectively, at each time step during the occupied period (time steps 13 to 24). Table 33 shows the actual calculations for each time step of the same sample day, in order to illustrate the methodology of deriving the DH values.

143

Table 33: Calculation of the DH values for each time step for the sample day. The summation of the values in the end represents the values of DHL, DHN, and DHU used in the precooling ratios.

Time step Ambient Temperature DHL DHN DHU 1 21.4 0 N/A N/A 2 20.1 0 N/A N/A 3 19.1 0.9 N/A N/A 4 18.5 1.5 N/A N/A 5 17.9 2.1 N/A N/A 6 17.5 2.5 N/A N/A 7 17.1 2.9 N/A N/A 8 17 3 N/A N/A 9 16.9 3.1 N/A N/A 10 16.8 3.2 N/A N/A 11 16.9 3.1 N/A N/A 12 18.4 1.6 N/A N/A 13 20.8 N/A 4.2 0 14 23.1 N/A 1.9 0 15 25 N/A 0 0 16 26.6 N/A 0 1.6 17 27.8 N/A 0 2.8 18 28.7 N/A 0 3.7 19 29.2 N/A 0 4.2 20 29.3 N/A 0 4.3 21 29 N/A 0 4 22 28.5 N/A 0 3.5 23 27.1 N/A 0 2.1 24 25 N/A 0 0 Total - 23.9 6.1 26.2

5.2.4.1 Precooling potential

For the days when precooling conditions are met, the precooling potential ratio (r1) can be defined as:

144

퐷퐻 (𝑖) ∑푛 퐿 푖=1 퐷퐻 (𝑖) 푟 = 푈 (7) 1 푛

The potential for precooling would be higher for days preceded by cooler nights

(temperature drops below 20° C overnight and hence large value of DHL), followed by relatively low cooling loads during the occupied times (temperature rises only a few degrees above 25° C and hence lower value of DHU). During very hot days (high value of DHU), the potential for precooling would be lower, as any overnight cooling would only be able to cover a minor part of the much larger cooling load during the daytime.

This is represented in the sum of fractions in the numerator of equation 7 for each hourly time step i. The ratio r1 is averaged across all precooling days by dividing the sum by the number of days per year that the precooling conditions were met (n) (as calculated from the weather analysis of historical data).

If the potential ratio is close to 1, that means that the cooling overnight (DHL) is just enough to cover the needs for cooling the day for the times that the temperature rises above 25°C (DHU). However, higher values of r1, are not necessarily useful, as they may result from relatively cold summer days (which would have a low value of

DHU), during which the building cooling requirements are low anyway.

5.2.4.2 Precooling utilisation

In addition to r1 (potential ratio), it is possible to calculate the utilisation ratio r2.

This refers to the ratio of precooling degree-hours during the night before (DHL) divided by the degree-hours that ambient temperature is below the upper boundary during the daytime (DHN):

퐷퐻 (𝑖) ∑푛 퐿 푖=1 퐷퐻 (𝑖) 푟 = 푁 (8) 2 푛

A higher utilisation ratio would indicate that most of the precooling was useful in mitigating cooling loads. This is because the denominator value of equation 8 (DHN) would be small during hot days that temperature stays over 25°C for most of the time. Hence any precooling overnight (proportional to DHL) would contribute towards reducing the following day’s cooling loads.

145

If on the other hand, the utilisation ratio is small, it implies that the precooling occurring the night before (DHL) was not making significant impact in reducing the day’s cooling load. This can be observed when the ambient temperature does not rise to high levels for long periods during the daytime (hence larger value of DHN). During these days, the cooling load requirements would be comparatively lower to a hot summer day. As described earlier in section 5.2.4.1, this may result in very high values (above 10-15) for r1. This can also be observed when the temperature does not drop below 20°C for long overnight (hence low value of DHL). As with r1, r2 is averaged across all precooling days by dividing by n.

5.2.4.3 Precooling frequency and annualised ratios Another important ratio that can be derived from analysis of the historical weather data for the location is the precooling frequency (h), which is defined as the ratio of days that the precooling conditions are met (n) over the number of days in a year. The precooling frequency is used to annualise both the precooling and utilisation ratios. Annualising is necessary as the ratios by themselves convey information only about precooling days (when the conditions of section 5.2.3 are met), and do not include any component of the frequency they occur throughout a year. Hence the annualised forms of both ratios (R) are calculated by multiplying each one with the precooling frequency ratio h.

푅1 = 푟1ℎ (9)

푅2 = 푟2ℎ (10)

The physical significance of equations 9 and 10 is that they allow the comparison of climatological conditions on an annual basis, as opposed to the ratios r1 and r2, which can be used for comparisons only for the days that the precooling conditions are met. This is necessary, as for some sites the precooling potential may be high, but the precooling conditions are not met often, and hence the overall potential would not be significant throughout a year.

5.2.4.4 Theoretical precooling value

While both ratios r1 and r2 can assist in making useful conclusions about the precooling potential at a given site individually, there may be certain challenges in interpreting the theoretical precooling value. An example would be a situation 146 where the potential may appear too high during summer days that are relatively cool and preceded by nights that are significantly cooler than usual. Another example may be a situation where the utilisation may appear too high during days that the diurnal differences in temperature are symmetrical either way of the boundaries, but remaining within a narrow zone around the comfort band. In such cases, the precooling may not have the significant impacts that the ratios suggest.

Hence, a value ratio can be calculated that incorporates dimensions of both the potential and utilisation of precooling. The theoretical value of precooling v at a particular site is calculated as the average ratio of the difference of degree hours the temperature falls below the lower boundary the night before (퐷퐻퐿) minus the degree hours that the temperature does not rise above the upper boundary the day after (퐷퐻푁) over the degree hours that the temperature rises above the upper boundary the day after (퐷퐻푈).

퐷퐻 (𝑖) − 퐷퐻 (𝑖) ∑푛 퐿 푁 푖=1 퐷퐻 (𝑖) 푣 = 푈 (11) 푛

A negative numerator in equation 11 indicates that DHN is larger than DHL for the days that precooling conditions are met. In simple terms, this occurs when temperatures during most of the summer days are relatively cool (below 25°C) making DHN large and at the same time the temperatures during the nights before remain relatively high, close to the lower boundary of 20°C, making the value of DHL small. The more negative the ratio v, the less value can be realised in precooling. On the contrary, a positive numerator shows there is a significant diurnal temperature differences and that the precooling that occurs overnight can be utilised the day after to reduce the cooling load. The bigger the ratio, the more value can be realised.

It should be noted that equation 11 shows the theoretical precooling value, based on the analysis of the historical weather trends. This can be regarded as the maximum value that may be extracted via the process of precooling, but in practice the actual value may be limited by the building’s characteristics and BEMS policies. Chapter 5 is concerned exclusively with the theoretical precooling value based on weather analysis, however a case study attempts to link the findings to real value in an existing building in Chapter 6. 147

5.3 Simulation results and discussion 5.3.1 Results summary The results of the simulations for the frequency, potential and utilisation ratios are summarised for all sites across Sydney in Table 34.

Table 34: Summary of results across five sites in Sydney for each simulation year

Site 1 - Bondi 2014 2013 2012 2011 2010 2006 2005 h (frequency) 0.06 0.08 0.05 0.04 0.07 0.09 0.06 n (Precool days) 21 28 19 16 26 34 22

r1 (potential) 24.8 41.0 7.9 22.6 17.6 10.3 10.4

R1 1.4 3.2 0.4 1.0 1.3 1.0 0.6

r2 (utilisation) 0.54 0.64 0.43 0.41 0.36 0.48 0.44

R2 0.03 0.05 0.02 0.02 0.03 0.04 0.03

Site 2 - Airport 2014 2013 2012 2011 2010 2006 2005 h (frequency) 0.14 0.16 0.14 0.09 0.13 0.17 0.17 n (Precool days) 52 57 51 34 48 63 60

r1 (potential) 23.5 8.8 14.1 10.0 14.4 5.8 9.1

R1 3.4 1.4 2.0 0.9 1.9 1.0 1.5

r2 (utilisation) 0.81 1.05 0.87 0.76 0.77 0.98 0.95

R2 0.12 0.17 0.12 0.07 0.10 0.17 0.16

Site 3 - Canterbury 2014 2013 2012 2011 2010 2006 2005 h (frequency) 0.17 0.20 0.17 0.12 0.19 0.21 0.23 n (Precool days) 60 72 61 43 70 76 84

r1 (potential) 19.0 9.8 18.6 2.8 13.7 9.7 25.0

R1 3.1 2.0 3.1 0.3 2.6 2.0 5.8

r2 (utilisation) 1.17 1.25 1.09 0.92 1 1.42 1.2

R2 0.19 0.25 0.18 0.11 0.19 0.3 0.28

Site 4 - Bankstown 2014 2013 2012 2011 2010 2006 2005 h (frequency) 0.21 0.24 0.20 0.15 0.24 0.26 0.28 n (Precool days) 78 88 73 54 87 94 101

r1 (potential) 11.0 19.7 14.7 6.7 16.5 7.7 9.1

R1 2.4 4.8 3.0 1.0 3.9 2 2.5

r2 (utilisation) 1.36 1.45 1.42 1.17 1.24 1.64 1.61

R2 0.29 0.35 0.28 0.17 0.3 0.42 0.45

Site 5 - Penrith 2014 2013 2012 2011 2010 2006 2005 h (frequency) 0.21 0.22 0.18 0.13 0.19 0.21 0.28 n (Precool days) 78 80 66 47 70 78 103

r1 (potential) 20.9 6.9 10.2 9.4 24.0 5.6 12.5

R1 4.5 1.5 1.9 1.2 4.6 1.2 3.5

r2 (utilisation) 1.76 1.87 1.74 1.3 1.35 2.01 2.01

R2 0.38 0.41 0.31 0.17 0.26 0.42 0.57 148

The results were compared to Typical Meteorological Year (TMY) Sydney data from the US Department of Energy (DoE, 2015) . The data was developed for the Australia Greenhouse Office for use in complying with the Building Code of Australia and can be extracted from EnergyPlus. The comparison was carried out for the sake of investigating if the use of the proposed tool instead of using weather data in historical archives from external sources produces a more useful analysis. Specifically, the TMY data is provided for larger geographical areas, instead of being localised. Hence if the precooling potential and value differ across the sites, the TMY data will be less capable of showing the variations. As with TMY data, there is a necessity for a large dataset of observations from different years. In building design and operation modelling that involves climate analyses, average values are used rather than individual yearly ones that may fluctuate depending on cyclical phenomena or climate irregularities.

Table 35 compares the final results over all years considered (2005-2006 and 2010- 2014) for the ratios across all sites from the simulations and TMY data. The ratios in Table 35 are mean values averaged over the seven simulation years. The individual year results can be found in Table 34.

Table 35: Summary of precooling potential and utilisation across the five sites in Sydney

Site 1 Site 2 Site 3 Site 4 Site 5 TMY h (frequency) 0.07 0.14 0.18 0.23 0.20 0.10 Precooling days 23.7 52.1 66.6 82.1 74.6 36.0 r1 (potential) 19.2 12.2 14.1 12.2 12.8 7.2 Annualised R1 1.3 1.7 2.7 2.8 2.6 0.7 r2 (utilisation) 0.47 0.88 1.15 1.41 1.72 0.54 Annualised R2 0.03 0.13 0.21 0.32 0.36 0.05

The frequency of occurrence of precooling days ranged between 0.04 (Site 1 in 2011) to 0.28 (Sites 4 and 5 in 2005). This represents a difference of 87 days per year that the precooling conditions are met between the minimum and maximum results.

The potential ratio (r1) ranged between 2.8 (Site 3 in 2011) and 41 (Site 1 in 2013).

For sites 2-5, the r1 was observed to fall within a narrower range of 8-20 for most simulation years. Site 1 had a higher r1 for most simulation years (and average). The

149 main reason behind this result is the low count of days that precooling conditions are met at that site. In simple terms, the simulations showed that the conditions were not met that often in Site 1, but when they were met there were days with rather high diurnal temperature differences. In turn, these may be attributed to macro and meso-scale weather phenomena, such as rapid pressure fronts passing by or summer storms. After the potential ratio has been annualised (R1), Site 1 had the lowest average (1.3) and absolute minimum (0.4 in 2012) and Site 4 had the highest average (2.8) and absolute maximum (4.8 in 2013). This confirms the previous observation for Site 1 regarding the occurrence of precooling days.

Regarding the utilisation ratio (r2), it ranged between 0.36 (Site 1 in 2010) and 2.01

(Site 5 in 2005 and 2006). Interestingly, for Site 1 in 2010 the lowest simulated r2 of

0.36 was accompanied by a simulated r1 of 17.6, which is relatively high. This implies that 2010 was a relatively cold summer for Site 1: the r1 being great as a result of a high value of DHL in the numerator, and the r2 being low as a result of a high value of DHN in the denominator. The opposite observation can be made for Site 5 and

2006, when the r1 was just 5.6 (much lower than the site average of 12.8) and the r2 was 2.01 (much higher than the site average of 2.01): the summer was relatively hotter, as the overnight temperatures were not that low during precooling days

(hence low values of DHL and r1), and the daytime temperatures were rather high for the following days (hence low values of DHN and high r2). This is also indicated in R2, which showed that 2005 had more precooling days than 2006 for Site 5 and hence higher R2, even if the r2 was identical for both years.

In the following sections (5.3.2 – 5.3.5), the results from the simulations for the five Sydney locations will be further discussed.

5.3.2 Precooling frequency The occurrence of precooling days in a year and the frequency h were consistent throughout the simulation years, which is an indication that certain sites are more appropriate than others when implementing precooling measures. For example in 2011 the precooling frequency dropped notably for all sites, showing that during that summer, diurnal differences were lower than average. Furthermore, the gradients of the precooling occurrence curves display a similarity across all sites in 150

Sydney, which was expected since the general climate over a year is similar for locations within a short range (of up to 50km) of each other (Figure 49).

Annual precooling occurrences 120

100

Precooling days in a a days yearin Precooling 20

0 2005 2006 2010 2011 2012 2013 2014

Site 1 (Bondi) Site 2 (Airport) Site 3 (Canterbury) Site 4 (Bankstown) Site 5 (Penrith)

Figure 49: Number of days that precooling conditions were met per year for each site Sites 3, 4 and 5 (Bankstown, Canterbury and Penrith) had the highest frequency ratios, as they are further inland. For sites 1 and 2, the proximity to the ocean (enormous thermal mass) acts as a thermal buffer and tends to smooth the diurnal temperature differences, hence they occur less often.

The average number of precooling days per year for each site can be seen in Figure 50 .

151

Annual average precooling occurences 90

80 82.1 70 74.6 60 66.6

50 52.1 40

30 36.0

Precooling days days peryear Precooling 20 23.7 10

0 Site 1 Site 2 Site 3 Site 4 Site 5 TMY

Figure 50: Average number of days that precooling conditions are met across the five Sydney sites The TMY data for frequency showed a flat rate of 36 precooling days across the metro Sydney area, however, as seen from the analysis based on the numerical simulations from real data, the frequency can vary significantly from that value. Another indication that TMY is not sufficient to predict precooling potential for sites even a few kilometres apart is that the distribution of days throughout the year that precooling conditions are met varies significantly across all sites. After comparing the dates of precooling days over the 7 years of the data for all sites, it was found that there is little overlap. The results showed that only a small fraction of precooling days were in fact common across sites.

Table 3536 shows the percentage of days that precooling conditions were met at the same time at different sites for each simulation year, as well as on average. The results were obtained by counting the number of days in each simulation year that the precooling conditions (as described in section 2.3) were met at the same time across any of the five sites. For instance, if for a particular day the precooling conditions were met for sites 2, 4 and 5 (but not for sites 1 and 3), this would count as a common occurrence across three sites. The annual relative frequency for each year and the overall average relative frequency were then obtained by dividing the number of each common occurrence counter by the total number of precooling days according to the simulations in all sites. This provides an indication of how often the

152 precooling conditions are met at the same time for the five sites that are within the same metropolitan area of Sydney.

Table 36: Comparison of distribution of precooling days

Common 2005 2006 2010 2011 2012 2013 2014 Average occurrences relative frequency

1 55.26% 56.90% 54.20% 41.65% 49.12% 53.91% 50.40% 51.63%

2 19.35% 20.61% 19.44% 25.85% 24.39% 20.12% 21.24% 21.57%

3 11.17% 10.64% 11.31% 15.58% 14.09% 12.02% 12.10% 12.42%

4 13.98% 10.94% 14.81% 14.77% 12.27% 13.40% 15.85% 13.73%

5 0.24% 0.91% 0.24% 2.15% 0.13% 0.55% 0.41% 0.65%

As seen in 36, it was found that over 50% of the precooling days were unique to each site. In simple terms, this implies that while the climate around the broader Sydney area is similar, diurnal temperature differences tend to vary due to local weather and landscape factors. 2011 and 2012 were the years with the lowest incidence of precooling days according to the simulations. During 2011 and less notably during 2012, there were more common occurrences compared to the rest of the simulation years. This is due to the fact that the high diurnal temperatures occurred on a limited number of days during these years in the Sydney area, and these days were shared more often between sites.

5.3.3 Precooling potential

As discussed in section 3, the ratio r1 can be thought of as the precooling potential. In essence it expresses the theoretical percentage of cooling load during the daytime

(which is proportional to the DHU) that may be covered by precooling the night before (which is proportional to the DHL, assuming precooling infrastructure is in place).

153

The average precooling potential can be seen in Figure 51. The value of r1 (before being normalised over a year) shows that the potential for precooling is significant across all sites. In fact the r1 range is between 12 and 19.25, which implies that on average there is significant precooling potential even for sites 1 and 2, where the frequency is low.

Precooling potential 25.0 3.0

2.5 20.0

2.0 15.0

1.5

10.0 1.0

5.0 Mean r1 ratioMean (dimensionless) 0.5 R1 (dimensionless) Annualised

0.0 0.0 Site 1 Site 2 Site 3 Site 4 Site 5 TMY

Mean Ratio 1 Annualised

Figure 51: Summary of precooling potential across sites in Sydney After annualising by multiplying the precooling potential ratio by the precooling frequency ratio, it was shown in that sites 3, 4 and 5 are associated with the highest precooling potential. This illustrates the value of annualising the proposed ratio in interpreting the results: for instance, site 1 (Bondi), has the highest r1, and the lowest R1 among the Sydney sites. This is due to the fact that at site 1 the precooling conditions do not occur that often in a typical year, however when they do the diurnal temperature differences are rather high (hence high r1). Site 2 has much lower r1, which implies that when the precooling conditions are met, the potential is not great – however, it occurs more frequently than in site 1 in a typical year. On the contrary, and while the average r1 for the inland sites (3, 4 and 5) is lower than in site 1, the overall yearly precooling potential is much more significant, as the conditions are met more often. As seen in Figure 52, the gradients over the simulation years are not as consistent as with precooling frequency. Regardless, 154 certain features are still evident, such as that all sites had low potential for precooling in 2011 (due to relatively low diurnal temperature differences in the region).

Annualised potential ratio 7

2 Ratio r1 (dimensionless) r1 Ratio

0 2005 2006 2010 2011 2012 2013 2014

Site 1 (Bondi) Site 2 (Airport) Site 3 (Canterbury) Site 4 (Bankstown) Site 5 (Penrith)

Figure 52: Annualised potential ratio R1 for the Sydney sites for each simulation year 5.3.4 Precooling utilisation The potential ratio by itself is not sufficient to make solid conclusions about the precooling value. While high values of ratio r1 imply that there is potential for precooling, it may be the result of a low number of degree-hours that actual cooling is needed (denominator of r1). In simple terms, this would refer to situations that cool nights are followed by relatively mild days, with only a few degree-hours over the upper comfort zone boundary. In such cases, the cooling requirements would be low and hence precooling would not make a significant contribution towards reducing energy and costs; however the ratio r1 would indicate a high potential for precooling.

