Forecasting for Concentrated Solar Thermal Power Plants in Australia Edward Weikin

Forecasting for Concentrated Solar Thermal Power Plants in

Australia

Edward Weikin Law

A thesis in fulfilment of the requirements for the degree of

Doctor of Philosophy

University of New South Wales

School of Photovoltaic and Renewable Energy Engineering

Faculty of Engineering

May 2017

THE UNIVERSITY OF NEW SOUTH WALES

Thesis/Dissertation Sheet

Surname or Family name: LAW

First name: Edward Other name/s: Weikin

Abbreviation for degree as given in the University calendar: PhD

School: School of Photovoltaic and Renewable Energy Engineering Faculty: Engineering

Title: Forecasting for concentrated solar thermal power plants in Australia

Abstract:

Up to 50% of electricity needs in Australia could be supplied by solar power. At these high levels of solar power generation, solar forecasting is necessary to manage the impact of solar variability. However, there has been little research on using solar forecasting in Australia. This study used modelling to investigate the benefits of using short-term and long-term solar forecasts to operate a concentrated solar thermal (CST) plant for a year at four sites that covered different climate zones within the Australian National Electricity Market.

Using 1-hour ahead short-term forecasts increased net value by $0.90-$2.07 million for a CST plant with storage, and by $0.76-$3.10 million for a CST plant without storage. It also improved reliability by reducing the equivalent forced outage rate by 21-38 percentage points for a CST plant with storage, and by 16-42 percentage points for a CST plant without storage. Using 1-hour forecasts achieved 59%-94% of the net value achievable if the 48-hour forecast were perfect. At each site, the highest net value and reliability were achieved by a CST plant with storage and using 1-hour forecasts, thus a CST plant should have both storage and short-term forecasts. If only one can be used, then a CST plant with storage and without 1-hour forecasts achieves higher net value, whereas a CST plant without storage and with 1-hour forecasts achieves higher reliability.

These results demonstrated that using short-term forecasts is beneficial for CST plants that operate in electricity markets that allow updated bids to be submitted at short-term time frames. The results can be used to estimate the return on investment in obtaining short-term forecasts for operating a CST plant. Furthermore, the research method can be adapted into a tool for estimating value to assist CST plant project planning.

Declaration relating to disposition of project thesis/dissertation

I hereby grant to the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole or in 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 property 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 Abstracts International (this is applicable to doctoral theses only).

…………………………………………………………… ……………………………………..……………… ……….……………………...…….… Signature Witness Signature Date

The University recognises that there may be exceptional circumstances requiring restrictions on copying or conditions on use. Requests for restriction for a period of up to 2 years must be made in writing. Requests for a longer period of restriction may be considered in exceptional circumstances and require the approval of the Dean of Graduate Research.

FOR OFFICE USE ONLY Date of completion of requirements for Award:

THIS SHEET IS TO BE GLUED TO THE INSIDE FRONT COVER OF THE THESIS

iii

Acknowledgements

“It takes a village to raise a child.” – An African proverb

“If I have seen further it is by standing on the shoulders of giants.” – Isaac Newton

First and foremost, I thank my mother, father, and sister for their love and support, without which I would not be where I am today.

I thank my supervisors, Robert Taylor and Merlinde Kay, for tirelessly giving me guidance and their time for the duration of my candidature. I appreciate the invaluable feedback provided by my co-supervisor, Graham Morrison, my progress review panellists, Alistair Sproul and Evatt

Hawke, and the peers who reviewed my journal and conference papers.

I am grateful to the Australian Renewable Energy Agency (ARENA) and the University of New

South Wales for awarding me ARENA and Australian Postgraduate Award scholarships.

My research could not have been conducted without the ground and satellite solar radiation data supplied by the Australian Bureau of Meteorology, the National Centers for

Environmental Prediction Reanalysis data provided by the US National Oceanic and

Atmospheric Administration Office of Oceanic and Atmospheric Research Earth System

Research Laboratory Physical Sciences Division, and The Air Pollution Model provided by the

Commonwealth Scientific and Industrial Research Organisation.

Finally, I have a word of thanks for my colleagues and friends, who each supported me in their own small ways.

Abstract

Using 1-hour ahead short-term forecasts increased net value by $0.90-$2.07 million for a CST plant with storage, and by $0.76-$3.10 million for a CST plant without storage. It also improved reliability by reducing the equivalent forced outage rate by 21-38 percentage points for a CST plant with storage, and by 16-42 percentage points for a CST plant without storage. Using 1- hour forecasts achieved 59%-94% of the net value achievable if the 48-hour forecast were perfect. At each site, the highest net value and reliability were achieved by a CST plant with storage and using 1-hour forecasts, thus a CST plant should have both storage and short-term forecasts. If only one can be used, then a CST plant with storage and without 1-hour forecasts achieves higher net value, whereas a CST plant without storage and with 1-hour forecasts achieves higher reliability.

List of Publications

Journal papers

Law, E.W., Kay, M., Taylor, R.A., 2016, Evaluating the benefits of using short-term direct normal irradiance forecasts to operate a concentrated solar thermal plant, Solar Energy 140, p.

93-108 http://dx.doi.org/10.1016/j.solener.2016.10.037

Law, E.W., Kay, M., Taylor, R.A., 2016, Calculating the financial value of a concentrated solar thermal plant operated using direct normal irradiance forecasts, Solar Energy 125, p. 267-281 https://dx.doi.org/10.1016/j.solener.2015.12.031

Law, E.W., Prasad, A.A., Kay, M., Taylor, R.A., 2014, Direct normal irradiance forecasting and its application to concentrated solar thermal output forecasting – A review, Solar Energy 108, p.

287-307 http://dx.doi.org/10.1016/j.solener.2014.07.008

Conference papers

Law, E.W., Kay, M., Taylor, R.A., 2014, Assessing the economic benefit of forecasting concentrated solar thermal energy output, Solar2014: The 52nd Annual Conference of the

Australian Solar Council, p. 131-140

vii

Table of Contents

Acknowledgements ...... v

Abstract ...... vi

List of Publications ...... vii

Table of Contents ...... 1

Nomenclature ...... 5

List of Figures ...... 8

List of Tables ...... 10

1. Introduction ...... 12

1.1. Research Aim, Motivation and Scope ...... 12

1.2. An Overview of the Australian National Electricity Market ...... 18

2. Concentrated Solar Thermal Power ...... 23

2.1. CST Plant Operation and Components ...... 23

2.1.1. Solar field collector ...... 24

2.1.2. Power block ...... 27

2.1.3. Storage ...... 28

2.1.4. Heat transfer fluid ...... 29

2.2. Studies on the Value of CST ...... 30

2.3. Studies on the Value of Using Forecasts to Operate CST ...... 34

2.4. Chapter Summary ...... 37

3. Methods to forecast direct normal irradiance ...... 39

3.1. Clear Sky Models ...... 40

3.1.1. CSM modelling accuracy ...... 41

3.1.2. Global-to-diffuse models ...... 42

3.2. Numerical Weather Prediction Models ...... 43

3.2.1. NWP forecast accuracy ...... 46

3.2.2. Improving forecast accuracy ...... 48

3.3. Time Series Analysis ...... 50

3.3.1. TSA forecast accuracy ...... 51

3.4. Cloud Motion Vectors ...... 53

3.4.1. CMV forecast accuracy ...... 56

3.4.2. Improving forecast accuracy ...... 58

3.4.3. Converting satellite image to solar irradiance ...... 59

3.5. Hybrid Forecasting Methods ...... 61

3.5.1. Hybrid DNI forecasting methods ...... 61

3.5.2. Other hybrid forecasting methods applied to global irradiance ...... 64

3.6. Chapter Summary ...... 66

4. Study Method ...... 69

4.1. Data ...... 69

4.1.1. DNI data ...... 70

4.1.2. Electricity price ...... 76

4.1.3. Site data comparison ...... 77

4.2. DNI Forecasting ...... 80

4.2.1. Methods for 48-hour forecasts ...... 81

4.2.2. Method for 1-hour forecasts ...... 83

4.2.3. Forecast example ...... 85

4.2.4. Forecast accuracy metrics ...... 87

4.3. CST Plant Model ...... 88

4.4. Financial Value Calculation Method ...... 92

4.5. Reliability Calculation Method ...... 94

4.6. Chapter Summary ...... 99

5. Results ...... 100

5.1. Site DNI Forecast Accuracy ...... 100

5.2. Financial Value ...... 102

5.3. Reliability ...... 113

5.4. Summary of Results ...... 117

6. Discussion ...... 118

6.1. Comparison with Past Studies ...... 118

6.2. Reflection on the Scope of the Study ...... 120

6.2.1. Local climate ...... 121

6.2.2. CST plant configuration ...... 121

6.2.3. Electricity market ...... 122

6.2.4. Value perspectives ...... 123

6.2.5. Short-term forecasts ...... 124

6.2.6. CST model ...... 124

6.3. Chapter Summary ...... 125

7. Conclusion ...... 127

References ...... 131

A. LKT model equations ...... 145

A.1. Solar field model ...... 145

A.1.1. Optical loss from factors dependent on solar position ...... 146

A.1.2. Optical loss from factors independent of solar position ...... 148

A.1.3. Heat loss ...... 149

A.1.4. Solar field thermal output ...... 150

A.2. Power block and storage model ...... 153

A.3. LKT model parameter values ...... 155

B. Numerical Results ...... 158

B.1. Financial Value Metrics ...... 158

B.2. EFOR ...... 160

B.3. Normalised Revenue and Net Value ...... 160

B.4. Total Energy ...... 161

Nomenclature

ACF autocorrelation function ACT Australian Capital Territory AEMC Australian Energy Market Commission AEMO Australian Electricity Market Operator AER Australian Energy Regulator AERONET Aerosol Robotic Network AEST Australian Eastern Standard Time AFSOL Aerosol-based Forecasts of Solar Irradiance for Energy Applications AI artificial intelligence ANN artificial neural network AOD aerosol optical depth AR autoregressive ARIMA autoregressive integrated moving average ARMA autoregressive moving average ARPS Advanced Multiscale Regional Prediction System AUSTELA Australian Solar Thermal Energy Association BOM Australian Bureau of Meteorology BRL Boland-Ridley-Lauret model CAISO California Independent System Operator CARDS Coupled Autoregressive and Dynamical System CEC Clean Energy Council CMV cloud motion vector CSM clear sky model CST concentrated solar thermal DNI direct normal irradiance ECMWF European Centre for Medium-range Weather Forecasts EFDH equivalent forced derated hour EFOR equivalent forced outage rate ERCOT Electricity Reliability Council of Texas ERSFDH equivalent reserve shutdown forced derated hour ESRA European Solar Radiation Atlas EUMETSAT European Organisation for the Exploitation of Meteorological Satellites FCAS frequency control ancillary services FOH forced outage hour GFS Global Forecast System GHI global horizontal irradiance GMC gross maximum capacity GOES-9 Geostationary Operational Environmental Satellite GW Gigawatt GWh Gigawatt-hour HCE heat collection element HIRLAM High Resolution Limited Area Model HTF heat transfer fluid IAM incidence angle modifier 5

IEA International Energy Agency IEEE Institute of Electrical and Electronics Engineers JMA Japanese Meteorological Agency libRantran library for radiative transfer LS-SVM least squares support vector machine MA moving average MACC Monitoring Atmospheric Composition and Climate MAE mean absolute error MAPE mean absolute percentage error MASS Mesoscale Atmospheric Simulation System MBD mean bias deviation MBE mean bias error MLP muti-layer perceptron MM5 Fifth-generation Pennsylvania State University National Center for Atmospheric Research Mesoscale Model MOS Model Output Statistics MRM Meteorological Radiation Model MSG Meteosat Second Generation MTSAT-1R Multi-Functional Transport Satellite MW Megawatt MWh Megawatt-hour NAM North American Model NDFD National Digital Forecast Database NEM Australian National Electricity Market NESDIS NOAA's Satellite and Information Service nMAE normalised MAE nMBE normalised MBE NOAA National Oceanographic and Atmospheric Agency NREL National Renewable Energy Laboratory nRMSE normalised RMSE NSW New South Wales NWP numerical weather prediction PACF partial autocorrelation function PV photovoltaic QLD Queensland RDPS Regional Deterministic Prediction System REST2 Reference Evaluation of Solar Transmittance - 2 Bands RG reserve generation RMSD root mean square deviation RMSE root mean square error RPS Renewable Portfolio Standard RTM radiative transfer model SA South Australia SAM System Advisor Model SAR seasonal autoregressive SH service hour 6

SM solar multiple SMA seasonal moving average SUNY State University of New York GOES satellite-based solar model SVM support vector machine SVR support vector regression TAPM The Air Pollution Model TAS Tasmania TDNN time delay neural network TES thermal energy storage THI total horizon invariant TSA time series analysis TW Terawatt TWh Terawatt-hour UIGF unconstrained intermittent generation forecast VIC Victoria WRF Weather Research and Forecasting WRF-CLDDA high-resolution direct-cloud-assimilating WRF

List of Figures

Figure 1-1: A simplified illustration of global horizontal irradiance (GHI), diffuse irradiance, direct irradiance, and direct normal irradiance (DNI)...... 13

Figure 1-2: How value may be assessed for DNI forecasts ...... 16

Figure 1-3: Transmission grid and distribution networks in the National Electricity Market (AER,

2015) ...... 19

Figure 2-1: The main components of a CST plant and how they operate together to convert DNI into electricity ...... 24

Figure 2-2: Simplified illustrations of the different collector technologies. The parabolic trough and linear Fresnel collectors are shown as widthwise cross-sections with the line of concentration extending into the page...... 26

Figure 4-1: Closed and open BOM stations for measuring one minute DNI (BOM, 2012a) ...... 70

Figure 4-2: Scatterplots of satellite DNI against ground DNI before and after bias correction for each site ...... 74

Figure 4-3: Example of (a) a 1-hour forecast used to update over-predictions by the 48-hour persistence forecast, and (b) the corresponding bids, for 3 January 2005 in Mt. Gambier ...... 85

Figure 4-4: Example of (a) a 1-hour forecast used to update under-predictions by the 48-hour persistence forecast, and (b) the corresponding bids, for 9 January 2005 in Mt. Gambier ...... 86

Figure 4-5: LKT model net electric output plotted against SAM net electric output (a) including power block start-up energy in the LKT model, left, and (b) excluding power block start-up energy in the LKT model, right ...... 92

Figure 4-6: Summary of method to calculate financial value metrics ...... 93

Figure 4-7: Chart of the process used to classify each time step for use in calculating the equivalent forced outage rate (EFOR) ...... 98

Figure 5-1: Total revenue earned in each scenario ...... 104

Figure 5-2: Total RG cost paid in each scenario ...... 105 8

Figure 5-3: Total dump cost in each scenario ...... 106

Figure 5-4: Total net value achieved in each scenario ...... 107

Figure 5-5: Net value achieved in each scenario as a fraction the net value of a perfect 48-hour forecast at the same site ...... 111

Figure 5-6: Net value achieved in each scenario divided by the corresponding amount of electricity generated ...... 112

Figure 5-7: EFOR achieved in each scenario ...... 115

Figure 5-8: EFOR against 48-hour forecast MAE and RMSE ...... 116

List of Tables

Table 2-1: Summary of previous studies that assessed the value of using DNI forecasts to operate a CST plant ...... 35

Table 3-1: Forecasting horizon with corresponding application and forecast method, adapted from Kleissl (2010) ...... 39

Table 3-2: Summary of DNI forecast accuracy from NWP ...... 47

Table 3-3: Summary of DNI forecast accuracy from TSA methods ...... 52

Table 3-4: Summary of DNI forecast accuracy from CMV ...... 57

Table 3-5: Summary of performance of methods which convert satellite images to DNI ...... 60

Table 3-6: Summary of DNI forecast accuracy from hybrid methods ...... 63

Table 3-7: Summary of best DNI forecast accuracy for appropriate forecast horizons ...... 68

Table 4-1: Gaps in ground measurements for each site in 2005 ...... 72

Table 4-2: Satellite measurements minutes past the hour in 2005 (BOM, 2012b) ...... 72

Table 4-3: Satellite grid-point and minutes past the hour for each site ...... 73

Table 4-4: Bias correction equation coefficients for each site ...... 73

Table 4-5: MBD and RMSD of satellite DNI at each site ...... 75

Table 4-6: Site climate and site average daily global irradiation over 1990-2011 and 2005 ...... 78

Table 4-7: Total DNI at each site over 2005 ...... 78

Table 4-8: Descriptive statistics for hourly electricity price ($/MWh) in each region for 2005 . 79

Table 4-9: Parameter values used for the LKT model ...... 89

Table 5-1: The accuracy of each forecast at each site ...... 101

Table 5-2: Mean daily DNI, mean and median 1-hour electricity price, and correlation between

DNI and electricity price for each site in 2005 ...... 108

Table 6-1: Comparison of RMSE between the present study and other studies ...... 119

Table A-1: Variables in the power block and storage model ...... 154

Table A-2: LKT model parameters ...... 155 10

Table B-1: Financial value metrics for a CST plant without storage ...... 158

Table B-2: Financial value metrics for a CST plant with storage ...... 159

Table B-3: EFOR for a CST plant without storage ...... 160

Table B-4: EFOR for CST a plant with storage ...... 160

Table B-5: Normalised revenue and net value for a CST plant without storage ...... 160

Table B-6: Normalised revenue and net value for a CST plant with storage ...... 161

Table B-7: Total energy values for a CST plant without storage ...... 161

Table B-8: Total energy values for a CST plant with storage ...... 162

1. Introduction

Solar power, or generating electricity from sunlight via photovoltaic (PV) and concentrated solar thermal (CST) technology, is attractive for reducing electricity sector greenhouse gas emissions (IEA, 2014; Teske et al., 2016). Solar power has the disadvantage of sudden large changes in solar intensity, known as solar variability or solar intermittency, due to events such as moving cloud cover. Solar variability may cause sharp changes in solar power electricity supply which cannot be ignored because an electricity network must balance electricity supply and demand for safe and reliable operation (NERC, 2009; AEMO, 2010a). An important tool for mitigating solar variability is solar forecasting at forecast horizons ranging from minutes ahead to months ahead for purposes such as committing generator units to the electricity market and network planning (NERC, 2010; Sayeef et al., 2012). The solar forecasts need to accurately predict the magnitude and timing of solar intensity variation. Regulators of electricity networks and markets may introduce regulation to encourage solar power plant operators to use forecasts and thereby contribute positively towards network safety and reliability. Assessing the regulation’s effectiveness requires knowing how much a solar forecast assists network safety and reliability, and how valuable the same solar forecast appears to a solar power plant under the regulation.

1.1. Research Aim, Motivation and Scope

The main aim of this study is to evaluate the value of using direct normal irradiance (DNI) forecasts to operate a CST power plant in the Australian National Electricity Market (NEM).

This study focuses on CST instead of PV because CST can use thermal storage, which is cheaper and more efficient than other forms of storage that may be used to add storage to PV

(Madaeni et al., 2012b). CST with storage was the primary energy storage choice in a study on

100% renewable energy in the NEM (AEMO, 2013a) because it was estimated to have the

12 minimum levelised capital and operation costs of all possible storage options (James and

Hayward, 2012). The availability of storage is useful because it can help to mitigate variable output from variable renewable energy resources by storing energy when it is plentiful to use when it is scarce (NERC, 2009; Sayeef et al., 2012; AEMO, 2013a). Thus, using DNI forecasts to operate a CST plant with storage enables two approaches to mitigating solar variability.

Figure 1-1: A simplified illustration of global horizontal irradiance (GHI), diffuse irradiance, direct irradiance, and direct normal irradiance (DNI)

Solar irradiance incident on a horizontal or tilted surface may be called global irradiance, and it consists of a direct component and a diffuse component. The direct component is incident directly from the sun, whereas the diffuse component is incident from all directions due to being scattered as solar irradiance passes through the atmosphere or clouds. The direct irradiance incident on a surface oriented normal to the path of the direct irradiance is called direct normal irradiance (DNI). Global irradiance incident on a horizontal surface is called

13 global horizontal irradiance (GHI). The relationship between GHI, DNI and diffuse irradiance is shown in E. 1-1. A simplified illustration of GHI, DNI, and the direct and diffuse components of global irradiance is shown in Figure 1-1.

퐷𝑖푓푓푢푠푒 + 퐷푁퐼 ∗ cos 휃푧 = 퐺퐻퐼 E. 1-1

where 휃푧 is the solar zenith angle.

DNI can be effectively concentrated because of its unidirectional nature whereas diffuse irradiance is multi-directional and cannot be effectively concentrated. CST technology generates electricity through concentrating solar irradiance, hence DNI forecasts are required.

In contrast, PV is a non-concentrating technology so it is able to generate electricity from both direct and diffuse irradiance. Consequently, PV uses global irradiance or GHI forecasts.

DNI forecasts are unlike GHI forecasts because, apart from forecasting different solar irradiance quantities, DNI is affected by changing sky conditions more strongly than GHI

(Marquez and Coimbra, 2011; Lara-Fanego et al., 2012b; Gueymard, 2012b). For example, when solar irradiance passes through clouds, the scattering reduces the direct irradiance intensity and increases the diffuse irradiance intensity. GHI consists of both direct irradiance and diffuse irradiance, so the decrease in direct irradiance intensity can be partially offset by an increase in diffuse irradiance intensity. Only direct irradiance contributes to DNI, thus changing cloud cover produces larger changes in DNI than in GHI. Consequently, the results in studies on forecasting GHI and PV output, which have been done extensively as shown in the review by Inman et al. (2013), should not be applied to forecasting DNI and CST output because of the different in how greatly GHI and DNI are affected by changing sky conditions.

However, the forecast methods used may be the same. Thus, separate research on forecasting

DNI and CST output is necessary.

As mentioned earlier, solar forecasts are useful for managing network safety and reliability, which is the concern of the network operator. Regulations can provide incentives for solar power plant operators to use solar forecasts for their operational decision making. As such, there are two parties directly interested in the value of using DNI forecasts to operate CST plants. Considering the perspective of each party reveals what value should be applied to using

DNI forecasts. Network reliability is the risk of losing generation or the expected generation lost over a time period (AEMO, 2010a). Thus, the risk of losing generation from a CST plant can be used to represent the network operator’s perspective of the value of operating that CST plant using DNI forecasts. In the NEM, the risk of losing generation from a generator is modelled by calculating the equivalent forced outage rate (AEMO, 2010c). A CST plant is a commercial asset, hence profitability or financial value is an important factor for driving CST plant deployment as shown in studies on CST deployment in Australia (Lovegrove et al., 2012) as well as overseas (Pitz-Paal et al., 2012; IEA, 2014). DNI forecasts that accurately predict CST plant electricity generation help the CST plant operator place bids in the electricity market that maximise revenue earned from electricity sales. The information can also help avoid penalties for failing to meet bids. The revenue earned and the penalties avoided can be used to quantify the value of DNI forecasts to the CST plant. An electricity market trades in units of electricity, so a DNI forecast must be converted into a generation forecast in order for its value to be assessed. The conversion can be done by means of a CST plant model. An illustration is shown in Figure 1-2 on how value can be evaluated for a DNI forecast.

Figure 1-2: How value may be assessed for DNI forecasts

There have been studies on the value of CST plants. However, the majority of them assumed that available DNI was perfectly known, and this is not the case in reality. Other studies have calculated value for CST plants operated using DNI forecasts. A detailed review of these studies is available in Chapter 2. The studies are set in US and Spanish electricity markets, which have different regulations to the NEM. Market regulations determine the forecast horizons that generators use to operate compliantly. Hence, a study that uses forecast horizons appropriate for an Australian electricity market is necessary for precisely evaluating the value of using DNI forecasts to operate CST plants in Australia.

Australia has excellent opportunities for generating electricity using CST because of its exceptional solar resource. The annual solar energy incident on flat locations within 25 km of

16 existing network infrastructure is estimated to exceed the Australian annual energy usage by a factor of 500 (Geoscience Australia and BREE, 2014). At present, very little of the available solar resource is being used. A report by the Clean Energy Council (CEC, 2016) on Australian renewable energy generation in 2015 stated that CST generated 27 GWh of electricity (0.08% of annual renewable energy production or 0.01% of annual electricity production) and PV generated 276 GWh of electricity (0.8% of annual renewable energy production or 0.12% of annual electricity production). The report also revealed that the installed solar power capacity was about 4,246 MW, of which 53.3 MW was CST. The NEM, which covers all but two states in

Australia and is thus the largest Australian electricity network and market, reached 3,700 MW of installed PV in 2015 (AER, 2015). This was 8% of total generation capacity in the NEM, and it supplied 2.7% of the total annual energy electricity demand. The report did not mention CST, despite the 53.3 MW noted by CEC (2016) being in the NEM.

Despite the currently low usage of solar power, it is expected to increase. The total NEM PV capacity is predicted to reach 12,861 MW (21% of total NEM capacity) and supply 16,427 GWh

(7.5% of total NEM energy demand) by 2024-25 (AEMO, 2015). A study on using 100% renewable energy in the NEM estimated that PV would contribute 22-49 GW (22-48% of total network capacity) and CST would contribute 11-18 GW (10-15% of total network capacity), depending on the assumptions used (AEMO, 2013a). Solar forecasting in Australian contexts is neither widely studied nor well understood (Sayeef et al., 2012; Elliston and MacGill, 2010), thus research on CST plants operated with DNI forecasts in Australia will help to make a smooth transition to higher CST capacities.

This study will demonstrate the value of using DNI forecasts to operate a CST plant in the NEM by calculating improvement in CST plant reliability and increase in CST plant financial value.

These results will be useful for assessing whether current regulations encourage DNI forecast usage by CST plants, and for guiding investment decisions relating to obtaining or improving

DNI forecasts for operating CST plants in Australia. Besides that, it may drive the deployment of CST plants to replace fossil fuel generators and thereby reduce electricity sector greenhouse gas emissions. This study is presented in the following structure. A brief description of the

NEM is in Section 1.2. A brief description of CST plants and a review of past studies that assessed the value of CST power plants are in Chapter 2. DNI forecast methods are reviewed in

Chapter 3. The specific forecast methods and forecast horizons used for this study are described in Chapter 4. Besides the forecast methods used in this study, Chapter 4 describes the data, the CST model, and the methods for calculating CST plant financial value and reliability. The equations that form the CST model are detailed in Appendix A. The results of this study are presented in Chapter 5 and discussed in Chapter 6. A closing summary is contained in Chapter 7.

1.2. An Overview of the Australian National Electricity Market

The NEM is the market for the largest electricity network in Australia, and it delivered 196 TWh of electricity in 2014-2015 from 46 GW of installed capacity (AER, 2015). The NEM covers the eastern and southern states of Queensland (QLD), New South Wales (NSW), Victoria (VIC),

Tasmania (TAS), South Australia (SA), and the Australian Capital Territory (ACT). The extent of the transmission and distribution network in the NEM is shown in Figure 1-3. The western and northern states of West Australia and the Northern Territory each have their own electricity networks and markets. Each state in the NEM forms one region, except for ACT which is included in the NSW region.

Figure 1-3: Transmission grid and distribution networks in the National Electricity Market (AER,

2015)

The NEM is a wholesale spot market where generators submit bids of how much electricity they are willing to sell for a given price. There is only one spot market, hence only a single electricity price is earned by generators for selling electricity. The spot market price, or

19 electricity price, is determined for each 5-minute dispatch interval using the merit order

(AEMO, 2010a). This means that the lowest cost generator bids are dispatched until demand is satisfied, subject to network safety and reliability constraints. The dispatched bid with the highest price is the electricity price for that interval. Consequently, electricity prices are high when electricity supply is low relative to demand and vice versa. The electricity price is bounded by a price cap and a price floor. The price cap is recalculated each year and has risen from $12,500/MWh before 1 July 2012 to $13,800/MWh for 1 July 2016 to 30 June 2017

(AEMC, 2016b). A market floor price of -$1,000/MWh was in place in 2014 and it was recommended to continue to apply to the NEM from 1 July 2016 onwards (AEMC Reliability

Panel, 2014). Every six 5-minute dispatch intervals are combined to form 30-minute trading intervals, and the revenue earned by a generator is determined by the electricity it sold and the average price of the six dispatch intervals during a trading interval (AEMO, 2010a). The electricity price in each region is determined separately (AEMO, 2010a) despite the regions being electrically connected to allow generation in one region to supply demand in another. A study by Christensen et al. (2012) showed that the 10th and 90th percentiles of the trading interval (30-minute) electricity prices throughout a day in QLD, NSW, VIC, and SA over the period 1 March 2001 to 30 June 2007 remained between $0/MWh and $100/MWh. Analysis of electricity prices from all NEM regions over the period 1 January 2006 to 6 September 2012 by

Higgs et al. (2015) showed that overall the NEM had a mean electricity price of $42.068/MWh, a median price of $29.526/MWh, and a standard deviation of $84.188/MWh. SA had the highest mean price of $48.965/MWh, whereas TAS had the highest median price of

$34.118/MWh. Both the lowest mean price of $38.446/MWh and lowest median price of

$26.420/MWh were in QLD. The highest standard deviation was $132.245/MWh in SA and the lowest was $44.770/MWh in TAS.

In the event of unexpected loss of generation or excess generation during a dispatch interval, the frequency on the network will decrease below or increase above its targeted value, respectively. In order to correct the deviations, frequency control ancillary services (FCAS) are used to restore the balance between supply and demand. The NEM also has network control ancillary services to regulate voltage and power flow, and system restart ancillary services to restore the network from blackouts. FCAS are divided into regulation FCAS that correct minor deviations, such as small deviations in generation, and contingency FCAS that correct major deviations, such as the loss of a generator (Duque et al., 2012). FCAS are also divided into raise

FCAS that act to increase the frequency and lower FCAS that act to lower the frequency.

Furthermore, contingency FCAS are divided into three response times of 6-second, 60-seconds, and 5-minutes. Each type of FCAS has its own market for determining the dispatch price, thus there are eight separate FCAS markets. The cost of dispatching FCAS is paid for by the market participants that contributed towards the deviation, as per the “causer pays” methodology

(AEMO, 2010b). The amount of FCAS on standby for dispatch is set to ensure that the NEM reliability standard can be met. The NEM reliability standard requires that the annual unserved energy, i.e. energy demand that the NEM fails to supply, be less than 0.002% of the annual total electricity supplied (AEMC Reliability Panel, 2014).

Forecasts are required by generators to participate in the NEM. All participating generators with a generating capacity greater than or equal to 30 MW are required to submit initial bids to the market in accordance with the deadlines of the Australian Energy Market Operator

(AEMO) spot market operations timetable. The spot market operations timetable specifies that a trading day is 4 a.m. to 4 a.m. Australian Eastern Standard Time (AEST), and the initial bids for a given trading day must be submitted by noon before the trading day (AEMO, 2011). Thus a forecast horizon of at least 40 hours ahead is required for determining initial bids. Electricity is traded in the NEM in 5-minute dispatch interval to balance supply and demand. Generators

21 are allowed to submit updated bids for a dispatch interval up to 5 minutes before that dispatch interval arrives (AEMO, 2010a). Thus, generators would benefit from forecasts as short as 5 minutes ahead. AEMO also uses longer forecast horizons up to two years ahead for planning purposes to ensure that enough generators are available for generating electricity while other generators are shut down for maintenance or modifications (AEMO, 2010a). The electricity sold by a generator will mostly depend on its initial bids and updated bids rather than these longer forecast horizons.

2. Concentrated Solar Thermal Power1

As established in the Introduction, two parties have direct interests in the value of using direct normal irradiance (DNI) forecasts to operate a concentrated solar thermal (CST) plant. One party is the plant operator who is interested in maximising CST plant financial value on behalf of owners and investors. The other party is the network operator who is interested in how the

CST plant can contribute towards maximising network security and reliability. Studies have analysed CST plant operation from these perspectives, and these studies are covered in this chapter. This chapter first briefly describes the basic operation of a CST plant and gives a brief overview of the technology options for the main components of a CST plant. Following that is a review of studies that assessed the value of using CST for power generation. The studies are divided according to whether DNI was assumed to be perfectly known or was forecasted.

2.1. CST Plant Operation and Components

The conversion of DNI to electricity by a CST plant involves two essential components which are a solar field for collecting and converting solar energy to thermal energy, and a power block to convert thermal energy to electricity. A CST plant may also have a storage system for storing thermal energy, or a backup heater system for supplying thermal energy from fossil fuels, or both, for generating electricity when DNI intensity is too low for the solar field to supply enough thermal energy to operate the power block. A solar field consists of mirrors spread over a large area that are oriented to focus incident DNI onto an absorber with a small area. A heat transfer fluid (HTF) flows through the absorber where it is heated by the

1 Some of the material presented in this chapter has been previously published in: Law, E.W., Prasad, A.A., Kay, M., Taylor, R.A., 2014, Direct normal irradiance forecasting and its application to concentrated solar thermal output forecasting – A review, Solar Energy, 108, p. 287-307 http://dx.doi.org/10.1016/j.solener.2014.07.008

Law, E.W., Kay, M., Taylor, R.A., Evaluating the benefits of using short-term direct normal irradiance forecasts to operate a concentrated solar thermal plant, Solar Energy 140, p. 93-108 http://dx.doi.org/10.1016/j.solener.2016.10.037 23 concentrated DNI. The flow rate of the HTF can be controlled to achieve a desired solar field outlet temperature or to prevent the HTF temperature falling outside a safe operating range.

The hot HTF can be sent directly to the power block for generating electricity. Alternatively, the hot HTF can deliver thermal energy to storage instead of the power block if storage is available. After delivering heat to the power block or storage, the cold HTF returns to the solar field for heating again. If the solar field is unable to heat the cold HTF to the temperature required by the power block because of low DNI intensity, then the cold HTF may be heated by storage or a backup heater if they are available. A simplified diagram of how a typical CST plant with storage and a backup heater operates is shown in Figure 2-1.

Figure 2-1: The main components of a CST plant and how they operate together to convert DNI into electricity

2.1.1. Solar field collector

The solar field may use parabolic trough collectors, linear Fresnel collectors, power towers, or parabolic dishes for concentrating DNI and obtaining thermal energy. Simplified illustrations of these collectors are presented in Figure 2-2. The collector technology is often used as the main characteristic for classifying CST plants, as shown in the reviews by Baharoon et al. (2015) and

Siva Reddy et al. (2013). In general, the efficiency of converting DNI to thermal energy is maximised by minimising the optical losses incurred while concentrating and absorbing DNI, and minimising the heat losses caused by operating above the ambient air temperature. 24

Parabolic trough collectors are currently the most widely used collector technology by generation capacity (Baharoon et al., 2015; Teske et al., 2016). A parabolic trough collector is a long trough that has a parabolic widthwise cross-section and a receiver at the trough’s focal line to absorb heat. A linear Fresnel collector concentrates DNI in a similar manner to a parabolic trough collector. The difference is that a linear Fresnel collector uses multiple parallel rows of independently rotatable mirrors, rather than a single parabolic trough mirror, to reflect DNI to a focal line. Parabolic trough and linear Fresnel collectors are called line concentrators because they concentrate DNI from a wide area to a line. The concentration reaches 70-100 times and produces outlet temperatures of 350-550°C (Teske et al., 2016). Line concentrators only rotate along a single axis, so they are best placed in a north-south orientation to track the movement of the sun from east to west.

