Thermal Modelling and Cost Analysis for Large-Scale Battery Energy Storage System (BESS) in Grid-Connected PV Plant

Md Mehedi Hasan Bachelor of Engineering (Electrical and Electronic Engineering)

A thesis submitted for the degree of Master of Philosophy at The University of Queensland in 2019 School of Information Technology and Electrical Engineering Abstract

With the growing renewable energy capacity at an exponential rate every year, energy storage system has become popular. Among available energy storage systems, battery energy storage system (BESS) is widely used technology for its high-power density and fast response in renewable plants such as photovoltaic (PV) farm. The indispensable operation of battery system is to store excess PV generation and utilise it effectively. Fluctuation in PV generation output is a usual incident due to the unpredictable nature of PV generation. Therefore, a storage system is required to smooth its output generation prior to export it to the grid. Despite a lot of advancement in electrochemical storage technology and cost reduction, battery degradation remains a big obstacle for its wider acceptance in large-scale PV plant. Battery degrades because of charging/discharging operation as well as during idle condition. Elevated temperature of battery cell accelerates battery degradation, which indicates an imminent loss of expensive investment on BESS. In addition, elevated temperature is responsible for adding to the cost of battery system operation in the plant by accelerating cooling system operation. A number of research studies have been carried out to model thermal characteristics and behaviour of battery. However, the majority of studies reported in the literature are conducted on the small-scale battery thermal management where thermal behaviour is investigated using complicated thermo-dynamical models. However, a detailed thermo-dynamical model is not feasible for a large- scale BESS, where thousands of battery cells are installed and thermally interacting with each other. To consider battery degradation cost in the EMS, a solution is required to estimate battery thermal behaviour for any given proﬁle without following any complex computational methods. An accurate battery cell temperature estimation can play a signiﬁcant role in ensuring BESS optimal operation considering its degradation. In this thesis, the work starts with investigating the thermal behaviour of the battery with respect to its charging/discharging operation along with other possible effective parameters. Identifying the most dominant factors for battery cell temperature is one of the primary steps. A comprehensive thermal model will be developed to estimate battery cell temperature with 24 hours horizon in advance

ii iii with high accuracy. Finally, an effective cost function will be developed considering dominant parameters. Different simulation strategies are implemented to assess the proposed models. This thesis will consider all pertinent matters and undergo rigorous assessments using real ﬁeld data to ensure practicality. Scalability of the solution will be kept in mind for future development along with the aforementioned features. Power industries and academia will be beneﬁted from the studies on thermal behaviour of battery, battery cell temperature estimation, and cost functions for effective BESS operations in a PV plant. Declaration by Author

This thesis is composed of my original work, and contains no material previously published or written by another person except where due reference has been made in the text. I have clearly stated the contribution by others to jointly-authored works that I have included in my thesis.

I have clearly stated the contribution of others to my thesis as a whole, including statistical as- sistance, survey design, data analysis, signiﬁcant technical procedures, professional editorial advice, ﬁnancial support and any other original research work used or reported in my thesis. The content of my thesis is the result of work I have carried out since the commencement of my higher degree by research candidature and does not include a substantial part of work that has been submitted to qualify for the award of any other degree or diploma in any university or other tertiary institution. I have clearly stated which parts of my thesis, if any, have been submitted to qualify for another award.

I acknowledge that an electronic copy of my thesis must be lodged with the University Library and, subject to the policy and procedures of The University of Queensland, the thesis be made available for research and study in accordance with the Copyright Act 1968 unless a period of embargo has been approved by the Dean of the Graduate School.

I acknowledge that copyright of all material contained in my thesis resides with the copyright holder(s) of that material. Where appropriate I have obtained copyright permission from the copyright holder to reproduce material in this thesis and have sought permission from co-authors for any jointly authored works included in the thesis.

iv Publications and Submitted Manuscripts

Publications included in this thesis

1. M. M. Hasan, S. A. Pourmousavi, F. Bai, and T. K. Saha. The impact of temperature on battery degradation for large-scale bess in pv plant. In Universities Power Engineering Conference (AUPEC), 2017 Australasian, pp. 16 (IEEE, 2017).

This paper is incorporated in Chapter 3.