The utilisation ratio r2 shows the overall utilisation of precooling potential. As seen in Figure 53, both the mean r2 and annualised R2 display a more consistent pattern than the respective r1 and R2 ratios.

155

Precooling utilisation 2.00 0.40

1.80 0.35 1.60 0.30 1.40 1.20 0.25 1.00 0.20

0.80 0.15 0.60 0.10

0.40

Mean r2 ratioMean (dimensionless) Annualised R2 R2 (dimensionless) Annualised 0.20 0.05 0.00 0.00 Site 1 Site 2 Site 3 Site 4 Site 5 TMY Axis Title

Mean Ratio 2 Annualised

Figure 53: Precooling utilisation summary for sites across Sydney This trend can be seen in Figure 54 as well. Over the years, there is a consistent assessment of precooling utilisation across all sites.

Annualised utilisation ratio 0.6

0.5

0.4

0.3

0.2

Ratio r2 (dimensionless) r2 Ratio 0.1

0 2005 2006 2010 2011 2012 2013 2014

Site 1 (Bondi) Site 2 (Airport) Site 3 (Canterbury) Site 4 (Bankstown) Site 5 (Penrith)

Figure 54: Annualised utilisation ratio (R2) for Sydney sites for each simulation year

156

5.3.5 Precooling value The results showing the theoretical precooling value for each year (v), as well as the average theoretical precooling value of seven years’ worth of simulations can be seen in Table 37.

Table 37: Summary of precooling value ratios for the five sites in Sydney

2005 2006 2010 2011 2012 2013 2014 Average value Site 1 -2.89 -1.73 -3.40 -4.03 -1.70 -1.25 -2.10 -2.44 Site 2 -0.42 -0.07 -0.21 -0.49 -0.64 -0.03 -0.34 -0.31 Site 3 0.16 0.23 0.04 -0.12 -0.14 0.01 0.09 0.04 Site 4 0.16 0.22 0.04 0.00 0.03 0.22 0.19 0.12 Site 5 0.29 0.35 0.01 0.03 0.11 0.27 0.23 0.18

Sites 4 (Bankstown) and 5 (Penrith) display a positive theoretical precooling value and more importantly, consistently positive results for the years considered. On the other hand, sites 1 (Bondi) and 2 (Airport) have negative average values, with individual years being negative as well. This implies that sites 4 and 5 possess the potential for a building to utilise natural precooling with meaningful savings, while the savings will be negligible for buildings in sites 1 and 2. Site 3 (Canterbury) has a positive average theoretical precooling value, however as seen in specific years with lower diurnal temperature differences (such as 2011-12) the savings are expected to be negligible. Hence, site 5 (Penrith) was the best performing site according to the analysis in this chapter and it is expected that whenever precooling conditions are met there is a theoretical average of 18% coverage of the cooling loads via precooling.

Of course the exact savings by the precooling process depend largely on the actual building characteristics and infrastructure. The results of the simulations in this chapter can serve as a preliminary assessment for the theoretical value that is realised at a specific location, regardless of building type. Simulations and estimations of actual savings in energy and peak reduction are carried out for a case study building in chapter 6.

Furthermore, the precooling potential, utilisation and value ratios are sensitive to the chosen boundary temperatures used in the precooling conditions (section 5.2.3). In the current analysis the boundaries were chosen to be 20 and 25°C respectively 157 for the lower and upper boundaries. In order to understand the sensitivity of the ratios to the selected temperatures the analysis has been carried out with different boundaries for site 2 and site 5 (one site with negligible value and one site with positive value) and the results are displayed in Table 38 and Table 39.

Table 38: Sensitivity of the value ratio (v) to the selection of boundary temperatures (Site 2 – Airport)

Boundary 2005 2006 2010 2011 2012 2013 2014 Average temperatures (°C) 18-23 -0.77 -0.43 -0.49 -1.02 -1.09 -0.33 -0.79 -0.70 19-24 -0.61 -0.22 -0.35 -0.77 -0.85 -0.19 -0.52 -0.50 20-25 -0.42 -0.07 -0.21 -0.49 -0.64 -0.03 -0.34 -0.31 21-26 -0.33 0.02 -0.18 -0.34 -0.58 0.05 -0.2 -0.22 22-27 -0.25 0.11 -0.06 -0.26 -0.41 0.11 -0.04 -0.11

Table 39: Sensitivity of the value ratio (v) to the selection of boundary temperatures (Site 5 – Penrith)

Boundary 2005 2006 2010 2011 2012 2013 2014 Average temperatures (°C) 18-23 0.05 0.06 -0.17 -0.15 -0.09 0.04 -0.03 -0.04 19-24 0.22 0.26 -0.06 -0.05 0.01 0.2 0.14 0.10 20-25 0.29 0.35 0.01 0.03 0.11 0.27 0.23 0.18 21-26 0.35 0.44 0.12 0.15 0.19 0.35 0.3 0.27 22-27 0.48 0.51 0.22 0.2 0.27 0.51 0.43 0.37

It can be seen that when both the upper and lower boundary temperatures are raised, the value ratio (v) increases as well. This is mainly due to fact that the denominator of the value ratio (DHU) decreases (less hours above the upper boundary). At the same the difference in the numerator of the value ratio (DHL-DHN) for a particular precooling day remains relatively unchanged, as there are overall more hours below the lower boundary overnight (DHL) and more hours below the upper boundary during the daytime (DHN) compared to the 20-25°C band. Accordingly the value decreases when both temperature boundaries are decreased. Practically, this shows that precooling is more useful when there is a higher upper boundary of temperatures. It should be noted however, that in these cases the frequency of precooling occurrences (and hence ratios R1 and R2) decreases as well, as the precooling conditions are met less often. Specifically for Site 2, the frequency dropped from 0.17 to 0.11 and for Site 5 the frequency dropped from 0.28 to 0.19 when the boundary temperature zones were changed from 20-25°C to 22-27°C. 158

5.3.6 Precooling dependence on local climate factors As stated in the beginning of the chapter, the development of the proposed tool may help in obtaining useful insight on the effects of climate patterns on different aspects of precooling processes. Certain trends in climate may cause “irregular” results to appear in the simulations, which can be realised as significant deviations of the precooling ratios from the average values (outliers). The precooling ratios as well as the frequency depend on the diurnal temperature differences. The magnitude of diurnal temperature differences (between maximum and minimum temperatures) is of particular importance and needs to be considered in addition to the actual observed values to characterise the potential of a site.

In general, it was found that if the diurnal temperature difference is bigger than normal, the precooling frequency and utilisation would improve. Accordingly the theoretical value of precooling would improve as well. The reason for that is that more natural cooling is available overnight and higher cooling loads the day after. To illustrate this effect, the mean monthly diurnal temperature differences for Site 5 (Penrith) were compared for the simulation years. The value was calculated as the mean maximum temperature for that month minus the respective mean minimum temperature. The comparison data was obtained from an independent source (Bureau of Meteorology (Bureau of Meteorology, 2015)) to eliminate model bias. The comparison was carried out for the months of the year that precooling is possible (summer). The results are displayed in Table 40.

Table 40: Comparison of mean monthly diurnal temperature differences for each simulation year and 20 year average for Penrith

Average 2014 2013 2012 2011 2010 2006 2005 Jan 12.3 13.4 12.8 11.1 12.3 12.7 10.1 11.5 Feb 10.8 10.6 10 9.4 11.9 10.6 11.8 10.7 Mar 10.9 9.9 11.6 9.6 10.2 10.8 10.5 8.5 Oct 13.8 14.8 17.6 15.2 11.7 12.1 15.1 12.7 Nov 12.3 13.8 12.7 12.3 12 10.7 15.1 10.5 Dec 12.3 11.3 13.7 12.4 9.7 11.2 13 15.7 Annual 12.07 12.30 13.07 11.67 11.30 11.35 12.60 11.60

Figure 55 shows the correlation between precooling frequency in site 5 and the magnitude of diurnal temperature differences for each year.

159

Correlation of precooling and diurnal differences 120 13.50

100 13.00

12.50 80 12.00 60 11.50 40

11.00 Precooling Precooling days in ayear 20 10.50

0 10.00

2005 2006 2010 2011 2012 2013 2014 Mean Mean annual diurnal temperature difference Precooling frequency Diurnal difference

Figure 55: Comparison of precooling days per year and the magnitude of the mean annual diurnal temperature difference It can be seen that for 2011, during which the frequency (as well as precooling utilisation) were lower than the rest of the simulation years, the mean diurnal difference for the precooling months was the lowest as well. This climate anomaly in 2011 was mainly due to December maximum temperatures being significantly lower than average (Bureau of Meteorology, 2015). It is worth noting that the absolute magnitude of the maximum and minimum temperatures, as well as the weighting of each month to precooling can be considered to improve the accuracy of estimating the correlation between frequency and diurnal differences.

The findings from the simulations confirm the expected results stated in the beginning of this chapter; high diurnal temperatures are indeed the primary factor for precooling potential and even within a relatively short distance, diurnal temperatures may vary by a notable extent. In the next section, the findings are generalised for various climatic regimes across Australia.

5.3.7 Simulations for different climates The precooling assessment tool analysis showed that the five sites across Sydney demonstrate different characteristics. Overall, inland sites within a similar climate were proven superior for precooling. For sites closer to the coast, diurnal temperature differences are moderated by the effects of the heat stored in the ocean. 160

Furthermore, breezes from the open sea may act to reduce the maximum temperatures at coastal sites during a hot day compared to an inland site. These effects were visible in the results for both the precooling frequency and value.

Hence, it is expected that the difference in results will be even more notable when applying the model in climates that are significantly different to Sydney. For the first part of the analysis in this section, five Australian cities were considered. The simulations for each site and the calculations for all ratios have been conducted with the same methods discussed in section 5.2 and involved the same simulation years as Sydney.

Melbourne, Victoria has a different climate compared to Sydney with higher temperatures during the summer and lower temperatures during the winter. The diurnal differences are expected to be larger, hence resulting in higher precooling potential and value.

On the other hand, Brisbane, Queensland has a subtropical climate with milder temperature gradients and warmer summer nights. Hence it is expected that the precooling potential will be lower compared to Sydney.

Canberra, is relatively close to Sydney, but the climate is significantly different, as it is located in the middle of a valley, 70km inland and an elevation of approximately 580m. Due to its continental climate, high vegetation coverage, low density and geographical features it is expected to demonstrate high precooling potential.

Alice Springs in the Northern Territory is a city located in a dry desert climate more than 1,000 km from the coast. Typically, due to high radiation of desert overnight and high solar exposure during the day, diurnal differences are expected to be rather high in such an environment.

For all cities the simulations were carried out at locations relatively close to the ocean, where applicable (less than 2 km), and compared with a similar location in Sydney (site 2 – Airport) (Table 41).

161

Table 41: Comparison of precooling ratios from the simulations in different climates in

Australia

Sydney Melbourne Brisbane Canberra Alice Springs

Climate type Temperate Temperate Humid Temperate Desert (Köppen oceanic oceanic Subtropical oceanic classification) Ratio h 0.14 0.18 0.01 0.19 0.34 Annual Precool 52 66 4 69 124 days Annualised R1 1.71 3.96 0.03 5.83 9.53 Annualised R2 0.13 0.59 0 0.45 0.91 Precooling value -0.31 1.26 -9.46 1.42 0.17

As seen in Table 41, the assumptions from considering the different climates and geographical locations of each capital are confirmed by the results of the simulations.

For Melbourne, the average precooling value of 1.26 implies that natural precooling is theoretically able to fully cover the cooling loads on certain days. As with Melbourne, Canberra has the potential for excellent precooling value, with theoretical 100% coverage of cooling loads around 19% of the year. The utilisation ratio for Canberra (0.45) is lower than Melbourne (0.59), since there are lower peak temperatures during the daytime and hence the denominator of R2 is higher. For both Melbourne and Canberra, the frequency, potential, utilisation and value appear to be higher than Sydney.

Brisbane does not display any realistic precooling potential with a negligible amount of precooling days and very low precooling value. This is primarily due to the relatively warm summers with very few days displaying significant diurnal temperature differences.

Finally, Alice Springs demonstrated some interesting results. The ratios for the precooling frequency, potential and utilisation are much higher than the rest of the sites, due to the significant magnitude of diurnal temperature differences. The high distance from the coast appears to exacerbate the diurnal differences. However, the precooling value is not as high since the daytime temperatures are much higher for longer periods of time, resulting in a high denominator in the value ratio (DHU). In

162 simple terms this means that precooling may be very successfully utilised (R2=0.91) to shift a part of the load overnight, but it will likely have a less significant impact (v=0.17) compared to the total cooling requirements, as the daytime weather is often very hot. Regardless, due to consistently high diurnal differences throughout the summer, the potential and frequency of precooling for Alice Springs are notably higher than the other sites.

It should be noted that for sites like Alice Springs, where the theoretical value of precooling was rather low due to high daytime temperatures may justify investment in sophisticated demand controllers in order to tap the high utilisation. This is beyond the scope of this thesis, but it may be associated with reduced carbon emissions from the displaced HVAC load.

The second part of the analysis attempts to generalise the findings and extract conclusions for various types of climates globally. The simulations for the eight sites and the calculations for all ratios have been conducted with the same methods discussed in section 5.2 and involved three simulation years (2011, 2012 and 2013). It should be noted that for the sites in the north hemisphere, the cooling periods were simulated between 15th May and 15th September instead of the 15th November to 15th March that were simulated in Australian and other southern hemisphere sites. Table 42 and Table 43 summarise the results.

Table 42: Comparison of precooling ratios from the simulations in different climates globally

Cape Town, S. Athens, Tokyo, Hong Kong, Africa Greece Japan China

Climate type Mediterranean Mediterranean Humid Humid (Köppen Subtropical Subtropical classification) Ratio h 0.21 0.04 0.16 0.01 Annual Precool 75 16 59 5 days Annualised R1 4.13 1.25 2.29 0.01

Annualised R2 0.68 0.10 0.36 0 Precooling value 1.08 -0.71 0.39 -11.50

163

Table 43: Comparison of precooling ratios from the simulations in different climates globally

Milan, Italy Dubai, UAE Johannesburg, Moscow, S. Africa Russia

Climate type Temperate Desert Subtropical Continental (Köppen Oceanic Highland classification) Ratio h 0.11 0.09 0.18 0.07 Annual Precool 40 34 64 25 days Annualised R1 1.65 1.14 3.96 2.94 Annualised R2 0.51 0.77 0.55 0.42 Precooling value 0.71 -1.21 1.39 0.05

A high resemblance in the results appears between the humid subtropical climates of Brisbane and Hong Kong: both sites have very low precooling potential, utilisation and value, as the diurnal temperature differences are relatively low, most summer nights are warm (hence very low DHL) and most summer days tend to be consistently hot (hence high DHU and low DHN). However, Tokyo’s climate which is still classified as humid subtropical, displays results from the simulations that are much more similar to Temperate Oceanic and quite different to Brisbane and Hong Kong. Specifically, there is a respectable number of precooling days, and the utilisation, potential and value ratios are all positive. The relatively favourable precooling conditions are very prominent during May and June, where a significant portion of cool nights were followed by hot days.

Perhaps the most notable difference between simulations for sites at the same climate classification was observed for Cape Town and Athens (both Mediterranean). Cape Town summer conditions were found to be rather favourable for precooling: the conditions were met often and there were plenty of cool nights

(high DHL) followed by relatively warm days (moderate DHU and DHN). It was found that while the diurnal temperature differences were not too extreme for Cape Town, the conditions were met often. Hence the value overall was found to be very positive. On the other hand, the summers of Athens appear to be notably hotter, with very few cool nights that would favour precooling. Hence the precooling ratios were all quite low. This significant difference may be attributed to Athens being a sub-class climate of “Hot Summer – Mediterranean”, while Cape Town is “Warm Summer –

164

Mediterranean”. The classification traits appear to coincide with the simulation results considering the subclasses specifically for summer conditions.

The desert climate of Dubai resembles the trends seen in the desert climate of Alice Springs, albeit with some differences. Firstly, the precooling value is relatively low for the same reasons as Alice Springs: the days are typically too hot (hence the denominator of the value ratio DHU is high) for precooling to make significant impact; however its utilisation (R2) is comparatively high. Unlike Alice Springs, the precooling potential (R1) and frequency (h) are relatively low, which is due to Dubai’s summers being on average much hotter than Alice Springs and the diurnal temperature differences being lower as it is a coastal site.

Comparing the Temperate Oceanic climates of Milan and Melbourne, there are certainly more similarities to be found. The values of all ratios according to the simulations are of comparable magnitude for both sites, however Melbourne has notably higher precooling frequency and potential. Interestingly, the daytime temperatures are relatively close for both cities (hence DHU and DHN are similar), with Milan’s values being slightly higher. However, a notable number of Melbourne summer nights are rather cool (due to the effects of synoptic scale cold fronts developing in the southern Arctic Ocean). On the other hand Milan is located at a comfortable distance from any such regular synoptic scale phenomena and as a result its summer nights are much warmer.

The site of Johannesburg displayed similar results to the Canberra site. Both are classified as subtropical (with different subclasses) and both are inland sites. The results showed a significant number of cool nights (high DHL) followed by relatively moderately hot days, hence both the utilisation and potential ratios are high. Additionally, since extremely hot days are not frequent, the value of precooling appears to be rather positive for Johannesburg, confirming the findings for inland sites with similarly mild summers from before.

Finally, a continental site was examined as well. Moscow, while located at a very high latitude may still see hot summer days. The conditions are not met that often, but the potential is very high (since the overnight temperatures tend to consistently be

165 low, hence high DHL). The precooling value is not as significant, as the numerator difference of DHL and DHN is quite low.

It should be noted that the analysis conducted in this section served the purpose of investigating if the trends observed in the initial Sydney and Australian simulations may be generalised. While general conclusion may indeed be made, as certain trends are visible in the comparisons, the simulation results are more site specific. As with Sydney, the results may vary significantly for different locations in any of the other cities.

5.4 Conclusions Precooling is recognised as a sustainable solution for the optimisation and management of energy performance of commercial buildings through the reduction of energy consumption and peak shaving. Precooling a building is a low cost solution as it utilises the diurnal temperature differences to manage energy demand with appropriate infrastructure. Actively managed HVAC operations can be applied regardless of the prevalent weather conditions, but nevertheless depend on the diurnal temperature differences as well. In literature, there are plenty of studies modelling the building performance and savings through such strategies in specific buildings, however there is a lack of studies related to the effects of the location and climate on precooling.

Hence, in this chapter a tool was proposed that is able to assess the potential of precooling at a given location. The tool is expected to be a useful option for building designers and energy managers as it may be integrated in decision making for implementing certain preconditioning related features or policies in any location.

The model is based on numerical simulations from TAPM and the development of four dimensionless ratios for easy comparison between various sites and climates. The results from the simulations suggest that locations with different characteristics across the same metropolitan area display different potential for natural precooling according to their proximity to the ocean, building density and landscape roughness. The findings also confirm that the precooling process is favoured by high diurnal temperature differences in the summer. Inland sites with high vegetation index and relatively low building density displayed the highest precooling potential and value, 166 due to high diurnal temperature differences that occur frequently in such locations. However, if the daytime temperatures get too high, the value may drop significantly (as precooling may only cover a small fraction of the daily cooling requirements).

Furthermore, analysing sites across Australia and globally with the same method it is obvious that the findings of the above paragraph may be generalised. Sites with similar climate conditions appear to demonstrate similar trends for the precooling frequency, potential, utilisation and value. That being said, the significance of the local conditions cannot be overstated: besides the climate classification, the local landscape, proximity to geographical features, such as mountains or large bodies of water and synoptic phenomena that affect the regularity and intensity of weather are all significant contributors to precooling analysis.