In contrast to line concentrators, power tower collectors and parabolic dish collectors concentrate DNI from a wide area to a single point, and are thus called point collectors. A power tower configuration consists of a receiver at the top of a tower in a field of heliostats or mirrors. The azimuth angle and tilt angle of each heliostat are changed throughout the day to reflect DNI incident on the heliostat to the top of the tower. A parabolic dish is a dish that is shaped as a parabola for any cross-section made through its centre. Due to its shape, the dish has dual axis tracking so that it can face towards the sun to reflect DNI to its focal point. Point collectors can concentrate DNI by 600-1000 times and achieve outlet temperatures of 800-

1000°C (Teske et al., 2016). Although parabolic dishes can achieve high concentration ratios, there are mechanical limits to the size of a parabolic dish that can be safely supported.

Consequently, many parabolic dishes are necessary to achieve high thermal output capacities.

The size of a solar field can be measured using the solar multiple (SM). The SM is the ratio of the thermal output of the solar field under design conditions to the thermal input required by the power block to generate its rated electrical output. If the SM is one, then the solar field will produce just enough thermal output for the power block to generate the rated electrical output when operating conditions match design conditions. If the SM exceeds one, then under design conditions the solar field thermal output will exceed the power block’s thermal requirements to operate at rated capacity. An SM greater than one is beneficial because, if the

26 incident DNI is less than that of the design conditions, the solar field may still supply sufficient thermal output for the power block to operate at rated capacity. Furthermore, if storage is available then any solar field thermal output that exceeds the power block’s requirements can be stored. This is beneficial because a greater proportion of the incident DNI can be converted into electricity and electricity can be generated when prices are high or over a longer period of time.

2.1.2. Power block

The power block of a CST plant generates electricity from thermal energy by using the same single stage Rankine or Brayton power cycles used by fossil fuel power plants (Siva Reddy et al.,

2013; Dunham and Iverson, 2014). This allows a CST plant to use a fossil fuel or biomass backup heater to supplement its solar field thermal output, or a fossil fuel plant to use solar field thermal output to offset its fossil fuel usage (IEA, 2014). Research by Dunham and Iverson

(2014) recommends using Rankine cycles when the power cycle inlet temperature is less than

600°C and using Brayton cycles when it is above 600°C. As mentioned earlier, the maximum solar field outlet temperature that line concentrators can achieve is 550°C whereas for point concentrators it is 1000°C. Hence, parabolic trough collectors and linear Fresnel collectors are combined with Rankine cycles, and power towers are combined with Brayton cycles. Although parabolic dish collectors are reported to reach temperatures as high as 400°C (Kaushika and

Reddy, 2000) or 750°C (Richter et al., 2009), commercial parabolic dish systems have only used

Stirling engines (Baharoon et al., 2015). Real CST plants have achieved power cycle efficiencies of 37%-42%, and research on advanced power cycles is aiming to achieve efficiencies of 50% or more (Dunham and Iverson, 2014). The size of a power block is usually given in terms of its rated electric output using Megawatts-electric (MW-e).

2.1.3. Storage

Thermal energy from the solar field may be stored by using sensible heat storage, latent heat storage, or thermochemical storage (Kuravi et al., 2013). In the case of sensible heat storage, the HTF that circulates in the solar field may be directly stored or it may be used to heat a separate storage material. Latent heat storage and thermochemical storage require a different storage material from the HTF in the solar field because these storage processes involve changes in the storage material. There are many potential storage materials with different thermal properties as shown by Kuravi et al. (2013) and Tian and Zhao (2013). Only sensible heat storage has been used by commercial CST plants (Teske et al., 2016), but latent heat storage and thermochemical storage can achieve higher energy densities, meaning they store more energy per unit volume (Kuravi et al., 2013; Tian and Zhao, 2013).

Some storage materials can operate at temperatures as low as 100°C while others are suited to operating temperatures as high as 2500°C, so storage material selection requires consideration of the solar field outlet temperature and the power block inlet temperature, in addition to other factors (Tian and Zhao, 2013). Sensible heat storage using molten salts or high temperature oils is suitable for temperatures in the range 120°C-850°C (Tian and Zhao,

2013), while solid materials such as fire bricks can be used for temperatures as high as 1200°C

(Kuravi et al., 2013). The phase change temperature of phase change materials for latent heat storage may be in the range 100°C-900°C (Tian and Zhao, 2013) or 307°C-714°C (Kuravi et al.,

2013). Thermochemical storage materials can operate in the temperature range of 180°C-

2500°C (Tian and Zhao, 2013) or 180°C -1000°C (Kuravi et al., 2013).

Storage size can be measured using units of thermal energy or hours of storage. The hours of storage is calculated by dividing the thermal energy capacity of storage by the thermal input of the power block to generate its rated electrical output. Thus, the number of hours of storage is

28 the number of hours that the power block can generate its rated output if it is supplied thermal energy from storage only and storage is initially fully charged.

2.1.4. Heat transfer fluid

There are a variety of HTF materials which can be broadly categorised into synthetic thermal oils, molten salts, liquid metals, pressurised gases, and water-steam (Vignarooban et al., 2015;

Benoit et al., 2016). A major factor for choosing a HTF is the CST system operating temperature because HTF materials have minimum temperature limits to prevent freezing and maximum temperature limits to prevent decomposing. Thermal oils must be used below 400°C

(Vignarooban et al., 2015; Benoit et al., 2016). Most molten salts have a maximum of 530°C-

600°C (Benoit et al., 2016), but some can be used in temperatures as high as 800°C-900°C

(Vignarooban et al., 2015). Among liquid metals, liquid sodium is stable up to 883°C whereas lead-bismuth eutectic mixture is stable up to 1533-1670°C (Vignarooban et al., 2015; Benoit et al., 2016). Pressurised gases and water-steam do not have upper temperature limits

(Vignarooban et al., 2015; Benoit et al., 2016). As such, only thermal oils are restricted to use with parabolic trough and linear Fresnel CST plants, whereas all other HTF materials could be used with line and point concentrators.

Besides the acceptable temperature range, other HTF properties to consider are density, specific heat capacity, thermal conductivity, viscosity, and cost. Overall, Vignarooban et al.

(2015) believe that molten salts are the best HTF for CST plants because they generally have relatively low melting points, high thermal stability, good thermal conductivity, and high heat capacity. However, molten salts are very corrosive to metal alloys, which Vignarooban et al.

(2015) acknowledge as the main barrier to widespread commercial use of molten salts as a

HTF. (Benoit et al., 2016) believe that the best HTF would be one that can be also be used for storage to minimise CST plant cost, which is a design approach that also favours molten salts.

2.2. Studies on the Value of CST

Studies have assessed the financial value of CST by either calculating the reduction in the network cost of generating electricity to satisfy demand (Mehos et al., 2015; Jorgenson et al.,

2014; Denholm et al., 2013; Jorgenson et al., 2013) or the revenue that the CST plant earns

(Madaeni et al., 2012b; Sioshansi and Denholm, 2010). The reduction in cost of generating electricity takes the point of view of the network and requires models of the CST plant and the network that the CST plant is part of. In contrast, the revenue that the CST plant earns takes the point of view of the CST plant owner and only requires a model of the CST plant. Using a

CST model means that changing parameters such as the solar field size and storage size will affect the calculated value. Using a network model means that the calculated value will be affected by parameters such as the total generation capacity of the network, the electricity demand, and the amount of renewable energy generation in the network. Thus the financial value calculated from the point of view of the network depends on more factors.

Other studies have focused on the reliability of a CST plant because it is of interest to the network for ensuring that electricity can be supplied securely and reliably. If low reliability prevents a CST plant generating electricity, the revenue earned will decrease, hence calculating revenue can implicitly account for the value of reliability to the CST plant. As the

CST plant’s financial value already accounts for the value of reliability, only the network’s point of view for the value of reliability will be considered here. The value of reliability has been represented in studies by the firm capacity added to the network (Madaeni et al., 2012a; b;

Madaeni et al., 2011) or the dollar value of the added firm capacity (Mehos et al., 2015;

Jorgenson et al., 2013). Firm capacity refers to the power plant capacity that is guaranteed to be available for generating electricity, and it is less than the power plant rated capacity. The dollar value of the firm capacity is the avoided cost of deploying new generation capacity to satisfy demand. Both approaches to quantify the value of reliability involve calculating the firm capacity added to the network by using a model of the network. The value of the added firm 30 capacity involves further calculations and assumptions of alternative new generation options to estimate the avoided cost of adding new generation capacity.

The study by Mehos et al. (2015) showed that a power tower CST plant could provide a value of $47-$119/MWh to an electric network by reducing the cost of generating electricity and by avoiding the need to construct additional power generation capacity. The value was lower when storage size was increased beyond 6 hours of storage and the solar field size was increased beyond a solar multiple of 1.3. Increasing storage size and increasing solar field size allowed a CST plant to generate more electricity, however the extra generation would eventually occur during periods of lower electricity prices because of the limited generation capacity of the power block. Electricity generated during periods of low prices was less valuable than electricity generated during periods of high prices, thus electricity generated during periods of lower prices reduced the value per MWh of the CST plant.

Mehos et al. (2015) also considered the first year power purchase agreement (PPA) price required by the CST plant. If the first year PPA price, which represents the cost of generation to the CST plant, was less than the value provided to the network, then the CST plant was more likely to be considered profitable and be constructed. This approach assumed that the CST plant was paid compensation equal to the value it provided to the network. Mehos et al. (2015) showed that different natural gas prices and investment tax credit percentages affected the likelihood of the CST plant being profitable. Overall, a power tower CST plant with a large solar field and three hours of storage was most likely to be profitable. Higher natural gas prices and higher investment tax credit percentages further improved the likelihood that the CST plant would be profitable. A study by Jorgenson et al. (2013) produced results with similar trends as

Mehos et al. (2015) for a power tower CST plant. The total value of reducing the cost of generating electricity and avoiding the need for adding more generation capacity was $55-

$105/MWh, which is similar to the results from Mehos et al. (2015). The small difference is due

31 to the use of different modelling parameters. Jorgenson et al. (2013) also used parabolic trough CST plants in the study and found that they required larger power block sizes to achieve the same value as power tower CST plants.

Denholm et al. (2013) assessed the value of CST in an electricity market with 33% renewable portfolio standard (RPS) scenario, which was the same as the RPS scenario used by Mehos et al.

(2015). Results from Denholm et al. (2013) showed that a parabolic trough CST plant with a solar multiple of 2 and 6 hours of storage that generated electricity and provided ancillary services could achieve a value of $80-135/MWh. The exact value depended on the ancillary services capacity provided, with higher values corresponding to higher ancillary services capacities. The study by Denholm et al. (2013) was taken further by Jorgenson et al. (2014) in a study that examined scenarios with 33% and 40% RPS. Jorgenson et al. (2014) showed that a power tower CST plant had a value of $94.6-$107/MWh in a 33% RPS scenario. The range was produced by varying network constraints and the power block size and storage size of the CST plant. The difference in results compared to Denholm et al. (2013) was due to different assumptions such as lower natural gas prices, lower carbon prices, the presence of other electricity storage services in the network, and different CST plant operation and cost assumptions. In a 40% RPS scenario, the range of values that the power tower CST plant achieved increased slightly to $96-$109/MWh.

Value from the CST plant’s point of view has been calculated by Madaeni et al. (2012b) and

Sioshansi and Denholm (2010) using the revenue earned by the CST plant. Both studies used a model developed by Sioshansi and Denholm (2010) to maximise revenue earned by selling electricity during periods of high prices. Results from both studies showed that increasing solar field size and storage size increased the value of a parabolic trough CST plant by increasing the revenue earned. The annual operating revenue found by Sioshansi and Denholm (2010) was

$14-$35 million, depending on the CST plant configuration and the solar resource of the

32 location. Revenue could be further increased by up to 8% depending on the amount of ancillary services provided. In the study by Madaeni et al. (2012b) the annual operating revenue was $10-$35 million, depending on the CST plant configuration and the solar resource of the location. The higher values were produced by CST plants with larger solar fields or more storage. A larger solar field provided more thermal energy to generate electricity, and a larger storage size allowed more thermal energy to be stored for generating electricity when the electricity price was higher. Madaeni et al. (2012b) also evaluated the financial feasibility of the CST plant by comparing the annual operating revenue and the cost of financing the CST plant. Not all CST plant configurations were considered financially feasible because larger components had higher capital cost that could not be covered by the increase in revenue. The study concluded that feasibility depended on the solar resource of the location and cost of financing the CST plant. The result from Sioshansi and Denholm (2010) and Madaeni et al.

(2012b) that increasing solar multiple increased value from the CST plant’s point of view contrasts with the finding that value from the network’s point of view reached a maximum value for a small solar multiple (Mehos et al., 2015; Jorgenson et al., 2014). This shows that it is important to state the perspective used to assess the value of a CST plant because not all perspectives find the same value in the same scenarios.

The firm capacity added to the network was shown to be higher if a parabolic trough CST plant had a large solar field and a large storage size (Madaeni et al., 2012b). The minimum added firm capacity was 80% of the rated capacity of a CST plant with a solar multiple of 1.5 and no storage. Larger solar multiples and storage sizes could increase the added firm capacity to 100% of the CST plant rated capacity. The added firm capacity was also found to depend on the solar resource of the location. In another study, Madaeni et al. (2012a) showed that the added firm capacity could be as low as 45% of the CST plant rated capacity if there was no storage, the solar multiple was 1, and the location had poor solar resource. Similar results were obtained in

33 another study by (Madaeni et al., 2011), which showed that the added firm capacity could be

45%-90% of rated capacity. Again, the value was shown to be affected by the amount of storage, the solar multiple, and the solar resource of the location.

2.3. Studies on the Value of Using Forecasts to Operate CST

The studies cited in the previous section calculated CST plant value by assuming that the CST plant was operated with perfect knowledge of available DNI, i.e. DNI was perfectly forecasted.

In reality, DNI is not perfectly forecasted. Forecast errors prevent real CST plants achieving the value calculated by using a perfect forecast. There are few studies that examined the value of

CST plants operated using DNI forecasts. These studies are summarised in Table 2-1. The low number of studies is likely because new CST generation capacity only started being added to electricity networks from 2007 onwards after a long period of no new added capacity covering

1991 to 2006 (Hinkley et al., 2013). This suggests that the study by Wittmann et al. (2008) is probably the earliest study on this topic. The study considered the use of 48-hour ahead DNI forecasts to operate a CST plant in the Spanish electricity market. The CST plant was modelled after Andasol-1, a parabolic trough plant with a rated output of 50 MW and 7.5 hours of storage. The DNI forecasts were produced using the European Centre for Medium-range

Weather Forecasts (ECMWF) numerical weather prediction (NWP) model, the Aerosol-based

Forecasts of Solar Irradiance for Energy Applications (AFSOL) NWP model, and persistence of

DNI. The analysis was conducted over two clear sky days and three cloudy days. Results showed that the persistence forecasts performed better than the NWP models by achieving revenue within 5% of the revenue from a perfect forecast in both clear sky and cloudy conditions. The unusual outcome of persistence being best was explained by the three consecutive cloudy days having similar DNI profiles. Besides that, using AFSOL achieved higher revenue than ECMWF on the clear sky days, whereas the opposite was true on the cloudy days.

Table 2-1: Summary of previous studies that assessed the value of using DNI forecasts to

operate a CST plant

Electricity Forecast Forecast market or Key result regarding value of using DNI Authors horizon method system forecasts to operate a CST plant Persistence of On clear days, AFSOL and persistence achieve DNI Spanish at least 99.6% of revenue earned using Wittmann et al. 48-h AFSOL electricity perfect forecasts. On cloudy days, ECMWF (2008) ECMWF market and persistence achieve at least 92.8% of Perfect revenue earned using perfect forecasts. Persistence of Using persistence forecasts can achieve at Sioshansi and ERCOT and 24-h DNI least 87% of annual profits achieved by using Denholm (2010) CAISO, USA Perfect a perfect forecast. Persistence of DNI Spanish Using NWP forecasts can reduce penalty Kraas et al. 48-h Ensemble of electricity payments by 47.6% compared to using (2013) HIRLAM and market persistence forecasts. ECMWF, post-processed Using day-ahead market prices with forecasts Spanish Channon and Persistence of reduces balancing costs compared to using 24-h electricity Eames (2014) DNI within-day market prices when the CST plant market solar multiple exceeds 2. Persistence of Using NWP forecasts can reduce relative Nonnenmacher DNI Various US 36-h reserve requirements by an average of 28.6% et al. (2016) RDPS, markets compared to using persistence forecasts. post-processed The Aerosol-based Forecasts of Solar Irradiance for Energy Applications (AFSOL), the European Centre for Medium- range Weather Forecasts (ECMWF), the High-resolution Limited Area Model (HIRLAM), and the Regional Deterministic Prediction System (RDPS) are numerical weather prediction (NWP) models. ERCOT is the Electricity Reliability Council of Texas and CAISO is the California Independent System Operator.

The study by Sioshansi and Denholm (2010) used 24-hour persistence of DNI to determine the

operating profit of a generic parabolic trough CST plant with different solar field sizes and

storage sizes. The operating profit was defined as the revenue less the variable cost of

generating electricity. Electricity market data from the California Independent Service

Operator (CAISO) and the Electricity Reliability Council of Texas (ERCOT) were used.

Simulations were conducted for one year in two locations. Results showed that persistence

forecasting achieved at least 87% of the operating profits achieved by a perfect forecast across

35 all solar field sizes and storage sizes in both locations. This result is somewhat similar to that in the study by Wittmann et al. (2008) in terms of the value achieved by persistence forecasting relative to the value achieved by a perfect forecast. The results from Sioshansi and Denholm

(2010) are more reliable because a 1-year study period includes more weather conditions than a 5-day study period and thus are less likely to be biased.

The study by Kraas et al. (2013) used 48-hour DNI forecasts to operate a CST plant in the

Spanish electricity market. In contrast to the previous two studies, Kraas et al. (2013) focused on the difference in penalty payments instead of the difference in revenue from using different forecast methods. Two forecast methods were used: one was a hybrid of forecasts from both the High Resolution Limited Area Model (HIRLAM) and the ECMWF with post-processing, and the other was persistence of DNI. The CST plant was modelled after Andasol-3, which has the same design as Andasol-1. CST plant operation simulated over 2.5 years showed that the total penalty payment incurred by using the hybrid forecast was 47.6% lower than that incurred by using persistence forecasting.

Nonnenmacher et al. (2016) also considered the effect of forecast accuracy on reserves instead of revenue. The reserve requirements from using the Regional Deterministic Prediction

System (RDPS) NWP model and were compared to those of a persistence forecast at the 36- hour horizon. Results showed that the RDPS forecasts reduced reserve requirements by 28.6% compared to the persistence forecasts. The result that NWP forecasts reduce the amount of reserves to purchase compared to persistence forecasts is consistent with the result from

Kraas et al. (2013).

Channon and Eames (2014) took a slightly different approach to the other studies by comparing the outcome of using different market prices rather than different forecast methods. The study used a 24-hour persistence forecast and took prices from the Spanish day- ahead market and the within-day market. The use of day-ahead market prices was found to 36 produce lower balancing costs because it required purchasing less balancing electricity, i.e. ancillary services or reserve generation. This only occurred when the solar multiple was greater than 2 and otherwise there was no significant difference in the outcomes from using either price.

The past studies demonstrate that the different forecast accuracy of different forecast methods will affect the financial value achieved by the CST plant. Better forecast accuracy helps the CST plant earn revenue and reduce reserve requirements. None of the past studies used forecasts below the 24-hour horizon. In contrast, studies on using global horizontal irradiance forecasts to operate photovoltaic plants have considered short-term forecasts between the 1-minute and 2-hour horizon. These forecasts were shown to be beneficial by reducing the cost of mitigating generation variability (Perez et al., 2014) and reducing the probability of causing imbalances (Kaur et al., 2016). This study will consider using short-term forecasts below the 24-hour horizon to operate a CST plant, and will thus be investigating an as yet unexplored area of the research field.

2.4. Chapter Summary

This chapter presented a brief description of the different options available for the solar field collector, the power block, the storage, and the heat transfer fluid that may be used by a CST plant. Only these components were described because they are the main components that determine how a CST plant can operate.

After that, studies that assessed the value of CST plants were reviewed. The value of a CST plant may be from the CST plant’s perspective or from the network’s perspective. The CST plant’s perspective considers only the revenue that is earned by selling electricity. The network’s perspective can consider the amount of firm capacity added to the network, the

37 money saved by adding firm capacity to the network, or the money saved by reducing the cost of supplying electricity. In general, the studies agreed that value depends on CST plant configuration and solar resource of the plant site. Studies from different perspectives showed that contrasting CST plant configurations would maximise value, thus highlighting the difference in value from each perspective. Studies showed that increasing the solar field size increased revenue earned by the CST plant but decreased the savings from adding firm capacity to the network. Thus it is important to consider the correct perspective when calculating the value of a CST plant.

In reality, CST plants rely on forecasts of available DNI to operate. Forecasts have errors and errors may incur penalties if the errors cause the CST plant to make it harder to balance electricity supply and demand on the network. This occurs if the CST plant generates less electricity than expected and requires reserve generation from ancillary services to maintain the balance of supply and demand. Thus the value of a CST plant calculated without considering real forecasts will likely overestimate the value. Studies that assessed value of CST plants by using DNI forecasts were reviewed. The studies showed that DNI forecasts that are more accurate are able to achieve higher value. However, there are few studies about using

DNI forecasts and none have considered forecast horizons less than 24 hours ahead. This study will consider forecasts below the 24-hour horizon and evaluate their effect on the financial value and reliability of CST plants with and without storage. This study also differs from existing studies by using Australian solar data and an Australian electricity market.

3. Methods to forecast direct normal irradiance2

As mentioned in the review of studies on CST value, DNI forecasts should be used when assessing the value of operating a CST plant because DNI is not perfectly known in advance and forecast errors will affect the value that can be achieved. Different forecast methods should be used for particular forecast horizons, as recommended in various studies (Schroedter-

Homscheidt et al., 2009; Kleissl, 2010; Schroedter-Homscheidt et al., 2012; Kostylev and

Pavlovski, 2011; Diagne et al., 2013; Coimbra et al., 2013). This is because the forecast accuracy depends on how well the method is able to accurately forecast the atmospheric phenomena that affect solar irradiance intensity and are significant over the time scale of the forecast horizon. For example, numerical weather prediction (NWP) models produce more accurate solar irradiance forecasts beyond 5 hours ahead than cloud motion vectors (CMVs) because NWP models consider cloud motion, formation and dissolution in their physical modelling whereas CMVs only account for cloud motion (Coimbra et al., 2013). A summary of the forecast horizons with corresponding applications and forecast methods is presented in

Table 3-1.

Table 3-1: Forecasting horizon with corresponding application and forecast method, adapted from Kleissl (2010)

Forecast horizon Application Forecast method 1 – 10 minutes Baseline Persistence Ground-based cloud motion 10 minutes – 1 hour Short-term ramps, regulation vectors Time series analysis models Satellite-based cloud motion 1 hour – 5 hours Load following vectors Numerical weather prediction 6 hours – 10 days Unit commitment models

2 The material presented in this chapter has been previously published in: Law, E.W., Prasad, A.A., Kay, M., Taylor, R.A., 2014, Direct normal irradiance forecasting and its application to concentrated solar thermal output forecasting – A review, Solar Energy, 108, p. 287-307 http://dx.doi.org/10.1016/j.solener.2014.07.008 New and rapid research in this field since the writing of this thesis may have produced information that replaces the literature reviewed in this thesis as the state of the art. 39

The accuracy of forecasting GHI at different forecast horizons has been extensively researched as demonstrated in the review by Inman et al. (2013). Although GHI forecasts are well researched and DNI can be statistically derived from GHI, the results should not be directly applied to DNI forecasts because sky conditions affect DNI more strongly than GHI. Clouds are the main cause of DNI intensity reduction when the sky is cloudy, otherwise aerosols are the main cause when the sky is clear of clouds (Wittmann et al., 2008; Gueymard, 2012b). The reduction in DNI intensity due to aerosols ranges from 30% to 100% depending on aerosol concentration whereas for GHI the reduction in intensity is considerably lower at about 10%

(Marquez and Coimbra, 2011; Lara-Fanego et al., 2012b; Schroedter-Homscheidt and Oumbe,

2013).

This chapter reviews DNI forecast methods. A brief description is provided for each method to demonstrate how it is different from other methods. The accuracy of each method at particular forecast horizons as reported in literature is presented.

3.1. Clear Sky Models

The solar irradiance at a particular location on the Earth’s surface typically follows an annual cycle. Simple clear sky models (CSMs) can model this cycle by using variables which determine solar position and extra-terrestrial solar irradiance (Reno et al., 2012). More complex CSMs can simulate attenuation of solar irradiance after transmission through the atmosphere by including atmospheric variables such as aerosol optical depth (AOD) or ozone thickness (Reno et al., 2012). The availability of high quality data for calibrating the model to the location is vital for high modelling accuracy (Reno et al., 2012; Ineichen, 2006; Badescu et al., 2012).

When selecting a CSM, modelling accuracy is important and it is equally important to consider ease of use, availability of input data and the ability of the model to output desired irradiance values (Ineichen, 2006). 40

Some CSMs directly output DNI, such as the Simplified Solis model (Ineichen, 2008a), the

Reference Evaluation of Solar Transmittance – 2 bands (REST2) model (Gueymard, 2008) and the Bird model (Bird and Hulstrom, 1981). If DNI is not a direct output, then global-to-diffuse models, also called global-to-beam models, may be used to separate the GHI into DNI and diffuse components by statistical means. Examples of such models are the DirInt model (Perez et al., 1991), the Skartveit and Olseth model (Skartveit et al., 1998), and the Boland-Ridley-

Lauret (BRL) model (Ridley et al., 2010). If either the diffuse or DNI component of GHI is determined, then the other can be calculated due to the relationship described in E. 1-1.

CSMs can assist solar irradiance forecasting by modelling clear sky irradiance based on forecast values of atmospheric composition. For example, Ruiz-Arias et al. (2012) used the Weather and

Research Forecasting (WRF) model to forecast pressure, humidity, and AOD to input to the

REST2. Besides that, CSMs may be used to interpolate NWP solar irradiance output from 3- hour resolution to 1-hour resolution as applied in studies by Breitkreuz et al. (2009), Lorenz et al. (2009) and Martin (2011). Correlations between cloud cover and reduction in clear sky irradiance have been identified by Kasten and Czeplak (1980), Gul et al. (1998), Chen et al.

(2007) and Myers (2007). Thus cloud cover data may be added to modify clear sky irradiance and estimate solar irradiance in overcast conditions.

3.1.1. CSM modelling accuracy

The REST2 model was shown to marginally have the lowest nRMSE of 27% in the comparative study by Ineichen (2006). The European Solar Radiation Atlas (ESRA) model (Rigollier et al.,

2000) and the Molineaux model (Molineaux et al., 1998) shared the same nRMSE as the REST2 model. The nRMSE of 5 other models tested in the study ranged from 28% to 34%. The REST2 model was also shown to have the largest nMBE of -25%. The result of the REST2 model having lowest nRMSE was reproduced in the comparative study by Gueymard (2012a). Results averaged over 5 test locations showed that the REST2 model had nRMSE of 1.42% while the

41 nRMSE range for 17 other models in the study was 2.64% to 43.18%. Gueymard (2012a) considered the Simplified Solis model, the Hoyt model (Hoyt, 1978), the Bird model, and the

Iqbal-C model (Iqbal, 1983) to also have good DNI modelling accuracy. The comparative study by Badescu et al. (2012) tested 54 CSMs using a variety of methods that analysed model sensitivity to inputs and seasons. No single CSM clearly outperformed the others. However the shortlist of best models for global and diffuse irradiance listed the ASHRAE 1972 model

(ASHRAE, 1972), the Biga model (Biga and Rosa, 1979), the ESRA model, and the REST2 model.

In all of the studies, the normalised error values were calculated through division by the mean

DNI of the test data for a particular location. The REST2 model was also found to be the most accurate at modelling GHI in a comparative study by Reno et al. (2012).

3.1.2. Global-to-diffuse models

A comparison of 17 global-to-diffuse models was conducted by Torres et al. (2010) and the models most able to match the measured diffuse irradiance data were the Skartveit and Olseth model, the DirInt model, and the BRL model. The RMSE were 46.57 W/m2, 45.53 W/m2, and

48.12 W/m2 respectively for the 3 models. The other 14 models had RMSE larger than 53

W/m2. The Erbs model (Erbs et al., 1982) was found to have similar nRMSE and nMBE to the

DirInt model and the Skarveit and Olseth model in a study by Ineichen (2008b). Vick et al.

(2012) compared the DirInt model against measured DNI and found that the annual percentage difference was at most ±4.28%. This was better than the Direct Insolation

Simulation Code (DISC) model (Maxwell, 1987) which had annual percentage difference of

±6.46% and worse than the DirIndex model (Perez et al., 2002) which had the result of ±2.41%.

A study by Ridley et al. (2010) showed that the BRL model had an average median absolute percentage error of 19.55% which outperformed 27.02% achieved by the Reindl model (Reindl et al., 1990), 21.96% achieved by the Skartveit and Olseth model, and 32.63% achieved by the

Perez model (Perez et al., 1992) in the Southern Hemisphere. In the Northern Hemisphere the

42 difference in performance was smaller with average median absolute percentage error of 8.22% for the BRL model, 9.24% for the Reindl model, 9.13% for the Skartveit and Olseth model, and

11.30% for the Perez model. The BRL model was found by Copper and Sproul (2012) to have an average RMSE of 113 W/m2 which was less than 118 W/m2, 121 W/m2, and 115 W/m2 achieved by the Skartveit and Olseth model, the Watanabe model (Watanabe et al., 1983) and the Zhang model (Zhang et al., 2002) respectively for Australian locations. The MBE of the BRL model (-27 W/m2) was neither the minimum nor the maximum value among all the MBE values.

Overall, the BRL model is shown to perform better than other global-to-diffuse models in Spain by Torres et al. (2010), in Australia by Copper and Sproul (2012), and in both hemispheres by

Ridley et al. (2010). The Skarveit and Olseth model is also a good global-to-diffuse model because it had slightly lower RMSE than the BRL model in the study by Torres et al. (2010) and came second to the BRL model and ahead of other models in the studies by Ridley et al. (2010) and Copper and Sproul (2012). The DirInt model is another model that may be good because it was shown to have similar performance to the Skartveit and Olseth model by Torres et al.

(2010) and Ineichen (2008b).

3.2. Numerical Weather Prediction Models

Solar irradiance forecasts may be obtained from the use of numerical weather prediction

(NWP) models. A description of the atmosphere at a given time is used as a starting point for

NWP models. Equations that physically describe horizontal momentum, vertical momentum, hydrostatic continuity and conservation of energy are used to calculate changes in the state of the atmosphere across fixed time steps. Thus NWPs model are able to forecast atmospheric variables by evaluating the equations over the necessary number of time steps to arrive at the desired forecast horizon. 43

NWP models may be generally separated into two categories according to the modelling domain spatial extent. Models in the first category cover the entire globe and are called global models. Examples of global models are the European Centre for Medium-Range Weather

Forecasts (ECMWF) model (ECMWF, 2012) and the Global Forecast System (GFS) (EMC, 2013).

Models in the other category cover only a region of the globe and are called regional or mesoscale models. Examples of mesoscale models are the Weather and Research Forecasting

(WRF) model (Skamarock et al., 2008), the Mesoscale Atmospheric Simulation System (MASS) model (Manobianco et al., 1996), the Advanced Multiscale Regional Prediction System (ARPS)

(Xue et al., 2000), the North American Model (NAM) (Janjic et al., 2010), and the Fifth- generation Pennsylvania State University – National Center for Atmospheric Research

Mesoscale Model (MM5) (Grell et al., 2008). Mesoscale models depend on global models for defining boundary conditions of the modelling domain. However, the smaller modelling domain allows the topography to be described at higher resolution in mesoscale models compared to global models and thus forecasts can be obtained at higher resolution (Perez et al., 2013). The typical time resolution of output for regional NWP models is 1 hour or 3 hours for global models (Diagne et al., 2013; Perez et al., 2013). NWP models are recommended to be used when the forecast horizon exceeds 5 hours (Perez et al., 2010a; Kleissl, 2010) due to the “spin up” time required to assimilate data and initialise (Coimbra et al., 2013).

Within each category, NWP models can vary according to design parameters such as horizontal spatial resolution, vertical spatial resolution, forecast horizon, output time step and parameterization to model atmospheric processes. Due to the different configurations, forecast accuracy for a particular location will vary among the NWP models as demonstrated for GHI forecast accuracy in the review by Perez et al. (2013) for sites in the US, Canada and

Europe using several global and regional NWP models. The nRMSE normalised by mean GHI for

1-day ahead forecasts ranged from 20% to 69% in US sites, 29% to 44% in Canadian sites, and

40% to 64% in central European sites.

NWP use radiative transfer models (RTMs) to calculate solar irradiance and they typically predict GHI not DNI (Ruiz-Arias et al., 2012; Larson, 2013). There are several methods to obtain

DNI from NWP model outputs. Breitkreuz et al. (2009) used a global-to-diffuse model and the ratio of forecasted GHI to GHI from a clear sky model (CSM) to derive DNI forecasts at 1 hour resolution from the 3 hour resolution output from the ECMWF. In the same study, hourly DNI forecasts were also obtained from a forecasting system called the Aerosol-based Forecasts of

Solar Irradiance for Energy Applications (AFSOL). AFSOL was made from the MM5, an emission database, a chemical transport model and the library for radiative transfer (libRadtran). Lara-

Fanego et al. (2012b) combined satellite measurements of aerosol and ozone with outputs from a WRF model initialised with GFS boundary conditions to derive DNI forecasts. Lara-

Fanego et al. (2012a) derived DNI forecasts from a WRF model initialised with ECMWF boundary conditions by using WRF outputs and a global-to-diffuse model. Ruiz-Arias et al.

(2011) used satellite measurements of ozone and AOD, WRF outputs, and a broadband CSM to obtain DNI forecasts. González et al. (2010) used the Meteorological Radiation Model (MRM) to derive DNI from the WRF model initialised with GFS boundary conditions.