Contributor Statement of contribution Md Mehedi Hasan Simulation and modelling (100%) Result interpretation and discussion (75%) Paper writing (75%) Seyyed Ali Pourmousavi Kani Result interpretation and discussion (10%) Paper writing (10%) Tapan Saha Result interpretation and discussion (10%) Paper writing (10%) Feifei Bai Result interpretation and discussion (5%) Paper writing (5%)

v vi

2. Md Mehedi Hasan, S. Ali Pourmousavi and Tapan K. Saha, “Battery Cell Temperature Estima- tion Model and Cost Analysis of a Grid-Connected PV-BESS Plant”. In IEEE Innovative Smart Grid Technologies Asia (ISGT ASIA 2019), (IEEE, 2019).

This paper is incorporated in Chapter 4 and 5.

Contributor Statement of contribution Md Mehedi Hasan Simulation and modelling (100%) Result interpretation and discussion (75%) Paper writing (75%) Seyyed Ali Pourmousavi Kani Result interpretation and discussion (15%) Paper writing (15%) Tapan Saha Result interpretation and discussion (10%) Paper writing (10%)

Submitted manuscripts included in this thesis

1. Md Mehedi Hasan, Ali Pourmousavi, Tapan K. Saha, “A data-driven approach to estimate battery cell temperature using NARX neural network model. (Submitted to Applied Energy).

This paper is incorporated in Chapter 4.

Contributor Statement of contribution Md Mehedi Hasan Simulation and modelling (100%) Result interpretation and discussion (75%) Paper writing (75%) Seyyed Ali Pourmousavi Kani Result interpretation and discussion (15%) Paper writing (15%) Tapan Saha Result interpretation and discussion (10%) Paper writing (10%) vii

Other publications during candidature “No other publications”

Contributions by others to the thesis “No contributions by others”

Statement of parts of the thesis submitted to qualify for the award of another degree “No works submitted towards another degree have been included in this thesis”

Research Involving Human or Animal Subjects “No animal or human subjects were involved in this research” Acknowledgements

A few words cannot be enough to express my appreciation for kind guidance and supports of a number of people. This work would not be happened without their supports. Firstly, I would like to express my gratitude and profound respect to Professor Tapan Saha for advising my MPhil work and thesis. His unforgettable help, understanding and supports as principal supervisor helped me to go forward smoothly during my candidature. It was an outstanding journey working with him, as I have not only learnt technical parts of my work but also a lot of lessons from him, which made my MPhil easier and it would be useful for future endeavour. Moreover, I am really grateful to my associate supervisor, Dr. Ali Pourmousavi for his continuous advice and help on my journey to MPhil thesis. I appreciate his time and encouragement, which assisted to generate a fruitful research outcome. I would like to acknowledge the financial support, I received from The University of Queensland through RTP Scholarship. I would like to convey my gratitude to Professor Tapan Saha for his efforts and all supports towards getting financial funding for attending conference. I would like to acknowledge the research facilities, I have received from UQ Gatton Solar Facility (UQ GSRF) to complete this work. I would like to express my appreciation to all of my colleagues from Power and Energy System (PES) Group of the University of Queensland. Advice and valuable time provided by the group members were undoubtedly beneficial throughout my journey to complete MPhil. I would also like to remember my enjoyable time being in the group. Last but not least, I would like to convey my appreciation to my wife and my family for their love and indispensable supports. Their unconditional encouragement in every step of my study helped me a lot to move forward smoothly.

viii Financial Support, Keywords and Classiﬁcations

Financial Support: “This research was supported by an Australian Government Research Training Program Scholar- ship”

Keywords: battery energy storage system, cell temperature, neural network, photovoltaic, cost.

Australian and New Zealand Standard Research Classiﬁcations (ANZSRC): ANZSRC code: 090607, Power and Energy Systems Engineering (excl. Renewable Power), 50% ANZSRC code: 090608, Renewable Power and Energy Systems Engineering (excl. Solar Cells), 50%

Fields of Research (FoR) Classiﬁcation: FoR code: 0906, Electrical and Electronic Engineering 100%

ix Table of Contents

Abstract ii

Declaration by Author iv

Publications and Submitted Manuscripts v

Acknowledgements viii

Financial Support, Keywords and Classiﬁcations ix

List of Figures xiii

List of Tables xvi

List of Abbreviations xvii

1 Introduction 1 1.1 Background and Motivation ...... 3 1.2 Objectives ...... 4 1.3 Thesis Overview ...... 5

2 Literature Review 7 2.1 Temperature Effects on BESS ...... 7 2.2 Battery Cell Temperature Estimation Methods ...... 10 2.3 Battery Degradation and Operational Cost ...... 13 2.4 Summary ...... 16