The proposed tool is unique in existing literature, as it is the first of its kind utilising numerical climate analysis to assess the potential of precooling, independently of the building characteristics. Any historical weather data could be used as input for the algorithm, however such data are most often not readily available for most locations. TMY or nearby weather station data could be used as an alternative, but the proposed tool provides a more accurate assessment of the precooling value, as it accounts for localised effects of the climate and landscape; it was shown that even sites separated by distances of a few kilometres have significantly difference precooling performance. The precooling assessment of a location via TAPM has the additional advantage of being part of a broader range of prediction and characterisation algorithms aiding in short term forecasting, ensemble forecasting and building parameter estimation.

167

6 Development of a building load predictive control algorithm based on enthalpy forecasts

 Ambient temperature and relative humidity predictions from the numerical model used to correlate enthalpy differences to cooling loads  Enthalpy predictions used to develop a decision algorithm for the dynamic load prediction and control over the course of a day  Three modules proposed: short-term weather forecasting, peak load forecasting with ensemble, preconditioning forecasting  Control algorithms are responsible for weather-dependent responses according to the forecasting components, as well as dynamic monitoring of cooling load  Simulations of integration of modules in a case study building show a reduction of cooling demand of up to 24% and peak loads of up to 21% 6.1 Model outline This chapter will describe the development of a number of control modules for the potential integration of the numerical weather forecasts in a system with MPC. The methodology is based on a number of predictive control algorithms that take into account the information from the tools developed in chapters 3-5, as well as several additional outputs from the NWP model. The proposed control algorithms can activate and make decisions that regulate a series of responses depending on the BEMS infrastructure.

Section 6.2 will describe the model inputs and justify its design. This includes numerical weather predictions according to the methodology developed in earlier chapters (section 6.2.1), the calculation of enthalpy of ambient air (section 6.2.2), a novel approach that correlates differences in enthalpy of ambient and interior air with cooling load in a building (section 6.2.3) and the description of additional (non- weather) inputs required (section 6.2.4). Section 6.3 analyses the algorithm responses according to a number of multi-level decision trees, as well as their tolerance thresholds. Following that, there will be a quantitative assessment of the control algorithms based on simulations with real data using a UNSW building as the case study site (section 6.4). The approach is based on the development of a time-series neural network with external inputs (which is the predictions of ambient enthalpy of air). Sections 6.4.2.2-6.4.2.4 present and discuss the simulation

168 results and comment on the value of the predictive algorithms compared to the existing load control method.

As discussed in section 2.3.3, building conditioning is affected by the weather conditions, as well as the building thermal mass and occupancy levels. This case study is carried out in order to investigate the effectiveness of localised weather forecasts and historical analyses according to the methods proposed in chapters 3- 5. Hence, the algorithms do not include any occupancy level predictions. Rather than functioning as a complete HVAC control system, the proposed algorithms presented in sections 6.2 and 6.3 can be regarded as individual modules that can be integrated in a holistic BEMS for a large commercial building, the design of which is beyond the scope of this thesis. Regardless, the proposed model is a novel addition to the existing literature for the evaluation of potential savings by implementing localised weather forecasting and analyses.

6.2 Predictive control algorithms 6.2.1 Weather forecasting inputs and horizons The proposed integrated predictive control system operates with weather forecasts inputs in two horizons: daily and weekly.

Daily horizon predictions form the basis of dynamic control and consist of the following modules:

1. Short-term numerical weather forecasting (ST), based on the methodology of chapter 3 for the development of the hybrid prediction models. ST simulations are run four times per day at the reference hours (1, 7, 13, 19) and receive hourly weather updates from onsite observations to feed in the persistence model (U). The hybridisation is based on the ARX model (section 3.2.4.3.), as it is associated with the highest accuracy with hourly persistence model updates. However, if no historical weather observations are available onsite for estimating the weights, the WF model may be used instead (section 3.2.4.1). As discussed in section 3.4, the extreme event correction algorithm is implemented when the ST predicts such an occurrence at reference hour 1. Furthermore, the ST uses NWP predictions instead of ARX predictions

169

when abrupt changes are predicted at reference hour 1, due to the improved accuracy of the numerical model in such instances. 2. Peak load forecasting with ensembles (PL), based on the methodology of chapter 4. PL simulations are run once per day at hour 1. The ensemble configuration and distribution threshold are as discussed in chapter 4. 3. Precooling control (PC), which takes into account the local climate (from the analysis discussed in chapter 5) and the building characteristics. This module is run once per day at hour 1.

Figure 56 shows the operation times of each module in the daily horizon control. The time steps correspond to each hour in a day, ie time step 0 stands for midnight, time step 1 for 1:00 AM and so on.

Figure 56: Timeline of operation of the modules in daily horizon control. The timeline shows when does each forecasting component is scheduled to run the simulations and develop predictions throughout a day Weekly horizons are useful for energy pricing optimisation and load exchange planning, as discussed in chapter 2. As the persistence assumption does not hold true for periods longer than a few hours, the numerical model alone is sufficient to produce weekly weather predictions. The predictions are run each Monday at hour 1, according to the ensemble parameterisation discussed in chapter 4 and with a 7 day horizon.

When the ST component runs at each reference hour and an abrupt change is predicted in temperature or relative humidity (section 3.3.4) for the next time block, the inputs from the numerical TAPM simulations are used instead, without any statistical post-processing. This measure is activated in accordance to the accuracy increase of TAPM over the hybrid prediction models discussed in section 3.4. Furthermore, the prediction correction algorithm for extreme heat events is activated as described in section 3.2.6.

170

6.2.2 Enthalpy of air Besides ambient temperature predictions that may be used directly as inputs to the response decision algorithms, the weather forecasts allow for the prediction of the enthalpy of air. As discussed in section 2.2.4.8, the enthalpy of air accounts for the moisture levels and is necessary for understanding and forecasting the heating and cooling loads. To calculate the enthalpy of air at any time step, both the ambient temperature (T) and relative humidity predictions (RH) are used. From the relative humidity predictions it is possible to calculate the humidity ratio (X) of air. Firstly, the water vapour pressure (Pv) at a given temperature (T) is calculated as the product of RH and the saturated water vapour pressure (Pvs,T) (Vaisala, 2013):

푃푣 = 푃푣푠,푇 × 푅퐻 (12)

The values of Pvs,T for the predicted temperatures are obtained from psychrometric tables (Engineering Toolbox, 2016b). X is then calculated as the ratio of water vapour pressure to the difference between atmospheric pressure (Pa) and water vapour pressure multiplied by 0.62198 kg of water vapour per kg of dry air (Vaisala, 2013):

푃 푘𝑔 푋 = 0.62198 푣 ( ) (13) 푃푎 − 푃푣 푘𝑔

Finally the enthalpy of air (h) can be calculated from the ambient temperature prediction (T) and the ratio X as (Vaisala, 2013):

푘퐽 ℎ = 1.006푇 + 푋(1.84푇 + 2501) ( ) (14) 푘𝑔

The values of specific heat for water vapour (1.006 KJ/kg C), evaporation heat (2,501 KJ/kg) and specific heat of air (1.006 KJ/kg C) are considered constant (Vaisala, 2013).

As discussed in section 2.2.4.8, the enthalpy of air is a major factor considered in predicting cooling loads, as it includes both the sensible and latent heat that needs to be managed. Hence, it is a useful input for the control algorithms and quantitative assessment of their effectiveness. It should be noted that the enthalpy of air as calculated from the prediction weather data does not account for any moisture

171 introduced via occupant respiration and other sources within the building. Regardless, the case study building’s control system makes decisions based on temperature and relative humidity, rather than temperature alone and hence the proposed algorithm aligns more closely to it.

Ambient temperature, relative humidity and enthalpy are calculated at hourly time steps according to the predictions derived from the tool in section 3.2.4.3. For the purpose of managing the control response algorithms, the predicted value of variable A from the ST model at time k is denoted as Ak, and the value predicted α time steps from time k as Ak+α. For instance the enthalpy of air at each time step would be denoted as h1, h2 and so on. The reference hours occur as described in section 3.2 and in these cases k=n. The enthalpy of air is used to correlate the predictions from the numerical model with the cooling load according to the methodology described in the following section 6.2.3.

6.2.3 Correlation of enthalpy differences of ambient and interior air with cooling load To simulate the savings from predictive control integration and compare them with the standard control, it is required to obtain an understanding of the ways the cooling load correlates to changes in the weather conditions. Thermal modelling of the building was considered, however there were a number of challenges with this option, including limited resources, time constraints and lack of accurate data.

An alternative, less complex (and novel) technique was used instead based on the calculation of the enthalpy of air, as described in section 6.2.2. As discussed in section 2.2.4, the concept of enthalpy is used broadly in existing literature in the field of load forecasting. The advantage of this approach is that it allows to directly investigate the effects of weather on cooling, as cooling loads have to handle both sensible and latent heat gained by the ambient conditions. Using the weather conditions to estimate enthalpy gains, does not take into account any heat gain components from internal sources (occupant perspiration, electrical devices) or the building’s thermal mass effects. Thermal modelling of these inputs with physical techniques (section 2.2.4) is possible, however a neural network approach was instead implemented in this section. The network allows for a relatively easier and

172 faster tool design for the correlation of cooling loads to the enthalpy changes, without the need to model the thermal interactions of the building envelope and internal heat sources in great detail.

In order to run these simulations and assess the correlation of enthalpy change and cooling load, a non-linear autoregressive neural network with external input (NARX) was developed with the help of the NARX time series tool in MATLAB. The network consists of 8 neurons and attempts to predict y(t) (cooling load) based on the enthalpy change time series (x(t)) and an autoregressive time series of past load values. In the proposed network, the enthalpy of ambient air was firstly calculated according to the methodology in section 6.2.2 from the ambient conditions (ha) as taken from the onsite weather observations for each hourly time step in the simulation period.

The enthalpy of air in the interior of the building was also calculated (hi) for the same time period. As stated earlier, due to the lack of indoor humidity data, it was assumed that the relative humidity stays within the range of 30-60% during the occupied time, a common control strategy in commercial buildings. In order to calculate indoor relative humidity at each time step, several assumptions had to be made. Firstly, it was assumed that the rate of change of indoor relative humidity is equal to the rate of change of ambient relative humidity plus an evaporation constant (k). The evaporation constant represents the indoor relative humidity gains as occupants sweat during their activities in the summer, as well as any moisture gains from appliances, kitchens or bathrooms in the building. It was assumed that the evaporation constant k is proportional to ha. As discussed in section 6.2.2, the enthalpy of air increases during hot, humid weather. During these conditions, evaporation through sweating was assumed to be higher in the building. This assumption was based on the fact that occupants entering the building or carrying out activities would sweat more easily if the weather is hot. The 1 proportionality constant was set to be equal to 푡ℎ of the positive percent 10 difference between ha at each time step and the enthalpy of air when T=25°C and relative humidity = 50%. For instance, if the ambient temperature is 30°C and the ambient relative humidity is 40%, the enthalpy of air would be roughly 14% higher than the enthalpy of air at 25°C and 50% rel. humidity (according to calculations as 173 stated in section 6.2.2). Hence the proportionality constant would be 1.4%. In the same example if the hourly rate of change of ambient relative humidity was found to be 5% per hour, the indoor relative humidity rate of change would be 5 + 1.4 = 6.4%.

At the beginning of each occupied period it was also assumed that the relative humidity indoors would be equal to the ambient relative humidity (as any differences should balance out during the unoccupied night-time period). If the ambient relative humidity exceeded 60% or was below 30% at the beginning of the day, the HVAC control would attempt to bring it to these limits instead. Knowing the rate of change of ambient RHa, it was then possible to calculate the interior RHi at time step t as:

푑푅퐻 푅퐻 (푡) = 푅퐻 (푡 − 1) + 푎 + 푘 (15) 푖 푖 푑푡

Finally, as the cooling load is proportional to the enthalpy gains from the environment, the enthalpy difference (Δh) between the ambient and interior air for each time step (ha-hi) was calculated. Since the enthalpy of the interior air depends on the set temperature and relative humidity conditions, this method allows one to correlate cooling load with the enthalpy of air and assess different cooling control strategies.

Figure 58 shows the proposed NARX architecture, that receives both load (y(t)) and external enthalpy difference inputs (x(t)) as time series. The network was developed with the help of the NARX time series tool in MATLAB and is able to correlate the two time series with a network of 8 layers of hidden neurons.

Figure 57: Time series nonlinear input-output neural network architecture for the correlation of enthalpy change and cooling load 174

The NARX described in this section was used to validate the predictive algorithm with a case study building, presented later in section 6.4.

6.2.4 Additional inputs In addition to weather related inputs from the NARX, the decision algorithm utilises the following inputs:

1. Building peak load (BP) at each daily time step for a past period (commonly month or year). BP is necessary, as peak charges in commercial building are issued on a rolling basis. This means that in most cases energy costs are calculated based on the highest peak load observed in a specific period before the time of calculation (most often one year or month). Knowing the magnitude of these costly yearly or monthly peaks will allow the control system to set upper boundaries and develop responses that act to keep the load below these boundaries. In the simulations presented in this chapter, monthly BP data inputs were used, according to the case study building’s energy plan. 2. Building dynamic load (DL) monitoring. DL inputs offer the necessary feedback to adjust the responses as they happen and correct the predictive control as needed. This refers to HVAC load monitoring where available. In the control response algorithm, the change of DL is also considered. Unless otherwise stated, the short term load change (within the last hour) is used as input. 3. Internal temperature (IT) at selected points within the building. IT monitoring is necessary, as it needs to be maintained within the pre-set comfort range zone. Furthermore, ambient temperature changes are typically attenuated and delayed (subject to the building envelope construction), so constant feedback to assess the effectiveness of control responses is necessary. IT may be obtained at multiple zones within a building and averaged in order to simplify the demand modelling. 4. Onsite distributed generation (DG) where available. Depending on the type of DG sources available in the building, and the management policies in place the generated energy may be used to either alleviate the demand or export to the grid. 175

6.3 Weather dependent response control 6.3.1 Dynamic internal temperature control algorithm The central theme of this weather dependent response system is to use weather forecasting information to dynamically regulate the IT. Unlike typical deterministic control setups, where the IT is set at a fixed constant level throughout the occupied period, the proposed predictive control module takes advantage of the whole range of the thermal comfort zone. The rationale of this type of response is to minimise cooling or heating requirements throughout the day. The algorithm described in this section is referring to cooling during the summer, however it may be modified for heating control in the winter.

The intra-day responses of dynamic temperature control are determined by the predicted values of ambient temperature from the ST and feedback loops from the DL monitoring. Specifically, the change of the DL time series is used as input in conjunction with weather predictions. The main purpose of this module is to maintain minimum cooling demand without compromising the occupant comfort. The thermal comfort zone boundaries may be defined as necessary by the BEMS policies. In section 6.4, there will be an overview of the effects of thermal comfort zone selection on energy savings.

The algorithm decisions and conditions described in this section occur at each time step. The decision tree that was developed for the purpose of dynamic internal temperature control contains three decision nodes.

There are three decision branches at the first node. The controller receives the input for ambient temperature prediction (T) from the ST model and compares it to the comfort range boundary temperatures: the lower boundary of the thermal comfort zone (LT) and the upper boundary of the thermal comfort zone (UT). Both LT and UT values may be set according to the climate characteristics, the building and activity type. As thermal comfort is subjective (Havenith et al., 2002), occupants in different climates may perceive the indoor environment differently, and typically it is up to the energy management policies to determine an acceptable range. For instance, if the activity is low (offices) the thermal comfort zone may be set higher than in a building with higher occupant activity (eg. warehouses, manufacturing

176 spaces). ASHRAE’s Standard 55 provides an empirical estimation method of quantifying comfort based on temperature and humidity boundaries on a psychrometric chart (ASHRAE, 2012) that can be used to select the levels in thermal control models. In this work, the LT and UT limits were selected to lie within the neutral thermal sensation zone in a psychrometric chart for indoor spaces with low physical activity and average occupant metabolic rates (according to Standard 55).

The three first-level decision clauses are:

1. 푇 ≤ 퐿푇 2. 퐿푇 ≤ 푇 ≤ 푈푇 3. 푇 ≥ 푈푇

The second decision node compares the predicted temperature at the next time step (T) against the maximum future temperature (FT) within the 6-hourly time block of forecasts. In simple terms, this node checks if there is an expected rise or drop in ambient temperature for the rest of the time block. The decision tree can then be written as:

1. 푇 ≤ 퐿푇 a. 푇 < 퐹푇 b. 푇 > 퐹푇 2. 퐿푇 ≤ 푇 ≤ 푈푇 a. 푇 < 퐹푇 b. 푇 > 퐹푇 3. 푇 ≥ 푈푇 a. 푇 < 퐹푇 b. 푇 > 퐹푇

Finally, a feedback loop from the DL input comes into effect. The controller checks if the value of DL changes between the two latest hourly time points is positive (net heat gains were positive over the last hour) or negative (net heat gains were negative). The complete decision tree can be written as:

1. 푇 ≤ 퐿푇 a. 푇 < 퐹푇

177

i. Δ퐷퐿 > 0 ii. Δ퐷퐿 < 0 b. 푇 > 퐹푇 i. Δ퐷퐿 > 0 ii. Δ퐷퐿 < 0 2. 퐿푇 ≤ 푇 ≤ 푈푇 a. 푇 < 퐹푇 i. Δ퐷퐿 > 0 ii. Δ퐷퐿 < 0 b. 푇 > 퐹푇 i. Δ퐷퐿 > 0 ii. Δ퐷퐿 < 0 3. 푇 ≥ 푈푇 a. 푇 < 퐹푇 i. Δ퐷퐿 > 0 ii. Δ퐷퐿 < 0 b. 푇 > 퐹푇 i. Δ퐷퐿 > 0 ii. Δ퐷퐿 < 0

Each third level decision branch is associated with a controller response which is implemented for the current hourly time step. The controller does not make any changes (however, keeps the thermal comfort zone constraints active), if the response has not changed for two consecutive time steps. This is meaningful for the responses that IT is allowed to change freely (in cases of positive changes in load). By changing freely from a lower value, the IT can relieve some of the cooling load during these time steps.

For decision group (1), the predicted ambient temperature is lower than the LT (relatively cool hours). During these times, and while the ambient temperature is relatively low, there are low cooling requirements from external sources. However, sensible heat is still gained from internal sources (electronics, occupants other machinery) and radiative heat may be gained from direct sunlight. The response is to set the internal temperature to the level of LT and maintain it at that level, unless 178 there is a predicted increase in temperature and increasing cooling demand. In that case, the temperature is allowed to increase if necessary, to alleviate some of the extra cooling load.

For decision group (2) the predicted ambient temperature is within the comfort zone, and is the most common zone in the Sydney climate during the cooling period. The response is to set the internal temperature at the level of T or LT and keep it constant if the future predicted values of ambient temperature are higher. This is expected to reduce some of the future cooling loads at the hotter time steps. If the peak block temperature has been reached, the response is to allow the IT to change freely within the comfort zone instead.

Finally, decision group (3) is the hottest part of the day, when the predicted ambient temperature exceeds the UT. If even hotter temperatures are predicted, the controller keeps the IT constant at the UT level. However, if the rate of change of DL is negative (decreasing cooling load), the controller lowers the IT to the mean level of the comfort zone (MT) in an attempt to prepare for shaving future peak loads.

A summary of the responses described above can be seen in Table 44.

Table 44: Dynamic temperature control responses for each decision branch

Branch Response Strategy 1ai IT=LT Allow free change within thermal comfort zone 1aii IT=LT Constant 1bi IT=LT Constant 1bii IT=LT Constant 2ai IT=T Constant 2aii IT=LT Constant 2bi IT=T Allow free change within thermal comfort zone 2bii IT=T Allow free change within thermal comfort zone 3ai IT=UT Constant 3aii IT=MT Constant 3bi IT=UT Constant 3bii IT=MT Constant

179

The controller also checks the PL module for triggers of future peak loads. If there are no significant potential peaks in the future, the controller maintains status quo and runs again from for the next time step. However, if there is a significant potential peak in the future, the controller activates peak load responses (discussed in section 6.3.2) which may supersede the responses in Table 44.

The algorithm for the dynamic temperature control described in this section can be seen in Figure 58.