As cloud cover and aerosols strongly affect DNI intensity, the ability of NWP to model their presence and effects is vital for accurate forecasting. When clouds are present, forecast error can be caused by inaccurate modelling of the vertical and horizontal distribution of clouds as well as inaccurate modelling of cloud layer thickness (Larson, 2013). For aerosols, forecast error can be due to inaccurate parameterization of the scattering and absorption by water vapour, ozone and aerosols molecules (Larson, 2013). In general, NWP models usually fail to predict the presence of clouds and thus forecast accuracy is biased towards over-predicting available DNI (Larson, 2013; Mathiesen et al., 2013). General factors which cause errors in

NWP model forecasts include inaccurate initial values, coarse spatial resolution, and large calculation time steps.

3.2.1. NWP forecast accuracy

A summary of DNI forecast accuracy from NWP is presented in Table 3-2. The annual average

DNI forecast accuracy in all sky conditions from the WRF model was shown to outperform persistence (Lara-Fanego et al., 2012a; Lara-Fanego et al., 2012b). However, persistence can sometimes outperform NWP. For example, when the daily clearness index was 0.5, the annual nRMSE for persistence was about 125% while it was about 150% for the WRF model (Lara-

Fanego et al., 2012b).

The results presented in Table 3-2 from Lara-Fanego et al. (2012a), Lara-Fanego et al. (2012b),

Wittmann et al. (2008), and González et al. (2010) are for all sky conditions. They suggest that in any sky condition, the DNI forecast nRMSE for 1-day ahead forecasts would be about 60% and for 2-days ahead forecasts it would be between 40% and 60% depending on the NWP model used. The results from Lara-Fanego et al. (2012b) suggest that DNI forecast accuracy may increase slightly when forecast horizon is extended from 1 to 3 days ahead.

Only clear sky conditions were used by Breitkreuz et al. (2009) and Ruiz-Arias et al. (2012) which are shown in Table 3-2. Results from clear sky conditions only were also reported by

Lara-Fanego et al. (2012b) and Wittmann et al. (2008). Lara-Fanego et al. (2012b) showed 1- day ahead clear sky DNI forecast nRMSE of 21-22% whereas Ruiz-Arias et al. (2012) showed a range of 6-20% depending on the availability of aerosol data to input. The lower nRMSE was achieved by using aerosol data.

Table 3-2: Summary of DNI forecast accuracy from NWP

Forecast nRMSE RMSE nMAE MAE nMBE MBE Author(s) NWP description horizon (%) (W/m2) (%) (W/m2) (%) (W/m2) 30.22 1 hour 23.21 ECMWF - use a Gala et al. 65.92 CSM to derive 3 hours (2013) DNI 45.30 243.09 1 day 205.03 61 304 22 109 WRF initialised 1 day with GFS - use 60 294 35 172 Lara-Fanego et satellite 62 311 24 117 2 days al. (2012b) measured aerosol 58 290 36 178 and ozone to 63 319 25 119 derive DNI 3 days 62 305 38 182 WRF initialised Lara-Fanego et with ECMWF - 1 day al. (2012a) use a RTM to derive DNI 60 35 WRF initialised 18.5 143 17.3 134 with ECMWF - no 1 day 19.3 163 18.8 159 aerosol input 12.8 113 12.2 108 8.5 66 7.9 61 Ruiz-Arias et al. WRF initialised 9.7 82 9.3 79 (2012) with ECMWF - 7 62 6.6 58 aerosol input 1 day from satellite 7 54 6.2 48 measurements 7.6 64 7.1 60 5.9 52 5.4 48 AFSOL 2 days 18.8 96 11.2 57 ECMWF - use Breitkreuz et al. Skarveit and (2009) 2 days Olseth CSM to derive DNI 31.2 159 -26 -134

AFSOL 2 days 47 208.6 15.6 69.4 Wittmann et al. ECMWF - use (2008) Skarveit and Olseth CSM to derive DNI 2 days 41.7 184.9 -23 -103.2 WRF initialised González et al. with GFS - use Not (2010) MRM to derive stated DNI 18.23 87.28

For 2-days ahead clear sky DNI forecast nRMSE, Wittmann et al. (2008) reported 17.4% for the

AFSOL and 28.6% for the ECMWF which is similar to the results of Breitkreuz et al. (2009) which were 18.8% and 31.2% respectively. Lara-Fanego et al. (2012b) reported 22-24% for 2- days ahead clear sky DNI forecast nRMSE. The errors reported by Wittmann et al. (2008) are slightly smaller but less data were used for testing so Lara-Fanego et al. (2012b) may be considered more representative of expected forecast error.

Seasonal separation of results was done by Lara-Fanego et al. (2012b) and González et al.

(2010). Both showed that the WRF initialised with GFS boundary conditions would have better

DNI forecast accuracy in summer compared to the annual average forecast accuracy while the other seasons were likely be worse. This could be explained by cloudy conditions being less common during summer, thus forecasts for summer are generally more accurate (Lara-Fanego et al., 2012b). The nRMSE for 1-3 days ahead DNI forecasts was 33-48% in summer, 66-102% in autumn, 69-86% in winter, and 53-59% in spring (Lara-Fanego et al., 2012b). The nMAE was

13.11% in summer, 19.4% in autumn, 19.82% in winter, and 20.63% in spring (González et al.,

2010).

3.2.2. Improving forecast accuracy

Breitkreuz et al. (2009) suggested that the ECMWF should use AOD values that are more accurate than those in climatology databases whereas the AFSOL should use better cloud modelling to increase DNI forecast accuracy. As part of the Monitoring Atmospheric

Composition and Climate (MACC) project, AOD forecasts were integrated into the ECMWF

(ECMWF/MACC) by adding a chemical weather prediction suite. Using data from the Aerosol

Robotic Network (AERONET), it was found that hourly AOD forecasts from the ECMWF/MACC were more accurate or equal to a 2-day persistence method in Europe and America but not in eastern Asia and western Africa (Schroedter-Homscheidt et al., 2013). This was explained by the ECMWF/MACC emissions database lacking data about the rapid industrialisation in Asia

48 and Africa. Ruiz-Arias et al. (2012) demonstrated that additional aerosol data input to WRF initialised with ECMWF boundary conditions increased the clear sky DNI forecast accuracy. As shown in Table 3-2, the nMBE and nRMSE were almost halved when aerosol data were used.

Mathiesen et al. (2013) assimilated cloud cover in the initial conditions of a high-resolution, direct-cloud-assimilating WRF (WRF-CLDDA) by altering the water vapour mixing ratio. Testing showed the WRF-CLDDA had better GHI forecast accuracy than the NAM during overcast or partially cloudy conditions but not during clear sky conditions. It was suspected that this was due to WRF-CLDDA failing to model cloud dissipation accurately. The improvement in GHI forecast accuracy may also be achieved by DNI forecast accuracy.

The spatial resolution of a NWP model affects its ability to resolve the presence of clouds

(Diagne et al., 2013). It may be expected that higher resolution would improve the ability to resolve clouds and thus increase forecast accuracy. Lara-Fanego et al. (2012a) found that higher spatial resolution did not improve DNI forecast accuracy from the WRF when sky conditions were cloudy or overcast, but there was an improvement for clear sky conditions. An improvement in WRF clear sky global irradiance forecast accuracy was observed by Ruiz-Arias et al. (2011) when spatial resolution was increased. This suggests that higher spatial resolution improve the ability to resolve the effects of topography on clear sky solar irradiance but do not improve the ability to resolve the presence of clouds (Lara-Fanego et al., 2012a). Mathiesen et al. (2013) reported that increasing resolution usually reduces forecast accuracy and noted that spatial averaging can increase forecast accuracy by reducing random error. Lara-Fanego et al.

(2012a) showed that spatial averaging could significantly reduce 2-day ahead DNI forecast nRMSE and nMBE by the WRF, particularly during cloudy conditions. The best results were obtained for a horizontal resolution of 27 km and spatial averaging over 100 km x 100 km, which is similar to the findings of Lorenz et al. (2009) for GHI forecasts using the ECMWF.

3.3. Time Series Analysis

Time series analysis (TSA) methods produce forecasts by statistically analysing trends in historical time series of either the forecast variable or other variables, named exogenous variables, which influence the forecast variable. Clearness index and relative air mass were identified as the most relevant exogenous variables for predicting DNI (López et al., 2005).

Solar irradiance data that have been collected for a long period of time relative to the temporal resolution, such as over several days at 1 minute resolution or several months at 1 hour resolution, may be used to produce forecasts using TSA methods.

The simplest TSA model is the persistence model which assumes that conditions in the future will be exactly the same as those in a previous time. Forecasting models are often compared against the persistence model to demonstrate better forecast accuracy and thus justify the higher complexity. Regressive TSA models analyse recent time series behaviour to produce forecasts. Regression models are usually divided into autoregressive (AR) models, moving average (MA) models, or a combination of both terms in either autoregressive moving average

(ARMA) models or autoregressive integrated moving average (ARIMA) models. A description of the differences between these models and the time series behaviour suited to each one is in the text by Box et al. (1994). Artificial intelligence (AI) TSA models are able to identify recurring patterns in the historical time series and use the information to produce better forecasts.

These models may be created using either an artificial neural network (ANN) or a support vector machine (SVM). An ANN model may adopt one of several different structures defined by the number of layers or the presence of feedback loops. General use of AI models is detailed for ANN by Haykin (1999) and for SVM by Vapnik (1998).

ANN models are usually favoured over ARMA-based models because ANN models can be directly applied to nonlinear time series and can be made without detailed knowledge or assumptions regarding the behaviour of the time series (Tymvios et al., 2008; Zhang et al.,

1998). It has been speculated that ANN models are more suited than ARMA-based models to forecast solar irradiance time series because of non-linearity in the time series (Ji and Chee,

2011; Sfetsos and Coonick, 2000), however evidence to show non-linearity in the deseasonalised solar irradiance time series has not been provided. In addition, ANN models have the ability to adapt by changing the magnitudes of weights and biases which optimizes them to model new changes in a time series (Haykin, 1999). A review of 32 studies that used

ANN to forecast GHI or diffuse irradiance conducted by Mellit (2008) concluded that ANN can be generalized and applied to any location even if the location has incomplete data.

A disadvantage in using ANN models is that the model training process may result in a poor

ANN model which is more likely to occur if the model structure is complicated or the data have a lot of noise (Reikard, 2009; Mellit et al., 2013). To overcome this problem, the available time series data may be separated into a training set, a validation set and a testing set (Tymvios et al., 2008). The training set and validation set are jointly used during the training process to ensure that the ANN model ignores most of the noise. After the training process is completed, the testing set is used to evaluate the ANN model. This allows differently structured ANN models to be compared against each other to determine which has the best performance.

However, this solution requires significantly more data to be available.

3.3.1. TSA forecast accuracy

As shown in Table 3-3, there are few results to use for drawing conclusions about DNI forecast accuracy using TSA methods. All of the studies used ANN with feed-forward structure as part of their analysis. Mellit et al. (2010) showed that the feed-forward ANN had a higher correlation coefficient than an adaptive alpha model. Mishra et al. (2008) showed that the feed-forward structure had lower nRMSE than the radial basis function structure. The nRMSE results from Alam et al. (2006) fall in the same range for a feed-forward ANN. Neither study specified what value was used to normalise the RMSE.

Table 3-3: Summary of DNI forecast accuracy from TSA methods

TSA Forecast nRMSE RMSE nMBE MBE Correlation Author(s) description horizon (%) (W/m2) (%) (W/m2) coefficient

Feed- 1 hour Mellit et forward ANN 0.9834 al. (2010) Adaptive 1 hour alpha model 2.74 0.751 0.9754 Feed- 0.8 Not stated forward ANN Mishra et 5.4 al. (2008) Radial basis 7 function Not stated ANN 29 2.75 -0.55 2.13 -0.76 Alam et al. Feed- Not stated 2.79 -1.28 (2006) forward ANN 1.85 -0.26 1.65 -0.67

Reported trends in GHI forecasting from TSA methods may be used to estimate the relative

performances of TSA methods in DNI forecasting. A comparative study by Reikard (2009)

investigated GHI forecasting accuracy from a regression model, a regression model with added

sinusoids, an ARIMA model without exogenous inputs, an ARIMA model with exogenous inputs,

an ANN model and a hybrid ARIMA-ANN. The ARIMA model with no exogenous inputs was

best at 1 to 4-hour forecast horizons with mean absolute percentage error (MAPE) ranging

from 35.2% to 54.0% (Reikard, 2009). The ARIMA model with exogenous inputs and the

ARIMA-ANN model had very similar performance with MAPE ranging from 35.3% to 54.4% and

35.3% to 54.1% respectively (Reikard, 2009). The performances of these models were better

than the ANN model which had MAPE ranging from 41.4% to 56.2%. This was attributed to the

ARIMA model being best at following the diurnal solar irradiance cycle by using 24 hour

differencing and at modelling nonlinear variability by using time varying coefficients. The same

three models also had the best forecast accuracy when sub-hourly forecast horizons of 5

minutes, 15 minutes, and 30 minutes were used (Reikard, 2009). This was explained by the

52 dominance of the diurnal cycle being replaced by sub-hourly patterns which these models were able to reproduce better than the other models (Reikard, 2009).

A comparison of GHI forecast accuracy by a least-squares SVM (LS-SVM) model and a radial basis function ANN model was made by Zeng and Qiao (2013). The average MAPE for the LS-

SVM model was about 22% to 33% for 1 to 3-hour ahead forecasts which was better than the average MAPE of about 24% to 35% for the ANN model (Zeng and Qiao, 2013). The lower forecast error of the LS-SVM was explained by its better ability to model nonlinear and time- varying time series.

The results from Reikard (2009) and Zeng and Qiao (2013) respectively suggest that an ARIMA model or a SVM model may be preferable to an ANN model to forecast DNI when the forecast horizon is less than 4 hours. However, each of the studies only evaluated a single ANN model.

There are different structures that may be used to make an ANN model (Haykin, 1999) and a review by Mellit (2008) did not find any one structure to be best at solar irradiance forecasting.

Therefore, future investigations in DNI forecast accuracy using TSA methods should consider

ARIMA models, SVM models and ANN models.

3.4. Cloud Motion Vectors

The cloud motion vector (CMV) method uses satellite or ground-based imaging instruments to remotely track the motion of clouds. Fujita (1968) pioneered the CMV method by successfully tracking cloud motion from geosynchronous satellite pictures. He also developed a stereoscopic technique to track cloud features using ground-based sky cameras (Bradbury and

Fujita, 1968). Most of the pioneering work and the operational status of cloud tracking with satellite imagery was documented by Menzel (2001).

The process for deriving CMVs from satellite images begins with using known features which remain stationary within a tolerance limit between successive images to correct for consistency (Menzel and Purdom, 1994). Next, a tracer is selected within the target image.

Tracers usually correspond to cloud features that are easily traceable, such as locations with highest pixel brightness and the computed local gradients around these locations. Tracers must be consistent and all prospective tracers are checked by spatial coherence analysis checks for consistency (Coakley and Bretherton, 1982). After that, the tracers are assigned an altitude based on the infrared window sampling method (Fritz and Winston, 1962), the CO2- slicing method (Menzel et al., 1983) or the H2O-intercept method (Schmetz et al., 1993). Finally, the tracers are tracked using a pattern matching algorithm which matches a feature from the target area in one image to a search area in the following image.

Three basic pattern matching procedures used for tracking cloud motion are cross-correlation, the sum of absolute differences, and the sum of squared differences (Marcello et al., 2009).

The most commonly used algorithm is the maximum cross-correlation technique (Leese et al.,

1970; Leese et al., 1971; Merril, 1989). This algorithm determines the maximum cross- correlation coefficient from successive images to calculate the displacement vector of a chosen feature. It performs well for single clouds with unique patterns but fails to discriminate between motions in multiple cloud layers. Fourier transforms can increase the algorithm computation speed.

Another technique is the automatic cloud tracking algorithm which applies the standard pattern correlation coefficient to images from geostationary satellites (Schmetz et al., 1993).

This technique allows correlation of more displacements due to the segmentation of target and search area. It has faster computation speed and detects more features. Errors in pattern matching can produce poor results and therefore quality control is performed using automated editing algorithms (Hayden and Velden, 1991; Hayden and Nieman, 1996), manual

54 checks, and comparative climatology from radiosondes (Schmetz et al., 1993). The final CMVs are reported as averages of two to three component vectors calculated from a sequence of images (Nieman et al., 1997; Schmetz et al., 1993).

Other pattern matching methods include correlation-relaxation labelling (Evans, 2006),

Bayesian estimation (Konrad and Dubois, 1992), neural-network (Cote and Tatnall, 1995), functional-analytic (Bannehr et al., 1996), hierarchical method using simulated annealing and fuzzy reasoning (Mandal et al., 2005) and scale space classification (Mukherjee and Acton,

2002) methods. Most of these techniques show improvement in cloud classification but are computationally expensive for calculating CMVs.

Solar irradiance forecasting can use CMVs to describe advection of present clouds and predict images of future cloud cover. In this method, cloud images are converted to solar irradiance forecasts using statistical or physical models. Statistical models are created from regression of satellite measurements and corresponding ground measured solar irradiance for a given area.

These models are computationally efficient and require neither meteorological data nor calibration, but lack generality due to shortage of ground-based measurements (Noia et al.,

1993a). Physical models use radiative transfer equations that describe transmission, absorption and scattering in the earth-atmosphere system to reduce clear sky irradiance (Noia et al., 1993b). These models do not need ground-based measurements and are frequently applicable depending on available meteorological data. The HELIOSAT method (Hammer et al.,

2003) is most commonly used for short-term global irradiance forecasts using satellite-based

CMVs (Hammer et al., 1999; Heinemann et al., 2006a; b; Lorenz et al., 2004).

Ground-based imaging instruments to forecast solar irradiance using cloud cover indices

(Marquez et al., 2013a; Yang et al., 2012; Crispim et al., 2008), automated cloud classifications

(Tapakis and Charalambides, 2013), and CMVs (Bosch et al., 2013; Chow et al., 2011; Marquez and Coimbra, 2013; Stefferud et al., 2012) were found to work well for specific sites and lack 55 spatial coverage. In contrast, satellite images have greater spatial and temporal resolution that ensures cloud structures are easily detectable (Zelenka et al., 1999).

CMVs are used for weather forecasting by several meteorological centres, including the

European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) (Holmlund et al., 2010), the National Oceanographic and Atmospheric Administration’s Satellite and

Information Service (NOAA/NESDIS) (Daniels et al., 2004), the Japan Meteorology Agency (JMA)

(Oyama and Shimoji, 2008) and the Australia Bureau of Meteorology (BOM) (Le Marshall et al.,

2012).

3.4.1. CMV forecast accuracy

The DNI forecast accuracy from CMVs is summarised in Table 3-4. Only the 5 minute and 10 minute forecast results from Marquez and Coimbra (2013) and Quesada-Ruiz et al. (2014) are included to compare against the results of Chu et al. (2013). Marquez and Coimbra (2013) evaluated DNI forecast accuracy from ground-based CMV using forecast horizons from 3 minutes to 15 minutes and found CMVs to outperform persistence at all forecast horizons. The method used by Marquez and Coimbra (2013) divided sky images into sectors with fixed sizes.

This method was modified by Quesada-Ruiz et al. (2014) to vary the size of sectors to increase the accuracy of estimating DNI for each sector. Both methods were tested and the RMSE results showed that the modified method performed better than the original method in all sky conditions and forecast horizons. The values in Table 3-4 were approximated from a graph and the lower RMSE is for low variability clear sky conditions whereas the higher RMSE is for high variability broken cloud conditions. Chu et al. (2013) used ground-based CMV to forecast DNI at 5 minute and 10 minute horizons.

Results from Chu et al. (2013) and Quesada-Ruiz et al. (2014) found that ground-based CMV forecasts of DNI up to 10-minutes ahead achieved better forecast accuracy than persistence when sky conditions had high variability and worse forecast accuracy than persistence when 56

sky conditions had low variability. This suggests that sky conditions should be considered

together with forecast horizon when choosing an appropriate DNI forecasting method, at least

for very short-term forecasts up to 10-minutes ahead.

Table 3-4: Summary of DNI forecast accuracy from CMV

Author(s) CMV type Forecast horizon RMSE (W/m2) MBE (W/m2) 226 208 5 minutes 267 257 341 Marquez and Coimbra 317 Ground-based 10 minutes (2013) 303 283 360 392 15 minutes 332 299 5 minutes 150.5 -10.7 (Chu et al., 2013) Ground-based 10 minutes 144 -8.3 ~10 Ground-based, 5 minutes ~290 variable sector ~10 size 10 minutes Quesada-Ruiz et al. ~300 (2014) ~11 5 minutes Ground based, ~370 fixed sector size ~11 10 minutes ~350

The RMSE values reported by Chu et al. (2013) are lower most likely because their testing

period of 15 September 2011 to 15 October 2011 contained days with both low and high

variability in sky conditions whereas Marquez and Coimbra (2013) used 4 days with only highly

variable sky conditions for testing. It is easier to forecast DNI when sky conditions have low

variability compared to when sky conditions have high variability. This is reflected in the results

from Quesada-Ruiz et al. (2014) which show RMSE to be an order of magnitude lower in clear

57 sky conditions with low variability compared to broken cloud conditions with high variability.

Marquez and Coimbra (2013) also acknowledged that uniform sky conditions reduce error averages.

In the absence of papers published about the use of satellite-based CMVs for DNI forecasting, results of GHI forecasting from satellite-based CMVs will be used to indicate potential forecast accuracy. Hammer et al. (2001) found that the nRMSE of 1-hour ahead GHI forecasts in north

Germany was about 28-35% depending on the number of pixels in the satellite image used to produce the forecast. Perez et al. (2010a) found that 1 hour ahead GHI forecast RMSE from satellite-based CMV had a range of 68-120 W/m2 across 7 locations in the US. A slightly lower

RMSE of about 60 W/m2 averaged from 274 locations in Germany was reported by Kühnert et al. (2013). RMSE was calculated for up to 5-hours ahead forecasts and the 5-hours ahead RMSE values were again close: 116-175 W/m2 (Perez et al., 2010a) and about 110 W/m2 (Kühnert et al., 2013). Both studies also showed that 1 to 5-hours ahead GHI forecasts from satellite-based

CMV had lower RMSE than persistence. In 1-hour ahead forecasts the RMSE of satellite-based

CMV up to 14 W/m2 lower than that of persistence and in 5-hours ahead forecasts the RMSE of satellite-based CMV was up to 40 W/m2 lower.

3.4.2. Improving forecast accuracy

Increasing solar irradiance forecast accuracy from ground and satellite imagery may be achieved by minimising the errors associated with parameters that influence the number and quality of CMVs. Such parameters include the size of the target and search windows (Lunnon and Lowe, 1992; Wade et al., 1992), the time interval between successive images (Purdom,

1996), the position of the vector on the earth disk and rotational motion (Kamachi, 1989).

Another crucial factor is the pattern matching algorithm used for cloud tracking. Area-based approaches, such as cross-correlation, compare and match the intensity patterns of the block- wise areas and are unable to accurately detect when the shape of a cloud changes. Local- 58 feature approaches outperform correlation methods due to enhanced sensitivity of changes in shape and brightness. However, local features are non-uniform and biased in spatial location.

Huang et al. (2012) proposed a robust hybrid approach using both area and feature-based techniques with ground-based sky imagers. This approach performed better for variable intensity patterns, multilayer clouds and different cloud shapes.

Forecast accuracy may also be increased by improving solar irradiance models. Better forecasts from the HELIOSAT-2 model (Rigollier et al., 2004) may be obtained by including cloud index median and air mass (Polo et al., 2011; Zarzalejo et al., 2009), improving the algorithm for calculating the cloud index (Dagestad and Olseth, 2007), and enhancing clear sky modules

(Mueller et al., 2004).

3.4.3. Converting satellite image to solar irradiance

The process of converting satellite images to solar irradiance contributes to the overall forecast error from satellite-based CMVs (Kühnert et al., 2013). Thus the satellite-to-irradiance conversion accuracy for DNI is reviewed. The satellite-to-irradiance methods reviewed are briefly described in the paragraph below and their conversion accuracies are summarised in

Table 3-5.

The DLR-SOLEMI method is described in detail by Schillings et al. (2004a). It uses visible and infrared images from Meteosat-7 to derive cloud index for input to a CSM and accounts for ozone, aerosols and water vapour. The Perez method (Perez et al., 2002) processes GOES 8

(now GOES-East) visible images for cloud cover as input to CSM and accounts for snow cover, aerosol, ozone and water vapour. The HELIOSAT-MSG method (Hammer et al., 2009) is a combination of the HELIOSAT method (Hammer et al., 2003), the SOLIS clear sky model, and an exponential parameterisation for direct irradiance. The Eissa method (Eissa et al., 2013) uses 6 thermal channels of the Meteosat Second Generation (MSG) to detect dust, clouds, and water vapour. These are input to an ANN ensemble along with solar zenith angle, solar time, day 59

number and eccentricity correction to derive atmospheric optical depth. The atmospheric

optical depth is input to the Beer-Bouguer-Lambert law to calculate DNI. The State University

of New York GOES satellite-based solar model (SUNY) version 1 method was developed from

work by Perez et al. (Perez et al., 2002; Perez et al., 2004). It operates similarly to the Perez

method with 3 major differences. Firstly, visible images from GOES-West are used in addition

to those from GOES-East. Secondly, the solar irradiance time series is calibrated. Thirdly,

anomalous pixel-to-pixel variation is corrected. The SUNY version 3 method is upgraded from

SUNY version 1 to use infrared images from GOES-West and GOES-East. A detailed description

of the upgrade is provided by Perez et al. (2010b). In brief, the infrared images are used to

obtain empirical thresholds for deriving solar irradiance over snow covered regions.

Table 3-5: Summary of performance of methods which convert satellite images to DNI

Method nRMSE RMSE nMAE MAE nMBE MBE Author(s) description (%) (W/m2) (%) (W/m2) (%) (W/m2) 21.67 137.9 10.5 97.7 10.5 66.8 Nonnenmacher 30.29 151.2 14.21 108.3 14.21 71 SUNY version 3 et al. (2014) 29.79 119.9 8.75 75.1 8.75 35.2 42.24 161.9 -6.39 92.4 -6.39 -24.5 Eissa et al. Eissa method (2013) 26.1 -6 Djebbar et al. SUNY version 1 82.3 164.3 2.5 4.6 (2012) SUNY version 3 67.2 133.7 14.3 28.5 Hammer et al. HELIOSAT-MSG (2009) method 31.3 1.1 Vignola et al. Perez method (2007) 40.9 200 2 Schillings et al. DLR-SOLEMI (2004a) method 36.1

Polo et al. (2008) reviewed most of the common methods to predict solar irradiance from

satellite images and concluded that the nRMSE of satellite-to-GHI conversion has a range of

17-25%. It is seen in Table 3-5 that the nRMSE range for satellite-to-DNI conversion is about

21.7-82.3%. Only the Eissa method and SUNY version 3 are able to match satellite-to-GHI 60 conversion accuracy. Both of these methods were evaluated in the most recently published papers in Table 3-5, suggesting that the latest satellite-to-DNI conversion methods may enable

DNI forecast accuracy from satellite-based CMVs to be similar to GHI forecast accuracy in terms of nRMSE.

3.5. Hybrid Forecasting Methods

There are advantages and disadvantages associated with each individual forecasting method.

Besides improving the forecast method algorithm, poor forecast accuracy by a method may be overcome by using another method that does not share the same disadvantages. This combination forms a hybrid method that may achieve higher forecast accuracy than either method could achieve alone. Although widely studied for GHI, the design and evaluation of hybrid methods is sparsely studied for DNI. The results of studies of hybrid GHI forecasting methods are included to indicate the possible effectiveness of hybrid DNI forecasting methods.

3.5.1. Hybrid DNI forecasting methods

A summary of hybrid methods applied to forecasting DNI and their reported accuracies are presented in Table 3-6. One common hybrid method is using Model Output Statistics (MOS), which is the use of statistical methods to post-process NWP outputs for the purpose of calculating outputs not directly provided by the NWP (e.g. deriving DNI) or correcting systematic biases in the output by comparing NWP output against historical ground measurements (Heinemann et al., 2006a). MOS has been shown to improve GHI forecast accuracy by the WRF, MASS and ARPS mesoscale NWP models in the US (Perez et al., 2013;

Perez et al., 2011). However, the DNI forecast accuracy of the ECMWF without MOS has been found to be more accurate in Spain compared to commercial MOS systems (Schroedter-

Homscheidt et al., 2012). NWP with MOS was shown to produce more accurate DNI forecasts than persistence for 2-days ahead forecasts (Kraas et al., 2013). Gerstmaier et al. (2012) 61 compared DNI forecast accuracy from 4 different forecast providers that used NWP with MOS.

The RMSE and MBE results from Gerstmaier et al. (2012) were mostly larger than those from

Kraas et al. (2013). This may be explained by the preference of forecast providers to use probability of overestimation or underestimation to describe their forecast accuracy

(Gerstmaier et al., 2012) and thus the MOS was not designed to minimise RMSE and MBE.

Methods other than MOS can be used for post-processing NWP outputs. The National Digital

Forecast Database (NDFD) is an NWP based on the GFS (Perez et al., 2010a). Meteorological data from the NDFD were post-processed by ANN to forecast DNI and shown to outperform persistence (Marquez and Coimbra, 2011) for 1-day ahead forecasts. For 2-day ahead to 6-day ahead forecasts, only nRMSE was calculated and results showed that nRMSE increased with longer forecast horizon (Marquez and Coimbra, 2011). For 2-days ahead forecasts, nRMSE was between 35% and 45% which is less than that from Kraas et al. (2013). When compared against the 1 to 3-days ahead forecast accuracy of an NWP without post-processing in all seasons and sky conditions (Lara-Fanego et al., 2012b), the results from Kraas et al. (2013) do not show any significant forecast accuracy improvement whereas those from Marquez and Coimbra (2011) do.

González et al. (2010) used ANN to post-process WRF outputs and results suggest that forecast accuracy is good but comparisons are difficult to make because RMSE and MBE were not calculated and the forecast horizon was not stated. The study showed that ANN post- processing improved DNI forecast accuracy in both all sky conditions and seasons compared to

WRF forecasts without post-processing (González et al., 2010).

Table 3-6: Summary of DNI forecast accuracy from hybrid methods

Method Forecast nRMSE RMSE nMAE MAE nMBE MBE Author(s) description horizon (%) (W/m2) (%) (W/m2) (%) (W/m2) 5 minutes 88.6 -0.2 Chu et al. CMV input to 10 (2013) ANN minutes 103.3 -1.9 ECMWF with 28.00 1 hour SVR post- 22.77 processed at 3 59.21 hour resolution 3 hour 44.15 then interpolated to 203.89 1 hour 1 day resolution using Gala et al. CSM 179.79 (2013) ECMWF with 27.71 1 hour SVR post- 22.57 processed at 58.76 daily resolution 3 hour 43.89 than interpolated to 218.05 1 hour 1 day resolution using CSM 176.02 ECMWF and 56 257 0.2 Kraas et al. HIRLAM with 2 days 77 347 -33 (2013) MOS post- processing 58 266 -38 Marquez 32.3 162 -0.5 -2.4 and NDFD with ANN 31.2 156 -1.3 -3.4 1 day Coimbra post-processing 33.2 161 -10.3 -7.1 (2011) 32 158 -8.3 -4.2 363 153 Gerstmaier NWP with MOS 312 -32 1 day et al. (2012) post-processing 352 136 316 104 González et WRF with ANN Not al. (2010) post-processing stated 16.61 79.52

SVM was used by Gala et al. (2013) to post-process outputs from the ECMWF at temporal

resolutions of either 3 hours or 1 day to produce hourly direct horizontal irradiance forecasts.

The difference in forecast accuracy from varying the ECMWF output temporal resolution was

insignificant. However, SVM post-processing did produce lower MAE compared to ECMWF 63 forecasts without post-processing (Gala et al., 2013). A greater reduction in MAE may be achieved if the algorithm used to derive direct horizontal irradiance from GHI output was optimized to the location (Gala et al., 2013). Again, the lack of reporting results in RMSE and

MBE makes comparisons against other forecast results difficult.

Post-processing can also be applied to non-NWP methods. The use of ANN to post-process ground-based CMV outputs was evaluated by Chu et al. (2013) for 5-minutes and 10-minutes ahead forecasts. The hybrid model outperformed persistence and CMV without post- processing during periods of very low to intermediate variability in sky conditions, but CMV without post-processing was most accurate for periods of high variability (Chu et al., 2013).

3.5.2. Other hybrid forecasting methods applied to global irradiance

A successful hybrid method that improved GHI forecast accuracy and may improve DNI forecast accuracy is ensemble modelling, i.e. averaging the results obtained from more than one NWP model initialised with the same inputs. Each NWP model may produce a slightly different forecast from the same inputs because of variation in model design. This is true for

DNI forecasting as shown by Breitkreuz et al. (2009) for the AFSOL and the ECMWF, and by

Gerstmaier et al. (2012) for a group of 4 NWP with MOS. The different forecast information from each NWP can be averaged to obtain an overall more accurate forecast, as shown by

Perez et al. (2013) for GHI forecasting. The success of ensemble modelling depends on the extent to which forecast errors of each NWP offset one another (Perez et al., 2013).

Another approach using multiple forecast methods would be to apply each forecasting method to conditions for which it is most accurate. Forecast conditions were separated into clear sky for ARMA modelling and cloudy sky conditions for ANN modelling in the study by Voyant et al.

(2013). Comparisons against solo ARMA and solo ANN models showed an improvement of 0.9% to 3% in forecast accuracy which was considered insufficient to justify the hybrid model’s complexity (Voyant et al., 2013). Further improvement may be achieved by using geographical 64 features and forecast data as inputs instead of only historical time series data (Voyant et al.,

2013).

Alternatively, the different methods may be used to model the linear and non-linear components in solar irradiance time series data separately before being combined to make the forecast. Ji and Chee (2011) demonstrated that using ARMA to model the linear component and time delay neural network (TDNN) to model the nonlinear component would produce more accurate forecasts than either an individual ARMA model or TDNN model for most of the time.

Benmouiza and Cheknane (2013) produced hourly GHI forecasts by using a hybrid of a MLP

ANN and a k-means clustering method that divided data into groups for pre-processing before input to the ANN. The hybrid method was better when compared against an ARMA model

(Benmouiza and Cheknane, 2013), but there was no comparison against a solo MLP ANN. The latter comparison would have been useful for indicating any improvement in forecast accuracy caused by adding the clustering method.