It should be noted that the IT control assumes homogeneous zone heat distribution. Zones within the building with different heating or cooling requirements (for instance computer rooms) are of course excluded from the control algorithm proposed in this section.

Figure 58: Dynamic internal temperature control algorithm for cooling based on the outputs of the ST model

180

6.3.2 Peak load control algorithm As discussed in chapter 4, the peak load prediction algorithm consists of predictions for the timing and relative magnitude of the peak according to the ensemble distribution. For the timing predictions of the potential peak periods, the outputs from the PL are sufficient and may be used as described in section 4.2.1 in the proposed MPC. For the significant peak detection and peak magnitude predictions, the MPC utilises inputs from the DL and BP data series. The module described in this section is activated for cooling peak load control, however in future work it may be modified to examine its applicability on heating peak load control as well. The responses implemented by the controller on the basis of weather predictions will in most cases supersede the responses described in section 6.3.1.

The peak load control response is activated when there is a chance of predicted significant peak cooling load (according to the algorithm in section 4.2.2). The response attempts to minimise the cooling load during the potential peak period. This may be achieved via a range of DR measures. The algorithm, unlike the dynamic control of section 6.3.1, initialises once the PL module runs at hour 1 of the day.

Firstly, there is a period prior to the potential peak period that the predictive controller attempts to lower the IT, in preparation for the increased cooling loads. The IT is set to the level of LT at least 2 hours before the potential peak period begins. This response occurs if the predicted temperature exceeds the statistical threshold discussed in section 4.2.1. The threshold was derived from the distribution of historical daily peak loads, which showed that in the case study building there was a band of “normal” daily peak loads as well as a number of instances with peak loads exceeding this band by up to 35%. The controller also considers the BP module, as very often commercial peak loads are charged on a rolling fashion according to the highest peak in a specific period in the past (usually between a month and a year). In the case study a monthly scenario has been used, according to the cost policy of the building. The controller activates only when the predicted load starts approaching the historical peak according to the BP input (a 10% safety margin has been used).

181

As shown in section 4.3.2, the ambient temperature was the factor with the highest correlation to the peak loads, and hence when the predicted temperature from the ensembles is higher than the statistical threshold, the IT will be set at LT 2 hours before the potential peak period . This measure accounts for the prediction error (1.25 hours on average) and allows a safety margin time to offset any significant peaks. If ambient temperature is predicted to be lower than the statistical threshold, this measure activates at the beginning of the potential peak period instead.

During the potential peak period, the controller receives inputs from the DL and compares them against increasing levels in relation to the base cooling load and the max load from the BP time series. When the DL exceeds certain thresholds, the controller allows the IT to increase until it reaches the UT level. According to the analysis in chapter 4, these levels are set when the cooling peak load increases by 22%, 28% and 33% from the base level (calculated from historical data). The rationale of this approach is to gradually ramp up the cooling load as the building heats up more during the potential peak period, by regulating the load’s rate of increase and taking advantage of the whole thermal comfort zone. Besides IT control, this module assumes reserving any onsite DG (from PV, wind, batteries and/or cogeneration) for peak shaving, instead of exporting to other buildings or the grid if there is a significant peak load predicted (three favourable factors predicted from the ensemble as described in section 4.2.2). Where possible, minimising the heat gains from the environment is another option during the potential peak period. Depending on the building infrastructure this may involve deploying artificial shading devices, closing windows to minimise heat transfers, or using curtains to block direct sunlight heat gains.

Dynamic load monitoring and response infrastructure in the BEMS is required at this stage to optimise the responses. Figure 59 shows a flowchart of the primary peak load control algorithm.

182

Figure 59: Peak load control algorithm with gradual IT rises during the potential peak period Assuming the peak load exceeds 133% of the base load, the controller adds an extra decision node to the algorithm. It compares the load to the maximum rolling peak load from the BP series (over the last year or month, depending on the tariff framework) and attempts to avoid exceeding that level, since it would result in substantial charges. Depending on the building infrastructure this may involve reducing, rescheduling or cancelling altogether loads from low priority and non- critical modules in the building, such as non-safety lighting in unused areas or areas with access to day lighting or rescheduling energy intensive laboratory equipment operations to another zone.

6.3.3 Preconditioning control (PC) algorithm The PC module initialises the simulations on hour 1 of each day in the cooling period (summer months). If the precooling conditions are met (section 5.2.3) in the predictions, the precooling controller attempts to set the IT equal to T at each time step overnight (via natural ventilation where available), while T is below the LT. If the predicted temperature rises above the LT (most likely in the early morning hours before the occupied period), the controller attempts to maintain the IT at the

183 levels of LT. Figure 60 shows the flowchart of the PC algorithm. The first decision node runs at hour 1 and the second decision node at each hourly step before the occupied period.

Figure 60: Precooling control algorithm at every time step before the occupied period 6.4 Case study 6.4.1 Building characteristics To assess the usefulness of the proposed control algorithms, a series of simulations were conducted. The simulations were based on data from a building in the UNSW campus (L5 building, totalling 12,400 m2 in area and spanning over 7 floors (Kaji- O'Grady, 2006). The building has highly exposed facades in all directions, with only a few low height dwellings on the North and South sides. The building contains offices, a few small classrooms and testing facilities as well as a small coffee shop.

The majority of the areas in the building are conditioned by a HVAC plant, which operates between temperature set points of 22-23°C in the summer (regardless of weather conditions). Hence, it is a prime candidate for illustrating the potential

184 savings with predictive weather control. Figure 61 shows the west view of the building.

Figure 61: Western facade of the L5 building (Gollings, 2006) Data from the building HVAC plant was obtained and analysed for a period of 3 years (2012-2014). The data for the cooling period alone were considered in the simulations of this section (November – March). Additionally, as with the case study in section 4.3, weekends and holidays were excluded from the analysis, as the loads during these days are not representative of the weather changes since the occupancy levels are very low. The internal temperature was averaged from half- hourly measurements from 3 sensors inside the building (L1, L3, L6), placed in open office areas. While there is no humidity data from inside the building, it was assumed to stay within the 30-60% comfort range throughout the working day. The assumption was based on existing practices in humidity control for commercial buildings (Yang and Wang, 2015, Li et al., 2014, ASHRAE, 2013). It should be noted that there is only limited control of RH (certain spaces only) in the case study building, however the details were not possible to obtain and integrate in the simulations presented in the following sections.

185

6.4.2 Application of the NARX in the case study building The time series of ambient enthalpy of air and interior enthalpy of air were developed for the simulation years (2012-2014) for the case study building according to the methodology presented in sections 6.2.2 and 6.2.3. The enthalpy difference (Δh) was then used to validate the model with load inputs from the same period for the case study building.

With 85% of the data in the load time series used for training, the correlation coefficient when testing was found to be approximately 0.82, indicating a strong positive correlation between cooling loads and Δh for each time step. Figure 62 shows the correlation coefficient, as well as the mean square errors for the three sets of data (training, validation and testing).

Figure 62: Summary of the time series neural network performance for the training, validation and testing data The results in Figure 62 suggest a strong correlation (0.82) between cooling load and enthalpy change. Figure 63 displays the scatter plots for all three sets of data, as well as the overall scatter plot between the data and target predictions.

186

Figure 63: NARX performance in correlating cooling load to enthalpy changes The scatter plots in Figure 63 show that the positive correlation is evident for all data sets. The data points on the right side of the plots (high cooling loads over 200 kW) are arguably the most important, since they represent loads in the middle of the day and are predicted with reasonable accuracy. The seemingly high spread in the left side of the plots (low cooling loads) is due to HVAC control policies overnight that make the correlation of enthalpy change and load weaker, since there is no need for cooling at these times. Furthermore, quite often the enthalpy of ambient air overnight does not affect the cooling needs at all, since the temperatures may drop to low levels. This can be observed as the outputs (predicted load) are typically much higher than the targets (actual load) for the NARX simulations in Figure 63.

The error distribution can be seen in Figure 64. The errors were relatively uniformly distributed, with 55% of the errors being overestimates, and 45% being underestimates.

187

Figure 64: Prediction error distribution. The majority (90%) of the errors are within 44 kW of the actual observed load The correlation error distribution in Figure 64 shows that for around 90% of the time steps in the neural network simulation, the prediction error was found to be within 44 kW from the actual observed load value. Hence an error of around 44 kW, represents an accuracy of more than 90% in predictions of peak loads. As discussed in section 4.3.2, the source of errors found for this UNSW building follow a similar trend to the TETB case study, due to the fact that in several instances the building’s cooling load is partially covered from the university central plant. For the sake of comparison of the relative magnitudes of the prediction errors, during the summer the base load during the night time (including cooling) of the building is in the range of 100-130 kW. However, the building load during the occupied times in the cooling period is typically in the vicinity of 450kW. Furthermore, the building peak daily loads often exceed 500 kW (cooling peak loads exceed 350 kW).

6.4.3 Estimation of savings with predictive control algorithms The NARX time series was used to estimate the savings with predictive control over the standard setup. The input time series function y(t), that represents the load was designed according to the temperature control and response algorithms that were described in section 6.3 and based on the predictions from the ST, PL and PC

188 modules that were described in section 6.1. Humidity was assumed to remain within the comfort range of 30-60% throughout the day and following the gradients of the exterior humidity as described in section 6.2.3. The simulations were carried out for a period of 3 years (2012-2014) only in the workdays of the cooling season (November – March). The comfort zone was initially set between 22-25 degrees. For the control during the night time (20:00-08:00), unless the PC module was activated, the cooling load and internal conditions of the predictive control were assumed to be aligned with the time series of load of standard control (real data).

Table 45 compares the performance of standard control (real data) and predictive control simulations in terms of cooling loads and peak load reduction during the cooling season of the simulation period. The energy values in Table 45 represent thermal energy (rather than electrical).

Table 45: Summary of cooling load and peak reduction with predictive weather control

Standard Predictive Reduction control control Average cooling demand* 246,592 kWh 217,300 kWh -24% Average daily peak cooling 381 kW 301 kW -21% load* Maximum peak cooling load 480 kW 430 kW -10% 2012* Maximum peak cooling load 470 kW 429 kW -9% 2013* Maximum peak cooling load 403 kW 361 kW -10% 2014* * working days from 1st November to 30th March

The simulations indicate energy savings of approximately 24% in cooling loads compared to the non-weather dependent control strategy over the simulation period of the cooling seasons (2012-2013). The main contributor to this reduction was the dynamic IT control module, which was discussed in section 6.3.1. Furthermore, the simulations resulted in an average 21% decrease of peak loads mainly as a result of the implementation of the PL and PC algorithm responses. The

189 predictive control was not as effective for the reduction of maximum loads (which impose substantial costs), however it still achieved reducing each year’s maximum peak load by around 10%. The reason for that is during very hot days (where the maximum peaks tend to occur), there is increased demand for cooling and less margin to take advantage of the thermal comfort zone.

The results from predictive control in Table 45 are the averages of 10 simulation runs in the NN. The annual load predictions from the simulations had a standard deviation of 9.9 MWh. The resulting distribution of annual cooling load and peak load simulation results for the 10 runs can be seen in Figure 65 and Figure 66 respectively. The figures show the averages for annual cooling demand and maximum peak load as derived from each individual simulation, as well as the average and compare it to the standard control of the case study building (real data).

Comparison of simulations for average annual cooling load

Average NN 217.3 NN10 215 NN9 211 NN8 205 NN7 220 NN6 218 NN5 222 NN4 236 NN3 214 NN2 202 NN1 230 Standard Control 247

0 50 100 150 200 250 300 Annual cooling demand (MWh)

Figure 65: Simulation result distribution for each NN iteration - predictions of average annual cooling load and comparison with standard control

190

COMPARISON OF SIMULATIONS FOR MAXIMUM PEAK LOAD NN1 NN2 NN3 NN4 NN5 NN6 NN7 NN8

NN9 NN10 Average NN Standard Control

480

470

462.00 462.00

460.0

458

452

448.00

445.00

440.0

437.00

437

436.00

433

430.40

430.00

429.30

418.00

410

404.00

403

402.00

401.00

384.00

379.0

378

376

375.00

372.00

365.00

361.00

360.90

356

350.00

346

329.00 MAXIMUM LOAD(KW) MAXIMUM PEAK

2012 2013 2014

Figure 66: Simulation result distribution for each NN iteration - predictions of maximum peak load and comparison with standard control To illustrate the ways that savings may be realised a sample two day period is analysed. The selected days (7th and 8th January 2013) demonstrate the effects of the dynamic IT control and peak load control algorithms during a normal summer day and a very hot summer day. The respective cooling load and building load over 48 hours with standard control can be seen in Figure 67.

191

Comparison of loads and ambient temperature 800 45

700 40

35 600 30 500 25 400

20 Load (kW) Load 300 15 (C) Temperature 200 10

100 5

0 0

0:00 2:00 4:00 6:00 8:00 0:00 2:00 4:00 6:00 8:00

14:00 20:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 10:00 12:00 16:00 18:00 22:00

Cooling Total Ambient

Figure 67: Comparison of total building and cooling loads (standard control) with ambient temperature in a period of 2 days (7-8th January 2013) For the first day (7/1) the peak detection algorithm (section 4.2.2) does not detect a significant peak, while for the following day (8/1) it does, as the ambient temperature, humidity and total CDH are predicted to be relatively higher than standard. Hence for the second day in addition to the dynamic IT responses, the PL algorithm control is activated. The second day (8/1) was associated with very high temperatures throughout the day and one of the highest peak loads for the year.

Figure 68 shows a comparison of how IT is controlled with both the standard setup and the predictive model. The predictive control attempts to minimise loads by taking advantage of the whole range of the thermal comfort zone in response to the anticipated conditions and expected peak zones and this is evident in the temperature changes during the occupied period.

192

Comparison of internal temperature controls 45

30 Temperature (C) Temperature

0:00 1:30 3:00 4:30 6:00 7:30 9:00 0:00 1:30 3:00 4:30 6:00 7:30 9:00

21:00 21:00 10:30 12:00 13:30 15:00 16:30 18:00 19:30 22:30 10:30 12:00 13:30 15:00 16:30 18:00 19:30 22:30

Ambient Standard control Predictive Control

Figure 68: Evolution of internal temperature with standard control and predictive control (simulated according to algorithms in section 6.3) for the period of 7-8th January 2013. The dashed lines represent the thermal comfort boundaries (22-25°C) As discussed above, the simulations showed that during very hot days with extreme peaks the savings potential decreases, as there is lower margin to benefit from changing IT in the thermal comfort zone. The total cooling demand for the two days was decreased by 24.5% and 14% respectively with the predictive control. The peak load was decreased by 13.5% and 8.4% respectively. Table 46 summarises the performance of each control method.

Table 46: Comparison of demand and peak loads for the sample two day period for each control method

7th January 8th January Standard control Total cooling demand (kWh) 3,432 4,645 Peak cooling load (kW) 372 465 Predictive control Total cooling demand (kWh) 2,590 3,997 Peak cooling load (kW) 322 426

193

The cooling load changes for each control method during the two day period can be seen in Figure 69.

Cooling load comparison between two control strategies 500 465 450 372 426

400

350 322 300

250

200 Cooling load (kW) loadCooling 150

100

0:00 1:30 3:00 4:30 6:00 7:30 9:00 0:00 1:30 3:00 4:30 6:00 7:30 9:00

13:30 21:00 10:30 12:00 13:30 15:00 16:30 18:00 19:30 21:00 22:30 10:30 12:00 15:00 16:30 18:00 19:30 22:30

Standard Control Predictive control

Figure 69: Cooling load comparison for a two day period (7th and 8th January 2012) between standard control and predictive control. The daily peak loads are indicated for both models. In Figure 69 it is possible to see the predictive control in action. According to the algorithms, the cooling load tends to be increased (sometimes higher than standard control) during the early times of the day. This allows the building to be kept cool and eliminate substantial demand from the afternoon times, when the IT is set to gradually increase towards the HT level.

6.4.4 Precooling control algorithm effects The building does not any have passive precooling policy in place, such as night-time ventilation via open windows or ducts. Some occupants may actually leave windows open (according to the campus management), however there is no consistent pattern or data that could be modelled for passive precooling. Hence all overnight precooling was assumed to occur via the HVAC system. This results in negligible cooling demand savings, as the energy used to cool down the building overnight is 194 offset by the energy that would be otherwise used during the following day. However, since higher demand occurs overnight during the off-peak times, the potential costs are much lower. Unfortunately, pricing data was deemed commercially sensitive and could not be used at the time the simulations were implemented; hence it was not possible to estimate the cost savings.

Nevertheless, PC may also assist with peak shaving, which was visible in certain cases during the simulations. The peak shaving process, as discussed in section 2.3.3.2, is the result of reduced cooling demand during the middle of the day, assuming the building was sufficiently preconditioned overnight. While the relative peak shaving is marginal in most cases (section 2.3.3.2), it contributes significantly to financial savings, as rolling maximum peaks impose very high energy costs in commercial building systems. It should be noted that for the location of the case study building, the precooling frequency and value ratios (as calculated according to the tool in chapter 5) were relatively low: 0.10 and -0.58 respectively. This implies that precooling has only limited potential for this site. Regardless, an illustration of the IT as set by the predictive control algorithm compared to the standard control can be seen in Figure 70.

195

Precooling control temperature profile 35

23 Temperature (C) Temperature 21

Ambient Standard Predictive

Figure 70: Comparison of internal temperature control with the standard or predictive setup for a sample day (1st November 2012) - precooling control is active in order to shift some of the load overnight in expectation for the following hot day. The dashed lines represent the thermal comfort zone boundaries (22-25°C) As discussed earlier, the predictive control demand was assumed to match the standard control (real data) for the non-occupied period overnight. However, this changes when the PC module is activated and the IT tracks the T predictions the night before. This can be seen in Figure 70, for the time steps that the T is predicted to be lower than the precooling boundary (20°C). Figure 71 shows the cooling load profile for the same day.

196

Precooling control load profile 450 426

400

350 322

300

250

200 Cooling load (kW) loadCooling 150

100

Standard control Predictive control

Figure 71: Load profile with standard and predictive control for 1st November 2012. The effects of PC are visible after the reference hour (01:00). While the energy savings with PC activated are insignificant - 3,416 kWh for standard control and 3,259 kWh for predictive control, translating to savings of around 4.5% - the peak daily load was reduced from 426 kW to 322 kW (approximately 24% reduction). As discussed earlier, an additional benefit of precooling is that part of the cooling load is shifted to the period prior to the occupied zone (where usually energy is billed at an off-peak rate). Specifically, 484 kWh or 15% of the total cooling demand was shifted to the period 01:00-08:00 with the PC algorithm. Table 47 shows a summary of the effects of the PC response algorithm in shifting the load to the non-occupied period, as well as shaving the peak loads for the 144 days over the 4 years that the precooling conditions were met.

197

Table 47: Comparison of daily load distribution for days that the precooling conditions are met over the simulation period in the simulation time series (2012-2014)

Standard Predictive Control Control Average ratio of daily load in occupied period 96.4% 88.2% Average ratio of daily load in non-occupied 3.6% 11.8% period Average daily peak cooling load 402 kW 318 kW Average daily cooling demand 3,025 kWh 2,880 kWh

6.4.5 Effects of thermal comfort zone boundaries An advantage of the response algorithms described in section 6.3 is that there is flexibility in selecting the thermal comfort zone boundaries (LT and UT). As discussed in section 2.3, there is no consensus in the literature for the optimal thermal comfort zone, as it depends on a number of factors that are challenging to quantify. Specifically, in large commercial buildings, occupants may have significantly different personal perception of “thermal comfort” according to their age, health, physical, metabolic and cultural traits. Additionally, the type of activity and building characteristics affect the thermal comfort zone width. While it is beyond the scope of this thesis to assess the aforementioned factors, a short analysis was carried out to illustrate the effects of changing the thermal comfort zone boundaries on cooling load and peaks.

In addition to the base scenario (22-25°C) of the simulations described in sections 6.4.2.1-6.4.2.3 four more scenarios were examined:

1. LT=22°C and UT=25.5°C 2. LT=22°C and UT=26°C 3. LT=22.5°C and UT=25°C 4. LT=23°C and UT=25°C

The first two scenarios examine situations with increased UT, which may be implemented in buildings with light occupant activity (for instance offices). The increased UT may compromise some of the occupant comfort, however it is expected

198 to reduce the cooling demand, as the enthalpy difference between the ambient and internal air would be lower.

In the latter scenarios, the LT is raised instead, while the UT is constant at 25°C. This strategy may be implemented for hotter days, to relieve some of the load without compromising too much of the occupant comfort.

Table 48 shows a summary of the simulation results for the same period (2012- 2014) for each scenario. The methodology to extract the results was identical to the one described in section 6.4.2.2, using the response algorithms of section 6.3 and the predictions from the ST, PL and PC modules.

Table 48: Comparison of savings for different control strategies according to the thermal comfort range boundaries

Predictive LT (°C) UT (°C) Average annual Average peak control cooling load reduction load reduction Scenario 1 22 25.5 31% 23% Scenario 2 22 26 37% 28% Scenario 4 22.5 25 26% 21% Scenario 5 23 25 27% 22% Base 22 25 24% 21% scenario

The results show that generally higher thermal comfort zones result in higher cooling savings. Raising the UT has a more profound effect to the savings. This can be explained as when there is a very hot day, the difference between the enthalpy of ambient and interior air is lower when the IT is set to higher temperatures compared to a lower temperature. As a result the cooling system needs less energy to cool the air down. However, there is a trade-off with comfort, as many occupants may find 26°C slightly uncomfortable. Raising the LT may still result in slightly better performance of the controller compared to the base scenario for the same reasons.

6.5 Conclusions Integrating the numerical weather forecasts in a BEMS may be realised in various ways. Chapter 6 demonstrated three such applications, which are by no means 199 exhaustive. The methodology described in this chapter is based on the principle of MPC. The predictive control responses proposed in section 6.3 are activated according to three novel algorithms. The design and conditions associated with each algorithm are linked to the capacity of predictions from the numerical tools described in chapters 3-5. To illustrate the potential savings of integrating the algorithms in an existing system, an extensive set of simulations was conducted for a case study building in UNSW campus. The simulations utilised historical load and weather data, in order to correlate predictions of enthalpy (from temperature and relative humidity) to the cooling load.

The results of the simulations showed promising reductions of up to 24% in annual cooling demand and up to 21% in daily peak loads. However, for very hot days when the peak loads are getting extremely high, there is lower margin for shaving. Precooling, when possible, may improve shaving potential as demonstrated in section 6.4.4. Furthermore, precooling allows to shift a significant part of the daily cooling load (roughly 12%) to the overnight periods, when energy is typically billed at off-peak rates. Finally, it was shown that raising both the lower and upper thresholds of the thermal comfort zone may result in higher saving potential, albeit at the potential reduction of occupant comfort.

A major limitation of this work is the lack of comprehensive data. For instance, internal temperatures were averaged from sensors within the building, however many parts of the building have different set temperatures. Hence to obtain a better understanding of the cooling load, data from more locations inside the building would be necessary. Additionally, there was no reliable humidity monitoring within the case study building and hence assumptions had to be made for the humidity range during the day from the environmental conditions. Finally, part of the cooling load of the case study building is handled externally from the campus plant. Hence the correlation model was less accurate than a building with HVAC load handled exclusively onsite. Despite the limitations, the response set indicates promising result, which may be reassessed in future work. Future work may also investigate more extensively the effects of changing the thermal comfort zone on energy savings, as well as attempt to optimise the control strategies of the decision algorithm according to their relative effectiveness in peak shaving or energy savings. 200

Furthermore, pricing data may be used as inputs in the predictive control algorithms to allow quantification of the energy cost savings.

7 Conclusion Efforts to improve energy management in buildings have been fuelled by a number of drivers: reducing energy costs, improving occupant comfort and minimising the carbon footprint. Information about the past, present and future weather conditions has been recognised as a critical input for any modern building energy system, as they directly or indirectly affect the building energy demand. Specifically, the heating and cooling loads, as well as the energy generation from onsite renewable energy sources are both determined by the ambient weather conditions. The energy costs of a building are also dependent on weather, both in terms of maximum rolling peak loads and annual energy demand. In the attempt to optimise energy management, the occupant comfort, which is arguably related to productivity, imposes certain constraints.

This thesis attempts to develop and assess the integration of localised weather forecasts and analyses in building energy management. The forecasts and analyses of past weather trends are developed with the help of a numerical prediction model. The advantages of numerical predictions include improved accuracy for horizons up to a few days ahead, application at any site and ease of implementation. Using localised numerical predictions in building energy management is a novel research endeavour and this thesis constitutes a major attempt to illustrate the advantages of this approach and incite further interest.

The existing literature in the fields of forecasting and energy management strongly suggests that weather prediction integration in energy management systems results in improved performance over a non-weather sensitive control system. Data driven forecasting techniques are common in predicting weather and loads and were reviewed extensively in chapter 2 of this thesis. It was also highlighted that localisation of numerical predictions is valuable, however the development of consistently accurate and computationally light weather forecasting models is challenging. Regardless, the literature suggests that integration of weather inputs

201 and adoption of predictive control in building energy management may reduce energy demand by up to 30% in most cases.

Based on the findings and the gaps identified in the literature review, chapter 3 described the development of a short term numerical weather forecasting tool. The proposed tool was developed to produce forecasts in 6-hourly blocks with hourly resolution for ambient temperature, relative humidity and wind speed at any location. It is based on the hybridisation of two components: the numerical predictions and statistical post-processing via a linear regression or an autoregressive model. The numerical component gradually downscales synoptic scale weather observations to a localised region. The statistical post-processing is designed according to the persistence assumption, which dictates stationarity of the conditions in the atmosphere over short periods of time. In this thesis, the persistence was used as both the reference model to determine the skill of the numerical and the hybrid models, as well as an input component with decreasing weighting for the hybrid models. The hybrid models showed notable improvements in skill over both individual base components up to 38% for temperature, 28% for relative humidity and 9% for wind speed respectively. More frequent updates of the persistence component inputs, improved the accuracy of the hybrid models even further. Specifically, when the update intervals of the reference component occurred twice as often (every 3 hours), the predictions improved by up to 50% compared to the original models. The hybrid models were adjusted to develop forecasts useful for building energy system management, such as the occurrence of a sudden change in weather or an extreme heat event. Appropriate forecasting correction algorithms were designed to improve accuracy in predictions under these conditions, by adjusting the weights of the numerical and/or the statistical component. It was also shown that localised NWP result to significant improvement in accuracy in comparison to forecasts generated at a distance from an external entity.

Peak cooling loads in buildings during the summer result into high energy costs and are considered as a significant reason for costly grid upgrades. Predictions about the occurrence of peak cooling loads can help in developing demand response measures in buildings for minimising costs and lowering the building’s carbon footprint. Chapter 4 proposed a novel approach for peak cooling load predictions based on the 202 generation of an ensemble of numerical weather forecasts with a day ahead horizon. The ensemble comprises of 15 prediction branches, with each individual branch being different to others in regards to its geometry and spatial resolution. The prediction model is based on two algorithms. The first algorithm detects a period during the day that a potential cooling peak load may occur, based on the distribution of the ensemble branch predictions. The second algorithm estimates the relative magnitude of the expected peak, based on a number of factors, such as seasonality, temperature, relative humidity and cooling degree hours. For the case study university building in Sydney, Australia, the model was able to detect the occurrence of all significant daily peak loads for a period of two years. Over 90% of the peak loads were observed within the potential peak period as predicted by the model. Furthermore, it was demonstrated that the relative peak load magnitude was correlated to the ambient weather conditions, which is useful for energy demand control.

Precooling a building during the summer was identified in chapter 2 as a promising response, which when managed properly, has the potential to contribute significantly to the reduction of building energy demand and cost. The principle behind precooling is using ventilation or the air conditioning system overnight at low cost to shift some of the cooling load away from the occupied period of the following day. The potential of precooling depends on the characteristics of the building as well as the diurnal weather differences at the location. The proposed tool in chapter 5 was used to assess the climatological component of precooling from real data by running simulations in the numerical model platform. The results were processed into a range of dimensionless ratios that may be used in building design, energy management and optimisation decisions, as well as site comparison. The analytical tool was applied to a number of locations within the Sydney metropolitan area, and five more Australian cities. It was shown that sites within the same climate separated by a few kilometres have significantly different responses to precooling due to the effects of (or lack thereof) proximity to the ocean, vegetation coverage, urban density and landscape. These effects cannot be accounted for in design analyses using typical meteorological year (TMY) data or data obtained from weather stations that are at a distance from the building. For cities in temperate

203 climates, such as Sydney and Melbourne, it was found that inland sites with high vegetation and low urban density are primary candidates for precooling.

Finally, to illustrate the value of integrating the forecasting and analytical tools chapter 6 described three predictive control algorithms. The algorithms are based on the short term hybrid forecasting methodology of chapter 3, the peak load prediction methodology of chapter 4 and the precooling potential assessment of chapter 5. The primary operation principle of the algorithms is to adjust the internal temperature of a building both in anticipation and in response to ambient weather conditions. Compared to standard HVAC control, the predictive controller proposed in this chapter takes advantage of the whole range of the thermal comfort zone. Three years’ worth of simulations were carried out for a case study university building to investigate the effectiveness of the proposed responses. The simulations were based on the strong positive correlation of cooling load over the summer period and the hourly enthalpy of air changes. The results showed a reduction of the cooling demand by 24% on average and the average daily peaks by 21%. However, the maximum cooling peaks for the season, which are more significant for the building energy costs were only reduced by 9% on average.

The work presented in this thesis aimed to both design tools for the generation of weather forecasts and trend analysis, as well as illustrate a range of applications for their effective integration in building energy management. Additionally, the analytical tools may be integrated in the building design stage in place or in parallel to modelling based on Typical Meteorological Year data or thermal modelling. The findings from the discussion in the thesis are expected to enhance the understanding of the value of integration of weather inputs in building energy management and control. Furthermore, the novel approach of implementing responses based on numerical predictions is expected to spark further interest in the field of forecasting and predictive control.

204

Appendix

Figure 72: Comparison of predictions for temperature using the overall ensemble mean vs the ensemble mean from branches within 1 standard deviation for a sample 11 day period 205

Figure 73: Comparison of predictions for relative humidity using the overall ensemble mean vs the ensemble mean from branches within 1 standard deviation for a sample 11 day period 206

Figure 74: Time series ensemble forecasting for temperature for a sample 11 period, showing the actual temperature (BOM in red), the ensemble z1 predicted temperature (in green) and the predictions of individual branches 207

Figure 75: Time series ensemble forecasting for temperature for a sample 11 period, showing the actual rel. humidity (BOM in red), the ensemble z1 predicted rel. humidity (in green) and the predictions of individual branches 208

Figure 76: Quartiles of ensemble branch predictions comparison to observed temperature (in purple) for a sample 11 day period 209

Figure 77: Quartiles of ensemble branch predictions comparison to observed rel. humidity (in purple) for a sample 11 day period 210

Bibliography ABDEL-AAL, R. E. 2004. Hourly temperature forecasting using abductive networks. Engineering Applications of Artificial Intelligence, 17, 543-556. ADAMOPOULOU, A. A., TRYFERIDIS, A. M. & TZOVARAS, D. K. 2016. A context-aware method for building occupancy prediction. Energy and Buildings, 110, 229- 244. ADAMS, M., DUC, H. & TRIEU, T. 2015. Impacts of land-use change on Sydney's future temperatures. Sydney: NSW Office of Environment & Heritage. AGÜERA-PÉREZ, A., PALOMARES-SALAS, J. C., GONZÁLEZ DE LA ROSA, J. J. & MORENO-MUÑOZ, A. 2013. Spatial persistence in wind analysis. Journal of Wind Engineering and Industrial Aerodynamics, 119, 48-52. AHMED, O. Prediction of building thermal load using recurrent neural network. Engineering Applications of Neural Networks. Proceeding of the 5th International Conference on Engineering Applications of Neural Networks (EANN'99), 13-15 Sept. 1999, 1999 Torun, Poland. Wydawnictwo Adam Marszalek, 99-104. AL-MESSABI, N., LI, Y., EL-AMIN, I. & GOH, C. Forecasting of photovoltaic power yield using dynamic neural networks. 2012 Annual International Joint Conference on Neural Networks, IJCNN 2012, Part of the 2012 IEEE World Congress on Computational Intelligence, WCCI 2012, June 10, 2012 - June 15, 2012, 2012 Brisbane, QLD, Australia. Institute of Electrical and Electronics Engineers Inc., IEEE Computational Intelligence Society (CIS); International Neural Network Society (INNS). AL-RABGHI, O. M. A. & AL-JOHANI, K. M. 1997. Utilizing transfer function method for hourly cooling load calculations. Energy Conversion and Management, 38, 319-332. AL-YAHYAI, S., CHARABI, Y., AL-BADI, A. & GASTLI, A. 2012. Nested ensemble NWP approach for wind energy assessment. Renewable Energy, 37, 150-160. ALESSANDRINI, S., DELLE MONACHE, L., SPERATI, S. & CERVONE, G. 2015. An analog ensemble for short-term probabilistic solar power forecast. Applied Energy, 157, 95-110. ANCELL, B. C., KASHAWLIC, E. & SCHROEDER, J. L. 2015. Evaluation of wind forecasts and observation impacts from variational and ensemble data assimilation for wind energy applications. Monthly Weather Review, 143, 3230-3245. ARGIRIOU, A. A., BELLAS-VELIDIS, I. & BALARAS, C. A. 2000. Development of a neural network heating controller for solar buildings. Neural Networks, 13, 811-820. ARGIRIOU, A. A., BELLAS-VELIDIS, I., KUMMERT, M. & ANDRE, P. 2004. A neural network controller for hydronic heating systems of solar buildings. Neural Networks, 17, 427-440. ARMSTRONG, A., OSTLE, N. J. & WHITAKER, J. 2016. Solar park microclimate and vegetation management effects on grassland carbon cycling. Environmental Research Letters, 11. ARMSTRONG, P. R., LEEB, S. B. & NORFORD, L. K. Control with building mass - Part I: Thermal response model. 2006 Winter Meeting of the American Society of Heating, Refrigerating and Air-Conditioning Engineers, ASHRAE, January 21,

211

2006 - January 25, 2006, 2006 Chicago, IL, United states. Amer. Soc. Heating, Ref. Air-Conditoning Eng. Inc., 449-461. ASANO, H., IMAMURA, E. & SOGAI, S. 1994. Optimal planning of cogeneration systems under time-of-use rates. Electrical Engineering in Japan (English translation of Denki Gakkai Ronbunshi), 114, 102-110. ASHRAE 2009. ASHRAE handbook: fundamentals, Atlanta, USA, American Society of Heating, Refrigerating and Air-Conditioning Engineering. ASHRAE 2012. Standard 55 - Thermal Environmental Conditions for Human Occupancy. Berkeley, CA, USA. ASHRAE 2013. Ventilation for Acceptable Indoor Air Quality, Atlanta, American Society of Heating, Refrigerating and Air-Conditioning Engineers. ASWANI, A., MASTER, N., TANEJA, J., KRIOUKOV, A., CULLER, D. & TOMLIN, C. Energy-efficient building HVAC control using hybrid system LBMPC. 2012. 496-501. BACHER, P., MADSEN, H. & NIELSEN, H. A. 2009. Online short-term solar power forecasting. Solar Energy, 83, 1772-1783. BACHER, P., MADSEN, H. & PERERS, B. Short-term solar collector power forecasting. 30th ISES Biennial Solar World Congress 2011, SWC 2011, August 28, 2011 - September 2, 2011, 2011 Kassel, Germany. International Solar Energy Society, 4579-4589. BEN-NAKHI, A. E. & MAHMOUD, M. A. 2004. Cooling load prediction for buildings using general regression neural networks. Energy Conversion and Management, 45, 2127-2141. BIYIK, E., GENC, S. & BROOKS, J. D. Model predictive building thermostatic controls of small-to-medium commercial buildings for optimal peak load reduction incorporating dynamic human comfort models: Algorithm and implementation. 2014 IEEE Conference on Control Applications, CCA 2014, October 8, 2014 - October 10, 2014, 2014 Juan Les Antibes, France. Institute of Electrical and Electronics Engineers Inc., 2009-2015. BOLATTURK, A. 2008. Optimum insulation thicknesses for building walls with respect to cooling and heating degree-hours in the warmest zone of Turkey. Building and Environment, 43, 1055-1064. BORGES, C. E., PENYA, Y. K. & FERNANDEZ, I. Optimal combined short-term building load forecasting. IEEE Power and Energy Society'sInnovative Smart Grid Technologies Asia 2011 Conference,ISGT Asia 2011, November 13, 2011 - November 16, 2011, 2011 Perth, WA, Australia. IEEE Computer Society, IEEE Power and Energy Society. BOX, G. E. P. & JENKINS, G. 1970. Time series analysis, forecasting and control, San Francisco, CA, USA, Holden-Day. BRAUN, J. E. Reducing energy costs and peak electrical demand through optimal control of building thermal storage. 1990 Annual Meeting of the American Society of Heating, Refrigerating and Air-Conditioning Engineers, Technical and Symposium Papers, June 10, 1990 - June 13, 1990, 1990 St. Louis, MO, USA. Publ by ASHRAE, 876-888. BRAUN, J. E. & CHATURVEDI, N. 2002. An inverse gray-box model for transient building load prediction. HVAC and R Research, 8, 73-99. BRAUN, J. E., MONTGOMERY, K. W. & CHATURVEDI, N. 2001. Evaluating the performance of building thermal mass control strategies. HVAC and R Research, 7, 403-428. 212

BRAUN, J. E. & ZHONG, Z. 2005. Development and evaluation of a night ventilation precooling algorithm. HVAC and R Research, 11, 433-458. BRILLER, D. L. Strategies for optimally managing building peak demand under climate-altered scenarios. World Energy Engineering Congress 2012, WEEC 2012, October 31, 2012 - November 2, 2012, 2012 Atlanta, GA, United states. The Fairmont Press, Inc., 857-879. BROVKIN, V. 2002. Climate-vegetation interaction. J. Phys. IV France, 12, 57-72. BUNDE, A. & HAVLIN, S. 2002. Power-law persistence in the atmosphere and in the oceans. Physica A: Statistical Mechanics and its Applications, 314, 15-24. BUREAU OF METEOROLOGY. 2015. Monthly mean temperatures [Online]. Bureau of Meteorology. Available: http://www.bom.gov.au/jsp/ncc/cdio/weatherData/av?p_nccObsCode=36 &p_display_type=dataFile&p_startYear=&p_c=&p_stn_num=067113 [Accessed 16 August 2015. BURGER, E. M. & MOURA, S. J. 2015. Gated ensemble learning method for demand- side electricity load forecasting. Energy and Buildings, 109, 23-34. BURN. 2012. The electricity grid: A history [Online]. Burn: An Energy Journal. Available: http://burnanenergyjournal.com/the-electricity-grid-a-history/ [Accessed 10th July 2016. CAI, T., DUAN, S. & CHEN, C. Forecasting power output for grid-connected photovoltaic power system without using solar radiation measurement. 2nd International Symposium on Power Electronics for Distributed Generation Systems, PEDG 2010, June 16, 2010 - June 18, 2010, 2010 Hefei, China. IEEE Computer Society, 773-777. CAI, X., HEJAZI, M. I. & WANG, D. 2011. Value of probabilistic weather forecasts: Assessment by real-time optimization of irrigation scheduling. Journal of Water Resources Planning and Management, 137, 391-403. CANDANEDO, J. A., DEHKORDI, V. R. & STYLIANOU, M. 2013a. Model-based predictive control of an ice storage device in a building cooling system. Applied Energy, 111, 1032-1045. CANDANEDO, J. A., PARADIS, E. & STYLIANOU, M. Building simulation weather forecast files for predictive control strategies. 4th Symposium on Simulation for Architecture and Urban Design, SimAUD 2013, Part of the 2013 Spring Simulation Multiconference, SpringSim 2013, April 7, 2013 - April 10, 2013, 2013b San Diego, CA, United states. The Society for Modeling and Simulation International, 22-27. CHEN, C., DUAN, S., CAI, T. & LIU, B. 2011. Online 24-h solar power forecasting based on weather type classification using artificial neural network. Solar Energy, 85, 2856-2870. CHEN, T.-Y. & ATHIENITIS, A. K. 1996. Ambient temperature and solar radiation prediction for predictive control of HVAC systems and a methodology for optimal building heating dynamic operation. ASHRAE Transactions, 102, 26- 35. CHEN, T. & YU, Z. 2009. A statistical method for selection of sequences of coincident weather parameters for design cooling load calculations. Energy Conversion and Management, 50, 813-821. CHEN, Z. & SOH, Y. C. 2016. Comparing occupancy models and data mining approaches for regular occupancy prediction in commercial buildings. 1-9.