Some studies have proposed combining satellite-based CMVs with other methods to forecast

GHI. Miller et al. (2012) proposed combining geostationary satellite-based cloud images with wind field data from the GFS to forecast the motion of clouds up to 3 hours ahead. The cloud image forecasts could be made from using individual pixels or from a group of pixels. Neither approach had its forecast accuracy evaluated. Lorenz et al. (2012) combined GHI forecasts from satellite-based CMVs and two NWP models in a linear regression model to produce forecasts up to 6 hours ahead. Using data from 290 weather stations in Germany, results showed that the hybrid model had the lowest RMSE for all forecasts between 1 and 6 hours ahead. At the 3 hour ahead forecast horizon, the hybrid model performed 40% better than using only the ECMWF and 27% better than using only satellite-based CMVs. Marquez et al.

(2013b) proposed using satellite-based CMVs and ground measurements as input to an ANN 65 model to forecast GHI up to 2 hours ahead. Using data from a period of mostly cloudy days in the US, the RMSE of the hybrid method was shown to be 50%-80% lower than that of a persistence model.

3.6. Chapter Summary

This chapter reviewed the reported accuracy of different DNI forecasting methods at various forecast horizons and in different sky conditions. The review allows DNI forecast accuracy to be summarised by identifying the current best DNI forecast accuracy of each forecast method for appropriate forecast horizons. This result is presented in Table 3-7. Different metrics are used to describe the forecast accuracy in Table 3-7 because not all studies used the same metrics as shown by the gaps in Table 3-2, Table 3-3, Table 3-4 , and Table 3-6. The nRMSE was the most common forecast accuracy metric reported for NWP models whereas for ground- based CMV methods it was RMSE. The DNI forecasting methods most accurate for each forecast horizon agree with the forecasting methods suggested in Table 3-1.

The robustness of this summary is limited by the use of different metrics in different studies because it causes difficulty in fairly comparing all the results from different studies. Hence, not all studies reviewed are able to be considered for inclusion in the summary. This could be solved by adopting a uniform set of metrics to evaluate forecast accuracy. A range of metrics were recommended by Hoff et al. (2013) and Zhang et al. (2015a), and the only metrics both recommendations had in common were RMSE and MBE. However, the MBE of forecasts can mostly be corrected by post-processing and the RMSE is criticised for being unable to relate the spread of forecast error to the variability of the test data (Coimbra et al., 2013). To replace them, a time horizon invariant (THI) metric calculated from forecast accuracy uncertainty and test data variability is proposed (Coimbra et al., 2013). Describing DNI forecast performance in terms of test data variability would also allow the results from different test sets to be 66 compared more fairly by accounting for different distribution of weather conditions in the testing data of different studies. This is important because all DNI forecasting methods have higher forecast accuracy in clear sky conditions compared to cloudy conditions, so a test data set with a high proportion of clear sky days will show higher average DNI forecast accuracy compared to a set with a high proportion of cloudy days.

Table 3-7: Summary of best DNI forecast accuracy for appropriate forecast horizons3

Forecast Best DNI forecast accuracy for given horizon Notes References method 5 m 10 m 1 h 1 d 2 d 3 d ~3 W/m2 ~3 W/m2 Clear sky conditions Persistence RMSE RMSE only Quesada-Ruiz et al. (2014) Ground- ~10-~370 ~10-~350 based CMV W/m2 RMSE W/m2 RMSE All sky conditions Quesada-Ruiz et al. (2014) Ground- based CMV 88.6 W/m2 103.3 W/m2 input to ANN RMSE RMSE All sky conditions Chu et al. (2013) 0.9834 Feed- correlation forward ANN coefficient All sky conditions Mellit et al. (2010) Satellite Nonnenmacher et al. (2014), Eissa image to DNI Depends on et al. (2013), Vignola et al. (2007) conversion 21.67%-42.24% nRMSE conversion process and Schillings et al. (2004b) 60%- 41.7%- 62%- All sky conditions Lara-Fanego et al. (2012a), Lara- 61% 62% 63% and depends of NWP Fanego et al. (2012b) and NWP nRMSE nRMSE nRMSE model Wittmann et al. (2008) 5.9%- 17.4%- Clear sky conditions Lara-Fanego et al. (2012b), Ruiz- 31.2% 31.2% only and depends on Arias et al. (2012), Breitkreuz et al. NWP nRMSE nRMSE NWP model (2009), Wittmann et al. (2008) All sky conditions NWP with 31.2%- 56%- and depends of NWP post- 33.2% 77% model and post- Marquez and Coimbra (2011) and processing nRMSE nRMSE processing method Kraas et al. (2013)

3 Note that RMSE is an imperfect comparison metric because the variability of meteorological conditions affects forecast accuracy. For example, RMSE is naturally lower on a clear sky day due to low DNI variability, and is higher on a cloudy day when DNI variability is high.

4. Study Method4

As mentioned in the introduction, the aim of this study is to evaluate the value of a CST power plant operated using DNI forecasts in the NEM. The value is determined by calculating the financial value and reliability achieved as a result of the CST plant’s operation. The CST plant’s operation in different scenarios is simulated by using a CST model. DNI data are required both for the simulations and for evaluating the accuracy of the forecast methods used. Electricity price data are needed for calculating the financial value. This chapter covers the DNI and electricity price data, the CST model, the procedures for calculating financial value and reliability, and the forecast methods used in this study.

4.1. Data

This study is set in the Australian National Electricity Market (NEM) because it is the largest electricity market in Australia. As described in Section 1.2, the NEM is divided into five regions and the electricity price in each region is determined independently. As such, one location is selected from each region for this study, except for TAS because it does not have favourable solar resource, transmission infrastructure, and distance to load for significant CST prospects

(Lovegrove et al., 2012).

4 Some of the material presented in this chapter has been previously published in: Law, E.W., Kay, M., Taylor, R.A., 2014, Assessing the economic benefit of forecasting concentrated solar thermal energy output, Solar2014: The 52nd Annual Conference of the Australian Solar Council, Australian Solar Council, Melbourne, Australia, p. 131-140

Law, E.W., Kay, M., Taylor, R.A., 2016, Calculating the financial value of a concentrated solar thermal plant operated using direct normal irradiance forecasts, Solar Energy, 125, p. 267-281 https://dx.doi.org/10.1016/j.solener.2015.12.031

4.1.1. DNI data

The best data for validating DNI forecasts are ground measured DNI. Several Australian Bureau of Meteorology (BOM) ground stations have measured DNI in the past and present. The locations of the closed and operating stations are shown in Figure 4-1.

Figure 4-1: Closed and open BOM stations for measuring one minute DNI (BOM, 2012a)

The DNI is measured at 1-second resolution using a pyrheliometer with a maximum field of view of 5° (BOM, 2012a). The measurements are quality controlled to ensure that there is 95% confidence of the measurement uncertainty being within 3% or 15 W/m2. Calculation of uncertainty follows the International Organisation for Standardisation/International

Electrotechnical Commission guidelines for expressing uncertainty in measurements. The DNI is archived at 1-minute resolution by BOM. This study averages the 1-minute DNI to 1-hour resolution over the previous hour because the CST model uses 1-hour time steps (more details in Section 0).

This study is set in the NEM, so stations in Western Australia (Broome, Learmonth, Geraldton, and Kalgoorlie-Boulder) and the Northern Territory (Darwin, Katherine, Tennant Creek, and

Alice Springs) are ignored. The SECVI network stations do not have publicly available data, and the Cobar, Longreach, Townsville, and Woomera stations only operated for approximately one year (BOM, 2016), so these stations are also ignored. Based on suitability among the remaining sites for a CST plant (Lovegrove et al., 2012), the sites chosen for this study are Mildura, VIC (-

34.25° N, 142.10° E), Mt. Gambier, SA (-37.75° N, 140.75° E), Rockhampton, QLD (-23.38° N,

150.48° E), and Wagga, NSW (-35.16° N, 147.45° E). The latest full year of measurements shared by these stations is 2005 hence the data for 2005 are used for this study.

Each station has gaps in ground measurements in the 2005 record as shown in Table 4-1.

These gaps are filled using satellite measurements which were obtained from BOM (BOM,

2012b). Satellite measurements for 1 January to 31 October in 2005 were made by the

Geostationary Operational Environmental Satellite (GOES-9) system operated by the US

National Oceanic and Atmospheric Administration (NOAA). Measurements for 1 November to

31 December in 2005 were made by the Multi-Functional Transport Satellite (MTSAT-1R) series operated by the Japanese Meteorological Agency (JMA). BOM obtained GHI from the satellite images by a method based on the two-band physical model (Weymouth and Le Marshall, 2001) and converted it to DNI using a modified form of the Boland-Ridley-Lauret model (Ridley et al.,

2010). Analysis by BOM showed that the satellite measurements have typical positional accuracy of 0.01°, mean bias difference (MBD) of -20 to +18 W/m2 and root mean square difference (RMSD) of around 210 W/m2, depending on location (BOM, 2012b). The satellite measurements will deviate from ground measurements because each satellite grid-point value covers a spatial resolution of 0.05° in latitude and longitude, and the satellite measurements were taken a number of minutes after the hour. The exact number of minutes depends on the latitude, the hour of measurement, and the satellite (BOM, 2012b). The information for the

relevant latitudes and for 2005 is presented in Table 4-2. The satellite grid-point and the exact

minutes past the hour for each site are presented in Table 4-3.

Table 4-1: Gaps in ground measurements for each site in 2005

Mildura Mt Gambier Rockhampton Wagga Number of hours with daylight 4650 4578 4687 4671 Number of gaps 410 136 33 25 Amount of gaps (%) 8.8 3.0 0.70 0.54 16 Jan 6am-10am 3-7 Feb 9 Feb 2pm-3pm 16 Feb 6am-9am 4-6 Jan 12 Feb 16 Feb 6am-9am 10 Jun 11am-3pm 9 Feb 4pm-7pm 16 Feb 7am-9am 19 Feb 6am-8am 11 Jun 8am-9am 16 Feb 7am-10am 25-28 Feb 12 Jun 12pm-5pm 25 Aug 7am-12pm Times 3 May 3pm-6pm 12 Nov 6am-8am 29 Sep 1pm-2pm 12 Sep 8am-9am 4 May 7am-11am 4 Oct 11am-1pm 15 Sep 3pm-5pm 6-31 Dec 8 Oct 22 Sep 9am-2pm 17 Dec 5am-11am 23 Sep 5pm-6pm 29 Sep 6am-8am

Table 4-2: Satellite measurements minutes past the hour in 2005 (BOM, 2012b)

Start date 1 July 2003 1 November 2005 End date 31 October 2005 30 June 2010 Latitude (°N) GOES-9 A GOES-9 B MTSAT-1R -20 42.0 30.0 48.3 -25 43.0 31.0 49.2 -30 43.9 31.9 50.1 -35 44.7 32.7 51.0 -40 45.5 33.5 51.7 A: UT hours 18, 19, 20, 21, 23, 00, 01, 02, 03, 05, 06, 07, 08, 09, 11 B: UT hours 22, 04, 10

Table 4-3: Satellite grid-point and minutes past the hour for each site

Actual Satellite grid-point Minutes past the hour Latitude Longitude Latitude Longitude GOES-9 GOES-9 MTSAT- Site (°N) (°E) (°N) (°E) A B 1R Mildura -34.25 142.10 -34.25 142.10 44.58 32.58 50.87 Mt. Gambier -37.75 140.75 -37.75 140.75 45.14 33.14 51.39 Rockhampton -23.38 150.48 -23.40 150.50 42.68 30.68 48.91 Wagga -35.16 147.45 -35.15 147.45 44.73 32.73 51.02

Bias correction is conducted to improve the fit between satellite DNI and ground DNI. The bias correction method is based on the multiple-linear regression method by Blanksby et al. (2013) for correcting bias in satellite measurements over Australia. The bias correction equation is shown in E. 4-1. The coefficients for each site are determined by applying the regression equation to the hours when ground and satellite measurements are available, excluding night time. The coefficients for each site are shown in Table 4-4. The bias correction of Mt. Gambier data produced some negative values, which were changed to zeros because DNI cannot be negative.

2 2 퐷푁퐼(푐표푟) = 푏1 ∗ 퐷푁퐼(푠푎푡) + 푏2 ∗ 푘푡(푠푎푡) + 푏3 ∗ 퐷푁퐼(푠푎푡) + 푏4 ∗ 푘푡(푠푎푡) E. 4-1

+ 푏5

where 퐷푁퐼(푐표푟) is the bias corrected DNI in W/m2, 퐷푁퐼(푠푎푡) is the satellite DNI in

2 W/m , 푘푡(푠푎푡) is the instantaneous clearness index calculated using satellite GHI, and

푏1 to 푏5 are equation coefficients.

Table 4-4: Bias correction equation coefficients for each site

Bias correction equation coefficients

Site 푏1 푏2 푏3 푏4 푏5 Mildura 0.7680 39.12 2.260E-04 -96.26 61.85 Mt Gambier 0.6017 78.58 4.034E-04 -312.4 85.45 Rockhampton 0.4683 -796.8 1.128E-05 1,588 148.1 Wagga 0.6902 -109.4 3.011E-04 -14.47 85.28

Figure 4-2: Scatterplots of satellite DNI against ground DNI before and after bias correction for each site 74

Scatterplots for each site to compare satellite DNI and corrected satellite DNI against ground

DNI are shown in Figure 4-2. The obvious change caused by bias correction is that the corrected satellite data no longer has zero DNI when ground DNI is non-zero, except for some instances at Mt. Gambier. The satellite DNI may be zero when ground DNI is non-zero because the satellite measurements occur 30-51 minutes after the hour as shown in Table 4-3. If sunset occurs after the hour but before the satellite measurement is made, then the ground measurements may show non-zero DNI while the satellite measures zero DNI. The bias correction reduces the magnitude of the mean bias difference (MBD) and root mean square bias difference (RMSD) at each site, as shown in Table 4-5. The MBD reduces to zero at all sites except at Mt. Gambier where it is reduced to near zero (0.315 W/m2). The RMSD reduces by

3.1%-9.0% for all sites. The RMSD values for corrected satellite DNI in this study are similar to the range of 143-204 W/m2 found by Blanksby et al. (2013).

Table 4-5: MBD and RMSD of satellite DNI at each site

Before correction After correction MBD RMSD MBD RMSD RMSD Site (W/m2) (W/m2) (W/m2) (W/m2) reduction (%) Mildura -21.8 160 0.000 155 -3.1 Mt Gambier 8.36 200 0.315 182 -9.0 Rockhampton -3.84 197 0.000 185 -6.1 Wagga 17.8 175 0.000 165 -5.7

Satellite measurements deviate from ground measurements, so filling gaps in the ground data with satellite measurements may introduce uncertainty. The uncertainty may be estimated by using E. 4-2 to calculate the weighted average deviation of the combined data. At one extreme, the data are all ground measurements with zero RMSD. At the other extreme, the data are all satellite measurements with RMSD for the respective location as per Table 4-5. The weightings to apply to the satellite RMSD for each location are given in the “Amount of gaps” row in Table

4-1. The weighted average deviation is 14 W/m2 for Mildura, 5.5 W/m2 for Mt. Gambier, 1.3

W/m2 for Rockhampton, and 0.99 W/m2 for Wagga. The 95% confidence of the ground measurement uncertainty is 15 W/m2 hence the uncertainty introduced by filling gaps with satellite measurements are negligible for Rockhampton and Wagga, relatively small for Mt.

Gambier, and almost equal to the ground measurement uncertainty for Mildura.

푊푒𝑖푔ℎ푡푒푑 푎푣푒푟푎푔푒 푑푒푣𝑖푎푡𝑖표푛 = 푃푠푎푡 ∗ 푅푀푆퐷푠푎푡 + (1 − 푃푠푎푡) ∗ 푅푀푆퐷푔푛푑 E. 4-2 where 푃푠푎푡 is the weighting to apply to the satellite measurements, 푅푀푆퐷푠푎푡 is the RMSD of satellite data, and 푅푀푆퐷푔푛푑 is the RMSD of ground data.

4.1.2. Electricity price

This study uses electricity price data archived by the Australian Energy Market Operator

(AEMO) for the year 2005 for the QLD, NSW, VIC and SA regions (AEMO, 2013b) so that the electricity price corresponds to the same period and region as the DNI. The electricity price is also averaged to 1-hour resolution. It is assumed that the presence of the CST plant does not significantly affect the electricity price. It is also assumed that the electricity price is known perfectly in advance to isolate the effect of DNI forecast accuracy. As regulations allow updated bids to be submitted up to 5 minutes before dispatch, generators can respond within

5 minutes to periods of high electricity prices by increasing output, so it is reasonable to assume that the CST plant has a perfect electricity price forecast.

A single electricity price is used for calculating revenue because the NEM only has one spot market instead of separate day ahead and real time markets with separate prices. The electricity price is only paid for electricity that the CST plant actually generates, not the amount bid into the market. If a CST plant in the NEM fails to meet a bid according to AEMO’s

“reasonable opinion” (AEMC, 2016a) then it must pay the reserve generation (RG) cost of

76 dispatching frequency control ancillary services (FCAS) to cover for it according to the “causer pays” approach (AEMO, 2010b). The RG cost is calculated based on the amount of RG capacity required to cover the CST plant’s generation shortfall, i.e. the difference between the generation quantity bid into the market and the amount of electricity actually supplied by the

CST plant. There are eight markets in the NEM for FCAS according to their correction type, their response time, and the magnitude of severity that they can correct. To simplify the process of determining the RG cost, this study assumes that another regular generator is dispatched to generate the electricity that the CST plant fails to supply. Consequently, the unit

RG cost is assumed to equal the electricity price in the same time step. This assumption means that the unit RG cost will be higher when demand is high and vice versa. This is because high demand requires more generation to be dispatched, so bids with higher prices will be dispatched and the electricity price increases. A high unit RG cost when demand is high is appropriate because more generators would be dispatched instead of being on standby to supply RG in case of lost generation during this time.

4.1.3. Site data comparison

The four sites in this study are located at different latitudes and in different NEM regions. The climate classification (BOM, 2014), the 1990-2011 average daily global irradiation (BOM, 2013), and the 2005 average daily global irradiation (BOM, 2012c) of each site are shown in Table 4-6.

The 1990-2011 and 2005 average daily global irradiation values were derived from NOAA and

JMA satellite data. Each site has a different climate, so the frequency and intensity of weather features at each site will be different. The average daily solar resource correlates with the site displacement from the equator. Rockhampton is closest to the equator (-23.38° N) and has the highest solar resource, whereas Mt. Gambier is the furthest (-37.75° N) and has the lowest solar resource. In 2005, the solar resource at Rockhampton and Wagga Wagga in 2005 was within 0.5% of the 1990-2011 average, whereas at Mildura and Mt. Gambier it was more than

0.5% lower. This suggests that it was cloudier than average in 2005 at Mildura and Mt.

Gambier.

Table 4-6: Site climate and site average daily global irradiation over 1990-2011 and 2005

Average daily global irradiation

2005 2005 Key 1990- difference to difference to climate Key climate sub- 2011 2005 1990-2011 1990-2011 Site group division class (MJ/m2) (MJ/m2) (MJ/m2) (%) Warm summer Mildura Grassland drought 18.793 18.568 -0.225 -1.20 No dry season, Mt. Gambier Temperate mild summer 15.569 15.234 -0.335 -2.15 Rockhampton Subtropical No dry season 20.157 20.073 -0.084 -0.42 Distinctly dry Wagga Temperate and hot summer 17.956 18.008 0.052 0.29

In contrast to the solar resource according to satellite GHI, the combined ground and corrected

satellite DNI data shows that Mildura has the highest DNI resource of the four sites, as can be

seen in Table 4-7. Furthermore, the difference in DNI resource between Wagga and

Rockhampton is smaller than the difference in GHI resource. Besides that, the 2005 mean daily

DNI is higher than the 2005 mean daily GHI at each site, except for Mt. Gambier. The resources

have different values because of the differences between GHI and DNI mentioned in the

introduction.

Table 4-7: Total DNI at each site over 2005

2005 total DNI 2005 mean daily 2005 mean hourly Site (MJ/m2) DNI (MJ/m2) DNI (W/m2) Mildura 8365 22.9 500 Mt. Gambier 5003 13.7 304 Rockhampton 7782 21.3 461 Wagga 7685 21.1 457

The mean, median, maximum, minimum, and standard deviation of the 30-minute trading

interval electricity prices in each region for 2005 are calculated and shown in Table 4-8. 78

Although the electricity prices are averaged to 1-hour to use in this study, the statistics are calculated from 30-minute prices so that they can be fairly compared against the long-term descriptive statistics by Higgs et al. (2015). The descriptive statistics for the 2005 electricity price data contrast those for the long-term price on a couple of points. First, SA has the lowest mean price ($25.17/MWh) and median price ($18.54/MWh) in 2005, but had the highest long- term mean price. Second, QLD has the highest median price ($27.30/MWh) in contrast to being shown to have the lowest long-term mean and median price. The Australian Energy

Regulator (AER) publishes annual state of the energy market reports that highlight any price abnormalities. Unfortunately, the archived reports only date back to 2007.

Table 4-8: Descriptive statistics for hourly electricity price ($/MWh) in each region for 2005

Site Region Mean Median Standard deviation All All 30.22 21.12 155.37 Mildura VIC 35.83 20.77 250.31 Mt. Gambier SA 25.17 18.54 140.06 Rockhampton QLD 33.60 27.30 100.61 Wagga NSW 26.29 21.30 63.97

The unusually high electricity prices in QLD may be due to higher than normal temperatures in

2005. It was warmer than average in all months in 2005, and QLD had particularly large positive anomalies (+5°C to +6°C) in April (BOM, 2006). Furthermore, Rockhampton, QLD is subtropical, so it is naturally warmer than the other sites which have temperate and grassland climates (see Table 4-6). Consequently, more air-conditioning may have been used in QLD, thereby increasing electricity demand and electricity price compared to other regions.

The abnormally low electricity prices in SA may be due to the introduction of wind energy. In

2005, 237.5 MW of wind power began generating in SA (Geoscience Australia and ABARE,

2010), and prior to this there was almost no wind power. Wind power is able to place very low 79 priced bids in the NEM due to near-zero operating costs. The addition of this wind power may have been large enough to cause SA electricity prices to decrease significantly in 2005.

4.2. DNI Forecasting

In this study, DNI forecasts are produced at two different horizons that are chosen based on

NEM regulations that a CST plant must comply with. The NEM regulations require generators to produce generation forecasts at least 40 hours ahead to submit initial bids to the market. As such, the first forecast horizon used in this study is 48 hours ahead to determine initial bids.

The 48-hour forecasts are produced every 24 hours when the simulation clock reaches midnight. DNI is always zero from midnight to 4 a.m., so 48-hour forecasts produced at midnight will satisfy this requirement. The 48-hour DNI forecasts will not affect the actual electricity price because the NEM electricity price is determined by balancing the actual electricity supply and demand, not the anticipated supply and demand. However, if there is a large amount of solar generation in the market, the 48-hour DNI forecasts can affect the forecast electricity price because a forecast of high DNI will lead to expectation of a large amount of generation from solar. Solar power can be bid into the market at low prices, so a forecast of high solar generation would produce a forecast of lower electricity prices due to the merit order effect. As noted in Section 4.1.2, the electricity price forecast is assumed to be perfect to focus on the DNI forecast accuracy. Three different methods are used to produce the 48-hour forecasts, which are described in Section 4.2.1.

The other horizon is a short-term horizon of 1 hour ahead to determine updated bids. The 1- hour forecasts are produced at the start of each hour between sunrise and sunset. This corresponds to the start of each time step because the CST model time steps are in 1-hour resolution. The 1 hour update period means that forecast information beyond 1 hour will be replaced by the next forecast, so short-term horizons beyond 1 hour ahead are ignored. 80

Besides that, the simulation time step is 1 hour in size, so the short-term horizon cannot be less than 1 hour because its effects would not be captured by the LKT model. Studies have shown that solar irradiance forecasts are more accurate at shorter forecast horizons. The root mean square error (RMSE) of GHI forecasts has been shown to increase as the forecast horizon increased from 1 hour ahead to 5 hours ahead (Kühnert et al., 2013), and from 1 hour ahead to

7 days ahead (Perez et al., 2010a). Both GHI and DNI forecasts have been shown to have lower

RMSE at the 1-day horizon than the 2-day and 3-day horizons (Lara-Fanego et al., 2012b). As short-term forecasts are more accurate a CST plant is more likely to meet its updated bids and avoid incurring RG costs. Previous studies have not investigated operating CST plants using DNI forecasts under the 24-hour horizon. The short-term forecast method used is described in

Section 4.2.2.

The use of 1-hour forecasts to produce updated bids to replace initial bids determined from

48-hour forecasts may imply that the 48-hour forecasts are unnecessary in this study. However, the 1-hour forecasts are only made during the day. Thus, the 48-hour forecasts may still influence the net value if bids are made to generate electricity at night, which may occur for a

CST plant with storage. Suppose that a 48-hour forecast produced at midnight predicts that the day ahead will be clear. If there is energy in storage and the electricity price is high enough, the CST plant may place initial bids to sell electricity before sunrise because it expects to recharge storage later in the day. These initial bids will not be replaced by updated bids because 1-hour forecasts will not be made before sunrise. Thus, the 48-hour forecasts still have a small effect on the financial value metrics.

4.2.1. Methods for 48-hour forecasts

Three methods are used to produce 48-hour DNI forecasts. A perfect forecast is also used to provide context for the value achieved by using imperfect forecasts. The first 48-hour forecast method is simple persistence of DNI. The 48-hour persistence method used is repetition of DNI

81 over a 48-hour cycle. In other words, at midnight, the DNI for today (0-24 hours ahead) is assumed to equal the incident DNI from the day before yesterday, and the DNI for tomorrow

(24-48 hours ahead) is assumed to equal the incident DNI from yesterday. Persistence is used in this study because previous studies that evaluated the use of forecast methods to operate a

CST plant also included persistence forecasts as a baseline to compare more complex forecast methods against (Kraas et al., 2013; Wittmann et al., 2008).

The second method is The Air Pollution Model (TAPM). TAPM is a prognostic regional meteorological and aerosol transport model (Hurley, 2008). It produces forecasts over a region by solving equations that physically describe atmospheric dynamics, so TAPM is similar to a mesoscale numerical weather prediction (NWP) model. NWP models are recommended when the forecast horizon exceeds 6 hours (Perez et al., 2010a). As TAPM is a regional model, it requires boundary conditions as an input. This study uses National Centre for Environmental

Prediction synoptic data (Kalnay et al., 1996) as input to TAPM for defining boundary conditions. TAPM forecasts hourly global horizontal irradiance (GHI) but not DNI. As DNI is required for the LKT model, the GHI is converted to DNI by using the Boland-Ridley-Lauret (BRL) model (Ridley et al., 2010) with updated model coefficients (Lauret et al., 2013). The BRL model is used because it has been shown to be more accurate than the Reindl model, the

Skartveit model and the Perez model in three Southern Hemisphere locations (Ridley et al.,

2010).

The third method is an autoregressive integrated moving average (ARIMA) model. ARIMA models produce forecasts by analysing regular cycles and disturbances in recent time series data. This study uses an ARIMA model to forecast DNI because ARIMA models are suited to cyclic or seasonal time series, such as the 24-hour diurnal cycle of solar irradiance. The ARIMA model represents a forecast method with complexity in between that of persistence and TAPM because it processes historical data more than persistence does and uses less computational

82 resources than TAPM. An ARIMA model for cyclical data is made by choosing the autoregressive (AR), moving average (MA), seasonal autoregressive (SAR), and seasonal moving average (SMA) terms, and the degree of differencing and seasonal differencing (Box et al., 1994). The choices are made by analysing the autocorrelation function (ACF) and the partial autocorrelation function (PACF) of a sample of data to be modelled. The ARIMA model in this study is made using the MATLAB command “arima” with inputs for a seasonal difference of 24, a non-seasonal difference of 1, and AR, MA, and SMA terms. The ARIMA model coefficients for a forecast are determined by analysing the DNI data in the two weeks preceding the point when the forecast is made. For example, the 48-hour forecast made at the start of 1 June estimates coefficients by using data from 18-31 May and at the start of 2 June the data cover

19 May to 1 June.

4.2.2. Method for 1-hour forecasts

Persistence is a suitable solar irradiance forecast method for short-term horizons, although its accuracy decreases as sky conditions become more variable. When considering all sky conditions, cloud motion vectors (CMVs) provide more accurate GHI forecasts when the forecast horizon exceeds one hour (Coimbra et al., 2013; Kühnert et al., 2013). However, persistence has been shown to have lower GHI forecast RMSE than satellite-based CMVs up to the 4-hour horizon (Perez et al., 2010a). Persistence of DNI has been used as a benchmark for

DNI forecasts at the 5-minute horizon (Queener, 2012), up to the 15-minute horizon (Marquez and Coimbra, 2013), and up to the 20-minute horizon (Quesada-Ruiz et al., 2014). In addition, persistence has also been used as a benchmark in two previous studies on the value of a CST plant operated using DNI forecasts, as mentioned in Section 4.2.1 while describing the 48-hour forecasts.

Persistence is used to produce short-term forecasts because it is the most simple forecast method. Previous studies have not used short-term forecasts to operate CST plants, so the use

83 of persistence will show whether the present research method is appropriate for evaluating more complex short-term forecast methods. The persistence approach for the short-term forecast is persistence of DNI from the most recent time step over the 1-hour forecast horizon.

The short-term forecasts are made at the start of each time step that is between sunrise and sunset. As persistence of night time DNI over day time is clearly incorrect, no short-term forecasts are made until after the time step when sunrise occurs. This means that the first hour of daylight does not have a short-term forecast because its preceding hour has a zero DNI value. Also, no short-term forecasts are made beyond sunset because DNI is known to be zero after sunset.

After the 1-hour forecast is made, the optimisation is run for the current time step to the time step at the end of the present 48-hour forecast. If the new optimised generation schedule is different from the initial bids, then updated bids are submitted to the market so that the CST plant’s bids match the new optimised generation schedule. This is because forecasts at shorter horizons are typically more accurate than forecasts at longer horizons, so it is expected that the CST plant is more likely to meet its updated bids than its initial bids.

4.2.3. Forecast example

(a) 1200

1000

800 600

DNI (W/m2) 400 200 0 2 4 6 8 10 12 14 16 18 20 22 24 Hour, t (AEST)

Real DNI 48-h persistence DNI forecast 1-h forecast made at 10am

(b) 60

50 40 30 20

10 Generation Generation (MWh) bid 0 2 4 6 8 10 12 14 16 18 20 22 24 Hour, t (AEST)

Initial bids Updated bids

Figure 4-3: Example of (a) a 1-hour forecast used to update over-predictions by the 48-hour persistence forecast, and (b) the corresponding bids, for 3 January 2005 in Mt. Gambier

An example of a 1-hour forecast used on 3 January 2005 in Mt. Gambier is presented in Figure

4-3. Although 1-hour forecasts were made every hour after sunrise, this example shows a 1- hour forecast made at the start of hour 11, or at 10am. As can be seen in Figure 4-3(a), the 48- hour persistence forecast over-predicted DNI for the day. Consequently, the CST generation optimisation suggested bids to generate electricity over hours 8-19 to maximise revenue as shown by the black bars in Figure 4-3(b). The 1-hour forecast at the start of hour 11 predicted low DNI for 10am-11am better than the 48-hour forecast and produced updated bids that 85 reduced the CST plant’s generation commitment. As shown by the striped bars in Figure 4-3(b), the bid for hour 11 was reduced to zero. This helps the CST plant avoid RG costs. Bids for hours

12-19 were not reduced to zero because the updated bids produced at the start of hour 11 still used the 48-hour forecast’s over-predictions for hour 12 onwards. As the day goes on, the 1- hour forecasts for each hour would produce updated bids of zero generation.

(a) 1200

1000

800 600

DNI (W/m2) 400 200 0 2 4 6 8 10 12 14 16 18 20 22 24 Hour, t (AEST)

Real DNI 48-h persistence DNI forecast 1-h forecast made at 10am

(b) 60

50 40 30 20

10 Generation (MWh)bid 0 2 4 6 8 10 12 14 16 18 20 22 24 Hour, t (AEST)

Initial bids Updated bids

Figure 4-4: Example of (a) a 1-hour forecast used to update under-predictions by the 48-hour persistence forecast, and (b) the corresponding bids, for 9 January 2005 in Mt. Gambier

Another example is shown in Figure 4-4, this time for a 1-hour forecast used on 9 January 2005 in Mt. Gambier. In contrast to the previous example, the 48-hour persistence forecast under-

86 predicted DNI for 9 January and led to no generation being bid into the NEM. The 1-hour forecast produced at hour 11 had a smaller under-prediction than the 48-hour forecast and produced updated bids that increased the CST plant’s generation. This helps the CST plant earn revenue. As the day goes on, the 1-hour forecasts for each hour would produce updated bids to generate electricity.

4.2.4. Forecast accuracy metrics

The forecast accuracy is assessed in this study using the mean absolute error (MAE), mean bias error (MBE), and root mean square error (RMSE). Other accuracy metrics have been suggested by Zhang et al. (2015b) for assessing forecasts of solar power plant output, but these three metrics are used because they are commonly used and widely known. The forecast skill metric used by Chu et al. (2013) is also included because it demonstrates forecast accuracy improvement while considering the underlying DNI variability. The equations for calculating each metric are shown below. Only daytime hours are considered for calculating the metrics because DNI is known to be zero at night.

The calculation of MBE follows:

∑(퐹표푟푒푐푎푠푡 푣푎푙푢푒 − 푎푐푡푢푎푙 푣푎푙푢푒) E. 4-3 푀퐵퐸 = 푁푢푚푏푒푟 표푓 푣푎푙푢푒푠

The calculation of MAE follows:

∑ √(퐹표푟푒푐푎푠푡 푣푎푙푢푒 − 푎푐푡푢푎푙 푣푎푙푢푒)2 E. 4-4 푀퐴퐸 = 푁푢푚푏푒푟 표푓 푣푎푙푢푒푠

The calculation of RMSE follows:

E. 4-5 ∑(퐹표푟푒푐푎푠푡 푣푎푙푢푒 − 푎푐푡푢푎푙 푣푎푙푢푒)2 푅푀푆퐸 = √ 푁푢푚푏푒푟 표푓 푣푎푙푢푒푠

The calculation of forecast skill follows:

푅푀푆퐸 1 − 𝑖푓 푅푀푆퐸 < 푅푀푆퐸 표푓 푝푒푟푠𝑖푠푡푒푛푐푒 푅푀푆퐸 표푓 푝푒푟푠𝑖푠푡푒푛푐푒 푆푘𝑖푙푙 = 푅푀푆퐸 표푓 푝푒푟푠𝑖푠푡푒푛푐푒 E. 4-6 − 1 𝑖푓 푅푀푆퐸 > 푅푀푆퐸 표푓 푝푒푟푠𝑖푠푡푒푛푐푒 { 푅푀푆퐸

4.3. CST Plant Model

The CST model in this study is adapted from the mixed integer linear programming model by

Sioshansi and Denholm (2010). In order to clearly differentiate which CST model is being referred to, the model in this study will be called the LKT model and the model it is adapted from will be called the Sioshansi and Denholm model. The Sioshansi and Denholm model is chosen because it simulates thermal energy storage and power block energy transfer behaviour and can quickly optimise their operation using linear programming. Both data input and results output are at 1-hour resolution. Although DNI variation can occur in intervals much less than 1 hour long, the 1-hour modelling time step is important for enabling tractability. It allows runs of multiple scenarios for optimising annual CST plant operation from DNI forecasts to be simulated within a day. A resolution of 1-hour is used in the CST value studies by

Wittmann et al. (2008), Kraas et al. (2013), Channon and Eames (2014), and Nonnenmacher et al. (2016). It is also used by the National Renewable Energy Laboratory’s System Advisor Model

(SAM), which was used in the CST value studies by Mehos et al. (2015), Jorgenson et al. (2014),

Denholm et al. (2013), and Jorgenson et al. (2013). Madaeni et al. (2012b) used the same model as Sioshansi and Denholm (2010).