213

CHENVIDYAKARN, T. & WOODS, A. 2005. Top-down precooled natural ventilation. Building Services Engineering Research and Technology, 26, 181-193. CHERKASSKY, V., ROY CHOWDHURY, S., LANDENBERGER, V., TEWARI, S. & BURSCH, P. Prediction of electric power consumption for commercial buildings. 2011 International Joint Conference on Neural Network, IJCNN 2011, July 31, 2011 - August 5, 2011, 2011 San Jose, CA, United states. Institute of Electrical and Electronics Engineers Inc., 666-672. CHO, H., LUCK, R., EKSIOGLU, S. D. & CHAMRA, L. M. 2009. Cost-optimized real-time operation of CHP systems. Energy and Buildings, 41, 445-451. CHOU, S. K. & CHANG, W. L. 1993. Development of an energy-estimating equation for large commercial buildings. International Journal of Energy Research, 17, 759-773. CHOW, S. K. H., LEE, E. W. M. & LI, D. H. W. 2012. Short-term prediction of photovoltaic energy generation by intelligent approach. Energy and Buildings, 55, 660-667. CHOWDHURY, B. H. & RAHMAN, S. FORECASTING SUB-HOURLY SOLAR IRRADIANCE FOR PREDICTION OF PHOTOVOLTAIC OUTPUT. Conference Record of the Nineteenth IEEE Photovoltaic Specialists Conference - 1987., 1987 New Orleans, LA, USA. IEEE, 171-176. CHUA, K. J. & CHOU, S. K. 2011. A performance-based method for energy efficiency improvement of buildings. Energy Conversion and Management, 52, 1829- 1839. CIGLER, J. & PRIVARA, S. Subspace identification and model predictive control for buildings. 2010 11th International Conference on Control Automation Robotics & Vision (ICARCV 2010), 7-10 Dec. 2010, 2010 Piscataway, NJ, USA. IEEE, 750-5. CLOUDT, C. L., GOMEZ, J. D., NISHIMOTO, T. K. & SHEPHARD, L. E. Coupling simulation tools and real-time data to improve building energy performance. 2013 IEEE Green Technologies Conference, GREENTECH 2013, April 4, 2013 - April 5, 2013, 2013 Denver, CO, United states. IEEE Computer Society, 373- 379. COLE, W. J., POWELL, K. M., HALE, E. T. & EDGAR, T. F. 2014. Reduced-order residential home modeling for model predictive control. Energy and Buildings, 74, 69-77. COLLAZOS, A., MARECHAL, F. & GAHLER, C. 2009. Predictive optimal management method for the control of polygeneration systems. Computers and Chemical Engineering, 33, 1584-1592. COOPERMAN, A., DIECKMANN, J. & BRODRICK, J. 2010. Using weather data for predictive control. ASHRAE Journal, 52, 130-132. DAUM, D. & MOREL, N. 2010. Assessing the total energy impact of manual and optimized blind control in combination with different lighting schedules in a building simulation environment. Journal of Building Performance Simulation, 3, 1-16. DAVID, M., ADELARD, L., GARDE, F. & BOYER, H. Weather data analysis based on typical weather sequences. application: Energy building simulation. Building Simulation 2005 - 9th International IBPSA Conference, BS 2005, August 15, 2005 - August 18, 2005, 2005 Montreal, Canada. International Building Performance Simulation Association, 239-246.

214

DAY, A. R., JONES, P. G. & MAIDMENT, G. G. 2009. Forecasting future cooling demand in London. Energy and Buildings, 41, 942-948. DAY, A. R. & KARAYIANNIS, T. G. 1998. Degree-days: comparison of calculation methods. Building Services Engineering Research and Technology, 19, 7-13. DAY, T. 2006. TM41 Degree Days: Theory and Application, London, CIBSE. DE FELICE, M., ALESSANDRI, A. & CATALANO, F. 2015. Seasonal climate forecasts for medium-term electricity demand forecasting. Applied Energy, 137, 435- 444. DE FELICE, M. & YAO, X. Neural networks ensembles for short-term load forecasting. Symposium Series on Computational Intelligence, IEEE SSCI 2011 - 2011 IEEE Symposium on Computational Intelligence Applications in Smart Grid, CIASG 2011, April 11, 2011 - April 15, 2011, 2011 Paris, France. IEEE Computer Society, 61-68. DE ROSA, M., BIANCO, V., SCARPA, F. & TAGLIAFICO, L. A. 2014. Historical trends and current state of heating and cooling degree days in Italy. Energy Conversion and Management, 90, 323-335. DEHGHAN, A., PRASAD, A. A., SHERWOOD, S. C. & KAY, M. 2014. Evaluation and improvement of TAPM in estimating solar irradiance in Eastern Australia. Solar Energy, 107, 668-680. DHAR, A., REDDY, T. A. & CLARIDGE, D. E. 1999a. Fourier series model to predict hourly heating and cooling energy use in commercial buildings with outdoor temperature as the only weather variable. Journal of Solar Energy Engineering, Transactions of the ASME, 121, 47-53. DHAR, A., REDDY, T. A. & CLARIDGE, D. E. 1999b. Generalization of the Fourier series approach to model hourly energy use in commercial buildings. Journal of Solar Energy Engineering, Transactions of the ASME, 121, 54-62. DOBBS, J. R. & HENCEY, B. M. 2014. Model predictive HVAC control with online occupancy model. Energy and Buildings, 82, 675-684. DOE, U. 2015. EnergyPlus Simulation Weather data: Australia [Online]. US Department of Energy. Available: http://apps1.eere.energy.gov/buildings/energyplus/cfm/weather_data3.cf m/region=5_southwest_pacific_wmo_region_5/country=AUS/cname=Austr alia [Accessed 3 August 2015. DONG, B., LAM, K. P. & NEUMAN, C. P. Integrated building control based on occupant behavior pattern detection and local weather forecasting. 12th Conference of International Building Performance Simulation Association Building Simulation 2011, BS 2011, November 14, 2011 - November 16, 2011, 2011 Sydney, NSW, Australia. International Building Performance Simulation Association, 193-200. DURMAYAZ, A., KADOGLU, M. & EN, Z. 2000. An application of the degree-hours method to estimate the residential heating energy requirement and fuel consumption in Istanbul. Energy, 25, 1245-1256. ECMWF. 2015. Global Reanalyses [Online]. [Accessed 17 March 2015 2015]. EICHHORN, J., SCHRODIN, R. & ZDUNKOWSKI, W. 1988. Three-dimensional numerical simulations of the urban climate. Contributions to Atmospheric Physics, 61, 187-203. ENGINEERING TOOLBOX. 2016a. Cooling Loads - Latent and Sensible Heat [Online]. Available: http://www.engineeringtoolbox.com/latent-sensible-cooling- load-d_245.html [Accessed 13 July 2016. 215

ENGINEERING TOOLBOX. 2016b. Water vapor and saturation pressure in humid air [Online]. The Engineering Toolbox. Available: http://www.engineeringtoolbox.com/water-vapor-saturation-pressure-air- d_689.html [Accessed 29th July 2016. ERICKSON, V. L. & CERPA, A. E. Occupancy based demand response HVAC control strategy. 2nd ACM Workshop on Embedded Sensing Systems for Energy- Efficiency in Buildings, BuildSys'10, November 2, 2010 - November 2, 2010, 2010 Zurich, Switzerland. Association for Computing Machinery, 7-12. ESCRIVA-ESCRIVA, G., ALVAREZ-BEL, C. & VALENCIA-SALAZAR, I. 2010. Method for modelling space conditioning aggregated daily load curves: Application to a university building. Energy and Buildings, 42, 1275-1282. FAN, C., XIAO, F. & WANG, S. 2014. Development of prediction models for next-day building energy consumption and peak power demand using data mining techniques. Applied Energy, 127, 1-10. FERGUSON, A. B. & WINN, C. B. Development of an optimal controller of a small commercial passive solar building. Solar Engineering 1989 - Proceedings of the Eleventh Annual ASME Solar Energy Conference, April 2, 1989 - April 5, 1989, 1989 San Diego, CA, USA. Publ by American Soc of Mechanical Engineers (ASME), 65-69. FERHATBEGOVIC, T., PALENSKY, P., FONTANELLA, G. & BASCIOTTI, D. Modelling and design of a linear predictive controller for a solar powered HVAC system. 2012 IEEE 21st International Symposium on Industrial Electronics, 28-31 May 2012, 2012 Piscataway, NJ, USA. IEEE, 869-74. FERNANDEZ, I., BORGES, C. E. & PENYA, Y. K. Efficient building load forecasting. 2011 IEEE 16th Conference on Emerging Technologies and Factory Automation, ETFA 2011, September 5, 2011 - September 9, 2011, 2011 Toulouse, France. Institute of Electrical and Electronics Engineers Inc., IEEE Industrial Electronics Society; University of Toulouse 3. FERRANO, F. J. & WONG, K.-F. V. Prediction of thermal storage loads using a neural network. 1990 Annual Meeting of the American Society of Heating, Refrigerating and Air-Conditioning Engineers, Technical and Symposium Papers, June 10, 1990 - June 13, 1990, 1990 St. Louis, MO, USA. Publ by ASHRAE, 723-726. FERREIRA, P. M., GOMES, J. M., MARTINS, I. A. C. & RUANO, A. E. 2012a. A neural network based intelligent predictive sensor for cloudiness, solar radiation and air temperature. Sensors (Switzerland), 12, 15750-15777. FERREIRA, P. M., SILVA, S. M. & RUANO, A. E. Model based predictive control of HVAC systems for human thermal comfort and energy consumption minimisation. 1st IFAC Conference on Embedded Systems, Computational Intelligence and Telematics in Control, CESCIT 2012, April 3, 2012 - April 5, 2012, 2012b Wurzburg, Germany. IFAC Secretariat, 236-241. FLORITA, A. R. & HENZE, G. P. 2009. Comparison of short-term weather forecasting models for model predictive control. HVAC and R Research, 15, 835-853. FOUCQUIER, A., ROBERT, S., SUARD, F., STEPHAN, L. & JAY, A. 2013. State of the art in building modelling and energy performances prediction: A review. Renewable and Sustainable Energy Reviews, 23, 272-288. FRAISSE, G., VIARDOT, C., LAFABRIE, O. & ACHARD, G. 2002. Development of a simplified and accurate building model based on electrical analogy. Energy and Buildings, 34, 1017-1031. 216

FRANK, S. M. & SEN, P. K. Estimation of electricity consumption in commercial buildings. 43rd North American Power Symposium, NAPS 2011, August 4, 2001 - August 6, 2011, 2011 Boston, MA, United states. IEEE Computer Society, Power and Energy Society (IEEE PES); Northeastern University College of Engineering. GAO, X., MALKAWI, A. M. & YI, Y. K. A new method for predicting mixed-use building energy: The use of simulation to develop statistical models. 13th Conference of International Building Performance Simulation Association, BS 2013, August 26, 2013 - August 28, 2013, 2013 Chambery, France. International Building Performance Simulation Association, 2349-2356. GOLLINGS, J. 2006. Oblique view of the front facade reveals the sophisticated abstraction of the structure and fenestration. Architecture AU. GONZALEZ, P. A. & ZAMARRENO, J. M. 2005. Prediction of hourly energy consumption in buildings based on a feedback artificial neural network. Energy and Buildings, 37, 595-601. GOUDA, M. M., DANAHER, S. & UNDERWOOD, C. P. 2002. Application of an artificial neural network for modelling the thermal dynamics of a building's space and its heating system. Mathematical and Computer Modelling of Dynamical Systems, 8, 333-44. GOULD, P. G., KOEHLER, A. B., ORD, J. K., SNYDER, R. D., HYNDMAN, R. J. & VAHID- ARAGHI, F. 2008. Forecasting time series with multiple seasonal patterns. European Journal of Operational Research, 191, 205-220. GRANT, J., ELTOUKHY, M. & ASFOUR, S. 2014. Short-term electrical peak demand forecasting in a large government building using artificial neural networks. Energies, 7, 1935-1953. GREENSFELDERA, E. M., HENZEA, G. P. & FELSMANNB, C. 2011. An investigation of optimal control of passive building thermal storage with real time pricing. Journal of Building Performance Simulation, 4, 91-104. GREYBUSH, S. J., HAUPT, S. E. & YOUNG, G. S. 2008. The regime dependence of optimally weighted ensemble model consensus forecasts of surface temperature. Weather and Forecasting, 23, 1146-1161. GUNAY, H. B., BURSILL, J., HUCHUK, B., O'BRIEN, W. & BEAUSOLEIL-MORRISON, I. 2014. Shortest-prediction-horizon model-based predictive control for individual offices. Building and Environment, 82, 408-419. HAGAN, M. T. & BEHR, S. M. 1987. The time series approach to short term load forecasting. IEEE Transactions on Power Systems, PWRS-2, 785-91. HAGRAS, H., PACKHAM, I., VANDERSTOCKT, Y., MCNULTY, N., VADHER, A. & DOCTOR, F. An intelligent agent based approach for energy management in commercial buildings. 2008 IEEE International Conference on Fuzzy Systems, FUZZ 2008, June 1, 2008 - June 6, 2008, 2008 Hong Kong, China. Institute of Electrical and Electronics Engineers Inc., 156-162. HAJIAH, A. & KRARTI, M. Optimal control of building storage systems using both ice storage and thermal mass - Part I: Simulation environment. 2012 Langford Lane, Kidlington, Oxford, OX5 1GB, United Kingdom. Elsevier Ltd, 499-508. HAVENITH, G., HOLMÉR, I. & PARSONS, K. 2002. Personal factors in thermal comfort assessment: clothing properties and metabolic heat production. Energy and Buildings, 34, 581-591.

217

HAZYUK, I., GHIAUS, C. & PENHOUET, D. 2012. Optimal temperature control of intermittently heated buildings using Model Predictive Control: Part II - Control algorithm. Building and Environment, 51, 388-394. HE, D.-S. & ZHANG, X. 2005. Analysis of air conditioning load prediction by modified seasonal exponential smoothing model. Tongji Daxue Xuebao/Journal of Tongji University, 33, 1672-1676. HENZE, G. P., DODIER, R. H. & KRARTI, M. Development of a predictive optimal controller for thermal energy storage systems. Proceedings of the 1998 ASHRAE Winter Meeting. Part 1 (of 2), January 18, 1998 - January 21, 1998, 1998 San Francisco, CA, USA. ASHRAE, 54. HENZE, G. P., FELSMANN, C. & KNABE, G. 2004. Evaluation of optimal control for active and passive building thermal storage. International Journal of Thermal Sciences, 43, 173-183. HENZE, G. P., KALZ, D. E., LIU, S. & FELSMANN, C. 2005. Experimental analysis of model-based predictive optimal control for active and passive building thermal storage inventory. HVAC and R Research, 11, 189-213. HENZE, G. P. & KRARTI, M. Impact of forecasting uncertainty on the performance of a predictive optimal controller for thermal energy storage systems. ASHRAE Annual Meeting, June 18, 1999 - June 23, 1999, 1999 Seattle, WA, USA. ASHRAE, PART 2/. HENZE, G. P., PFAFFEROTT, J., HERKEL, S. & FELSMANN, C. 2007. Impact of adaptive comfort criteria and heat waves on optimal building thermal mass control. Energy and Buildings, 39, 221-235. HIBBERD, M. 2011. TAPM 4 Modelling of Wagerup Meteorology and 25A Dispersion. HINNENTHAL, J. & CLAUSS, G. 2010. Robust Pareto-optimum routing of ships utilising deterministic and ensemble weather forecasts. Ships and Offshore Structures, 5, 105-114. HOFFMAN, A. J. Peak demand control in commercial buildings with target peak adjustment based on load forecasting. Proceedings of the 1998 IEEE International Conference on Control Applications. Part 2 (of 2), September 1, 1998 - September 4, 1998, 1998 Trieste, Italy. IEEE, 1292-1296. HOLT, C. C. 2004. Forecasting seasonals and trends by exponentially weighted moving averages. International Journal of Forecasting, 20, 5-10. HOU, Z., LIAN, Z., YAO, Y. & YUAN, X. 2006. Cooling-load prediction by the combination of rough set theory and an artificial neural-network based on data-fusion technique. Applied Energy, 83, 1033-1046. HOWARD, J. & HOFF, W. Forecasting building occupancy using sensor network data. 2nd Int. Workshop on Big Data, Streams and Heterogeneous Source Mining: Algorithms, Systems, Programming Models and Applications, BigMine 2013 - Held in Conj. with SIGKDD 2013 Conf., August 11, 2013 - August 11, 2013, 2013 Chicago, IL, United states. Association for Computing Machinery, 87-94. HU, J. & KARAVA, P. 2014. Model predictive control strategies for buildings with mixed-mode cooling. Building and Environment, 71, 233-244. HUANG, C. Y., CHEN, H. J., CHAN, C. C., CHOU, C. P. & CHIANG, C. M. Thermal model based power-generated prediction by using meteorological data in BIPV system. 1st International Conference on Smart Grid and Clean Energy Technologies, ICSGCE 2011, September 27, 2011 - September 30, 2011, 2011 Chengdu, China. Elsevier Ltd, 531-537.