The LKT model is made to resemble the parabolic trough CST plant Andasol-1 because parabolic trough technology is presently the most mature and most common CST plant technology (IEA, 2014). The model parameters are determined by reference to the System

Advisor Model (SAM) version 2014.1.14 (NREL, 2014) project file of Andasol-1 made available by the Australian Solar Thermal Energy Association (AUSTELA, 2014). The parameter values

88 used are presented in Table 4-9. The equations used in the LKT model are presented in

Appendix A.

Solar field size can be defined using the solar multiple (SM), which is the ratio of designed solar field thermal output to rated power block thermal input. A solar field with a SM of 1 operating under design conditions produces the thermal energy required by the power block to generate its rated electric output. Storage size is defined using hours of storage, which is the maximum thermal energy that can be stored divided by the power block rated thermal energy input. This means it represents the number of hours that the power block can be operated at rated capacity when supplied only by storage and storage is initially full. Andasol-1 has a SM of 1.756,

7.5 hours of storage and its rated capacity is 50 MW.

Table 4-9: Parameter values used for the LKT model

Parameter Value Units Coefficient to convert power block Electric megawatt-hour per energy input to net electric output 0.3563 thermal megawatt-hour Design DNI for sizing solar field 700 W/m2 Power block maximum energy input (percentage of rated value) 100 % Power block minimum energy input (percentage of rated value) 25 % Power block minimum operating time 1 Hour Power block rated thermal energy input 144.357 Thermal megawatt-hours Storage charge-discharge cycle loss (percentage of discharge energy) 2.28 % Storage hourly energy loss (percentage of energy in storage) 0.031 % Storage maximum charge rate 130.247 Thermal megawatt-hours Storage maximum discharge rate 127.347 Thermal megawatt-hours

Two configurations for the LKT model are used in this study. The only difference between the two configurations is that one has 7.5 hours of storage whereas the other has no storage. This is because a review identified 37 plants that only have a solar field (i.e. no storage and no fossil

89 fuel backup), 39 plants that have some amount of storage, and 24 plants that are combined with coal or natural gas (Baharoon et al., 2015). All other parameters are the same. The capital cost of storage is ignored for simplification, thus the different costs due to different sizes are not considered. Operating and maintenance costs are also ignored except for reserve generation payments.

For this study, the Sioshansi and Denholm model is modified. Firstly, storage is allowed to both charge and discharge in each time step in the Sioshansi and Denholm model whereas storage is restricted to only one mode of operation in each time step in the LKT model. Both models have separate variables for charging and discharging because losses across the storage heat exchanger are accounted for by modifying the discharge energy. The Sioshansi and Denholm model allows the variables for charging and discharging to reach their maximum values in the same time step. The LKT model is based on Andasol-1, which uses a two-tank storage configuration that is in parallel with the solar field and power block. Hence when storage is being charged it cannot be discharged and vice versa. Thus the LKT model restricts storage to one mode of operation in each time step. This restriction is also present in SAM and the model by Usaola (2012).

Secondly, the energy balance between the power block, storage and solar field is changed from an inequality constraint in the Sioshansi and Denholm model to an equality constraint in the LKT model by adding a variable for dumped thermal energy. This variable is added because it is required for calculating the dump cost. Thirdly, the power block start-up thermal energy term in the Sioshansi and Denholm model is omitted in the LKT model. The net electric output from the LKT model is compared against that from SAM 2013.1.15 in Figure 4-5. SAM

2013.1.15 is used because it has been validated against Andasol-1 (NREL, 2013). The plot in

Figure 4-5(a) shows net electric output from the LKT model with power block start-up thermal energy. Many data points lie far from the y = x line, indicating that including the power block

90 start-up thermal energy causes a bias in the LKT model net electric output. When the power block start-up thermal energy term is removed, the data points are closer to the y = x line with smaller deviation, as shown in Figure 4-5(b). Thus, the power block start-up thermal energy is excluded from the LKT model.

Finally, the Sioshansi and Denholm model used SAM to convert DNI to solar field thermal output because it did not update forecasts. In contrast, the DNI forecasts in this study will be updated. The specific update times along with the reasons for choosing them are explained in

Section 4.2. Each DNI forecast update will require updating the solar field thermal output forecast. Using SAM will be time-consuming because SAM simulations cannot be restricted to only simulating the updated DNI forecast time frame. Instead, this study uses the solar field model developed by Patnode (2006). Patnode’s solar field model requires the solar field inlet temperature to calculate thermal output. The LKT model does not record temperatures for the power block or storage so this study assumes that the solar field inlet temperature equals the design solar field inlet temperature. The equations for the solar field are in Appendix A.

(a) 60 (b) 60

y = x y = x

e) e)

- - 50 50

40 40

30 30

20 20

10 10 LKT LKT model net electric output(MWh LKT LKT model net electric output(MWh 0 0 0 20 40 60 0 20 40 60 SAM net electric output (MWh-e) SAM net electric output (MWh-e)

4.4. Financial Value Calculation Method

The financial value calculation method in this study requires forecast DNI, actual DNI, electricity price and a CST model that can optimise the operation of the power block and storage. The procedure for calculating financial value is presented in Figure 4-6. Forecast DNI is converted to forecast solar field thermal output by the solar field component of the CST model.

Similarly, actual solar field thermal output is obtained from actual DNI. The forecast solar field thermal output and electricity price are used to optimise the operation of the power block and storage to maximise revenue. It is assumed that the optimum operation defines the dispatch targets that the CST plant commits to achieving. A CST plant with storage may have dispatch targets at night if that is when the highest prices occur. Actual solar field thermal output is used to determine whether the CST plant is able to achieve the dispatch targets. Generated

92 electricity earns revenue in proportion to the amount generated and the electricity price at that time.

Figure 4-6: Summary of method to calculate financial value metrics

Over-forecasting DNI leads the CST plant operator to believe that more electricity can be generated than possible. Consequently, the CST plant may not be able to meet its dispatch target. If the CST plant fails to meet its dispatch target, then it will incur a penalty reserve generation (RG) cost of paying for ancillary services to supply the electricity that it could not generate. Thermal energy stored in storage may be used to compensate for the over-forecast error, but it may be insufficient to completely avoid RG costs.

In the opposite case, when DNI is under-forecasted, the solar field produces more thermal energy than required to meet its dispatch target. Thus the dispatch target can be easily

93 achieved. If the dispatch target is less than the CST plant’s rated generating capacity, then the

CST plant curtails its output to only generate the amount of electricity in its bid and avoid paying penalties for causing the network to receive excess electricity supply. Excess thermal energy may be charged to storage. However, storage capacity is limited, so excess thermal energy will have to be dumped when storage is full. Dumped thermal energy could have been used to generate electricity and earn revenue, thus it represents a loss of potential revenue. It has been indicated that accurate solar forecasts are more valuable during peak demand than when demand is low (Luoma et al., 2014). Similarly, thermal energy dumped when electricity prices are high is a greater loss of potential revenue than when electricity prices are low. Thus the electricity price at the time of dumping is used to calculate dump cost. The electricity price applies to electrical energy, so the dump cost is calculated using the electricity that could have been generated from the dumped thermal energy.

The net value is calculated by subtracting the RG cost from revenue. The revenue intrinsically contains a penalty for under-forecasting because dumped thermal energy cannot be used to generate electricity and thereby does not contribute towards the calculated revenue. The dump cost is thus excluded from the net value calculation to avoid double counting the cost of under-forecasting. The dump cost is still calculated because it directly assigns a dollar value to under-forecast errors whereas revenue does not. The revenue, RG cost, dump cost and net value form a set of financial value metrics.

4.5. Reliability Calculation Method

A bid by a generator in the electricity market is a commitment to supply a specific amount of electricity during a specific dispatch interval for a minimum price. If a generator is dispatched, it may be instructed to be either fully dispatched or partially dispatched. A fully dispatched generator supplies electricity equal to its bid, whereas a partially dispatched generator 94 supplies an amount of electricity that is less than its bid. This study assumes that the CST plant is always fully dispatched. This assumption is reasonable because the CST plant can place low priced bids in the market due to its zero fuel cost. The market aims to minimise the cost of supplying electricity to satisfy demand, so the CST plant is very likely to be fully dispatched.

This avoids dispatching a generator that wants a higher price, thereby minimising the electricity price for the dispatch interval.

The reliability of a generator is its ability to follow the dispatch instructions it receives from the network operator. A generator that fully satisfies its dispatch instructions more often is considered more reliable. Improving CST plant reliability by using DNI forecasts is valuable from the perspective of AEMO, the NEM network operator. Higher reliability allows AEMO to have less reserve generation on standby to ensure it can meet the NEM reliability standard.

Generator reliability can be determined by following the internationally used standard for calculating generator reliability published by the Institute of Electrical and Electronics

Engineers (IEEE), the IEEE Standard Definitions for Use in Reporting Electric Generating Unit

Reliability, Availability and Productivity (IEEE, 2007). As this case study is set in the NEM, the

Australian Energy Market Operator (AEMO) method for calculating the equivalent forced outage rate (EFOR), which is “broadly consistent” with the IEEE standard (AEMO, 2010c), is used instead. The equations for calculating EFOR are as follows:

퐹푂퐻 + 퐸퐹퐷퐻 E. 4-7 퐸퐹푂푅 = ∗ 100 퐹푂퐻 + 푆퐻 + 퐸푅푆퐹퐷퐻

∑(푑푒푟푎푡푒푑 ℎ표푢푟푠 𝑖푛 푠푒푟푣𝑖푐푒 ∗ 푀푊 푙표푠푠) E. 4-8 퐸퐹퐷퐻 = 퐺푀퐶

∑(푑푒푟푎푡푒푑 ℎ표푢푟푠 𝑖푛 푟푒푠푒푟푣푒 푠ℎ푢푡푑표푤푛 ∗ 푀푊 푙표푠푠) E. 4-9 퐸푅푆퐹퐷퐻 = 퐺푀퐶

where EFOR is equivalent forced outage rate, FOH is forced outage hours, SH is service

hours, EFDH is equivalent forced derated hours, ERSFDH is equivalent reserve

shutdown forced derated hours, and GMC is gross maximum capacity.

According to the IEEE standard (IEEE, 2007), the GMC is also known as the dependable capacity and it changes according to seasonal variations. AEMO simplifies this by only using the winter capacity, i.e. the maximum output assuming winter ambient temperature conditions (AEMO,

2010c). This study assumes that the winter capacity is equal to the CST plant rated capacity (50

MW). If a generator has a partial outage that is less than 5% of its winter capacity, AEMO allows the generator to not report the partial outage although AEMO prefers that it is reported

(AEMO, 2010c).

This study only considers outages caused by insufficient thermal energy to generate expected electric output. Outages caused by other factors, such as mechanical failure of components, are ignored. Thus the EFOR calculated in this study will only show the effect of forecast accuracy on CST plant reliability. This will allow a fair comparison of CST plant reliability achieved by different forecast methods. It is assumed that forecast information does not affect the probability of other factors causing full or partial outages. As these other causes of outages are ignored, the calculated EFOR will not represent the real plant reliability.

Each simulation time step contributes towards calculating EFOR depending on three pieces of information for the time step: whether a bid is placed by the CST plant, what AEMO expects of the CST plant, and what the CST plant is able to accomplish. AEMO’s expectations need to be considered because AEMO produces an unconstrained intermittent generation forecast (UIGF)

– its own forecast of electricity generation from intermittent sources – to help ensure power system security and reliability of supply (AEMC, 2016a). AEMO uses a UIGF to determine whether the CST plant can be dispatched as a reserve generator if necessary when there is no bid placed for the CST plant to generate electricity. For simplicity, this study assumes that both 96

AEMO and the CST plant operator have the same forecast of CST plant output. Although this assumption is not necessarily true, it is realistic because a UIGF considers the plant availability and maximum generation as informed by the CST plant operator. The UIGF is used by AEMO, not the CST plant operator, so it does not affect the bids placed in the market by the CST plant.

The process of classifying each time step is summarised in Figure 4-7. If the CST plant places a bid to generate electricity, then AEMO expects the CST plant to meet its bid. If the bid is met in full, then one hour is added to the number of service hours (SH). If the CST plant generates electricity but does not meet its bid, then the time step contributes towards the number of equivalent forced derated hours (EFDH), regardless of the size of the difference. The amount of energy not supplied is added to the numerator for calculating EFDH. If the CST plant fails to generate any electricity to meet the bid, then one hour is added to the number of forced outage hours (FOH).

Figure 4-7: Chart of the process used to classify each time step for use in calculating the equivalent forced outage rate (EFOR)

If the CST plant does not place a bid, then AEMO either expects the CST plant to be available for generating electricity if required (also known as reserve shutdown state) or to be unavailable. If AEMO expects the CST plant to be unavailable, then the time step is not used to calculate EFOR. If AEMO expects the CST plant to be available if required, then AEMO will expect a certain output level that is determined using the forecast solar field thermal output and the thermal energy in storage. If the CST plant could generate the expected output if called upon, then the time step is not used to calculate EFOR. If the CST could generate, but not fully meet, the expected output if called upon, then the time step contributes towards the number of equivalent reserve shutdown forced derated hours (ERSFDH). The amount of energy that would not be supplied is added to the numerator for calculating ERSFDH. If the CST plant could not generate any electricity, then one hour is added to the number of FOH. 98

4.6. Chapter Summary

This chapter covered the four sites chosen for this study. The sites were compared to show the differences in their DNI resource and electricity prices. The DNI data consisted of ground and satellite data because of gaps in the ground data record. The satellite data were processed to reduce bias. A CST plant model was obtained for this study by combining a solar field model and a power block and storage model from past studies. The procedures for calculating financial value and EFOR were explained. The methods to forecast DNI at the 48-hour and 1- hour horizons were described. An example was given to demonstrate how the forecasts are used. The results produced by using the data and processes outlined in this chapter are presented in the next chapter.

5. Results

The scenarios investigated in this study differ by the site, the 48-hour forecast method, whether the CST plant has storage or not, and whether 1-hour forecasts are used or not. The accuracy of the forecasts, and the financial value and reliability of the CST plant operated with forecasts are determined for each scenario. The forecast accuracy is quantified using the commonly known mean bias error (MBE), mean absolute error (MAE), and root mean square error (RMSE). The financial value metrics are the revenue, reserve generation (RG) cost, dump cost, and net value. The reliability metric is the equivalent forced outage rate (EFOR). The numerical values for financial value and reliability results are not presented in this chapter and are instead available in Appendix B.

5.1. Site DNI Forecast Accuracy

The MBE, MAE, and RMSEs of each 48-hour forecasts and each 1-hour forecasts at each site are shown in Table 5-1. The MBE of TAPM is very large for all sites, ranging from 257 to 380

W/m2. The MBE of ARIMA and 48-hour persistence are much smaller, ranging from -3.08 to

14.6 W/m2 and -1.41 to -0.00948 W/m2, respectively. This shows that TAPM tended to greatly overestimate available DNI, which is similar to previous studies which have shown that TAPM underestimates cloud cover and thereby overestimates GHI in Australia (Hibberd, 2011;

Dehghan et al., 2014). The forecast skill scores show that the ARIMA forecasts improve upon the 48-hour persistence forecasts at all sites because all the skill scores are positive. In contrast, the TAPM forecasts have negative skill scores at all sites, meaning that TAPM forecasts are worse than the 48-hour persistence forecasts. On average over all sites, TAPM has the highest

MAE (357 W/m2) and RMSE (470 W/m2) and is 14.3% worse than 48-hour persistence. ARIMA has the lowest MAE (265 W/m2) and RMSE (345 W/m2) and is 14.4% better than 48-hour

100 persistence. This shows that, on average, TAPM produced the least accurate 48-hour DNI forecasts and ARIMA produced the most accurate forecasts in this study.

The 1-hour persistence forecast has small MBE values of -4.82 to 1.60 W/m2 across all sites.

The lowest MAE (138 W/m2) and RMSE (204 W/m2) of the 1-hour persistence forecast are in

Mt. Gambier and its highest MAE (164 W/m2) and RMSE (227 W/m2) are in Rockhampton. This suggests that the hour-to-hour DNI variability was lowest in Mt. Gambier and highest in

Rockhampton in 2005. The MAE and RMSE values of the 1-hour persistence forecast are lower than those of the 48-hour forecast methods, which is expected because solar irradiance forecasts have been shown to be more accurate at short-term horizons (Perez et al., 2010a).

Table 5-1: The accuracy of each forecast at each site

Site All Mildura Mt. Gambier Rockhampton Wagga MBE (W/m2) ARIMA 4.16 -0.0798 14.6 -3.08 5.18 48-h persistence -0.668 -0.401 -0.856 -0.00948 -1.41 TAPM 325 257 380 352 311 1-h persistence -1.82 -4.82 -2.25 -1.80 1.60

MAE (W/m2) ARIMA 265 253 269 257 281 48-h persistence 288 276 314 258 303 TAPM 357 305 432 361 333 1-h persistence 150 152 138 164 147

RMSE (W/m2) ARIMA 345 335 346 335 365 48-h persistence 403 398 426 363 424 TAPM 470 415 542 467 457 1-h persistence 216 221 204 227 214

Forecast skill ARIMA 0.144 0.158 0.188 0.0771 0.139 48-h persistence 0 0 0 0 0 TAPM -0.143 -0.0410 -0.214 -0.223 -0.072

101

5.2. Financial Value

For each scenario, the revenue is shown in Figure 5-1, the RG cost is shown in Figure 5-2, the dump cost is shown in Figure 5-3, and the net value is shown in Figure 5-4. Taking all scenarios into consideration, the use of 1-hour forecasts increases net value by $0.54-3.71 million, thus the use of 1-hour forecasts is financially beneficial. This shows that increasing financial value by using forecasts at shorter horizons around 1-hour ahead is a viable alternative to increasing value by improving the accuracy of forecasts at 24 hours or more ahead (Wittmann et al., 2008;

Sioshansi and Denholm, 2010; Kraas et al., 2013; Channon and Eames, 2014; Nonnenmacher et al., 2016).

The majority of the net value increase from using 1-hour forecasts is due to lower RG cost rather than higher revenue. In all scenarios involving ARIMA or 48-hour persistence, the decrease in RG cost contributes towards 57%-99% of the net value increase. In scenarios involving TAPM, the decrease in RG cost contributes towards 93%-177% of the net value increase. The contribution of RG cost reduction exceeds 100% because the use of short-term forecasts causes revenue to decrease in some of the scenarios involving TAPM. This can be explained by considering the MBE values of TAPM and the 1-hour forecasts. The large positive

MBE of TAPM means that it leads to initial bids with many offers to generate electricity. The magnitude of the MBE of the 1-hour forecasts is lower by a factor of about 100. This means that a large proportion of the updated bids made using 1-hour forecasts are to reduce offers to generate electricity made using TAPM. Consequently, the amount of electricity sold and the revenue earned decrease. Overall, these results show that the main financial benefit of using

1-hour forecasts is from reducing RG costs rather than increasing revenue.

The dump cost decreases by $0.02-$2.37 million when 1-hour forecasts are used in all scenarios involving ARIMA and 48-hour persistence. In contrast, the scenarios involving TAPM have the dump cost increase by $0.08-0.82 million. As mentioned earlier, TAPM causes the CST

102 plant to make many offers to generate electricity, and the 1-hour forecasts reduce the offers made using TAPM to generate electricity. Another consequence of this is that the CST plant is more likely to have excess thermal energy from the solar field that it cannot store or immediately use for generating electricity. As a result, the amount of dumped thermal energy increases. This shows that, unless the 48-hour forecast has a large positive MBE, 1-hour forecasts can help reduce the amount of revenue lost through dumping thermal energy.

When 1-hour forecasts are used, the net values achieved by all 48-hour forecasts are similar at all locations except at Wagga, where the 48-hour persistence forecast achieved about $1 million less than TAPM and ARIMA in net value. TAPM achieved similar net value as ARIMA and

48-hour persistence even though its forecast accuracy was worse. This shows that when short- term forecasts are used, the net value that a given CST plant at a given location achieves is mostly dependent on short-term forecast accuracy. That said, as TAPM has worse forecast accuracy than 48-hour persistence in this study, it is unlikely to be used for operational purposes in reality. As such, all TAPM results will be mentioned in a separate paragraph in the respective section and be excluded from the conclusions drawn from the study.

103

No storage With storage 6 Mildura 6 Mildura 5 5 4 4 3 3 2 2

1 1 Revenue ($ million) ($ Revenue 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

6 Mt. Gambier 6 Mt. Gambier 5 5 4 4 3 3 2 2

1 1 Revenue ($ million) ($ Revenue 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

7 Rockhampton 7 Rockhampton 6 6 5 5 4 4 3 3 2 2

1 1 Revenue ($ million) ($ Revenue 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

10 Wagga 10 Wagga 9 9 8 8 7 7 6 6 5 5 4 4 3 3 2 2

Revenue ($ million) ($ Revenue 1 1 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-1: Total revenue earned in each scenario

104

Mildura No storage Mildura With storage 2 2

1 1 RG costRG million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Mt. Gambier Mt. Gambier 4 4

3 3

2 2

1 1 RG costRG million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect Rockhampton Rockhampton 2 2

1 1 RG costRG million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Wagga Wagga 3 3

2 2

1 1 RG costRG million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-2: Total RG cost paid in each scenario

105

Mildura No storage Mildura With storage 4 4

3 3

2 2

1 1

Dump costDump million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Mt. Gambier Mt. Gambier 4 4

3 3

2 2

1 1

Dump costDump million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Rockhampton Rockhampton 4 4

3 3

2 2

1 1

Dump costDump million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

8 Wagga 8 Wagga 7 7 6 6 5 5 4 4 3 3 2 2

1 1 Dump costDump million) ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-3: Total dump cost in each scenario

106

No storage With storage 6 Mildura 6 Mildura 5 5 4 4 3 3 2 2

1 1 Net value ($ Net million) value ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

6 Mt. Gambier 6 Mt. Gambier 5 5 4 4 3 3 2 2

1 1 Net value ($ Net million) value ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

7 Rockhampton 7 Rockhampton 6 6 5 5 4 4 3 3 2 2

1 1 Net value ($ Net million) value ($ 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

10 Wagga 10 Wagga 9 9 8 8 7 7 6 6 5 5 4 4 3 3 2 2

Net value ($ Net million) value ($ 1 1 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-4: Total net value achieved in each scenario

107

When the 48-hour forecast is perfect, the net value achieved in Wagga is $9.7 million for a CST plant with storage, which is much higher than the $5.4-6.5 million at the other sites. For a CST plant without storage, the net value achieved in Wagga is $6.8 million whereas it is $3.5-4.7 million at the other sites. The lowest net values are achieved in Mt. Gambier for both CST plant configurations. The net value achieved depends on the forecast accuracy, DNI resource, and electricity price. As only the perfect 48-hour forecast scenarios are being considered, the forecast accuracy may be ignored. Information on the DNI resource and 1-hour electricity price of each site is presented in Table 5-2.

The mean daily DNI resource at Wagga (21.1 MJ/m2) is less than that at Mildura (22.9 MJ/m2), and the lowest mean daily DNI by far is at Mt. Gambier (13.7 MJ/m2). Wagga has the highest mean electricity price ($35.8/MWh) and Mt. Gambier has the second highest ($33.6/MWh).

The median electricity price at Wagga ($20.9/MWh) is lower than that of Mildura ($21.4/MWh) and Mt. Gambier ($27.3/MWh). The correlations between electricity prices and DNI at each site are very weakly positive, and the correlation at Wagga (0.060) is lower than that at

Mildura (0.068) and Mt. Gambier (0.086). Overall, this suggests that a CST plant can achieve high financial value if the mean electricity price is high, even if the DNI resource is not the best.

However, high mean electricity price cannot offset poor DNI resource.

Table 5-2: Mean daily DNI, mean and median 1-hour electricity price, and correlation between

DNI and electricity price for each site in 2005

2005 mean 2005 mean 2005 median daily DNI 1-hour price 1-hour price DNI-price Site (MJ/m2) ($/MWh) ($/MWh) correlation Mildura 22.9 26.3 21.4 0.068 Mt. Gambier 13.7 33.6 27.3 0.086 Rockhampton 21.3 25.2 18.5 0.055 Wagga 21.1 35.8 20.9 0.060

108

The net value in each scenario is shown in Figure 5-5 as a fraction of the net value achieved by a perfect 48-hour forecast at the same site. In scenarios that do not use 1-hour forecasts, the fractions range from 0.012 to 0.79 for scenarios involving ARIMA and 48-hour persistence. In comparison, the fractions range from 0.59 to 0.94 in scenarios that use 1-hour forecasts with either ARIMA or 48-hour persistence. Although the two ranges overlap, the fractions in the scenarios that use 1-hour forecasts are always higher than the fractions in the corresponding scenarios that do not use 1-hour forecasts. This is in agreement with the result that using 1- hour forecasts increases the net value achieved. Overall, using 1-hour forecasts achieves a greater proportion of the financial value achievable, ranging from 59% to 94%, if the 48-hour forecast were perfect.

The net value achieved in each scenario may be divided by the corresponding amount of electricity generated to obtain the net value per unit electricity sold. The results are shown using $/MWh in Figure 5-6. When 1-hour forecasts are used, the net value per MWh increases by $3.3/MWh to $43/MWh across all scenarios involving ARIMA and 48-hour persistence.

An important result to notice is that the use of 1-hour forecasts increases the net value per

MWh to be 85%-100% of the net value per MWh if the 48-hour forecast were perfect for all scenarios involving ARIMA and 48-hour persistence. This shows that the use of 1-hour forecasts can increase the net value per MWh to its theoretical maximum despite the net value still falling short of its theoretical maximum. This means that using 1-hour forecasts allows a CST plant to maximise or nearly maximise the value per unit electricity sold.

Adding storage to a CST plant without storage and no 1-hour forecasts increases net value by

$2.0-3.5 million over all sites that use either ARIMA or 48-hour persistence, whereas using 1- hour forecasts increases net value by $0.76-3.1 million. The net value increase from adding storage is the larger amount in all scenarios involving ARIMA and 48-hour persistence. This suggests that adding 7.5 hours of storage provides a greater increase in net value than using 1- 109 hour forecasts. However, it should be noted that the capital cost and operations and maintenance costs for storage would be significantly higher than the cost of producing and using 1-hour persistence forecasts.

In scenarios involving TAPM, adding 1-hour forecasts increases net value by $0.54-2.2 million.

The net value as a fraction of the net value achieved by a perfect 48-hour forecast at the same site ranges from 0.053 to 0.70 when 1-hour forecasts are not used and from 0.69 to 0.96 when

1-hour forecasts are used. The net value per MWh increases by $9/MWh to $40/MWh in scenarios without storage and by $11/MWh to $40/MWh for scenarios with storage. Adding storage increases net value by $0.92-1.7 million.

In summary, the results involving ARIMA and 48-hour persistence show that using 1-hour forecasts is financially beneficial because they increase net value by $0.76-3.1 million by reducing RG cost. It enables the CST to attain 59% to 94% of the net value achieved by a perfect 48-hour forecast at the same site (perfect 48-hour forecast achieves $3.5-9.7 million).

It also allows a CST plant to maximise or nearly maximise the value per unit electricity sold by increasing the net value per MWh to be 85%-100% of the net value per MWh achieved by a perfect 48-hour forecast at the same site (perfect 48-hour forecast achieves $33/MWh to

$61/MWh). The results also show that if a CST plant has neither storage nor 1-hour forecasts, then adding 7.5 hours of storage has a greater effect on increasing net value than adding 1- hour forecasts. However, the cost of adding 7.5 hours of storage would be greater than the cost of adding 1-hour persistence forecasts.

110

No storage With storage Mildura Mildura 1 1 0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

Net Net value (% of perfect) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Mt. Gambier Mt. Gambier 1 1 0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

Net Net value (% of perfect) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Rockhampton Rockhampton 1 1 0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

Net Net value (% of perfect) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Wagga Wagga 1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

Net Net value (% of perfect) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-5: Net value achieved in each scenario as a fraction the net value of a perfect 48-hour

forecast at the same site

111

Mildura No storage Mildura With storage 40 40

30 30

20 20

10 10 Net Net value ($/MWh) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

60 Mt. Gambier 60 Mt. Gambier 50 50 40 40 30 30 20 20

10 10 Net Net value ($/MWh) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Rockhampton Rockhampton 40 40

30 30

20 20

10 10 Net Net value ($/MWh) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

70 Wagga 70 Wagga 60 60 50 50 40 40 30 30 20 20

10 10 Net Net value ($/MWh) 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-6: Net value achieved in each scenario divided by the corresponding amount of electricity generated

112

5.3. Reliability

The EFOR for all scenarios is shown in Figure 5-7. It shows that using 1-hour forecasts reduces

EFOR by 16-42 percentage points across all scenarios involving ARIMA and 48-hour persistence.

This mirrors the reduction in RG cost caused by using 1-hour forecasts as shown in the previous section. Lower EFOR represents an increase in plant reliability hence the use of 1- hour forecasts is beneficial.

The EFOR of a CST plant without storage and no 1-hour forecasts in Mt. Gambier is almost 20 percentage points greater than the EFOR of any other CST plant at any site. This is because the

48-hour persistence forecast is the least accurate in Mt. Gambier, according to the MAE and

RMSE values shown in Table 5-1. The ARIMA forecasts in Mt. Gambier are the second least accurate after Wagga. The poor 48-hour forecast accuracy at Mt. Gambier means that the CST plant is less likely to be able to meet its initial bids. Consequently, the CST plant will have more generation outages and have a high EFOR.

The EFOR is plotted against 48-hour forecast MAE and RMSE in Figure 5-8. Results show that when 1-hour forecasts are not used, increasing the 48-hour forecast accuracy reduces the

EFOR by up to 20 percentage points. When 1-hour forecasts are used, the same increase in 48- hour forecast accuracy only reduces the EFOR by up to 2 percentage points. As mentioned earlier, the use of 1-hour forecasts reduces the influence of 48-hour forecast accuracy, and hence the reduction in EFOR is much smaller. This shows that if 1-hour forecasts are not available, then it is worthwhile to increase 48-hour forecast accuracy to increase CST plant reliability. Otherwise, it would be better to increase 1-hour forecast accuracy.

A noticeable feature is that the EFOR from highest to lowest (least reliable to most reliable) at each site is a CST plant without storage and no 1-hour forecasts, followed by a CST plant with storage and no 1-hour forecasts, then a CST plant without storage and using 1-hour forecasts, and finally a CST plant with storage and using 1-hour forecasts. This shows that a CST plant 113 without storage and no 1-hour forecasts achieves a larger improvement in reliability by using

1-hour forecasts instead of adding storage.

In scenarios involving TAPM, the EFOR reduction through using 1-hour forecasts is 25-55 percentage points. The EFOR reduction exceeds that of other scenarios because TAPM has the lowest forecast accuracy of the 48-hour forecast methods.

In summary, the results involving ARIMA and 48-hour persistence show that using 1-hour forecasts improve CST plant reliability by reducing EFOR by 16-42 percentage points. When 1- hour forecasts are used, the operation of the CST plant depends almost entirely on the accuracy of the 1-hour forecasts. The results also show that if a CST plant has neither storage nor 1-hour forecasts, then adding 1-hour forecasts produces a greater reduction in EFOR than adding 7.5 hours of storage.

114

Mildura No storage Mildura With storage 40 40

30 30

20 20

EFOR EFOR (%) 10 10

0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

80 Mt. Gambier 80 Mt. Gambier 70 70 60 60 50 50 40 40 30 30 EFOR EFOR (%) 20 20 10 10 0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Rockhampton Rockhampton 50 50

40 40

30 30

20 20 EFOR EFOR (%) 10 10

0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect

Wagga Wagga 50 50

40 40

30 30

20 20 EFOR EFOR (%) 10 10

0 0 ARIMA Persist TAPM Perfect ARIMA Persist TAPM Perfect 48-h forecast used to make initial bids 48-h forecast used to make initial bids

No 1-h forecast With 1-h forecast

Figure 5-7: EFOR achieved in each scenario

115

No 1-h forecasts CST without storage With 1-h forecasts 80 30

25 60 20 40 EFOR EFOR (%) 15

20 10 250 300 350 400 450 250 300 350 400 450 MAE (W/m2) MAE (W/m2)

80 30

25 60 20 40 EFOR EFOR (%) 15

20 10 300 350 400 450 500 550 300 350 400 450 500 550 RMSE (W/m2) RMSE (W/m2) CST with storage 80 8

6 60 4 40 EFOR EFOR (%) 2

20 0 250 300 350 400 450 250 300 350 400 450 MAE (W/m2) MAE (W/m2)

80 8

6 60 4 40 EFOR EFOR (%) 2

20 0 300 350 400 450 500 550 300 350 400 450 500 550 RMSE (W/m2) RMSE (W/m2)

Mildura Mt. Gambier Rockhampton Wagga

Figure 5-8: EFOR against 48-hour forecast MAE and RMSE

116

5.4. Summary of Results

In summary, the results of this study show that using 1-hour forecasts increases the financial value and reliability of a CST plant, regardless of whether the CST plant has storage or not. The majority of the financial value gain is obtained from reducing RG costs. Although the financial value is not increased to match the total net value of a perfect 48-hour forecast, the financial value per unit electricity sold is increased to match or nearly match that of a perfect 48-hour forecast. Financial gain through increasing revenue is achieved by reducing the loss of potential revenue (dump cost), unless the 48-hour forecast has a large positive MBE.

If 1-hour forecasts are not used, results show that increasing the 48-hour forecast accuracy significantly improves plant reliability. Contrasting results show that the financial value may increase or decrease when the 48-hour forecast accuracy improves. If 1-hour forecasts are used, then increasing the 48-hour forecast accuracy has a minimal effect on financial value and plant reliability due CST plant operation becoming mostly dependent on the 1-hour forecasts.

The availability of storage affects the increase in financial value and reliability. Results show that if a CST plant has no storage and does not use 1-hour forecasts, then adding 7.5 hours of storage provides a larger increase in financial value than using 1-hour forecasts. In contrast, using 1-hour forecasts produces a larger increase in reliability than adding 7.5 hours of storage.

In the next chapter, the results are compared against those from previous studies. The applicability of the results is considered to determine its limits and to identify further research to overcome the limits.

117

6. Discussion5

In the previous chapter, this study showed that it is more valuable to use both 1-hour and 48- hour forecasts instead of just 48-hour forecasts to operate a CST plant in the NEM. The value was demonstrated by increased financial value and reliability, which was achieved by the 1- hour forecasts producing updated bids that used the available DNI more effectively than the initial bids from the 48-hour forecasts. The value was achieved by CST plants with and without storage. In this chapter, the results are compared against those from past studies that examined using forecasts to operate CST plants. After that, this study is reflected upon to determine the limits of its applicability, and to identify research directions to further develop its findings and to overcome its limitations.

6.1. Comparison with Past Studies

The 48-hour DNI forecasts in this study are less accurate than those in the studies by Lara-

Fanego et al. (2012b), Breitkreuz et al. (2009), Wittmann et al. (2008) and Kraas et al. (2013) as shown in Table 6-1. The RMSE values in the other studies are between 15.5% and 249% lower than the RMSE values in this study. However, it should be noted that only clear sky conditions were considered by Breitkreuz et al. (2009), which is why the RMSE values from that study are extremely lower than the RMSE values in this study. The study by Wittmann et al. (2008) only considered five days of data, so the comparison may not be representative of the real difference in forecast accuracy. All sky conditions over at least one year were considered by

Lara-Fanego et al. (2012b) and Kraas et al. (2013). The RMSE values in these two studies are

5 The material presented in this chapter has been previously published in: Law, E.W., Kay, M., Taylor, R.A., 2016, Calculating the financial value of a concentrated solar thermal plant operated using direct normal irradiance forecasts, Solar Energy, 125, p. 267-281 https://dx.doi.org/10.1016/j.solener.2015.12.031

Law, E.W., Kay, M., Taylor, R.A., Evaluating the benefits of using short-term direct normal irradiance forecasts to operate a concentrated solar thermal plant, Solar Energy 140, p. 93-108 http://dx.doi.org/10.1016/j.solener.2016.10.037 118 between 15.5% and 74.3% lower than the RMSE values in this study. This may be because their forecast methods were more accurate, or their forecast methods were more suited to forecasting the weather associated with the climate of their locations compared to the forecast methods and locations in this study.

Table 6-1: Comparison of RMSE between the present study and other studies

RMSE lower value RMSE upper value Value Percentage difference Value Percentage difference to (W/m2) to present study (%) (W/m2) present study (%) Present study 335 0 542 0 Lara-Fanego et al. (2012b) 290 -15.5 311 -74.3 Breitkreuz et al. (2009) 96 -249 159 -241 Wittmann et al. (2008) 185 -81.1 209 -159 Kraas et al. (2013) 257 -30.4 347 -56.2

The 1-hour DNI forecast MAE values in this study (138-164 W/m2) are higher than those achieved by Gala et al. (2013) (23-28 W/m2). There are not many studies to compare 1-hour

DNI forecast accuracy with because previous studies mostly used horizons under 1 hour ahead

(Marquez and Coimbra, 2013; Chu et al., 2013; Quesada-Ruiz et al., 2014). Overall, the accuracy of the forecast methods in this study can be better, thus it may be possible to realise benefits greater than those reported in this study.

Turning to the value of using forecasts to operate CST plants, Wittmann et al. (2008) showed that 48-hour persistence and NWP forecasts achieved 67.3%-99.9% of the revenue of a perfect forecast in clear sky conditions, and 74.0%-94.7% in cloudy sky conditions. In comparison, this study showed that 48-hour persistence, NWP, and time series analysis forecasts achieved

51.7%-99.7% of the revenue achieved by a perfect forecast in all sky conditions. Although the results from Wittmann et al. (2008) were from a total of five days of data, the results from both studies are similar. 119

The study by Kraas et al. (2013) showed that 48-hour forecasts from a NWP ensemble with post-processing reduced penalty payments by 47.6% compared to a 48-hour persistence forecast. The study by Nonnenmacher et al. (2016) showed that 36-hour NWP forecasts reduced relative reserve requirements by an average of 28.6% compared to a 36-hour persistence forecast. In comparison, the ARIMA and TAPM forecasts in this study changed the

48-hour persistence RG cost by -4.7% to 199% and the EFOR by -10% to 67%. This was for scenarios that did not use 1-hour forecasts. The results in this study showed increases and reductions in RG cost and EFOR because ARIMA and TAPM were not more accurate than 48- hour persistence at all times and sites, which was acceptable for this study’s main aim of evaluating the impact of using 1-hour forecasts. As such, the results from this study do not contradict those of previous studies. Besides the difference in forecast accuracy, this study only penalised under-generation and assumed the CST plant curtailed output to avoid over- generation. In contrast, Kraas et al. (2013) and Nonnenmacher et al. (2016) considered both over-generation and under-generation.

Sioshansi and Denholm (2010) showed that 24-hour persistence achieved at least 87% of the annual profits achieved by using a perfect forecast. Their results applied to a range of CST plant configurations, including configurations with no storage. In comparison, this study showed that using only 48-hour forecasts achieved 1.2%-79% of the net value of a perfect forecast, and using 48-hour and 1-hour forecasts achieved 59%-94%. The results by Sioshansi and Denholm

(2010) appear better because the reported annual profits only accounted for revenue and variable operation costs, and ignored penalties for under-generation and over-generation.

6.2. Reflection on the Scope of the Study

The exact financial value and reliability achieved by using forecasts to operate a CST plant depends on factors such as the forecast accuracy, local climate, CST plant configuration, and 120 electricity market and network characteristics. As such, the greater the difference between the parameters of a scenario and those of this study, the less relevant the results of this study are to that scenario. The following discussion reflects upon the parameters of this study to establish the limits of its relevance and to identify how the limits may be overcome.

6.2.1. Local climate

The local climate affects the frequency of different weather patterns which in turn affects the

DNI resource. For example, a location that is often cloudy will have lower DNI resource than a location that is often clear. It also affects the accuracy of forecast methods because forecast accuracy depends on how well the method is able to predict the weather events that affect

DNI. DNI forecast accuracy is known to be higher in clear sky conditions than in cloudy conditions (Lara-Fanego et al., 2012b), so a given forecast method will be more accurate in a location that often has clear skies compared with a location that often has clouds.

Australia is divided into six key climate groups, and these six groups are further sub-divided into a total of 27 climate classes. The four sites in this study cover three of the six key climate groups and four of the 27 climate classes, as mentioned in Section 4.1.3. Although this study only covers the temperate, grassland and subtropical climate groups, these climate groups represent all the climates within the NEM. The central part of Australia, where the NEM does not extend to, is classified as desert. The northern parts of Australia that are classified as equatorial or tropical are all north of Cairns. The NEM only extends as far north as Cairns along the eastern coast. Thus, the results of this study may be relevant throughout the NEM.

6.2.2. CST plant configuration

There are various options for the solar field, power block and storage as reviewed in Section

2.1 hence there are many possible CST plant configurations to be studied. This study only considered a parabolic trough solar field, a Rankine cycle power block, and Hitec solar salt in a two-tank setup. Furthermore, the solar multiple was fixed at 1.756 and rated generation 121 capacity was fixed at 50 MW. The storage size was alternated between 7.5 hours of storage and zero storage. If different options or sizes were used for these components, then the CST plant may operate differently even if other parameters, such as DNI resource and forecast accuracy, remained the same. Thus, the results of this study only represent a narrow range of

CST plants.

Future research may consider a greater range of CST plant configurations to build deeper understanding of how different component options and sizes affect the value achieved. In particular, this study showed that adding 7.5 hours of storage increased net value more than by using 1-hour forecasts. As there are a range of possible storage sizes, investigating different storage sizes may yield a critical storage size below which using 1-hour forecasts is more valuable.

6.2.3. Electricity market

Electricity markets around the world differ by their market structure, such as the spatial and temporal resolution employed for determining electricity prices (Baritaud et al., 2016), or the deadline for submitting initial bids and updated bids (Botterud et al., 2010; Furió, 2011). The market structure restricts how the CST plant may operate, thus the optimum operation schedule in one market may be suboptimal in another.

When the electricity price is determined by market forces instead of tariffs, the generation mix affects the electricity price because the electricity price depends on the relative amounts of supply and demand. For example, in a market with many large capacity solar power plants

(CST or PV) without storage, the high output during clear skies will lower the daytime electricity price. It has been shown that the solar penetration affects the value of GHI forecasting for PV plants (Martinez-Anido et al., 2016). Lower electricity prices will reduce the revenue and net value of a CST plant without storage. A CST plant with storage will be able to shift its generation to evening or night periods with higher electricity prices, but storage 122 capacity and maximum charging rate will limit the amount of generation it can shift. Thus the revenue and net value of a CST plant with storage may also reduce, but it will not be as large as the reduction for a CST plant without storage. The data from 2005 used in this study are prior to large uptake of solar in the NEM (AER, 2014), so the financial value would be different if the data had higher solar penetration instead. Consequently, the financial value results of this study are most relevant to the NEM in 2005, and electricity markets with similar structure to the NEM and similar solar penetration as the NEM had in 2005.

Studies on the value of using DNI forecasts to operate CST plants in electricity markets other than the NEM may adapt the research method in this study. The first modification would be to select the long-term and short-term forecast horizons that suit the local electricity market under consideration. This will change the number of forecasts per day and the frequency of running optimisations to determine bids. Further modifications may be necessary to ensure that the forecast horizons are compatible with the time step of the CST model used. When the bids are finalised, real DNI can be used to determine whether the CST plant is able to meet its bids. After that, the procedure of calculating financial value from the relevant electricity prices and penalty rates, and calculating CST plant reliability from analysing met and unmet bids, can follow that used in this study.

6.2.4. Value perspectives

The value of using short-term forecasts to operate a CST plant was represented in this study by financial value and reliability because these reflect the interests of the CST plant operator and the network operator, respectively. The review of studies on the value of CST in Chapter 2 showed that different perspectives can favour different CST plant configurations for maximising value. A similar outcome was found in this study, where higher net value was achieved by a CST plant with storage and without 1-hour forecasts, in contrast to higher reliability being achieved by a CST plant without storage and with 1-hour forecasts. It may be

123 possible to bring both perspectives to favour the same CST plant configuration, for example by setting penalties such that the higher net value is also achieved by a CST plant without storage and with short-term forecasts.

6.2.5. Short-term forecasts

This study used 1-hour persistence as the short-term forecast method. There are other methods that outperform persistence at forecasting DNI at short-term horizons, and they can forecast DNI at horizons below 1 hour ahead, as shown in Chapter 3. Previous studies have not investigated the use of short-term forecasts to operate a CST plant, so the results of this study, although limited in scope, demonstrated that future research should consider the use of short- term forecasts.

Future research that uses multiple short-term forecast methods could determine the effect of short-term forecast accuracy on value. Kaur et al. (2016) showed that forecasts with higher accuracy up to the 75-minute horizon provided greater increases to PV plant reliability. If value steadily increases with forecast accuracy, it would justify continuing to improve short-term forecast methods. Alternatively, if value approaches an asymptote then a particular forecast accuracy value may be identified as a practical maximum accuracy to aim for. Besides that, short-term forecasts under the 1-hour horizon may be considered because the NEM allows updated bids to be submitted up to 5 minutes before dispatch. This research cannot be conducted using the LKT model because it needs a CST model with time steps at higher resolution than 1-hour.

6.2.6. CST model

A limitation was imposed on this study by the 1-hour time step of the LKT model. This prevents simulations taking full advantage of the NEM’s allowance for updated bids to be submitted up to 5 minutes before dispatch. Besides that, the 1-hour resolution cannot reflect events that occur in less than an hour and affect CST plant operation in the simulation results, such as fast 124 moving clouds. However, the fact that this study considers updated bids submitted 1 hour before dispatch is a vast improvement over previous studies that did not consider short-term forecasts. Another limitation of the LKT model is that it approximates CST plant operation by only considering the conservation of energy between components. It does not check temperatures and ignores the conservation of mass flow, both of which dictate the possible operation of a real CST plant. Thus the simulated operation may not necessarily be achievable in reality. These limitations may be overcome by using a CST model that has higher resolution time steps and checks mass flow rate and temperatures to simulate operation. These changes will increase the number of calculations for simulating operation, so the simulations will take longer to complete.

6.3. Chapter Summary

The 48-hour and 1-hour DNI forecast methods used in this study are less accurate compared to those reported in past studies when all sky conditions over at least one year is considered. This study produced financial value results that are similar to those of past studies where the use of long-term forecasts was compared against a perfect forecast. The reliability results were different to those of past studies where the use of long-term forecasts was compared against a long-term persistence forecast. The results differed because the other studies compared state of the art NWP forecasts against persistence whereas this study did not. Another reason for the difference was that the other studies considered both over-generation and under- generation errors whereas this study assumed that the CST plant avoided over-generation errors through curtailing output and was thus only penalised for under-generation errors.

A reflection on the scope of this study concluded that the results are restricted to two CST plant configurations, but they are applicable throughout the NEM. Possibilities for further research include changing the type or size of CST plant components to investigate different 125 configurations, changing the short-term forecast method to investigate different forecast accuracies, and changing the CST model to enable investigation in short-term horizons less than 1 hour ahead or to more accurately simulate CST plant operation.

In the next chapter, the entire study is summarised.

126

7. Conclusion6

Global CST generation capacity has been growing because it can replace fossil fuel generators in the electricity sector to reduce greenhouse gas emissions. Solar variability can make it problematic to use CST power, but the magnitude of the problem can be reduced by using DNI forecasts. CST generation capacity in Australia is expected to grow due to Australia’s excellent solar resource. However, DNI forecasting in Australia has not been widely researched and there are no studies on using DNI forecasts to operate CST plants in Australia. The largest electricity market and network in Australia is the NEM, hence this study aimed to evaluate the value of using DNI forecasts to operate a CST plant in the NEM.

A review of studies on the value of CST plants showed that the value could be calculated from the CST plant’s perspective in terms of the financial value of selling electricity. The value could also be calculated from the network operator’s perspective in terms of how the CST plant reduces the cost of supplying electricity or how reliable its electricity supply is. Results across studies showed that a CST configuration that maximises value from the CST plant’s perspective may not necessarily maximise value from the network operator’s perspective. The majority of studies assumed that the CST plant operated with perfect knowledge of available DNI. In reality, a CST plant operates using DNI forecasts, and forecast errors can lead to penalties that reduce the value. Studies on the value of using DNI forecasts to operate a CST plant showed that higher value is achieved by more accurate forecasts. None of the studies had considered forecasts at horizons less than 24 hours ahead. The forecasts used were at the 24-hour, 36- hour and 48-hour horizons.

6 Some of the material presented in this chapter has been previously published in: Law, E.W., Kay, M., Taylor, R.A., 2016, Calculating the financial value of a concentrated solar thermal plant operated using direct normal irradiance forecasts, Solar Energy, 125, p. 267-281 https://dx.doi.org/10.1016/j.solener.2015.12.031

Different DNI forecasting methods are recommended in the literature for particular horizons because they are most accurate at predicting the events that affect DNI over those horizons.

Persistence is useful at horizons that are several minutes ahead because DNI usually doesn’t change significantly in that time and the forecast can be produced quickly. CMVs are best at horizons from several minutes to several hours ahead because they can predict cloud cover better than persistence by extrapolating the movement of currently observed clouds. NWP models are best at horizons beyond several hours ahead because at these time frames future cloud cover is more dependent on cloud formation and dispersion rather than the movement of currently observed clouds. Despite the recommendations in the literature, there has not been a review of DNI forecast accuracy. Thus, a review was conducted for this study. The review confirmed that the forecast methods are indeed most accurate at their recommended horizons and summarised the best forecast accuracy reported in the literature at horizons from 5 minutes to 3 days ahead.

This study worked towards achieving its aim by simulating the use of 1-hour and 48-hour DNI forecasts to operate a CST plant in the NEM. The 48-hour forecasts were used to determine initial bids to place in the NEM, while the 1-hour forecasts were used to determine updated bids, which are accepted up to 5 minutes before dispatch. The financial value to the CST plant operator was demonstrated through calculating the net value, which was the revenue earned from selling electricity less the RG cost of causing generation shortfalls in the electricity market.

Another calculated cost was the dump cost which represented the loss of potential revenue from dumping thermal energy that could not be stored or used for generating electricity. The reliability value to the network operator was demonstrated by calculating the EFOR, which measures the likelihood that the CST plant causes a generation shortfall when attempting to meet its next bid. A CST plant with 7.5 hours of storage and a CST plant with no storage were considered because the number of commercial CST plants with no storage (37 plants) almost

128 equals the number with some amount of storage (39 plants). Four sites in the NEM were used to include operation in different climate zones and with different electricity prices. One year of data at each site was used to include seasonal differences in DNI resource and weather.

Results showed that there was value in using 1-hour forecasts together with 48-hour forecasts to operate a CST plant with or without storage in the NEM. Using 1-hour forecasts to operate a

CST plant with storage increased net value by $0.90-2.07 million and reduced EFOR by 16-42 percentage points compared to not using 1-hour forecasts. For a CST plant without storage, the results were a net value increase of $0.76-$3.10 million and an EFOR reduction of 16-42 percentage points. Between 57%-99% of the increase in net value was found to come from reduced RG costs, unless the 48-hour forecast had a large MBE in which case the RG cost reduction contributed 93%-177%. The rest of the increase in net value was due to higher revenue, which was achieved by reducing the loss of potential revenue (dump cost). The increases in financial value and reliability were due to 1-hour forecasts being more accurate overall than 48-hour forecasts. This meant the CST plant wasted less thermal energy and was more likely to meet its updated bids in the electricity market.

A CST plant with storage that used 1-hour forecasts achieved the highest net value at all sites, so it is valuable for a CST plant to have both short-term forecasts and storage. Besides that, results showed that a CST plant without storage that used 1-hour forecasts was more reliable, but a CST plant with storage that did not use 1-hour forecasts was more financially valuable.

This demonstrated that different perspectives on value could lead to favouring different CST plant configurations.

The net value achieved by using 1-hour forecasts was compared against that by using a perfect

48-hour forecast because a perfect 48-hour forecast achieves the theoretical maximum net value. The comparison showed that, for a CST plant with storage, using 1-hour forecasts achieved 76%-94% of the theoretical maximum net value, and the achievement was 59%-87% 129 for a CST plant without storage. When the net value was normalised by dividing by the generated electricity, results showed that using 1-hour forecasts increased the normalised net value to 85%-100% of that of a perfect 48-hour forecast. This suggests that while using short- term forecasts may not maximise total net value, it may maximise the net value per unit electricity generated.

Overall, this study aimed to evaluate the value of using DNI forecasts to operate a CST plant in the NEM. A review of CST value in the literature found that no studies used short-term forecasts below the 24-hour horizon. A review was also conducted on DNI forecasting, and it was used to produce a summary of the best DNI forecast accuracy reported in the literature at horizons from 5 minutes to 3 days ahead. The study simulated CST plant operation in the NEM.

Results showed that a CST plant should use short-term forecasts to operate in the NEM because they increase financial value and reliability. This is true regardless of whether the CST plant has storage or not. The results thus fulfilled the aim of the study. The results also support the development and improvement of short-term DNI forecasts by showing that they are beneficial to use to operate a CST plant in the NEM. Future studies to gain further understanding of the value of using short-term forecasts to operate CST plants can use the method in this study and modifying it to include a different range of CST plant configurations and short-term forecast methods. More extensive modifications may be made to apply the method to calculate financial value and reliability in other electricity markets. This study has contributed towards the field of DNI forecasting and CST plant value by providing a snapshot of the best DNI forecast accuracy at the 5-minute to 3-day horizons, and producing the first results that prove it is valuable for a CST plant to use short-term forecasts less than 24 hours ahead to operate within an electricity market that allows short-term updates to market bids.

130

References

AEMC, 2016a, National Electricity Rules Version 82, Australian Energy Market Commission, Australia AEMC, 2016b, Schedule of reliability settings, Australan Energy Market Commission. http://www.aemc.gov.au/getattachment/5e81890c-0a96-43bb-b8c0- 9cc9532c4bc5/Schedule-of-Reliability-Settings-(MPC-and-CPT-(4).aspx AEMC Reliability Panel, 2014, Reliability Standard and Reliability Settings Review 2014, Final Report, Australian Energy Market Commission http://www.aemc.gov.au/Markets- Reviews-Advice/Reliability-Standard-and-Settings-Review-2014 AEMO, 2010a, An introduction to Australia's National Electricity Market, Australian Energy Market Operator. http://www.aemo.com.au/About-the-Industry/Energy- Markets/National-Electricity-Market AEMO, 2010b, Guide to ancillary services in the National Electricity Market, Australian Energy Market Operator, pp. 14. http://www.aemo.com.au/Electricity/Market- Operations/Ancillary-Services AEMO, 2010c, Guidebook for forced outage data recording: definitions and assumptions, Australian Energy Market Operator. http://www.aemo.com.au/Electricity/Policies- and-Procedures/Reserve-Management/Forced-Outage-Data-Working-Group AEMO, 2011, Spot market operations timetable, Australian Energy Market Operator. http://www.aemo.com.au/Electricity/Market-Operations/Dispatch/Spot-Market- Operations-Timetable AEMO, 2013a, 100 per cent Renewables Study Modelling Outcomes, Australian Energy Market Operator, Australia, pp. 111. http://www.climatechange.gov.au/reducing- carbon/aemo-report-100-renewable-electricity-scenarios AEMO, Aggregate Price and Demand Data Files, [Web page], Australian Energy Market Operator. (accessed 12 September 2013). AEMO, 2015, 2015 National Electricity Forecasting Report, Australian Energy Market Operator. http://www.aemo.com.au/Datasource/Archives/Archive1391 AER, 2014, State of the energy market 2014, Australian Energy Regulator. https://www.aer.gov.au/publications/state-of-the-energy-market-reports/state-of- the-energy-market-2014 AER, 2015, State of the energy market 2015, Australian Energy Regulator. http://www.aer.gov.au/publications/state-of-the-energy-market-reports/state-of-the- energy-market-2015 Alam, S., Kaushik, S.C., Garg, S.N., 2006, Computation of beam solar radiation at normal incidence using artificial neural network, Renewable Energy 31, p. 1483-1491 http://dx.doi.org/10.1016/j.renene.2005.07.010 ASHRAE, 1972, Handbook of fundamentals, American Society of Heating, Refrigerating and Air- Conditioning Engineers, Atlanta, GA. AUSTELA, The NREL System Advisor Model for Australian CSP Stakeholders (SAM), [Web page], Australian Solar Thermal Energy Association. (accessed 6 March 2014). Badescu, V., Gueymard, C.A., Cheval, S., Oprea, C., Baciu, M., Dumitrescu, A., Iacobescu, F., Milos, I., Rada, C., 2012, Computing global and diffuse solar hourly irradiation on clear sky. Review and testing of 54 models, Renewable and Sustainable Energy Reviews 16, p. 1636-1656 http://dx.doi.org/10.1016/j.rser.2011.12.010 Baharoon, D.A., Rahman, H.A., Omar, W.Z.W., Fadhl, S.O., 2015, Historical development of concentrating solar power technologies to generate clean electricity efficiently – A

131

review, Renewable and Sustainable Energy Reviews 41, p. 996-1027 http://dx.doi.org/10.1016/j.rser.2014.09.008 Bannehr, L., Rohn, M., Warnecke, G., 1996, A functional analytic method to derive displacement vector fields from satellite image sequences, International Journal of Remote Sensing 17, p. 383-392 Baritaud, M., Spruck, S., Verstraeten, C., Wittenstein, M., Muller, S., Razali, N.M., Scott, J., Vailles, C., 2016, Re-powering markets: market design and regulation during the transition to low-carbon power systems, International Energy Agency, pp. 246. http://www.iea.org/publications/freepublications/publication/re-powering-markets- market-design-and-regulation-during-the-transition-to-low-carbon-power- systems.html Benmouiza, K., Cheknane, A., 2013, Forecasting hourly global solar radiation using hybrid k- means and nonlinear autoregressive neural network models, Energy Conversion and Management 75, p. 561-569 http://dx.doi.org/10.1016/j.enconman.2013.07.003 Benoit, H., Spreafico, L., Gauthier, D., Flamant, G., 2016, Review of heat transfer fluids in tube- receivers used in concentrating solar thermal systems: Properties and heat transfer coefficients, Renewable and Sustainable Energy Reviews 55, p. 298-315 http://dx.doi.org/10.1016/j.rser.2015.10.059 Biga, A.J., Rosa, R., 1979, Contribution to the study of the solar radiation climate of Lisbon, Solar Energy 23, p. 61-67 http://dx.doi.org/10.1016/0038-092X(79)90044-6 Bird, R.E., Hulstrom, R.L., 1981, A simplified clear sky model for direct and diffuse insolation on horizontal surfaces, Solar Energy Research Institute, Golden, Colorado. http://rredc.nrel.gov/solar/models/clearsky/ Blanksby, C., Bennett, D., Langford, S., 2013, Improvement to an existing satellite data set in support of an Australia solar atlas, Solar Energy 98, Part B, p. 111-124 http://dx.doi.org/10.1016/j.solener.2012.10.026 BOM, Annual Climate Summary - 2005, [Web page], Bureau of Meteorology. (accessed 8 August 2016). BOM, 2012a, Australian one minute solar radiation data, Bureau of Meteorology. www.bom.gov.au/climate/data/oneminsolar/IDCJAC0022_product_notes_v1.2.doc BOM, 2012b, Gridded hourly solar direct normal irradiance metadata, Bureau of Meteorology. http://www.bom.gov.au/climate/data- services/docs/IDCJAD0027_metadata_gridded_hourly_DNI.pdf BOM, 2012c, Gridded hourly solar global horizontal irradiance metadata, Bureau of Meteorology. http://www.bom.gov.au/climate/data- services/docs/IDCJAD0026_metadata_gridded_hourly_GHI.pdf BOM, Average daily solar exposure, [Web page], Bureau of Meteorology. (accessed 8 August 2016). BOM, Bureau of Meteorology Climate Zones, [Web page], Bureau of Meteorology. (accessed 5 August 2016). BOM, One minute solar data, [Web page], Bureau of Meteorology. (accessed 31 July 2016). Bosch, J.L., Zheng, Y., Kleissl, J., 2013, Deriving cloud velocity from an array of solar radiation measurements, Solar Energy 87, p. 196-203 http://dx.doi.org/10.1016/j.solener.2012.10.020 Botterud, A., Wang, J., Miranda, V., Bessa, R.J., 2010, Wind Power Forecasting in U.S. Electricity Markets, The Electricity Journal 23, p. 71-82 http://dx.doi.org/10.1016/j.tej.2010.03.006

132

Box, G.E., Jenkins, G.M., Reinsel, G.C., 1994, Time Series Analysis: Forecasting and Control, 3 ed., Prentice-Hall, Inc, USA. Bradbury, D.L., Fujita, T.T., 1968, Computation of height and velocity of clouds from dual, whole-sky, time-lapse picture sequences, University of Chicago. Breitkreuz, H., Schroedter-Homscheidt, M., Holzer-Popp, T., Dech, S., 2009, Short-Range Direct and Diffuse Irradiance Forecasts for Solar Energy Applications Based on Aerosol Chemical Transport and Numerical Weather Modeling, Journal of Applied Meteorology and Climatology 48, p. 1766-1779 http://dx.doi.org/10.1175/2009jamc2090.1 CEC, 2016, Clean Energy Australia Report 2015, Clean Energy Council. https://www.cleanenergycouncil.org.au/policy-advocacy/reports/clean-energy- australia-report.html Channon, S.W., Eames, P.C., 2014, The cost of balancing a parabolic trough concentrated solar power plant in the Spanish electricity spot markets, Solar Energy 110, p. 83-95 http://dx.doi.org/10.1016/j.solener.2014.08.036 Chen, R., Kang, E., Ji, X., Yang, J., Wang, J., 2007, An hourly solar radiation model under actual weather and terrain conditions: A case study in Heihe river basin, Energy 32, p. 1148- 1157 http://dx.doi.org/10.1016/j.energy.2006.07.006 Chow, C.W., Urquhart, B., Lave, M., Dominguez, A., Kleissl, J., Shields, J., Washom, B., 2011, Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed, Solar Energy 85, p. 2881-2893 http://dx.doi.org/10.1016/j.solener.2011.08.025 Christensen, T.M., Hurn, A.S., Lindsay, K.A., 2012, Forecasting spikes in electricity prices, International Journal of Forecasting 28, p. 400-411 http://dx.doi.org/10.1016/j.ijforecast.2011.02.019 Chu, Y., Pedro, H.T.C., Coimbra, C.F.M., 2013, Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning, Solar Energy 98, Part C, p. 592-603 http://dx.doi.org/10.1016/j.solener.2013.10.020 Coakley, J.A., Bretherton, F.P., 1982, Cloud cover from high‐resolution scanner data: Detecting and allowing for partially filled fields of view, Journal of Geophysical Research: Oceans (1978–2012) 87, p. 4917-4932 Coimbra, C.F.M., Kleissl, J., Marquez, R., 2013, Chapter 8 - Overview of Solar-Forecasting Methods and a Metric for Accuracy Evaluation, in: Solar Energy Forecasting and Resource Assessment, Academic Press, Boston, pp. 171-194 http://dx.doi.org/10.1016/B978-0-12-397177-7.00008-5 Copper, J.K., Sproul, A.B., 2012, Comparative study of mathematical models in estimating solar irradiance for Australia, Renewable Energy 43, p. 130-139 http://dx.doi.org/10.1016/j.renene.2011.11.050 Cote, S., Tatnall, A., 1995, A neural network-based method for tracking features from satellite sensor images, International Journal of Remote Sensing 16, p. 3695-3701 Crispim, E.M., Ferreira, P.M., Ruano, A.E., 2008, Prediction of the solar radiation evolution using computational intelligence techniques and cloudiness indices, International Journal of Innovative Computing, Information and Control 4, p. 1121-1133 Dagestad, K.-F., Olseth, J.A., 2007, A modified algorithm for calculating the cloud index, Solar energy 81, p. 280-289 Daniels, J., Velden, C., Bresky, N., Irving, A., 2004, Status of the NOAA/NESDIS operational satellite wind product system: Recent improvements, new products, product quality, and future plans, Proceedings of the 7th International Winds Workshop, p. 1-17 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, p. 668-680 http://dx.doi.org/10.1016/j.solener.2014.06.018 Denholm, P., Wan, Y.-H., Hummon, M., Mehos, M., 2013, An Analysis of Concentrating Solar Power with Thermal Energy Storage in a California 33% Renewable Scenario, National 133

Renewable Energy Laboratory. http://www.brightsourceenergy.com/stuff/contentmgr/files/0/bff7246d1744f938b5a 3e756febd14d6/attachment/nrel_33_.pdf Diagne, M., David, M., Lauret, P., Boland, J., Schmutz, N., 2013, Review of solar irradiance forecasting methods and a proposition for small-scale insular grids, Renewable and Sustainable Energy Reviews 27, p. 65-76 http://dx.doi.org/10.1016/j.rser.2013.06.042 Djebbar, R., Morris, R., Thevenard, D., Perez, R., Schlemmer, J., 2012, Assessment of SUNY Version 3 Global Horizontal and Direct Normal Solar Irradiance in Canada, Energy Procedia 30, p. 1274-1283 http://dx.doi.org/10.1016/j.egypro.2012.11.140 Duffie, J.A., Beckman, W.A., 2013, Solar Engineering of Thermal Processes, 4th ed., John Wiley & Sons, Inc., USA. Dunham, M.T., Iverson, B.D., 2014, High-efficiency thermodynamic power cycles for concentrated solar power systems, Renewable and Sustainable Energy Reviews 30, p. 758-770 http://dx.doi.org/10.1016/j.rser.2013.11.010 Duque, J., Stephens, A., Hoarau, M., Biddle, P., Mathur, S., 2012, Frequency control ancillary services, Australian Energy Market Operator. ECMWF, IFS documentation, [Web page]. (accessed 14 September 2013). Eissa, Y., Marpu, P.R., Gherboudj, I., Ghedira, H., Ouarda, T.B.M.J., Chiesa, M., 2013, Artificial neural network based model for retrieval of the direct normal, diffuse horizontal and global horizontal irradiances using SEVIRI images, Solar Energy 89, p. 1-16 http://dx.doi.org/10.1016/j.solener.2012.12.008 Elliston, B., MacGill, I., 2010, The potential role of forecasting for integrating solar generation into the Australian National Electricity Market, Solar2010, the 48th AuSES Annual Conference, Canberra, ACT, Australia, pp. 11. EMC, Global Forecasting System, [Web page], Environmental Modeling Center. (accessed 14 September 2013). Erbs, D.G., Klein, S.A., Duffie, J.A., 1982, Estimation of the diffuse radiation fraction for hourly, daily and monthly-average global radiation, Solar Energy 28, p. 293-302 http://dx.doi.org/10.1016/0038-092X(82)90302-4 Evans, A.N., 2006, Cloud motion analysis using multichannel correlation-relaxation labeling, Geoscience and Remote Sensing Letters, IEEE 3, p. 392-396 Fritz, S., Winston, J.S., 1962, Synoptic use of radiation measurements from satellite Tiros II, Monthly Weather Review 90, p. 1-9 Fujita, T.T., 1968, Present status of cloud velocity computations from ATS-1 and ATS-3, COSPAR Space Res 9, p. 557-570 Furió, D., 2011, A survey on the Spanish Electricity Intraday Market, Estudios de Economía Aplicada 29, p. 1-19 Gala, Y., Fernández, Á., Díaz, J., Dorronsoro, J., 2013, Support Vector Forecasting of Solar Radiation Values, in: J.-S. Pan, M. Polycarpou, M. Woźniak, A.P.L.F. Carvalho, H. Quintián, E. Corchado (Eds.) Hybrid Artificial Intelligent Systems, Springer Berlin Heidelberg, pp. 51-60 10.1007/978-3-642-40846-5_6 Geoscience Australia, ABARE, 2010, Australian Energy Resource Assessment, Canberra, Australia. Geoscience Australia, BREE, 2014, Australian Energy Resource Assessment - Second Edition, Department of Industry, Canberra. http://www.ga.gov.au/corporate_data/79675/79675_AERA.pdf Gerstmaier, T., Bührer, M., Röttger, M., Gombert, A., Hansen, C.W., Stein, J.S., 2012, How predictable is DNI? An evaluation of day ahead DNI forecasts from four different