218

HUANG, H., CHEN, L., MOHAMMADZAHERI, M., HU, E. & CHEN, M. Multi-zone temperature prediction in a commercial building using artificial neural network model. 2013 10th IEEE International Conference on Control and Automation, ICCA 2013, June 12, 2013 - June 14, 2013, 2013 Hangzhou, China. IEEE Computer Society, 1896-1901. HUANG, J., RITSCHARD, R. L., BULL, J. C. & CHANG, L. CLIMATIC INDICATORS FOR ESTIMATING RESIDENTIAL HEATING AND COOLING LOADS. ASHRAE Transactions: Technical and Symposium Papers Presented at the 1987 Winter Meeting., 1987 New York, NY, USA. ASHRAE, 72-111. HURLEY, P. 2008a. TAPM V4. Part 1: Technical Description. CSIRO Marine and Atmospheric Research Paper No. 25: CSIRO. HURLEY, P. 2008b. TAPM V4. User Manual, CSIRO Marine and Atmospheric Research Internal Report No. 5, CSIRO. HURLEY, P. Development and Verification of TAPM. 29th NATO/CCMS International Technical Meeting on Air Pollution Modelling and Its Application, 24-28 Sept. 2007, 2009 Dordrecht, Netherlands. Springer Netherlands, 208-16. HURLEY, P. J., PHYSICK, W. L. & LUHAR, A. K. 2005. TAPM: a practical approach to prognostic meteorological and air pollution modelling. Environmental Modelling & Software, 20, 737-52. HUURMAN, C., RAVAZZOLO, F. & ZHOU, C. 2012. The power of weather. Computational Statistics and Data Analysis, 56, 3793-3807. JETCHEVA, J. G., MAJIDPOUR, M. & CHEN, W.-P. 2014. Neural network model ensembles for building-level electricity load forecasts. Energy and Buildings, 84, 214-223. JI, W. & CHEE, K. C. 2011. Prediction of hourly solar radiation using a novel hybrid model of ARMA and TDNN. Solar Energy, 85, 808-817. KAJI-O'GRADY, S. 2006. L5 [Online]. Architecture AU. Available: http://architectureau.com/articles/l5/ [Accessed 20 July 2016. KANG, S. J. & PARK, H. Sensor network based optimal energy flow control in buildings. 2013 IEEE Global Communications Conference, GLOBECOM 2013, December 9, 2013 - December 13, 2013, 2013 Atlanta, GA, United states. Institute of Electrical and Electronics Engineers Inc., 539-544. KANG, S. J., PARK, J., OH, K.-Y., NOH, J. G. & PARK, H. 2014. Scheduling-based real time energy flow control strategy for building energy management system. Energy and Buildings, 75, 239-248. KARAVA, P., ATHIENITIS, A. K., STATHOPOULOS, T. & MOURIKI, E. 2012. Experimental study of the thermal performance of a large institutional building with mixed-mode cooling and hybrid ventilation. Building and Environment, 57, 313-326. KARMACHARYA, S., PUTRUS, G., UNDERWOOD, C. & MAHKAMOV, K. Thermal modelling of the building and its HVAC system using Matlab/Simulink. 2012 2nd International Symposium on Environment-Friendly Energies and Applications (EFEA 2012), 25-27 June 2012, 2012 Piscataway, NJ, USA. IEEE, 202-6. KAWASHIMA, M., DORGAN, C. E. & MITCHELL, J. W. Hourly thermal load prediction for the next 24 hours by Arima, Ewma, LR, and an artificial neural network. Proceedings of the 1995 ASHRAE Annual Meeting, January 29, 1995 - February 1, 1995, 1995 Chicago, IL, USA. ASHRAE, 186-200.

219

KAWASHIMA, M., DORGAN, C. E. & MITCHELL, J. W. 1996. Optimizing system control with load prediction by neural networks for an ice-storage system. ASHRAE Transactions, 102, 1169-1178. KEENEY, K. R. & BRAUN, J. E. Application of building precooling to reduce peak cooling requirements. Proceedings of the 1997 ASHRAE Winter Meeting, Jan 26 - 29 1997, 1997 Philadelphia, PA, United states. ASHRAE, 463-469. KELLER, B. & COSTA, A. M. S. 2011. A Matlab GUI for calculating the solar radiation and shading of surfaces on the earth. Computer Applications in Engineering Education, 19, 161-170. KENDRICK, C., OGDEN, R., WANG, X. & BAICHE, B. 2012. Thermal mass in new build UK housing: A comparison of structural systems in a future weather scenario. Energy and Buildings, 48, 40-49. KHALED, N. & KRARTI, M. Impact of precooling control on reducing electrical peak demand for commercial buildings in Tuniisia. 2007 Energy Sustainability Conference, June 27, 2007 - June 30, 2007, 2007 Long Beach, CA, United states. American Society of Mechanical Engineers, 565-572. KICINGER, R., KROZEL, J., STEINER, M. & PINTO, J. Airport capacity prediction integrating ensemble weather forecasts. AIAA Infotech at Aerospace Conference and Exhibit 2012, June 19, 2012 - June 21, 2012, 2012 Garden Grove, CA, United states. American Institute of Aeronautics and Astronautics Inc., American Institute of Aeronautics and Astronautics (AIAA). KIM, Y.-S. & SREBRIC, J. Improvement of building energy simulation accuracy with occupancy schedules derived from hourly building electricity consumption. 2015 ASHRAE Winter Conference, January 24, 2015 - January 28, 2015, 2015 Chicago, IL, United states. Amer. Soc. Heating, Ref. Air-Conditoning Eng. Inc., 353-360. KIMBARA, A., KUROSU, S., ENDO, R., KAMIMURA, K., MATSUBA, T. & YAMADA, A. On- line prediction for load profile of an air-conditioning system. Proceedings of the 1995 ASHRAE Annual Meeting, June 24, 1995 - June 28, 1995, 1995 San Diego, CA, USA. ASHRAE, 198-207. KINTNER-MEYER, M. & EMERY, A. F. 1995. Optimal control of an HVAC system using cold storage and building thermal capacitance. Energy and Buildings, 23, 19- 31. KOSSAK, B. & STADLER, M. 2015. Adaptive thermal zone modeling including the storage mass of the building zone. Energy and Buildings, 109, 407-417. KOTEY, N. A., COLLINS, M. R. & WRIGHT, J. L. 2009. A simplified method for calculating the effective solar optical properties of a venetian blind layer for building energy simulation. Journal of Solar Energy Engineering, Transactions of the ASME, 131, 0210021-0210029. KRARTI, M., HENZE, G. P. & BELL, D. Planning horizon for a predictive optimal controller for thermal energy storage systems. ASHRAE Annual Meeting, June 18, 1999 - June 23, 1999, 1999 Seattle, WA, USA. ASHRAE, PART 2/. KREIDER, J. F. & WANG, X. A. Improved artificial neural networks for commercial building energy use prediction. 1992 ASME-JSES-KSES International Solar Energy Conference Part 1 (of 2), April 5, 1992 - April 9, 1992, 1992 Maui, HI, USA. Publ by ASME, 361-366. KRESE, G., PREK, M. & BUTALA, V. 2011. Incorporation of latent loads into the cooling degree days concept. Energy and Buildings, 43, 1757-1764.

220

KRUGER, E. & FERNANDES, L. Error analysis of temperature predictions for the indoor temperature in low-cost houses. Healthy Buildings: Creating a Healthy Indoor Environment for People, HB 2006, June 4, 2006 - June 8, 2006, 2006 Lisboa, Portugal. International Society of Indoor Air Quality and Climate, 167-172. KRÜGER, E., GONZÁLEZ CRUZ, E. & GIVONI, B. 2010. Effectiveness of indirect evaporative cooling and thermal mass in a hot arid climate. Building and Environment, 45, 1422-1433. KUMMERT, M., ANDRE, P. & NICOLAS, J. Optimal heating control in a passive solar commercial building. EuroSun 2000. ISES-Europe Solar Congress, June 2000, 2000 UK. Elsevier, 103-16. KUSIAK, A., LI, M. & ZHANG, Z. 2010. A data-driven approach for steam load prediction in buildings. Applied Energy, 87, 925-933. KWAK, Y., SEO, D., JANG, C. & HUH, J.-H. 2013. Feasibility study on a novel methodology for short-term real-time energy demand prediction using weather forecasting data. Energy and Buildings, 57, 250-260. KWOK, S. S. K. 2011. A study of the importance of occupancy to building cooling load in prediction by intelligent approach. Energy Conversion and Management, 52, 2555-2564. KWOK, S. S. K., YUEN, R. K. K. & LEE, E. W. M. 2011. An intelligent approach to assessing the effect of building occupancy on building cooling load prediction. Building and Environment, 46, 1681-1690. KYUNGTAE, Y., LUCK, R., MAGO, P. J. & HEEJIN, C. 2012. Building hourly thermal load prediction using an indexed ARX model. Energy and Buildings, 54, 225-33. LAM, J. C. 1999. Climatic influences on the energy performance of air-conditioned buildings. Energy Conversion and Management, 40, 39-49. LANDMAN, S., ENGELBRECHT, F. A., CHRISTIEN, E., DYSON, L. & WILLEM, L. 2012. A short-range weather prediction system for South Africa based on a multi- model approach. Water SA, 38, 765-774. LANZA, P. A. G. & COSME, J. M. Z. 2001. A short-term temperature forecaster based on a novel radial basis functions neural network. International Journal of Neural Systems, 11, 71-7. LAYBERRY, R. L. 2008. Degree days for building energy management - Presentation of a new data set. Building Services Engineering Research and Technology, 29, 273-282. LAZOS, D., SPROUL, A. B. & KAY, M. 2014. Optimisation of energy management in commercial buildings with weather forecasting inputs: A review. Renewable and Sustainable Energy Reviews, 39, 587-603. LAZOS, D., SPROUL, A. B. & KAY, M. 2015. Development of hybrid numerical and statistical short term horizon weather prediction models for building energy management optimisation. Building and Environment, 90, 82-95. LEE, K.-H. & BRAUN, J. E. 2004. Development and application of an inverse building model for demand response in small commercial buildings. Proceedings of SimBuild, 1-12. LEE, K.-H. & BRAUN, J. E. Reducing peak cooling loads through model-based control of zone temperature setpoints. 2007 American Control Conference, ACC, July 9, 2007 - July 13, 2007, 2007 New York, NY, United states. Institute of Electrical and Electronics Engineers Inc., 5070-5075.

221

LEHMANN, B., GYALISTRAS, D., GWERDER, M., WIRTH, K. & CARL, S. 2013. Intermediate complexity model for Model Predictive Control of Integrated Room Automation. Energy and Buildings, 58, 250-262. LETHERMAN, K. M. & AL-AZAWI, M. M. J. 1986. PREDICTIONS OF THE HEATING AND COOLING ENERGY REQUIREMENTS IN BUILDINGS USING THE DEGREE HOURS METHOD. Building and Environment, 21, 171-176. LI, C., LI, Z., LI, M., MEGGERS, F., SCHLUETER, A. & LIM, H. B. Energy efficient HVAC system with distributed sensing and control. 2014 IEEE 34th International Conference on Distributed Computing Systems, ICDCS 2014, June 30, 2014 - July 3, 2014, 2014 Madrid, Spain. Institute of Electrical and Electronics Engineers Inc., 429-438. LI, D. H. W. & LAM, J. C. 2001. Development of solar heat gain factors database using meteorological data. Building and Environment, 36, 469-483. LI, N., KWAK, J.-Y., BECERIK-GERBER, B. & TAMBE, M. Predicting HVAC energy consumption in commercial buildings using multiagent systems. 30th International Symposium on Automation and Robotics in Construction and Mining, ISARC 2013, Held in Conjunction with the 23rd World Mining Congress, August 11, 2013 - August 15, 2013, 2013 Montreal, QC, Canada. Canadian Institute of Mining, Metallurgy and Petroleum, 987-994. LI, Q., MENG, Q., CAI, J., YOSHINO, H. & MOCHIDA, A. 2009a. Applying support vector machine to predict hourly cooling load in the building. Applied Energy, 86, 2249-2256. LI, Q., MENG, Q., CAI, J., YOSHINO, H. & MOCHIDA, A. 2009b. Predicting hourly cooling load in the building: A comparison of support vector machine and different artificial neural networks. Energy Conversion and Management, 50, 90-96. LI, X., DING, L. & LI, L. A Novel Building Cooling Load Prediction Based On SVR And SAPSO. 2010 International Symposium on Computer, Communication, Control and Automation (3CA), 5-7 May 2010, 2010a Piscataway, NJ, USA. IEEE, 528-32. LI, X., DING, L., LI, Y., XU, G. & LI, J. Hybrid Genetic Algorithm and Support Vector Regression in Cooling Load Prediction. 2010 3rd International Conference on Knowledge Discovery and Data Mining (WKDD 2010), 9-10 Jan. 2010, 2010b Los Alamitos, CA, USA. IEEE Computer Society, 527-31. LI, X. & MALKAWI, A. 2016. Multi-objective optimization for thermal mass model predictive control in small and medium size commercial buildings under summer weather conditions. Energy, 112, 1194-1206. LIU, S. & HENZE, G. P. 2007. Evaluation of reinforcement learning for optimal control of building active and passive thermal storage inventory. Journal of Solar Energy Engineering, Transactions of the ASME, 129, 215-225. LOCKERBIE, J. 2016. The use of thermal mass [Online]. Available: http://catnaps.org/islamic/approach3.html [Accessed 22 May 2016. LOUTZENHISER, P. G., MANZ, H., FELSMANN, C., STRACHAN, P. A., FRANK, T. & MAXWELL, G. M. 2007. Empirical validation of models to compute solar irradiance on inclined surfaces for building energy simulation. Solar Energy, 81, 254-267. LOVEDAY, D. L. & CRAGGS, C. 1993. Stochastic modelling of temperatures for a full- scale occupied building zone subject to natural random influences. Applied energy, 45, 295-312.

222

LU, N., TAYLOR, T., JIANG, W., CORREIA, J., LEUNG, L. R. & WONG, P. C. The temperature sensitivity of the residential load and commercial building load. 2009 IEEE Power and Energy Society General Meeting, PES '09, July 26, 2009 - July 30, 2009, 2009 Calgary, AB, Canada. IEEE Computer Society. LUO, Q. & ARIYUR, K. B. Building thermal network model and application to temperature regulation. 2010 IEEE International Conference on Control Applications, CCA 2010, September 8, 2010 - September 10, 2010, 2010 Yokohama, Japan. Institute of Electrical and Electronics Engineers Inc., 2190- 2195. LUO, Q., ARIYUR, K. B. & MATHUR, A. Real time energy management: Cutting the carbon footprint and energy costs via hedging, local sources and active control. 2009 ASME Dynamic Systems and Control Conference, DSCC2009, October 12, 2009 - October 14, 2009, 2010 Hollywood, CA, United states. American Society of Mechanical Engineers, 157-164. MA, Q. & LI, Q.-Q. Demand shedding with thermal mass in office buildings. 2010 8th World Congress on Intelligent Control and Automation, WCICA 2010, July 7, 2010 - July 9, 2010, 2010 Jinan, China. Institute of Electrical and Electronics Engineers Inc., 4060-4064. MACARTHUR, J., MATHUR, A. & ZHAO, J. 1989. On-line recursive estimation for load profile prediction. ASHRAE transactions, 95, 621-628. MAILHOT, J., BELAIR, S., CHARRON, M., DOYLE, C., JOE, P., ABRAHAMOWICZ, M., BERNIER, N. B., DENIS, B., ERFANI, A., FRENETTE, R., GIGUERE, A., ISAAC, G. A., MCLENNAN, N., MCTAGGART-COWAN, R., MILBRANDT, J. & TONG, L. 2010. Environment Canada's experimental numerical weather prediction systems for the Vancouver 2010 winter olympic and paralympic games. Bulletin of the American Meteorological Society, 91, 1073-1085. MASSANA, J., POUS, C., BURGAS, L., MELENDEZ, J. & COLOMER, J. 2015. Short-term load forecasting in a non-residential building contrasting models and attributes. Energy and Buildings, 92, 322-330. MATHIESEN, P. J. 2013. Advanced Numerical Weather Prediction Techniques for Solar Irradiance Forecasting: Statistical, Data-Assimilation, and Ensemble Forecasting. ProQuest LLC. MATHIEU, J. L., PRICE, P. N., KILICCOTE, S. & PIETTE, M. A. 2011. Quantifying Changes in Building Electricity Use, With Application to Demand Response. IEEE Transactions on Smart Grid, 2, 507-18. MAY-OSTENDORP, P., HENZE, G. P., CORBIN, C. D., RAJAGOPALAN, B. & FELSMANN, C. 2011. Model-predictive control of mixed-mode buildings with rule extraction. Building and Environment, 46, 428-437. MISHRA, A. K. & RAMGOPAL, M. 2014. Thermal comfort in undergraduate laboratories — A field study in Kharagpur, India. Building and Environment, 71, 223-232. MITTERMAIER, M. P. 2008. The Potential Impact of Using Persistence as a Reference Forecast on Perceived Forecast Skill. Weather and Forecasting, 23, 1022- 1031. MORGAN, S. & MONCEF, K. Field testing of optimal controls of passive and active thermal storage. 2010 ASHRAE Winter Conference, January 23, 2010 - January 27, 2010, 2010 Orlando, FL, United states. Amer. Soc. Heating, Ref. Air-Conditoning Eng. Inc., 134-146.

223

MTEMBO, V., TAYLOR, G. A. & EKWUE, A. A novel econometric model for peak demand forecasting. 49th International Universities Power Engineering Conference, UPEC 2014, September 2, 2014 - September 5, 2014, 2014 Cluj- Napoca, Romania. IEEE Computer Society. MUSTAFARAJ, G., CHEN, J. & LOWRY, G. 2010. Development of room temperature and relative humidity linear parametric models for an open office using BMS data. Energy and Buildings, 42, 348-356. MUSTAFARAJ, G., LOWRY, G. & CHEN, J. 2011. Prediction of room temperature and relative humidity by autoregressive linear and nonlinear neural network models for an open office. Energy and Buildings, 43, 1452-1460. NAGAI, T. A method for revising temperature and humidity prediction using additional observations and weather forecasts. Building Simulation 2007, BS 2007, September 3, 2007 - September 6, 2007, 2007 Beijing, China. International Building Performance Simulation Association, 245-252. NAGAI, T., REDDY, A. & MASSIE, D. Optimization method for minimizing annual energy, peak energy demand, and annual energy cost through use of building thermal storage. 2002 ASHRAE Winter Meeting, January 13, 2002 - January 16, 2002, 2002 Atlantic City, NJ, United states. Amer. Soc. Heating, Ref. Air- Conditoning Eng. Inc., 43-53. NIELSEN, T. S., JOENSEN, A., MADSEN, H., LANDBERG, L. & GIEBEL, G. 1998. A new reference for wind power forecasting. Wind Energy, 1, 29-34. OKAMOTO, S. A case study of energy saving by CCHP technology in a hospital. ASME 2010 International Mechanical Engineering Congress and Exposition, IMECE 2010, November 12, 2010 - November 18, 2010, 2010 Vancouver, BC, Canada. American Society of Mechanical Engineers, 759-767. OLDEWURTEL, F., JONES, C. N. & MORARI, M. A tractable approximation of chance constrained stochastic MPC based on affine disturbance feedback. 47th IEEE Conference on Decision and Control, CDC 2008, December 9, 2008 - December 11, 2008, 2008 Cancun, Mexico. Institute of Electrical and Electronics Engineers Inc., 4731-4736. OLDEWURTEL, F., PARISIO, A., JONES, C. N., GYALISTRAS, D., GWERDER, M., STAUCH, V., LEHMANN, B. & MORARI, M. 2012. Use of model predictive control and weather forecasts for energy efficient building climate control. Energy and Buildings, 45, 15-27. OLDEWURTEL, F., ULBIG, A., PARISIO, A., ANDERSSON, G. & MORARI, M. Reducing peak electricity demand in building climate control using real-time pricing and model predictive control. 2010 49th IEEE Conference on Decision and Control, CDC 2010, December 15, 2010 - December 17, 2010, 2010 Atlanta, GA, United states. Institute of Electrical and Electronics Engineers Inc., 1927- 1932. OLIVEIRA-LIMA, J. A., MORAIS, R., MARTINS, J. F., FLOREA, A. & LIMA, C. 2016. Load forecast on intelligent buildings based on temporary occupancy monitoring. Energy and Buildings, 116, 512-521. PAKANEN, J. & KARJALAINEN, S. 2009. A state machine approach in modelling the heating process of a building. Energy and Buildings, 41, 528-533. PEDERSEN, L. 2007. Use of different methodologies for thermal load and energy estimations in buildings including meteorological and sociological input parameters. Renewable and Sustainable Energy Reviews, 11, 998-1007.