134

providers, AIP Conference Proceedings 1477, p. 305-308 doi:http://dx.doi.org/10.1063/1.4753892 González, J., Serrano, E., Wiesenberg, R., 2010, DNI forecasting using coupled WRF and MRM, enhanced with a neural network for CSP applications, Proceedings of SolarPACES Conference. Perpignan, France, p. Grell, G.A., Dudhia, J., Stauffer, D.R., 2008, A Description of the Fifth-Generation Penn State/NCAR Mesoscale Model (MM5), NCAR Technical Note, National Centre for Atmospheric Research, Boulder, USA. Gueymard, C.A., 2008, REST2: High-performance solar radiation model for cloudless-sky irradiance, illuminance, and photosynthetically active radiation – Validation with a benchmark dataset, Solar Energy 82, p. 272-285 http://dx.doi.org/10.1016/j.solener.2007.04.008 Gueymard, C.A., 2012a, Clear-sky irradiance predictions for solar resource mapping and large- scale applications: Improved validation methodology and detailed performance analysis of 18 broadband radiative models, Solar Energy 86, p. 2145-2169 http://dx.doi.org/10.1016/j.solener.2011.11.011 Gueymard, C.A., 2012b, Temporal variability in direct and global irradiance at various time scales as affected by aerosols, Solar Energy 86, p. 3544-3553 http://dx.doi.org/10.1016/j.solener.2012.01.013 Gul, M.S., Muneer, T., Kambezidis, H.D., 1998, Models for obtaining solar radiation from other meteorological data, Solar Energy 64, p. 99-108 http://dx.doi.org/10.1016/S0038- 092X(98)00048-6 Hammer, A., Heinemann, D., Hoyer, C., Kuhlemann, R., Lorenz, E., Müller, R., Beyer, H.G., 2003, Solar energy assessment using remote sensing technologies, Remote Sensing of Environment 86, p. 423-432 http://dx.doi.org/10.1016/S0034-4257(03)00083-X Hammer, A., Heinemann, D., Hoyer, C., Lorenz, E., 2001, Satellite based short-term forecasting of solar irradiance—comparison of methods and error analysis, The 2001 EUMETSAT Meteorological Satellite Data User’s Conference, p. 677-684 Hammer, A., Heinemann, D., Lorenz, E., Lückehe, B., 1999, Short-term forecasting of solar radiation: a statistical approach using satellite data, Solar Energy 67, p. 139-150 Hammer, A., Lorenz, E., Kemper, A., Heinemann, D., Beyer, H., Schumann, K., Schwandt, M., 2009, Direct normal irradiance for CSP based on satellite images of Meteosat Second Generation, Proceedings of 15th SolarPACES Conference, Berlin, Germany, p. Hayden, C., Velden, C., 1991, Quality control and assimilation experiments with satellite derived wind estimates. Preprints, 9th Conference on Numerical Weather Prediction, p. 19-23 Hayden, C.M., Nieman, S.J., 1996, A primer for tuning the automated quality control system and for verifying satellite-measured drift winds, US Department of Commerce, National Oceanic and Atmospheric Administration, National Environmental Satellite, Data, and Information Service. Haykin, S., 1999, Neural Networks: A Comprehensive Foundation, 2nd ed., Prentice Hall. Heinemann, D., Lorenz, E., Girodo, M., 2006a, Forecasting of solar radiation, Solar energy resource management for electricity generation from local level to global scale. Nova Science Publishers, New York p. Heinemann, D., Lorenz, E., Girodo, M., 2006b, Solar irradiance forecasting for the management of solar energy systems, Energy and Semiconductor Research Laboratory, Energy Meteorology Group, Oldenburg University p. Hibberd, M., 2011, TAPM 4 modelling of Wagerup meteorology and 25A dispersion, Commonwealth Scientific and Industrial Research Organisation, Australia, pp. 96. https://publications.csiro.au/rpr/pub?list=BRO&pid=csiro:EP107076&sb=RECENT&n=1 4&rpp=25&page=178&tr=5040&dr=all&dc4.browseYear=2011

135

Higgs, H., Lien, G., Worthington, A.C., 2015, Australian evidence on the role of interregional flows, production capacity, and generation mix in wholesale electricity prices and price volatility, Economic Analysis and Policy 48, p. 172-181 http://dx.doi.org/10.1016/j.eap.2015.11.008 Hinkley, J.T., Hayward, J.A., Curtin, B., Wonhas, A., Boyd, R., Grima, C., Tadros, A., Hall, R., Naicker, K., 2013, An analysis of the costs and opportunities for concentrating solar power in Australia, Renewable Energy 57, p. 653-661 http://dx.doi.org/10.1016/j.renene.2013.02.020 Hoff, T.E., Perez, R., Kleissl, J., Renne, D., Stein, J., 2013, Reporting of irradiance modeling relative prediction errors, Progress in Photovoltaics: Research and Applications 21, p. 1514-1519 http://dx.doi.org/10.1002/pip.2225 Holmlund, K., van de Berg, L., Gustafsson, J., de Smet, A., Carranza, M., Dew, G., Borde, R., 2010, Current status of EUMETSAT operational winds, Proceedings of the 10th International Winds Workshop, p. 22-26 Hoyt, D.V., 1978, A model for the calculation of solar global insolation, Solar Energy 21, p. 27- 35 http://dx.doi.org/10.1016/0038-092X(78)90113-5 Huang, H., Yoo, S., Yu, D., Huang, D., Qin, H., 2012, Correlation and local feature based cloud motion estimation, Proceedings of the Twelfth International Workshop on Multimedia Data Mining, p. 1-9 Hurley, P., 2008, TAPM V4 Part 1: Technical Description, CSIRO. http://www.cmar.csiro.au/research/tapm/docs/tapm_v4_technical_paper_part1.pdf IEA, 2014, Technology Roadmap: Solar Thermal Electricity - 2014 edition, International Energy Agency, France, pp. 52. http://www.iea.org/publications/freepublications/publication/technology-roadmap- solar-thermal-electricity---2014-edition.html IEEE, 2007, IEEE Standard Definitions for Use in Reporting Electric Generating Unit Reliability, Availability, and Productivity, IEEE Std 762-2006 (Revision of IEEE Std 762-1987) p. 1- 75 http://dx.doi.org/10.1109/IEEESTD.2007.335902 Ineichen, P., 2006, Comparison of eight clear sky broadband models against 16 independent data banks, Solar Energy 80, p. 468-478 http://dx.doi.org/10.1016/j.solener.2005.04.018 Ineichen, P., 2008a, A broadband simplified version of the Solis clear sky model, Solar Energy 82, p. 758-762 http://dx.doi.org/10.1016/j.solener.2008.02.009 Ineichen, P., 2008b, Comparison and validation of three global-to-beam irradiance models against ground measurements, Solar Energy 82, p. 501-512 http://dx.doi.org/10.1016/j.solener.2007.12.006 Inman, R.H., Pedro, H.T.C., Coimbra, C.F.M., 2013, Solar forecasting methods for renewable energy integration, Progress in Energy and Combustion Science 39, p. 535-576 http://dx.doi.org/10.1016/j.pecs.2013.06.002 Iqbal, M., 1983, An introduction to solar radiation, Access Online via Elsevier. James, G., Hayward, J., 2012, AEMO 100% Renewable Energy Study, CSIRO, Newcastle, Australia. https://publications.csiro.au/rpr/download?pid=csiro:EP126455&dsid=DS2 Janjic, Z., Gall, R., Pyle, M.E., 2010, Scientific Documentation for the NMM Solver, NCAR Technical Note, National Center for Atmospheric Research. http://nldr.library.ucar.edu/repository/collections/TECH-NOTE-000-000-000-845 Ji, W., Chee, K.C., 2011, Prediction of hourly solar radiation using a novel hybrid model of ARMA and TDNN, Solar Energy 85, p. 808-817 http://dx.doi.org/10.1016/j.solener.2011.01.013 Jorgenson, J., Denholm, P., Mehos, M., 2014, Estimating the value of utility-scale solar technologies in California under a 40% renewable portfolio standard, National

136

Renewable Energy Laboratory, Golden, Colorado. http://www.nrel.gov/docs/fy14osti/61685.pdf Jorgenson, J., Denholm, P., Mehos, M., Turchi, C., 2013, Estimating the performance and economic value of multiple concentrating solar power technologies in a production cost model, National Renewable Energy Laboratory, Golden, Colorado, pp. 37. http://www.nrel.gov/docs/fy14osti/58645.pdf Kalnay, E., Kanamitsu, M., Kistler, R., Collins, W., Deaven, D., Gandin, L., Iredell, M., Saha, S., White, G., Wollen, J., Zhu, Y., Leetmaa, A., Reynolds, R., Chelliah, M., Ebisuzaki, W., Higgins, W., Janowiak, J., Mo, K., Ropelewski, C., Wang, J., Jenne, R., Joseph, D., 1996, The NCEP/NCAR 40-year reanalysis project, Bulletin of the American Meteorological Society 77, p. 437-470 Kamachi, M., 1989, Advective surface velocities derived from sequential images for rotational flow field: Limitations and applications of maximum cross‐correlation method with rotational registration, Journal of Geophysical Research: Oceans (1978–2012) 94, p. 18227-18233 Kasten, F., Czeplak, G., 1980, Solar and terrestrial radiation dependent on the amount and type of cloud, Solar Energy 24, p. 177-189 http://dx.doi.org/10.1016/0038-092X(80)90391- 6 Kaur, A., Nonnenmacher, L., Pedro, H.T.C., Coimbra, C.F.M., 2016, Benefits of solar forecasting for energy imbalance markets, Renewable Energy 86, p. 819-830 http://dx.doi.org/10.1016/j.renene.2015.09.011 Kaushika, N.D., Reddy, K.S., 2000, Performance of a low cost solar paraboloidal dish steam generating system, Energy Conversion and Management 41, p. 713-726 http://dx.doi.org/10.1016/S0196-8904(99)00133-8 Kleissl, J., 2010, Current State of the Art in Solar Forecasting, California Renewable Energy Forecasting, Resource Data and Mapping, University of California. Konrad, J., Dubois, E., 1992, Bayesian estimation of motion vector fields, IEEE Transactions on Pattern Analysis and Machine Intelligence 14, p. 910-927 Kostylev, V., Pavlovski, A., 2011, Solar Power Forecasting Performance–Towards Industry Standards, 1st International Workshop on the Integration of Solar Power into Power Systems, Aarhus, Denmark, p. Kraas, B., Schroedter-Homscheidt, M., Madlener, R., 2013, Economic merits of a state-of-the- art concentrating solar power forecasting system for participation in the Spanish electricity market, Solar Energy 93, p. 244-255 http://dx.doi.org/10.1016/j.solener.2013.04.012 Kühnert, J., Lorenz, E., Heinemann, D., 2013, Chapter 11 - Satellite-Based Irradiance and Power Forecasting for the German Energy Market, in: Solar Energy Forecasting and Resource Assessment, Academic Press, Boston, pp. 267-297 http://dx.doi.org/10.1016/B978-0- 12-397177-7.00011-5 Kuravi, S., Trahan, J., Goswami, D.Y., Rahman, M.M., Stefanakos, E.K., 2013, Thermal energy storage technologies and systems for concentrating solar power plants, Progress in Energy and Combustion Science 39, p. 285-319 http://dx.doi.org/10.1016/j.pecs.2013.02.001 Lara-Fanego, V., Ruiz-Arias, J.A., Pozo-VÃzquez, D., Gueymard, C.A., Tovar-Pescador, J., 2012a, Evaluation of DNI forecast based on the WRF mesoscale atmospheric model for CPV applications, 8TH INTERNATIONAL CONFERENCE ON CONCENTRATING PHOTOVOLTAIC SYSTEMS, 1477, p. 6 http://dx.doi.org/10.1063/1.4753895 Lara-Fanego, V., Ruiz-Arias, J.A., Pozo-Vázquez, D., Santos-Alamillos, F.J., Tovar-Pescador, J., 2012b, Evaluation of the WRF model solar irradiance forecasts in Andalusia (southern Spain), Solar Energy 86, p. 2200-2217 http://dx.doi.org/10.1016/j.solener.2011.02.014

137

Larson, V.E., 2013, Chapter 12 - Forecasting Solar Irradiance with Numerical Weather Prediction Models, in: Solar Energy Forecasting and Resource Assessment, Academic Press, Boston, pp. 299-318 http://dx.doi.org/10.1016/B978-0-12-397177-7.00012-7 Lauret, P., Boland, J., Ridley, B., 2013, Bayesian statistical analysis applied to solar radiation modelling, Renewable Energy 49, p. 124-127 http://dx.doi.org/10.1016/j.renene.2012.01.049 Le Marshall, J., Seecamp, R., Xiao, Y., Gregory, P., Jung, J., Steinle, P., Skinner, T., Tingwell, C., Le, T., 2012, The Operational Generation and Assimilation of Continuous (Hourly) Atmospheric Motion Vectors Winds with 4DVar, Proceedngs of 11th International Winds Workshop p. Leese, J.A., Novak, C.S., Clark, B.B., 1971, An automated technique for obtaining cloud motion from geosynchronous satellite data using cross correlation, Journal of Applied Meteorology 10, p. 118-132 Leese, J.A., Novak, C.S., Ray Taylor, V., 1970, The determination of cloud pattern motions from geosynchronous satellite image data, Pattern Recognition 2, p. 279-292 López, G., Batlles, F.J., Tovar-Pescador, J., 2005, Selection of input parameters to model direct solar irradiance by using artificial neural networks, Energy 30, p. 1675-1684 http://dx.doi.org/10.1016/j.energy.2004.04.035 Lorenz, E., Hammer, A., Heinemann, D., 2004, Short term forecasting of solar radiation based on satellite data, EUROSUN2004 (ISES Europe Solar Congress), Freiburg, Germany, p. Lorenz, E., Hurka, J., Heinemann, D., Beyer, H.G., 2009, Irradiance Forecasting for the Power Prediction of Grid-Connected Photovoltaic Systems, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 2, p. 2-10 Lorenz, E., Kühnert, J., Heinemann, D., 2012, Short term forecasting of solar irradiance by combining satellite data and numerical weather predictions, Proceedings of 27th European Photovoltaic Solar Energy Conference, p. 4401-4440 Lovegrove, K., Watt, M., Passey, R., Pollock, G., Wyder, J., Dowse, J., 2012, Realising the potential of concentrating solar power in Australia, Australian Solar Institute. http://www.australiansolarinstitute.com.au/reports/.aspx Lunnon, R., Lowe, D., 1992, Spatial scale dependency of errors in satellite cloud track winds, Advances in Space Research 12, p. 127-131 Luoma, J., Mathiesen, P., Kleissl, J., 2014, Forecast value considering energy pricing in California, Applied Energy 125, p. 230-237 http://dx.doi.org/10.1016/j.apenergy.2014.03.061 Madaeni, S.H., Sioshansi, R., Denholm, P., 2011, Capacity value of concentrating solar power plants, National Renewable Energy Laboratory, Golden, Colorado. http://www.nrel.gov/docs/fy11osti/51253.pdf Madaeni, S.H., Sioshansi, R., Denholm, P., 2012a, Estimating the Capacity Value of Concentrating Solar Power Plants: A Case Study of the Southwestern United States, Power Systems, IEEE Transactions on 27, p. 1116-1124 10.1109/TPWRS.2011.2179071 Madaeni, S.H., Sioshansi, R., Denholm, P., 2012b, How Thermal Energy Storage Enhances the Economic Viability of Concentrating Solar Power, Proceedings of the IEEE 100, p. 335- 347 http://dx.doi.org/10.1109/JPROC.2011.2144950 Mandal, A.K., Pal, S., De, A.K., Mitra, S., 2005, Novel approach to identify good tracer clouds from a sequence of satellite images, Geoscience and Remote Sensing, IEEE Transactions on 43, p. 813-818 Manobianco, J., Taylor, G.E., Zack, J.W., 1996, Workstation-Based Real-Time Mesoscale Modeling Designed for Weather Support to Operations at the Kennedy Space Center and Cape Canaveral Air Station, Bulletin of the American Meteorological Society 77, p. 653-672 10.1175/1520-0477(1996)077<0653:WBRTMM>2.0.CO;2

138

Marcello, J., Eugenio, F., Marqués, F., 2009, Cloud motion estimation in seviri image sequences, Geoscience and Remote Sensing Symposium, 2009 IEEE International, IGARSS 2009, 3, p. III-642-III-645 Marquez, R., Coimbra, C.F.M., 2011, Forecasting of global and direct solar irradiance using stochastic learning methods, ground experiments and the NWS database, Solar Energy 85, p. 746-756 http://dx.doi.org/10.1016/j.solener.2011.01.007 Marquez, R., Coimbra, C.F.M., 2013, Intra-hour DNI forecasting based on cloud tracking image analysis, Solar Energy 91, p. 327-336 http://dx.doi.org/10.1016/j.solener.2012.09.018 Marquez, R., Gueorguiev, V.G., Coimbra, C.F., 2013a, Forecasting of Global Horizontal Irradiance Using Sky Cover Indices, Journal of Solar Energy Engineering-Transactions of the ASME 135, p. Marquez, R., Pedro, H.T.C., Coimbra, C.F.M., 2013b, Hybrid solar forecasting method uses satellite imaging and ground telemetry as inputs to ANNs, Solar Energy 92, p. 176-188 http://dx.doi.org/10.1016/j.solener.2013.02.023 Martin, L., 2011, Global and direct normal irradiance forecast using GFS and statistical techniques. Martinez-Anido, C.B., Botor, B., Florita, A.R., Draxl, C., Lu, S., Hamann, H.F., Hodge, B.-M., 2016, The value of day-ahead solar power forecasting improvement, Solar Energy 129, p. 192-203 http://dx.doi.org/10.1016/j.solener.2016.01.049 Mathiesen, P., Collier, C., Kleissl, J., 2013, A high-resolution, cloud-assimilating numerical weather prediction model for solar irradiance forecasting, Solar Energy 92, p. 47-61 http://dx.doi.org/10.1016/j.solener.2013.02.018 Maxwell, E., 1987, A quasi-physical model for converting hourly global horizontal to direct normal insolation, Solar Energy Research Institute, Golden, Colorado, pp. 68. http://rredc.nrel.gov/solar/pubs/PDFs/TR-215-3087.pdf Mehos, M., Jorgenson, J., Denholm, P., Turchi, C., 2015, An Assessment of the Net Value of CSP Systems Integrated with Thermal Energy Storage, Energy Procedia 69, p. 2060-2071 http://dx.doi.org/10.1016/j.egypro.2015.03.219 Mellit, A., 2008, Artificial Intelligence technique for modelling and forecasting of solar radiation data: a review, International Journal of Artificial Intelligence and Soft Computing 1, p. 52-76 10.1504/IJAISC.2008.021264 Mellit, A., Eleuch, H., Benghanem, M., Elaoun, C., Pavan, A.M., 2010, An adaptive model for predicting of global, direct and diffuse hourly solar irradiance, Energy Conversion and Management 51, p. 771-782 http://dx.doi.org/10.1016/j.enconman.2009.10.034 Mellit, A., Pavan, A.M., Benghanem, M., 2013, Least squares support vector machine for short- term prediction of meteorological time series, Theor Appl Climatol 111, p. 297-307 10.1007/s00704-012-0661-7 Menzel, W., Smith, W., Stewart, T., 1983, Improved cloud motion wind vector and altitude assignment using VAS, Journal of Climate and Applied meteorology 22, p. 377-384 Menzel, W.P., 2001, Cloud tracking with satellite imagery: From the pioneering work of Ted Fujita to the present, Bulletin of the American Meteorological Society 82, p. 33-47 Menzel, W.P., Purdom, J.F., 1994, Introducing GOES-I: The first of a new generation of geostationary operational environmental satellites, Bulletin of the American Meteorological Society 75, p. 757-781 Merril, R., 1989, Advances in the automated production of wind estimates from geostationary satellite imagery, 4th Conference on Satellite Meteorology and Oceanography, p. 246- 249 Miller, S.D., Rogers, M.A., Heidinger, A., Laszlo, I., Sengupta, M., 2012, Cloud Advection Schemes for Short-Term Satellite-Based Insolation Forecasts, Proceedings of the World Renewable Energy Forum, p.

139

Mishra, A., Kaushika, N.D., Zhang, G., Zhou, J., 2008, Artificial neural network model for the estimation of direct solar radiation in the Indian zone, International Journal of Sustainable Energy 27, p. 95-103 10.1080/14786450802256659 Molineaux, B., Ineichen, P., O?Neill, N., 1998, Equivalence of Pyrheliometric and Monochromatic Aerosol Optical Depths at a Single Key Wavelength, Appl. Opt. 37, p. 7008-7018 10.1364/AO.37.007008 Mueller, R., Dagestad, K.-F., Ineichen, P., Schroedter-Homscheidt, M., Cros, S., Dumortier, D., Kuhlemann, R., Olseth, J., Piernavieja, G., Reise, C., 2004, Rethinking satellite-based solar irradiance modelling: The SOLIS clear-sky module, Remote sensing of Environment 91, p. 160-174 Mukherjee, D.P., Acton, S.T., 2002, Cloud tracking by scale space classification, Geoscience and Remote Sensing, IEEE Transactions on 40, p. 405-415 Myers, D.R., 2007, General cloud cover modifier for clear sky solar radiation models, 6652, p. 665208-665208-665212 NERC, 2009, Accommodating high levels of variable generation, North American Electric Reliability Corporation (NERC). http://www.nerc.com/files/ivgtf_report_041609.pdf NERC, 2010, Variable generation power forecasting for operations, North American Electric Reliability Corporation (NERC). http://www.nerc.com/docs/pc/ivgtf/Task2- 1%285.20%29.pdf Nieman, S.J., Menzel, W.P., Hayden, C.M., Gray, D., Wanzong, S.T., Velden, C.S., Daniels, J., 1997, Fully automated cloud-drift winds in NESDIS operations, Bulletin of the American Meteorological Society 78, p. 1121-1133 Noia, M., Ratto, C., Festa, R., 1993a, Solar irradiance estimation from geostationary satellite data: I. Statistical models, Solar Energy 51, p. 449-456 Noia, M., Ratto, C., Festa, R., 1993b, Solar irradiance estimation from geostationary satellite data: II. Physical models, Solar Energy 51, p. 457-465 Nonnenmacher, L., Kaur, A., Coimbra, C.F.M., 2014, Verification of the SUNY direct normal irradiance model with ground measurements, Solar Energy 99, p. 246-258 http://dx.doi.org/10.1016/j.solener.2013.11.010 Nonnenmacher, L., Kaur, A., Coimbra, C.F.M., 2016, Day-ahead resource forecasting for concentrated solar power integration, Renewable Energy 86, p. 866-876 http://dx.doi.org/10.1016/j.renene.2015.08.068 NREL, 2013, System Advisor Model (SAM) Case Study: Andasol-1, National Renewable Energy Laboratory. NREL, System Advisor Model Version 2014.1.14 (SAM 2014.1.14), [Web page], National Renewable Energy Laboratory. (accessed 12 February 2014). Oyama, R., Shimoji, K., 2008, Status of and future plans for JMA’s Atmospheric motion vectors, Proceedings of 9th International Winds Workshop, p. Patnode, A.M., 2006, Simulation and performance evaluation of parabolic trough solar power plants, College of Engineering, University of Wisconsin, USA. Perez, R., Beauharnois, M., Hemker, K., Kivalov, S., Lorenz, E., Pelland, S., Schlemmer, J., Van Knowe, G., 2011, Evaluation of numerical weather prediction solar irradiance forecasts in the US, 2011 ASES Annual Conference, p. Perez, R., Hoff, T., Dise, J., Chalmers, D., Kivalov, S., 2014, The Cost of Mitigating Short-term PV Output Variability, Energy Procedia 57, p. 755-762 http://dx.doi.org/10.1016/j.egypro.2014.10.283 Perez, R., Ineichen, P., Kmiecik, M., Moore, K., Renne, D., George, R., 2004, Producing satellite- derived irradiances in complex arid terrain, Solar Energy 77, p. 367-371 http://dx.doi.org/10.1016/j.solener.2003.12.016

140

Perez, R., Ineichen, P., Maxwell, E., Seals, R., 1992, Dynamic global-to-direct irradiance conversion models, ASHRAE Transactions, p. 354-369 Perez, R., Ineichen, P., Maxwell, E., Seals, R., Zelenka, A., 1991, Dynamic models for hourly global-to-direct irradiance conversion, 1991 Solar world congress: Proceedings of the biennial congress of the international solar energy society, 1, p. 951-956 Perez, R., Ineichen, P., Moore, K., Kmiecik, M., Chain, C., George, R., Vignola, F., 2002, A new operational model for satellite-derived irradiances: description and validation, Solar Energy 73, p. 307-317 http://dx.doi.org/10.1016/S0038-092X(02)00122-6 Perez, R., Kivalov, S., Schlemmer, J., Hemker Jr, K., Renné, D., Hoff, T.E., 2010a, Validation of short and medium term operational solar radiation forecasts in the US, Solar Energy 84, p. 2161-2172 http://dx.doi.org/10.1016/j.solener.2010.08.014 Perez, R., Kivalov, S., Zelenka, A., Schlemmer, J., Hemker Jr, K., 2010b, Improving the performance of satellite-to-irradiance models using the satellite’s infrared sensors, Proc. of American Solar Energy Society’s Annual Conference, p. Perez, R., Lorenz, E., Pelland, S., Beauharnois, M., Van Knowe, G., Hemker Jr, K., Heinemann, D., Remund, J., Müller, S.C., Traunmüller, W., Steinmauer, G., Pozo, D., Ruiz-Arias, J.A., Lara-Fanego, V., Ramirez-Santigosa, L., Gaston-Romero, M., Pomares, L.M., 2013, Comparison of numerical weather prediction solar irradiance forecasts in the US, Canada and Europe, Solar Energy 94, p. 305-326 http://dx.doi.org/10.1016/j.solener.2013.05.005 Pitz-Paal, R., Amin, A., Bettzuge, M.O., Eames, P., Flamant, G., Fabrizi, F., Holmes, J., Kribus, A., Laan, H.v.d., Martinez, C.L., 2012, Concentrating solar power in Europe, the Middle East and North Africa: a review of development issues and potential to 2050, Journal of Solar Energy Engineering 134, p. Polo, J., Zarzalejo, L., Cony, M., Navarro, A., Marchante, R., Martin, L., Romero, M., 2011, Solar radiation estimations over India using Meteosat satellite images, Solar Energy 85, p. 2395-2406 Polo, J., Zarzalejo, L.F., Ramirez, L., 2008, Solar Radiation Derived from Satellite Images, in: V. Badescu (Ed.) Modeling Solar Radiation at the Earth Surface, Springer-Verlag, The Netherlands, Price, H., 2003, A parabolic trough solar power plant simulation model, ASME 2003 International Solar Energy Conference, p. 665-673 Purdom, J., 1996, Detailed cloud motions from satellite imagery taken at thirty second, one and three minute intervals, Proceedings of Third International Winds Workshop, p. 137-146 Queener, B.D., 2012, Intra-hour Direct Normal Irradiance solar forecasting using genetic programming. Quesada-Ruiz, S., Chu, Y., Tovar-Pescador, J., Pedro, H.T.C., Coimbra, C.F.M., 2014, Cloud- tracking methodology for intra-hour DNI forecasting, Solar Energy 102, p. 267-275 http://dx.doi.org/10.1016/j.solener.2014.01.030 Reikard, G., 2009, Predicting solar radiation at high resolutions: A comparison of time series forecasts, Solar Energy 83, p. 342-349 http://dx.doi.org/10.1016/j.solener.2008.08.007 Reindl, D.T., Beckman, W.A., Duffie, J.A., 1990, Diffuse fraction correlations, Solar Energy 45, p. 1-7 http://dx.doi.org/10.1016/0038-092X(90)90060-P Reno, M.J., Stein, J.S., Hansen, C.W., 2012, Global horizontal irradiance clear sky models : implementation and analysis, pp. Medium: ED; Size: 68 p. http://www.osti.gov/bridge/servlets/purl/1039404/ Richter, C., Teske, S., Short, R., 2009, Concentrating solar power global outlook 09, Greenpeace International, SolarPACES, ESTELA, pp. 88.

141

http://www.greenpeace.org/international/en/publications/reports/concentrating- solar-power-2009/ Ridley, B., Boland, J., Lauret, P., 2010, Modelling of diffuse solar fraction with multiple predictors, Renewable Energy 35, p. 478-483 http://dx.doi.org/10.1016/j.renene.2009.07.018 Rigollier, C., Bauer, O., Wald, L., 2000, On the clear sky model of the ESRA — European Solar Radiation Atlas — with respect to the heliosat method, Solar Energy 68, p. 33-48 http://dx.doi.org/10.1016/S0038-092X(99)00055-9 Rigollier, C., Lefèvre, M., Wald, L., 2004, The method Heliosat-2 for deriving shortwave solar radiation from satellite images, Solar Energy 77, p. 159-169 http://dx.doi.org/10.1016/j.solener.2004.04.017 Ruiz-Arias, J.A., Gueymard, C.A., Dudhia, J., Pozo-Vãzquez, D., 2012, Improvement of the Weather Research and Forecasting (WRF) model for solar resource assessments and forecasts under clear skies, Wolrd Renewable Energy Forum. Ruiz-Arias, J.A., Pozo-Vázquez, D., Lara-Fanego, V., Santos-Alamillos, F.J., Tovar-Pescador, J., 2011, A High-Resolution Topographic Correction Method for Clear-Sky Solar Irradiance Derived with a Numerical Weather Prediction Model, Journal of Applied Meteorology and Climatology 50, p. 2460-2472 10.1175/2011JAMC2571.1 Sayeef, S., Heslop, S., Cornforth, D., Moore, T., Percy, S., Ward, J., Berry, A., Rowe, D., 2012, Solar intermittency: Australia's clean energy challenge, Australian Solar Institute. http://www.australiansolarinstitute.com.au/reports/.aspx Schillings, C., Mannstein, H., Meyer, R., 2004a, Operational method for deriving high resolution direct normal irradiance from satellite data, Solar Energy 76, p. 475-484 http://dx.doi.org/10.1016/j.solener.2003.07.038 Schillings, C., Meyer, R., Mannstein, H., 2004b, Validation of a method for deriving high resolution direct normal irradiance from satellite data and application for the Arabian Peninsula, Solar Energy 76, p. 485-497 http://dx.doi.org/10.1016/j.solener.2003.07.037 Schmetz, J., Holmlund, K., Hoffman, J., Strauss, B., Mason, B., Gaertner, V., Koch, A., Van De Berg, L., 1993, Operational cloud-motion winds from Meteosat infrared images, Journal of Applied Meteorology 32, p. 1206-1225 Schroedter-Homscheidt, M., Gesell, G., Kluser, L., Breitkreuz, H., Oumbe, A., Killius, N., Holzer- Popp, T., 2012, Solar resource forecasting. Schroedter-Homscheidt, M., Hoyer-Klick, C., Rikos, E., Tselepsis, S., Pulvermuller, B., 2009, Nowcasting and forecasting of solar irradiance for solar energy electricity grid integration, SolarPACES, Berlin, Germany. Schroedter-Homscheidt, M., Oumbe, A., 2013, Validation of an hourly resolved global aerosol model in answer to solar electricity generation information needs, Atmospheric Chemistry and Physics 13, p. 3777-3791 Schroedter-Homscheidt, M., Oumbe, A., Benedetti, A., Morcrette, J.-J., 2013, Aerosols for concentrating solar electricity production forecasts: Requirement quantification and ECMWF/MACC aerosol forecast assessment, Bulletin of the American Meteorological Society 94, p. 903-914 10.1175/BAMS-D-11-00259.1 Sfetsos, A., Coonick, A.H., 2000, Univariate and multivariate forecasting of hourly solar radiation with artificial intelligence techniques, Solar Energy 68, p. 169-178 http://dx.doi.org/10.1016/S0038-092X(99)00064-X Sioshansi, R., Denholm, P., 2010, The Value of Concentrating Solar Power and Thermal Energy Storage, Sustainable Energy, IEEE Transactions on 1, p. 173-183 http://dx.doi.org/10.1109/tste.2010.2052078

142

Siva Reddy, V., Kaushik, S.C., Ranjan, K.R., Tyagi, S.K., 2013, State-of-the-art of solar thermal power plants—A review, Renewable and Sustainable Energy Reviews 27, p. 258-273 http://dx.doi.org/10.1016/j.rser.2013.06.037 Skamarock, W.C., Klemp, J.B., Dudhia, J., Gill, D.O., Barker, D.M., Duda, M.G., Huang, X.-Y., Wang, W., Powers, J.G., 2008, A Description of the Advanced Research WRF Version 3, NCAR Technical Note, National Centre for Atmospheric Research, Boulder, USA. Skartveit, A., Olseth, J.A., Tuft, M.E., 1998, An hourly diffuse fraction model with correction for variability and surface albedo, Solar Energy 63, p. 173-183 http://dx.doi.org/10.1016/S0038-092X(98)00067-X Stefferud, K., Kleissl, J., Schoene, J., 2012, Solar forecasting and variability analyses using sky camera cloud detection & motion vectors, Power and Energy Society General Meeting, 2012 IEEE, p. 1-6 10.1109/pesgm.2012.6345434 Tapakis, R., Charalambides, A.G., 2013, Equipment and methodologies for cloud detection and classification: A review, Solar Energy 95, p. 392-430 http://dx.doi.org/10.1016/j.solener.2012.11.015 Teske, S., Leung, J., Crespo, L., Bial, M., Dufour, E., Richter, C., 2016, Solar thermal electricity global outlook 2016, European Solar Thermal Electricity Association (ESTELA), Greenpeace International, SolarPACES, pp. 114. http://www.solarpaces.org/press- room/news/item/98-new-solar-thermal-electricity-report The Dow Chemical Company, 1997, Dowtherm A Heat Transfer Fluid: Product Technical Data, The Dow Chemical Company, USA, pp. 30. Tian, Y., Zhao, C.Y., 2013, A review of solar collectors and thermal energy storage in solar thermal applications, Applied Energy 104, p. 538-553 http://dx.doi.org/10.1016/j.apenergy.2012.11.051 Torres, J.L., De Blas, M., García, A., de Francisco, A., 2010, Comparative study of various models in estimating hourly diffuse solar irradiance, Renewable Energy 35, p. 1325-1332 http://dx.doi.org/10.1016/j.renene.2009.11.025 Tymvios, F.S., Michaelides, S.C., Skouteli, C.S., 2008, Estimation of Surface Solar Radiation with Artificial Neural Networks, in: V. Badescu (Ed.) Modeling Solar Radiation at the Earth Surface, Springer-Verlag, The Netherlands, Usaola, J., 2012, Operation of concentrating solar power plants with storage in spot electricity markets, Renewable Power Generation, IET 6, p. 59-66 http://dx.doi.org/10.1049/iet- rpg.2011.0178 Vapnik, V.N., 1998, Statistical learning theory, Wiley, New York. Vick, B.D., Myers, D.R., Boyson, W.E., 2012, Using direct normal irradiance models and utility electrical loading to assess benefit of a concentrating solar power plant, Solar Energy 86, p. 3519-3530 http://dx.doi.org/10.1016/j.solener.2012.03.010 Vignarooban, K., Xu, X., Arvay, A., Hsu, K., Kannan, A.M., 2015, Heat transfer fluids for concentrating solar power systems – A review, Applied Energy 146, p. 383-396 http://dx.doi.org/10.1016/j.apenergy.2015.01.125 Vignola, F., Harlan, P., Perez, R., Kmiecik, M., 2007, Analysis of satellite derived beam and global solar radiation data, Solar Energy 81, p. 768-772 http://dx.doi.org/10.1016/j.solener.2006.10.003 Voyant, C., Muselli, M., Paoli, C., Nivet, M.-L., 2013, Hybrid methodology for hourly global radiation forecasting in Mediterranean area, Renewable Energy 53, p. 1-11 http://dx.doi.org/10.1016/j.renene.2012.10.049 Wade, G., Lau, K., Wood, N., 1992, Clustering applied to cloud wind determination, International Journal of Remote Sensing 13, p. 3135-3155 Watanabe, T., Urano, Y., Hayashi, T., 1983, Procedures for separating direct and diffuse insolation on a horizontal surface and prediction of insolation on tilted surfaces, Transactions p.

143

Weymouth, G., Le Marshall, J., 2001, Estimation of daily surface solar exposure using GMS-5 stretched-VISSR observations: the system and basic results, Australian Meteorological Magazine 50, p. 263-278 Wittmann, M., Breitkreuz, H., Schroedter-Homscheidt, M., Eck, M., 2008, Case studies on the use of solar irradiance forecast for optimised operation strategies of solar thermal power plants, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 1, p. 18-27 http://dx.doi.org/10.1109/JSTARS.2008.2001152 Xue, M., Droegemeier, K.K., Wong, V., 2000, The Advanced Regional Prediction System (ARPS) – A multi-scale nonhydrostatic atmospheric simulation and prediction model. Part I: Model dynamics and verification, Meteorol Atmos Phys 75, p. 161-193 10.1007/s007030070003 Yang, D., Jirutitijaroen, P., Walsh, W.M., 2012, Hourly solar irradiance time series forecasting using cloud cover index, Solar Energy 86, p. 3531-3543 http://dx.doi.org/10.1016/j.solener.2012.07.029 Zarzalejo, L.F., Polo, J., Martín, L., Ramírez, L., Espinar, B., 2009, A new statistical approach for deriving global solar radiation from satellite images, Solar Energy 83, p. 480-484 Zelenka, A., Perez, R., Seals, R., Renné, D., 1999, Effective Accuracy of Satellite-Derived Hourly Irradiances, Theor Appl Climatol 62, p. 199-207 10.1007/s007040050084 Zeng, J., Qiao, W., 2013, Short-term solar power prediction using a support vector machine, Renewable Energy 52, p. 118-127 http://dx.doi.org/10.1016/j.renene.2012.10.009 Zhang, G., Patuwo, B.E., Hu, M.Y., 1998, Forecasting with artificial neural networks: The state of the art, International Journal of Forecasting 14, p. 35-62 http://dx.doi.org/10.1016/S0169-2070(97)00044-7 Zhang, J., Florita, A., Hodge, B.-M., Lu, S., Hamann, H.F., Banunarayanan, V., Brockway, A.M., 2015a, A suite of metrics for assessing the performance of solar power forecasting, Solar Energy 111, p. 157-175 http://dx.doi.org/10.1016/j.solener.2014.10.016 Zhang, J., Hodge, B.-M., Lu, S., Hamann, H.F., Lehman, B., Simmons, J., Campos, E., Banunarayanan, V., Black, J., Tedesco, J., 2015b, Baseline and target values for regional and point PV power forecasts: Toward improved solar forecasting, Solar Energy 122, p. 804-819 http://dx.doi.org/10.1016/j.solener.2015.09.047 Zhang, Q., Huang, J., Siwei, L., 2002, Development of typical year weather data for Chinese locations, ASHRAE Transactions 108, p. 17

144

A. LKT model equations

This appendix presents the LKT model equations for simulating CST plant operation. The LKT model can be divided into two main sub-models – a parabolic trough solar field model that is based on the work of Patnode (2006) and a power block and storage model that is based on the work of Sioshansi and Denholm (2010). Section A.1 describes the solar field model equations, Section A.2 describes the power block and storage model equations, and Section

A.3 summarises the meaning, the value and the source of the value of each parameter in this study.

A.1. Solar field model

The solar field model describes a parabolic trough solar field with collectors oriented north to south to track the sun from east to west. As the collectors can only track the sun along one axis,

DNI is often not normally incident on the collectors and there are optical losses dependent on the incidence angle of DNI that need to be calculated. These losses are row shading, end loss, and reflection and absorption by the cover. Row shading occurs when the shadow of one row is cast upon an adjacent row because the sun is low in the sky. End loss occurs when non- normally incident DNI is not concentrated onto part of the absorber tube. Reflection and absorption by the cover are accounted for by calculating the incidence angle modifier (IAM).

There are also optical losses that are independent of DNI incidence angle. These optical losses instead depend on the optical properties and manufacturing inaccuracies of the parabolic trough collector mirrors, glass envelope, and receiver tube materials. Besides optical losses, the solar field will have heat losses because the temperature of the solar field will rise above the ambient air temperature as it operates.

145

A.1.1. Optical loss from factors dependent on solar position

The equations for calculating the position of the sun were obtained from Duffie and Beckman

(2013).

Angle of the day, B (radians):

푁 − 1 E. A-1 퐵 = 2 ∗ 휋 ∗ 365

where N is the number of the day of the year, N=1 corresponds to 1 January

Equation of time, EoT (minutes):

퐸표푇 = 229.18 ∗ (0.000075 + 0.001868 ∗ cos(퐵) − 0.032077 ∗ sin(B) E. A-2

− 0.014615 ∗ cos(2 ∗ B) − 0.04089 ∗ sin (2 ∗ B)

Solar time (hours):

퐿표푛푆푡푑 − 퐿표푛퐿표푐 퐸표푇 E. A-3 푆표푙푎푟푇𝑖푚푒 = 푆푡푑푇𝑖푚푒 + + 15 60

where StdTime is local standard time in hours LonStd is the standard longitude used for local standard time in degrees LonLoc is the site longitude in degrees

There is no need daylight savings adjustment in E. A-3 because the equation uses standard time. Daylight saving adjustment would only be necessary if the equation used local clock time instead of local standard time.

Hour angle, ω (radians):

휋 휔 = (푆표푙푎푟푇𝑖푚푒 − 12) ∗ E. A-4 12

Declination angle, δ (radians):

휋 284 + 푁 E. A-5 훿 = ∗ 23.45 ∗ sin (2 ∗ 휋 ∗ ) 180 365

146

Zenith angle, 휃푧 (radians):

휃푧 = acos(푐표푠(훿) ∗ 푐표푠(퐿푎푡) ∗ 푐표푠(휔) + 푠𝑖푛(훿) ∗ 푠𝑖푛(퐿푎푡)) E. A-6

where Lat is the site latitude in radians

Patnode (2006) stated that the collectors cannot be rotated enough to collect sunlight when the solar elevation is below 10 degrees (zenith angle above 80 degrees or about 1.39626 radians). Hence, solar field thermal output should not be calculated under these conditions.

Incidence angle (radians):

2 2 2 E. A-7 휃푖 = acos (√cos (휃푧) + cos (훿) sin (휔)

Row shading (dimensionless):

퐿푠푝푎푐푒 ∗ 푐표푠(휃푧) E. A-8 푅표푤푆ℎ푎푑𝑖푛푔 = min { , 1} 푊 ∗ 푐표푠(휃푖)

where 퐿푠푝푎푐푒 is the spacing between adjacent rows in metres W is the collector aperture width in metres

End loss (dimensionless):

tan(휃푖) E. A-9 퐸푛푑퐿표푠푠 = 1 − 푓 ∗ 퐿푆퐶퐴

where f is the collector focal length in metres 퐿푆퐶퐴 is the length of a solar collector assembly in metres

Incidence angle modifier (dimensionless):

180 180 2 E. A-10 푓 ∗ cos(휃 ) + 푓 ∗ 휃 ∗ + 푓 ∗ (휃 ∗ ) 0 푖 1 푖 휋 2 푖 휋 퐼퐴푀 = cos(휃푖)

where 푓0, 푓1 and 푓2 are coefficients for the incidence angle modifier equation

147

A.1.2. Optical loss from factors independent of solar position

When the CST plant first starts operating, the solar field is likely to consist of only one type of collector. As time passes, damaged collectors may be replaced by next generation collectors and as such the solar field can consist of several types of collectors. The overall optical properties of the solar field are assumed to be the weighted average of optical properties of each type of collector. For simplicity, this study assumes that there is only one type of collector in the solar field.

Solar field optical efficiency (dimensionless):

푁푢푚퐶표푙 E. A-11 휂푓푖푒푙푑 = ∑ 퐶표푙퐹푟푎푐푖 ∗ 푇푟푘퐴푐푐푖 ∗ 퐺푒표퐴푐푐푖 ∗ 푀𝑖푟푅푒푓푖 ∗ 푀𝑖푟퐶푙푛푖 푖=1

where NumCol is the number of different collector types in the solar field 퐶표푙퐹푟푎푐푖 is the fraction of collectors in the field that are the i-th type 푇푟푘퐴푐푐푖 is the twisting and tracking accuracy of the i-th collector type 퐺푒표퐴푐푐푖 is the geometric accuracy of the i-th collector type 푀𝑖푟푅푒푓푖 is the mirror reflectivity of the i-th collector type 푀𝑖푟퐶푙푛푖 is the mirror cleanliness of the i-th collector type

Heat collection element efficiency (dimensionless):

푁푢푚퐻퐶퐸 E. A-12 휂퐻퐶퐸 = ∑ 퐻퐶퐸푓푟푎푐푖 ∗ 퐻퐶퐸푑푢푠푡푖 ∗ 퐵푒푙푆ℎ푎푑푖 ∗ 퐸푛푣푇푟푎푛푠푖 ∗ 퐻퐶퐸푎푏푠푖 푖=1

∗ 퐻퐶퐸푚𝑖푠푐푖

where NumHCE is the number of HCE types in the solar field 퐻퐶퐸푓푟푎푐푖 is the fraction of HCEs in the field that are the i-th type 퐻퐶퐸푑푢푠푡푖 is the adjustment due to losses caused by dust on the cover of the i-th HCE type 퐵푒푙푆ℎ푎푑푖 is the adjustment due to losses caused by shading by bellows of the i-th HCE type 퐸푛푣푇푟푎푛푠푖 is the transmissivity of the cover of the i-th HCE type 퐻퐶퐸푎푏푠푖 is the absorptivity of the HCE selective coating of the i-th HCE type 퐻퐶퐸푚𝑖푠푐푖 is the adjustment due to losses caused by miscellaneous factors for the i-th HCE type

148

A.1.3. Heat loss

To calculate heat loss, Patnode (2006) developed a simplified equation in place of dynamic heat balance equations to reduce the computational overhead for solving equations. This reduces the computational overhead required for calculating heat loss by solving. The simplified equation calculates heat loss from DNI and bulk heat transfer fluid (HTF) temperature.

The solar field inlet and outlet temperatures are not known prior to this step. The solar field inlet temperature may be assumed to equal the Andasol-1 design solar field inlet temperature of 293°C because the CST plant would have been optimised for operation under design conditions. The solar field outlet temperature depends on the solar field thermal output, so it will initially need to be guessed and then found through iteration. The initial guess for the solar field outlet temperature is 393°C, which is the design solar field outlet temperature for

Andasol-1.

Heat loss equation per unit aperture area of solar field (W/m2):

퐻푒푎푡퐿표푠푠퐴 + 퐻푒푎푡퐿표푠푠퐵 E. A-13 퐻푒푎푡퐿표푠푠 = 푇표 − 푇푖

푎1 푎2 푎3 퐻푒푎푡퐿표푠푠퐴 = 푎 (푇 − 푇 ) + (푇2 − 푇2) + (푇3 − 푇3) + (푇4 − 푇4) E. A-14 표 표 푖 2 표 푖 3 표 푖 4 표 푖

푏1 E. A-15 퐻푒푎푡퐿표푠푠퐵 = 퐷푁퐼 ∗ [푏 ∗ (푇 − 푇 ) + (푇3 − 푇3)] 0 표 푖 3 표 푖

where 푇표 is the HTF temperature at the solar field outlet in °C 푇푖 is the HTF temperature at the solar field inlet in °C 푎0, 푎1, 푎2, 푎3, 푏0, and 푏1 are coefficients for the heat loss equation The heat loss equation coefficients can be varied to calculate heat loss from different HCE types. The solar field heat loss is calculated by weighing heat loss from each HCE type in the solar field.

149

Solar field heat loss per unit aperture area of solar field (W/m2):

푁푢푚퐻퐶퐸 E. A-16 퐻푒푎푡퐿표푠푠푖 푆퐹퐻퐿 = ∑ 퐻퐶퐸푓푟푎푐 ∗ 푖 푊 푖=1

Heat loss also occurs for the pipes that lead to and from the solar field and connect the rows together. The equation for pipes heat loss was empirically derived by Price (2003). It is a function of the difference between the average solar field temperature and the ambient air temperature.

Temperature difference (°C):

(푇표 − 푇푖) E. A-17 푇 = − 푇 푑푖푓푓 2 푎푚푏

Piping heat loss per unit aperture area of solar field (W/m2):

2 −7 3 푃𝑖푝푒퐻퐿 = 0.01693 ∗ 푇푑푖푓푓 − 0.0001683 ∗ 푇푑푖푓푓 + 6.78 ∗ 10 ∗ 푇푑푖푓푓 E. A-18

A.1.4. Solar field thermal output

As mentioned earlier, the solar field outlet temperature for calculating heat loss depends on the solar field thermal output, which in turn depends on heat loss, so the solar field outlet temperature needs to be found through iteration.

Absorbed thermal energy per unit aperture area of solar field (W/m2):

푄푎푏푠 = 퐷푁퐼 ∗ cos(휃푖) ∗ 퐼퐴푀 ∗ 푅표푤푆ℎ푎푑표푤 ∗ 퐸푛푑퐿표푠푠 ∗ 휂푓푖푒푙푑 ∗ 휂퐻퐶퐸 E. A-19

∗ 푆퐹푎푣푎𝑖푙

where SFavail is the fraction of the solar field that is operational and tracking the sun

150

Solar field output per unit aperture area solar field aperture (W/m2):

푆퐹푄. 표푢푡. 푝푎 = 푆퐹푓표푐푢푠 ∗ 푄푎푏푠 − 푆퐹퐻퐿 − 푃𝑖푝푒퐻퐿 E. A-20

where 푆퐹푓표푐푢푠 is the fraction of the solar field that is not defocused

Solar field thermal output (W):

푆퐹푄. 표푢푡 = 푆퐹푄. 표푢푡. 푝푎 ∗ 푊 ∗ 퐿푆퐶퐴 ∗ 푁푆퐶퐴 E. A-21

where 푁푆퐶퐴 is the total number of solar collector assemblies (SCAs) in the solar field

Besides iterating to solve for the solar field outlet temperature, the HTF mass flow rate must be checked to ensure that it remains within set limits. The mass flow rate can be calculated by using the solar field thermal output and the HTF enthalpies at the solar field inlet and outlet.

The HTF enthalpies will depend on the type of HTF used. The HTF used for Andasol-1 is

Dowtherm A, so the equation coefficients shown will apply to Dowtherm A. The data used to empirically derive the relationship between enthalpy and temperature was obtained from The

Dow Chemical Company (1997).

HTF enthalpy at solar field inlet (J/kg):

2 ℎ푖 = −16999 + 1500.7 ∗ 푇푖 + 1.4534 ∗ 푇푖 E. A-22

HTF enthalpy at solar field outlet (J/kg):

2 ℎ표 = −16999 + 1500.7 ∗ 푇표 + 1.4534 ∗ 푇표 E. A-23

Mass flow rate (kg/s):

푆퐹푄. 표푢푡 E. A-24 푚푓푙표푤 = ℎ표 − ℎ푖

The calculated mass flow rate depends on the outlet temperature to obtain HTF enthalpy, as shown in E. A-23. If the calculated mass flow rate falls between the minimum limit and maximum limit, then both the mass flow rate is accepted. If the mass flow rate exceeds the

151 upper limit, then the mass flow rate is set to the maximum value. Preventing the mass flow rate exceeding its upper limit will cause the solar field outlet temperature to increase, which may damage the HTF and other components if the temperature greatly exceeds the design solar field outlet temperature. To prevent this, thermal energy is dumped by defocusing sections of the solar field. This is achieved by reducing the value of the variable 푆퐹푓표푐푢푠 in E.

A-20 below 1. A new value for 푆퐹푓표푐푢푠 can be estimated by using E. A-25. If the mass flow rate is less than the lower limit, then the mass flow rate is set to the minimum value.

ℎ표 − ℎ푖 E. A-25 푆퐹 = 푚푓푙표푤 ∗ 푓표푐푢푠 푚푎푥 푆퐹푄. 표푢푡

where 푚푓푙표푤푚푎푥 is the maximum mass flow rate in kg/s

After an acceptable mass flow rate is obtained, a new solar field outlet temperature can be determined by calculating the HTF enthalpy rise in the solar field and using an equation that relates HTF temperature and HTF enthalpy. As mentioned, the HTF used by Andasol-1 is

Dowtherm A, so the equation coefficients are specific to Dowtherm A.

New HTF enthalpy at solar field outlet (J/kg):

푆퐹푄. 표푢푡 E. A-26 ℎ = ℎ + 표푢푡−푛푒푤 푖푛 푚푓푙표푤

New solar field outlet temperature (°C):

−10 2 푇표푢푡−푛푒푤 = 14.474 + 0.0006 ∗ ℎ표푢푡−푛푒푤 − 10 ∗ ℎ표푢푡−푛푒푤 E. A-27

If 푇표푢푡 and 푇표푢푡−푛푒푤 converge to within 0.01 °C, then the solar field outlet temperature is found. If the found solar field outlet temperature exceeds the maximum allowable HTF temperature, then one SCA is defocused at a time until the limit is not exceeded. The change in

푆퐹푓표푐푢푠 caused by defocusing the SCA is shown in E. A-28.

1 E. A-28 푆퐹푓표푐푢푠−푛푒푤 = 푆퐹푓표푐푢푠−표푙푑 − 푁푆퐶퐴 ∗ 푆퐹푎푣푎푖푙

152

After 푆퐹푓표푐푢푠−푛푒푤 is calculated, the procedure for finding the solar field outlet temperature is repeated until the solar field outlet temperature is less than the maximum allowable HTF temperature.

A.2. Power block and storage model

The power block and storage model is designed to simulate energy transfer between a two- tank thermal energy storage (TES) system and a wet-cooled Rankine cycle power block. The model optimises the TES system charge and discharge schedule and the power block generation by using forecast solar field thermal output and forecast electricity prices. The equations which form the CST model are given in Equations E. A-29 to E. A-39. The variables used are described in Table A-1.

푇 E. A-29 푀𝑖푛𝑖푚𝑖푠푒 ∑ −퐸푃(푡) ∗ 푥5(푡) + 푥10(푡) 푡=1

푥1(푡) − (1 − 푇퐸푆푙표푠푠) ∗ 푥1(푡 − 1) − 푥2(푡) + 푥3(푡) = 0 E. A-30

푥5(푡) − 푃퐵푐표푒푓푓 ∗ 푥4(푡) = 0 E. A-31

푥2(푡) − 푇퐸푆푒푓푓 ∗ 푥3(푡) + 푥4(푡) + 푥10(푡) = 푆퐹표푢푡(푡) E. A-32

푃퐵푚𝑖푛 ∗ 푃퐵푡 ∗ 푥6(푡) − 푥4(푡) ≤ 0 E. A-33

푥4(푡) − 푃퐵푚푎푥 ∗ 푃퐵푡 ∗ 푥6(푡) ≤ 0 E. A-34

푥6(푡) − 푥6(푡 − 1) − 푥7(푡) ≤ 0 E. A-35

푥7(푡) + 푥7(푡 − 1) + ⋯ + 푥7(푡 − 푃퐵표푝푚𝑖푛) − 푥6(푡) ≤ 0 E. A-36

푥2(푡) − 푇퐸푆푐 ∗ 푥8(푡) ≤ 0 E. A-37

푥3(푡) − 푇퐸푆푑 ∗ 푥9(푡) ≤ 0 E. A-38

푥8(푡) + 푥9(푡) ≤ 1 E. A-39

where EP(t) is the electricity price at time step t ($/MWh-e) SFout(t) is the solar field thermal output at time step t (MWh-t) TESc is the TES charging capacity (MWh-t)

153

TESd is the TES discharging capacity (MWh-t) TESeff is the TES round-trip efficiency (%) TESloss is the hourly TES heat loss rate (%) PBcoeff is the power block thermal energy to net electric output conversion efficiency (%) PBmax is the power block maximum operating capacity (%) PBmin is the power block minimum operating capacity (%) PBopmin is the minimum power block up time (hours) PBt is the power block rated thermal energy input (MWh-t)

Table A-1: Variables in the power block and storage model

Lower Upper Variable Quantity represented Units bound bound

푥1 Storage level at the end of the time step MWh-t 0 TEScap*TESc 푥2 Energy charged to storage in the time step MWh-t 0 TESc 푥3 Energy discharged from storage in the time step MWh-t 0 TESd 푥4 Energy delivered to the power block in the time step MWh-t 0 PBt 푥5 Net electric output in the time step MWh-e 0 Infinity Binary variable to check whether the power block is 푥6 operating during the time step - 0 1 Binary variable to check whether the power block 푥7 started during the time step - 0 1 Binary variable to check whether storage is being 푥8 charged during the time step - 0 1 Binary variable to check whether storage is being 푥9 discharged during the time step - 0 1 푥10 Thermal energy dumped during the time step MWh-t 0 Infinity

E. A-29 is the optimisation objective function. It aims to maximise electric output when the

price is high and minimise the dumping of thermal energy. E. A-30 governs the balance of

energy entering and leaving TES. E. A-31 states the conversion of power block thermal energy

input to net electric output. E. A-32 is the energy balance between solar field thermal energy

output, power block thermal energy input, dumped thermal energy, and TES charging and

discharging. E. A-33 and E. A-34 provide boundaries for the values which power block thermal

energy input may take. E. A-35 relates the binary variables for power block start up and

operation to ensure that they are assigned correct values. E. A-36 directs the optimisation to

operate the power block for at least the desired minimum up time each time the power block

154

is started. E. A-37 and E. A-38 are constraints on the maximum TES charge and discharge rates

respectively while obeying the condition imposed by E. A-39 that TES cannot simultaneously

charge and discharge in the same time step.

Each variable has upper and lower bounds as shown in Table A-1. The upper bound for 푥1 is

calculated by multiplying the number of hours of storage with the maximum TES charge

capacity because it resembles the amount of energy that can be discharged from TES after

accounting for the round-trip efficiency (Sioshansi and Denholm, 2010). The round-trip

efficiency is in E. A-32. The upper bounds for 푥2 and 푥3 are the maximum TES charging and

discharging capacities respectively. The upper bound for 푥4 is the maximum power block

thermal input. The upper bound for 푥5 is assigned a value of infinity but 푥5 is limited by the

power block thermal energy input due to the relationship in E. A-31. The binary variables 푥6,

푥7, 푥8 and 푥9 all share the common upper bound value of 1. It is assumed that there is no limit

on the ability to dump thermal energy so the upper bound of 푥10 is infinite. All of the variables

cannot be negative, thus the lower bounds are all zero.

A.3. LKT model parameter values

Parameters were chosen such that the LKT model resembles Andasol-1 with 1.7564 solar

multiple and 50 MW rated electric output.

Table A-2: LKT model parameters

Parameter Quantity represented Value Units Reference (Patnode, a0 Coefficient for the heat loss equation -9.463033 W/m2 2006) (Patnode, a1 Coefficient for the heat loss equation 3.029616E-01 W/(m2.degC) 2006) (Patnode, a2 Coefficient for the heat loss equation -1.386833E-03 W/(m2.degC2) 2006) (Patnode, a3 Coefficient for the heat loss equation 6.929243E-06 W/(m2.degC3) 2006)

155

(Patnode, b0 Coefficient for the heat loss equation 7.649610E-02 - 2006) (Patnode, b1 Coefficient for the heat loss equation 1.128818E-07 degC-1 2006) Adjustment due to losses caused by (Patnode, BelShad shading by bellows 0.97 - 2006) (Patnode, EnvTrans Transmissivity of the HCE cover 0.96 - 2006) (AUSTELA, f Collector focal length 2.11 m 2014) Coefficient for the incidence angle (Patnode, f0 modifier equation 1 - 2006) Coefficient for the incidence angle (Patnode, f1 modifier equation 8.84E-04 - 2006) Coefficient for the incidence angle (Patnode, f2 modifier equation 5.369E-05 - 2006) (Patnode, GeoAcc Geometric accuracy of the collector 0.98 - 2006) (Patnode, HCEabs Absorptivity of the selective coating 0.95 - 2006) Adjustment due to losses caused by (Patnode, HCEdust dust 0.98 - 2006) Adjustment due to losses caused by (Patnode, HCEmisc miscellaneous factors 0.96 - 2006) (AUSTELA, LSCA Length of a SCA 142.17 m 2014) (AUSTELA, Lspace Spacing between adjacent rows 15 m 2014) Maximum HTF flow rate in the solar (AUSTELA, mflowmax field 1,872 kg/s 2014) Minimum HTF flow rate in the solar (AUSTELA, mflowmin field 156 kg/s 2014) (Patnode, MirRef Mirror reflectivity of the collector 0.935 - 2006) (Patnode, MirCln Mirror cleanliness of the collector 0.951 - 2006) (AUSTELA, NSCA Total number of SCAs in the solar field 624 - 2014) Number of collector types in the solar NumCol field 1 - - NumHCE Number of HCEs in the solar field 1 - - Power block thermal energy to net (AUSTELA, PBcoeff electric output conversion efficiency 35.63 % 2014) (AUSTELA, Pbopmin Minimum power block up time 1 hours 2014) Power block maximum operating (AUSTELA, Pbmax capacity 100 % 2014) Power block minimum operating (AUSTELA, Pbmin capacity 25 % 2014) PBt Power block rated thermal energy 144.357 MWh-t (AUSTELA, 156

input 2014) Fraction of solar field operational and (Patnode, SFavail tracking the sun 0.99 - 2006) (AUSTELA, TESc Storage charging capacity 130.247 MWh-t 2014) (AUSTELA, TESd Storage discharging capacity 127.347 MWh-t 2014) (AUSTELA, TESeff Storage round-trip efficiency 100 % 2014) (Sioshansi and Denholm, TESloss Hourly storage heat loss rate 0.031 % 2010) (AUSTELA, Tin_design Solar field design inlet temperature 293 degC 2014) (AUSTELA, Tout_design Solar field design outlet temperature 393 degC 2014) Twisting and tracking accuracy of the (Patnode, TrkAcc collector 0.99 - 2006) (AUSTELA, W Collector aperture width 5.75 m 2014) HCE means heat collection element. HTF means heat transfer fluid. SCA means solar collector assembly.

157

B. Numerical Results

This appendix presents the numerical values for financial value and reliability results. The

scenarios that did not use 1-hour forecasts are labelled “no STF” and those that used 1-hour

forecasts are labelled “with STF”.

B.1. Financial Value Metrics

Table B-1: Financial value metrics for a CST plant without storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant without storage Revenue ($ million) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 3.4 3.4 2.4 2.7 4.0 4.2 5.7 6.2 48-h persistence 3.1 3.3 1.8 2.4 3.7 4.1 4.0 5.1 TAPM 3.9 3.4 3.3 2.8 4.7 4.3 6.8 6.3 Perfect 4.0 4.0 3.5 3.5 4.7 4.7 6.8 6.8

RG cost ($ million) ARIMA 1.0 0.26 1.8 0.42 1.0 0.38 2.3 0.25 48-h persistence 0.99 0.24 1.7 0.39 0.89 0.36 2.2 0.23 TAPM 1.2 0.28 3.1 0.45 1.4 0.39 2.6 0.27 Perfect 0 0 0 0 0 0 0 0

Dump cost ($ million) ARIMA 3.2 3.1 3.1 2.7 3.5 3.1 6.1 5.5 48-h persistence 3.5 3.1 3.7 2.7 3.7 3.1 7.8 5.5 TAPM 2.7 3.1 2.2 2.7 2.7 3.1 5.0 5.5 Perfect 2.6 2.6 2.0 2.0 2.7 2.7 5.0 5.0

Net value ($ million) ARIMA 2.4 3.1 0.6 2.3 2.9 3.9 3.4 5.9 48-h persistence 2.1 3.0 0.0 2.0 2.8 3.7 1.8 4.9 TAPM 2.6 3.2 0.2 2.4 3.3 4.0 4.2 6.0 Perfect 4.0 4.0 3.5 3.5 4.7 4.7 6.8 6.8

158

Table B-2: Financial value metrics for a CST plant with storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant with storage Revenue ($ million) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 5.2 5.5 4.6 5.0 5.8 6.1 8.5 8.7 48-h persistence 5.1 5.3 4.5 4.7 5.8 6.1 7.1 7.4 TAPM 5.5 5.5 4.7 4.9 6.4 6.3 8.7 8.7 Perfect 5.9 5.9 5.4 5.4 6.5 6.5 9.7 9.7

RG cost ($ million) ARIMA 0.89 0.019 1.2 0.052 0.86 0.030 1.7 0.041 48-h persistence 0.90 0.027 1.2 0.053 0.71 0.026 1.8 0.031 TAPM 1.7 0.058 3.6 0.11 1.8 0.055 2.8 0.055 Perfect 0 0 0 0 0 0 0 0

Dump cost ($ million) ARIMA 1.4 1.4 1.6 1.3 0.85 0.70 3.3 3.0 48-h persistence 1.4 1.4 1.7 1.3 1.4 1.1 3.4 2.9 TAPM 0.77 1.1 0.50 0.61 0.19 0.27 1.4 2.2 Perfect 0.76 0.76 0.44 0.44 0.18 0.18 1.4 1.4

Net value ($ million) ARIMA 4.4 5.5 3.4 5.0 4.9 6.0 6.8 8.6 48-h persistence 4.2 5.3 3.3 4.7 5.1 6.0 5.4 7.4 TAPM 3.8 5.5 1.1 4.8 4.5 6.2 5.9 8.6 Perfect 5.9 5.9 5.4 5.4 6.5 6.5 9.7 9.7

159

B.2. EFOR

Table B-3: EFOR for a CST plant without storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant without storage EFOR (%) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 31 10 60 26 33 15 38 12 48-h persistence 33 10 67 25 31 15 38 12 TAPM 38 12 75 27 40 15 44 13 Perfect 0 0 0 0 0 0 0 0

Table B-4: EFOR for CST a plant with storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant with storage EFOR (%) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 22 0.91 39 2.9 25 1.5 29 1.4 48-h persistence 22 0.81 40 2.5 22 1.0 27 1.2 TAPM 32 1.8 61 6.1 37 2.4 37 2.1 Perfect 0 0 0 0 0 0 0 0

B.3. Normalised Revenue and Net Value

Table B-5: Normalised revenue and net value for a CST plant without storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant without storage Revenue ($/MWh) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 34 34 56 52 39 39 64 64 48-h persistence 34 34 54 52 39 39 51 57 TAPM 33 34 49 50 37 39 60 63 Perfect 33 33 48 48 37 37 60 60

Net value ($/MWh) ARIMA 24 31 13 44 29 36 38 61 48-h persistence 23 32 1 44 30 36 23 55 TAPM 22 31 3 42 26 35 37 60 Perfect 33 33 48 48 37 37 60 60

160

Table B-6: Normalised revenue and net value for a CST plant with storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant with storage Revenue ($/MWh) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 35 35 63 58 39 37 63 58 48-h persistence 34 35 62 59 38 36 53 52 TAPM 32 34 49 53 35 36 54 55 Perfect 34 34 57 57 37 37 61 61

Net value ($/MWh) ARIMA 29 35 46 58 33 37 50 58 48-h persistence 28 35 45 58 33 36 40 52 TAPM 22 33 12 52 25 36 36 55 Perfect 34 34 57 57 37 37 61 61

B.4. Total Energy

Table B-7: Total energy values for a CST plant without storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant without storage Generation (GWh) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 100 100 43 53 102 109 90 97 48-h persistence 90 95 33 47 95 104 78 90 TAPM 118 102 66 56 125 112 112 101 Perfect 122 122 72 72 126 126 114 114

RG purchased (GWh) ARIMA 34 8 44 11 39 14 40 9 48-h persistence 32 7 39 10 31 13 36 8 TAPM 43 9 81 12 54 14 57 10 Perfect 0 0 0 0 0 0 0 0

Dumped (GWh) ARIMA 93 91 64 49 87 77 88 77 48-h persistence 103 91 74 49 94 77 99 77 TAPM 75 91 39 49 63 77 64 77 Perfect 71 71 33 33 62 62 63 63

161

Table B-8: Total energy values for a CST plant with storage

Site Mildura Mt. Gambier Rockhampton Wagga CST plant with storage Generation (GWh) no STF with STF no STF with STF no STF with STF no STF with STF ARIMA 150 159 73 86 150 165 134 148 48-h persistence 149 152 72 81 155 166 135 143 TAPM 174 164 96 93 181 175 162 157 Perfect 171 171 95 95 176 176 160 160

RG purchased (GWh) ARIMA 31 1 30 1 35 1 35 1 48-h persistence 29 1 30 1 28 1 33 1 TAPM 56 2 96 3 82 2 67 2 Perfect 0 0 0 0 0 0 0 0

Dumped (GWh) ARIMA 36 28 27 16 33 19 37 24 48-h persistence 37 29 28 17 28 17 36 24 TAPM 14 19 6 7 5 8 11 13 Perfect 14 14 4 4 5 5 10 10

162