224

PEDERSEN, L., STANG, J. & ULSETH, R. 2008. Load prediction method for heat and electricity demand in buildings for the purpose of planning for mixed energy distribution systems. Energy and Buildings, 40, 1124-1134. PENG, X., RONGXIN, Y., CARRIE, B. & DONGEUN, K. 2009. Demand Shifting with Thermal Mass in Large Commercial Buildings in a California Hot Climate Zone. CEC/LBNL. PENG XU, P. E. Case study of demand shifting with thermal mass in two large commercial buildings. 2009 ASHRAE Annual Conference, June 20, 2009 - June 24, 2009, 2009 Louisville, KY, United states. Amer. Soc. Heating, Ref. Air- Conditoning Eng. Inc., 586-598. PENYA, Y. K., BORGES, C. E., AGOTE, D. & FERNANDEZ, I. Short-term load forecasting in air-conditioned non-residential Buildings. 2011 IEEE International Symposium on Industrial Electronics, ISIE 2011, June 27, 2011 - June 30, 2011, 2011 Gdansk, Poland. IEEE Computer Society, 1359-1364. PEREZ-LOMBARD, L., ORTIZ, J. & POUT, C. 2008. A review on buildings energy consumption information. Energy and Buildings, 40, 394-398. PETERSON, J. L. & HUNN, B. D. Use of indirect evaporative cooling to reduce peak electric demand in new office buildings. ASHRAE Transactions: Technical Papers presented at the 1985 Winter Meeting., 1985 Chicago, IL, USA. ASHRAE, 329-341. PUSAT, S. & EKMEKCI, I. 2016. A study on degree-day regions of Turkey. Energy Efficiency, 9, 525-532. RABL, A. & NORFORD, L. K. 1991. Peak load reduction by preconditioning buildings at night. International Journal of Energy Research, 15, 781-798. RACKES, A. & WARING, M. S. 2014. Using multiobjective optimizations to discover dynamic building ventilation strategies that can improve indoor air quality and reduce energy use. Energy and Buildings, 75, 272-280. RAFTERY, P., KEANE, M. & COSTA, A. 2011. Calibrating whole building energy models: Detailed case study using hourly measured data. Energy and Buildings, 43, 3666-3679. RANAWEERA, D. K., KARADY, G. G. & FARMER, R. G. 1996. Effect of probabilistic inputs on neural network-based electric load forecasting. IEEE Trans Neural Netw, 7, 1528-32. RASOULI, M., GAOMING, G., SIMONSON, C. J. & BESANT, R. W. 2013. Uncertainties in energy and economic performance of HVAC systems and energy recovery ventilators due to uncertainties in building and HVAC parameters. Applied Thermal Engineering, 50, 732-42. REDDY, T. A. & CLARIDGE, D. E. 1994. Using synthetic data to evaluate multiple regression and principal component analyses for statistical modeling of daily building energy consumption. Energy and Buildings, 21, 35-44. REMPEL, A. R., GATES, K., REMPEL, A. W. & SHAW, B. Oregon sunspace redesign / build: New priorities for thermal mass. 42nd ASES National Solar Conference 2013, SOLAR 2013, Including 42nd ASES Annual Conference and 38th National Passive Solar Conference, April 16, 2013 - April 20, 2013, 2013 Baltimore, MD, United states. American Solar Energy Society, 199-206. REMPEL, A. R., REMPEL, A. W., GATES, K. R. & SHAW, B. 2016. Climate-responsive thermal mass design for Pacific Northwest sunspaces. Renewable Energy, 85, 981-993.

225

REN, M. J. & WRIGHT, J. A. 2002. Adaptive diurnal prediction of ambient dry-bulb temperature and solar radiation. HVAC and R Research, 8, 383-401. RIOS-MORENO, G. J., TREJO-PEREA, M., CASTANEDA-MIRANDA, R., HERNANDEZ- GUZMAN, V. M. & HERRERA-RUIZ, G. 2007. Modelling temperature in intelligent buildings by means of autoregressive models. Automation in Construction, 16, 713-722. RIVARD, H., JIN, Y. & ZMEUREANU, R. 2005. On-line building energy prediction using adaptive artificial neural networks. Energy and Buildings, 37, 1250-9. ROTH, K., DIECKMANN, J. & BRODRICK, J. 2009. Using off-peak precooling. ASHRAE Journal, 51, 80-83. RUANO, A. E., CRISPIM, E. M., CONCEICAO, E. Z. E. & LUCIO, M. M. J. R. 2006. Prediction of building's temperature using neural networks models. Energy and Buildings, 38, 682-694. RUMELHART, D. E., HINTON, G. E. & WILLIAMS, R. J. 1986. Learning representations by back-propagating errors. Nature, 323, 533-536. RUUD, M. D., MITCHELL, J. W. & KLEIN, S. A. Use of building thermal mass to offset cooling loads. 1990 Annual Meeting of the American Society of Heating, Refrigerating and Air-Conditioning Engineers, Technical and Symposium Papers, June 10, 1990 - June 13, 1990, 1990 St. Louis, MO, USA. Publ by ASHRAE, 820-829. RYU, S. H. & MOON, H. J. 2016. Development of an occupancy prediction model using indoor environmental data based on machine learning techniques. Building and Environment, 107, 1-9. SANTAMOURIS, M., PAPANIKOLAOU, N., LIVADA, I., KORONAKIS, I., GEORGAKIS, C., ARGIRIOU, A. & ASSIMAKOPOULOS, D. N. 2001. On the impact of urban climate on the energy consumption of buildings. Solar Energy, 70, 201-16. SEEM, J. E. & BRAUN, J. E. Adaptive methods for real-time forecasting of building electrical demand. ASHRAE Winter Meeting - Technical Papers, January 19, 1991 - January 23, 1991, 1991 New York, NY, USA. Publ by ASHRAE, 710-721. SHI, L. & WANG, J. Application of artificial neural network to predict the hourly cooling load of an office building. 2009 International Joint Conference on Computational Sciences and Optimization, CSO 2009, April 24, 2009 - April 26, 2009, 2009 Sanya, Hainan, China. IEEE Computer Society, 528-530. SHINKAWA, T. & NOBE, T. Time series data from internal load factors in offices. Healthy Buildings: Creating a Healthy Indoor Environment for People, HB 2006, June 4, 2006 - June 8, 2006, 2006 Lisboa, Portugal. International Society of Indoor Air Quality and Climate, 265-268. SIROKY, J., OLDEWURTEL, F., CIGLER, J. & PRIVARA, S. 2011. Experimental analysis of model predictive control for an energy efficient building heating system. Applied Energy, 88, 3079-3087. SOLEIMANI-MOHSENI, M., THOMAS, B. & FAHLEN, P. 2006. Estimation of operative temperature in buildings using artificial neural networks. Energy and Buildings, 38, 635-640. SPINDLER, H. C. & NORFORD, L. K. 2009a. Naturally ventilated and mixed-mode buildings-Part II: Optimal control. Building and Environment, 44, 750-761. SPINDLER, H. C. & NORFORD, L. K. 2009b. Naturally ventilated and mixed-mode buildings—Part I: Thermal modeling. Building and Environment, 44, 736- 749.

226

STADLER, M., KLOESS, M., GROISSBOCK, M., CARDOSO, G., SHARMA, R., BOZCHALUI, M. C. & MARNAY, C. 2013. Electric storage in California's commercial buildings. Applied Energy, 104, 711-722. STEINFELD, J., BRUCE, A. & WATT, M. 2011. Peak load characteristics of Sydney office buildings and policy recommendations for peak load reduction. Energy and Buildings, 43, 2179-2187. STOBIE, J. G. & AVJIAN, R. M. Using ensemble weather forecasts for air traffic management. 54th Air Traffic Control Association Annual Conference 2009, November 2, 2009 - November 4, 2009, 2009 Washington, DC, United states. Air Traffic Control Association, Inc., 97-101. SUN, Y., WANG, S. & HUANG, G. 2010a. A demand limiting strategy for maximizing monthly cost savings of commercial buildings. Energy and Buildings, 42, 2219-2230. SUN, Y., WANG, S. & HUANG, G. 2010b. Model-based optimal start control strategy for multi-chiller plants in commercial buildings. Building Services Engineering Research and Technology, 31, 113-129. SUN, Y., WANG, S. & XIAO, F. 2013a. Development and validation of a simplified online cooling load prediction strategy for a super high-rise building in Hong Kong. Energy Conversion and Management, 68, 20-27. SUN, Y., WANG, S., XIAO, F. & GAO, D. 2013b. Peak load shifting control using different cold thermal energy storage facilities in commercial buildings: A review. Energy Conversion and Management, 71, 101-114. SUZUKI, R., HAYASHI, Y. & FUJIMOTO, Y. Determination method of optimal planning and operation for residential PV system and storage battery based on weather forecast. 2012 IEEE International Conference on Power and Energy, PECon 2012, December 2, 2012 - December 5, 2012, 2012 Kota Kinabalu, Malaysia. IEEE Computer Society, 343-347. TAMIZH-MANI, G., ISHIOYE, J.-P., VOROPAYEV, A. & KANG, Y. Photovoltaic performance models: An evaluation with actual field data. Reliability of Photovoltaic Cells, Modules, Components, and Systems, August 11, 2008 - August 13, 2008, 2008 San Diego, CA, United states. SPIE, The International Society for Optical Engineering (SPIE). TAYLOR, J. W. 2003. Short-term electricity demand forecasting using double seasonal exponential smoothing. Journal of the Operational Research Society, 54, 799-805. TAYLOR, J. W. 2010. Exponentially weighted methods for forecasting intraday time series with multiple seasonal cycles. International Journal of Forecasting, 26, 627-46. TAYLOR, J. W. 2012. Short-term load forecasting with exponentially weighted methods. IEEE Transactions on Power Systems, 27, 458-464. TAYLOR, J. W. & BUIZZA, R. 2002. Neural network load forecasting with weather ensemble predictions. IEEE Transactions on Power Systems, 17, 626-632. TAYLOR, J. W. & SNYDER, R. D. 2012. Forecasting intraday time series with multiple seasonal cycles using parsimonious seasonal exponential smoothing. Omega, 40, 748-57. THATCHER, M. & HURLEY, P. 2010. A customisable downscaling approach for local- scale meteorological and air pollution forecasting: Performance evaluation for a year of urban meteorological forecasts. Environmental Modelling & Software, 25, 82-92. 227

TODHUNTER, P. E. & TERJUNG, W. H. 1988. Intercomparison of three urban climate models. Boundary-Layer Meteorology, 42, 181-205. TYAGI, R., BLACK, J. W. & PETERSEN, J. Optimal scheduling of demand response events using options valuation methods. 2011 IEEE PES General Meeting: The Electrification of Transportation and the Grid of the Future, July 24, 2011 - July 28, 2011, 2011 Detroit, MI, United states. IEEE Computer Society. UNIVERSITY OF NEW SOUTH WALES FACILITIES MANAGEMENT. 2015. Tyree Energy Technologies Building: sustainable energy research [Online]. Available: http://www.facilities.unsw.edu.au/tyree-energy-technologies- building-sustainable-energy-research [Accessed 23 December 2015. VAISALA. 2013. Humidity conversion formulas [Online]. Helsinki, Finland: Vaisala. Available: http://www.vaisala.com/Vaisala%20Documents/Application%20notes/Hu midity_Conversion_Formulas_B210973EN-F.pdf [Accessed 16th July 2016. VAKILOROAYA, V., SAMALI, B., MADADNIA, J. & HA, Q. P. Component-wise optimization for a commercial central cooling plant. 37th Annual Conference of the IEEE Industrial Electronics Society, IECON 2011, November 7, 2011 - November 10, 2011, 2011 Melbourne, VIC, Australia. IEEE Computer Society, 2769-2774. VAN PAASSEN, A. H. C. & LUO, Q. X. Weather data generator to study climate change on buildings. 2002. Chartered Inst. of Building Services Engineers, 251-258. WAHAB, H. A., DUKE, M., CARSON, J. K. & ANDERSON, T. Studies of control strategies for building integrated solar energy system. 2011 IEEE 1st Conference on Clean Energy and Technology, CET 2011, June 27, 2011 - June 29, 2011, 2011 Kuala Lumpur, Malaysia. IEEE Computer Society, 342-347. WAIDE, P. A. & NORTON, B. 1995. Degree-hour steady-state temperature index. Building Services Engineering Research and Technology, 16, 107-113. WAN AHMAD, W. K. A. & AHMAD, S. Arima model and exponential smoothing method: A comparison. 20th National Symposium on Mathematical Sciences: Research in Mathematical Sciences: a Catalyst for Creativity and Innovation, 18-20 Dec. 2012, 2013 USA. American Institute of Physics, 1312-21. WANG, F., WANG, L., ZHOU, H., VALERIANO, O. C. S., KOIKE, T. & LI, W. 2012. Ensemble hydrological prediction-based real-time optimization of a multiobjective reservoir during flood season in a semiarid basin with global numerical weather predictions. Water Resources Research, 48. WANG, S. & XU, X. 2006. Simplified building model for transient thermal performance estimation using GA-based parameter identification. International Journal of Thermal Sciences, 45, 419-432. WANG, Z., CLARKE, A. A., MOYNE, J. R. & TILBURY, D. M. Utilizing intra-day prediction modification strategies to improve peak power shaving using energy storage systems for smart buildings. ASME 2014 Dynamic Systems and Control Conference, DSCC 2014, October 22, 2014 - October 24, 2014, 2014 San Antonio, TX, United states. American Society of Mechanical Engineers, Dynamic Systems and Control Division. WEN, J. & SMITH, T. F. 2007. Development and validation of online models with parameter estimation for a building zone with VAV system. Energy and Buildings, 39, 13-22. WESTPHAL, F. S. & LAMBERTS, R. The use of simplified weather data to estimate thermal loads of non-residential buildings. 2004. Elsevier Ltd, 847-854. 228

WORKSAFE VICTORIA 2008. Workplace amenities and work environment - Compliance Code. Victoria State Government. WORLD METEORLOGICAL ORGANIZATION 2012. Guidelines on Ensemble Prediction Systems and Forecasting. Geneva, Switzerland: World Meteorological Organization. XINMEI HUANG, Y. M., GRAHAM MILLS 2008. Verification of mesoscale NWP forecasts of abrupt wind changes. CAWCR Techical Report No.008. Melbourne: CSIRO & BOM. XU, P. & HAVES, P. Case study of demand shifting with thermal mass in two large commercial buildings. 2006 Winter Meeting of the American Society of Heating, Refrigerating and Air-Conditioning Engineers, ASHRAE, January 21, 2006 - January 25, 2006, 2006 Chicago, IL, United states. Amer. Soc. Heating, Ref. Air-Conditoning Eng. Inc., 572-580. XU, P. H., P.; PIETTE, M. A.; BRAUN, J. E. 2004. Peak Demand Reduction from Pre- Cooling with Zone Temperature Reset in an Office Building. 2004 ACEEE Summer Study on Energy Efficiency in Buildings, Pacific Grove, CA. XUEMEI, L., LIXING, D., YUYUAN, D. & LANLAN, L. Hybrid support vector machine and ARIMA model in building cooling prediction. 2010 International Symposium on Computer, Communication, Control and Automation, 3CA 2010, May 5, 2010 - May 7, 2010, 2010a Tainan, Taiwan. IEEE Computer Society, 533-536. XUEMEI, L., MING, S., LIXING, D., GANG, X. & JIBIN, L. 2010b. Particle swarm optimization-based LS-SVM for building cooling load prediction. Journal of Computers, 5, 614-621. YANG, C. & XIE, L. A novel ARX-based multi-scale spatio-temporal solar power forecast model. 2012 North American Power Symposium, NAPS 2012, September 9, 2012 - September 11, 2012, 2012 Champaign, IL, United states. IEEE Computer Society. YANG, I.-H. & KIM, K.-W. 2004. Prediction of the time of room air temperature descending for heating systems in buildings. Building and Environment, 39, 19-29. YANG, R. & WANG, L. Control strategy optimization for energy efficiency and comfort management in HVAC systems. 2015 IEEE Power and Energy Society Innovative Smart Grid Technologies Conference, ISGT 2015, February 18, 2015 - February 20, 2015, 2015 Washington, DC, United states. Institute of Electrical and Electronics Engineers Inc. YANG, Z., GHAHRAMANI, A. & BECERIK-GERBER, B. Iterative reassignment algorithm: Leveraging occupancy based hvac control for improved energy efficiency. Winter Simulation Conference, WSC 2015, December 6, 2015 - December 9, 2015, 2016 Huntington Beach, CA, United states. Institute of Electrical and Electronics Engineers Inc., 3449-3460. YANG, Z., WANG, W., CAO, D., ZHANG, J. & CHEN, X. 2007. Ensemble chaotic forecasting method in short term load forecasting. Dianli Xitong Zidonghua/Automation of Electric Power Systems, 31, 33-37. YAO, Y., LIAN, Z., LIU, S. & HOU, Z. 2004. Hourly cooling load prediction by a combined forecasting model based on analytic hierarchy process. International Journal of Thermal Sciences, 43, 1107-1118.

229

YOKOYAMA, R., WAKUI, T. & SATAKE, R. 2009. Prediction of energy demands using neural network with model identification by global optimization. Energy Conversion and Management, 50, 319-327. YOSHIDA, H. & TERAI, T. 1991. ARMA type weather model for air-conditioning, heating and cooling load calculation. Energy and Buildings, 16, 625-634. YOSHIDA, H. & TERAI, T. Modeling of weather data by time series analysis for air- conditioning load calculators. ASHRAE Winter Meeting, January 25, 1992 - January 29, 1992, 1992 Anaheim, CA, USA. Publ by ASHRAE, 328-345. YOUSEFI, A., MYERS, T., MITCHELL, J. S. B., KOSTITSYNA, I. & SHARMA, R. Robust airspace design methods for uncertain traffic and weather. 2013 IEEE/AIAA 32nd Digital Avionics Systems Conference, DASC 2013, October 5, 2013 - October 10, 2013, 2013 East Syracuse, NY, United states. Institute of Electrical and Electronics Engineers Inc., 1D21-1D211. YU, F. W., CHAN, K. T. & SIT, R. K. Y. 2012. Climatic influence on the design and operation of chiller systems serving office buildings in a subtropical climate. Energy and Buildings, 55, 500-507. YUN, K., CHO, H., LUCK, R. & MAGO, P. J. Real-time combined heat and power operational strategy using a hierarchical optimization algorithm. 2011 55 City Road, London, EC1Y 1SP, United Kingdom. SAGE Publications Ltd, 403- 412. ZAVALA, V. M. 2013. Real-time optimization strategies for building systems. Industrial and Engineering Chemistry Research, 52, 3137-3150. ZAVALA, V. M., CONSTANTINESCU, E. M. & ANITESCU, M. Economic impacts of advanced weather forecasting on energy system operations. Innovative Smart Grid Technologies Conference, ISGT 2010, January 19, 2010 - January 21, 2010, 2010 Gaithersburg, MD, United states. IEEE Computer Society, IEEE; IEEE PES - Power and Energy Society. ZAVALA, V. M., CONSTANTINESCU, E. M., KRAUSE, T. & ANITESCU, M. 2009. On-line economic optimization of energy systems using weather forecast information. Journal of Process Control, 19, 1725-1736. ZHANG, Y. & HANBY, V. I. Short-term prediction of weather parameters using online weather forecasts. Building Simulation 2007, BS 2007, September 3, 2007 - September 6, 2007, 2007 Beijing, China. International Building Performance Simulation Association, 1411-1416. ZHAO, H.-X. & MAGOULES, F. 2012. A review on the prediction of building energy consumption. Renewable and Sustainable Energy Reviews, 16, 3586-3592. ZHOU, G., YANG, Y. & XU, H. 2011. Energy performance of a hybrid space-cooling system in an office building using SSPCM thermal storage and night ventilation. Solar Energy, 85, 477-485. ZHOU, Q., WANG, S., XU, X. & XIAO, F. 2008. A grey-box model of next-day building thermal load prediction for energy-efficient control. International Journal of Energy Research, 32, 1418-1431. ZONG, Y., MIHET-POPA, L., KULLMANN, D., THAVLOV, A., GEHRKE, O. & BINDNER, H. W. Model predictive controller for active demand side management with PV self-consumption in an intelligent building. 2012 3rd IEEE PES Innovative Smart Grid Technologies Europe, ISGT Europe 2012, October 14, 2012 - October 17, 2012, 2012 Berlin, Germany. IEEE Power and Energy Society; Technische Universitat Berlin - SENSE Lab.

230

ZORAS, S., TRIANTAFYLLOU, A. G. & HURLEY, P. J. 2007. Grid sensitivity analysis for the calibration of a prognostic meteorological model in complex terrain by a screening experiment. Environmental Modelling & Software, 22, 33-39.

231