3 Battery Cell Temperature and Degradation Study 18 3.1 Introduction ...... 18

x TABLEOF CONTENTS xi

3.2 PV Integrated BESS under Study ...... 19 3.2.1 Physical Layout ...... 19 3.2.2 Operational Rules ...... 20 3.2.3 Cooling Mechanisms ...... 22 3.3 Data for Analysis ...... 23 3.3.1 BESS Monitoring and Data Logging System ...... 23 3.3.2 Data Selection for Battery Cell Temperature Thermal Behaviour Investigation 24 3.4 Analysis of Battery Cell Temperature Behaviour ...... 25 3.4.1 Discharging Operation of BESS ...... 25 3.4.2 Charging Operation of BESS ...... 28 3.4.3 Ambient Temperature Effects during BESS Operations ...... 31 3.5 Temperature-Dependent Battery Degradation Study ...... 33 3.6 Summary ...... 35

4 Battery Cell Temperature Modelling 36 4.1 Introduction ...... 36 4.1.1 Thermal Behaviour Estimation Approaches ...... 37 4.1.2 Signiﬁcance of Data-Driven Approach ...... 37 4.2 Data-Driven Battery Cell Temperature modelling ...... 39 4.3 System Under Study ...... 43 4.4 Simulation Study ...... 44 4.4.1 Sensitivity analysis on the number of neurons and time delay ...... 45 4.4.2 The universal NARX model ...... 46 4.4.3 Seasonal NARX Model ...... 48 4.5 Comparison between linear and non-linear model ...... 54 4.5.1 Battery Cell Temperature Estimation with ARIMAX model ...... 54 4.6 Summary ...... 61

5 Development of Cost Function 62 5.1 Introduction ...... 62 5.2 Cost Function Development ...... 63 5.2.1 PV plant cost with and without BESS ...... 63 5.2.2 Battery Degradation Cost ...... 65 5.2.3 Cooling Cost ...... 67 TABLEOF CONTENTS xii

5.2.4 Cost of imported energy from the grid and demand charge ...... 68 5.2.5 Demand Charge ...... 69 5.2.6 FiT Beneﬁt ...... 69 5.3 Simulation Study ...... 69 5.3.1 Cost Analysis ...... 70 5.4 Summary ...... 72

6 Conclusion and Future Work 74 6.1 Conclusion ...... 74 6.2 Future Work ...... 76

References 78 List of Figures

1.1 Australian electricity generation from renewable sources ...... 2 1.2 Residential, commercial and large-scale PV capacity installed by states (a) September 2018, (b) by month ...... 3

2.1 Thermal images at the start (a), intermediate (b) and at the end (c) of discharge during discharge of 4.5C ...... 13

3.1 Single line diagram of Gatton PV plant with BESS ...... 20 3.2 Battery arrangement and bank level management system ...... 21 3.3 Cooling system arrangement for BESS ...... 23 3.4 Current and ambient temperature effect on battery cell temperature during a typical discharging event ...... 26 3.5 Peak current and temperature rising slope: A linear relationship for discharging events 27 3.6 Peak current and peak temperature: A linear relationship for discharging events . . . 27 3.7 Temperature rising slope vs charge of discharging events ...... 27 3.8 Temperature rising delay for each discharging event ...... 28 3.9 Absolute temperature change during each discharging event ...... 28 3.10 Current and ambient temperature effect on battery cell temperature during a typical charging event ...... 29 3.11 Relationship between peak current and temperature rising slope for charging events . 29 3.12 Relationship between peak current and peak temperature for charging events . . . . . 30 3.13 Temperature rising delay for every charging incident ...... 31 3.14 Absolute temperature change during each charging event ...... 31 3.15 Temperature rising slope vs charge of charging events ...... 32 3.16 Ambient temperature effects on battery cell temperature ...... 33

xiii LISTOF FIGURES xiv

3.17 Average extra degradation rates for the actual temperature in each charge/discharge event...... 34 3.18 Comparison of degradation rates between Scenario I and Scenario II ...... 34

4.1 NARX structural model based on MLP neural network ...... 40 4.2 Series-parallel architecture for NARX during training ...... 41 4.3 Parallel architecture for NARX during testing and actual simulation ...... 41 4.4 Schematic diagram of the BESS conﬁguration in the PV plant ...... 44 4.5 Average error of different NARX model with different number of neurons in the hidden layer ...... 46 4.6 Sensitivity analysis results on the cell temperature feedback delay ...... 46 4.7 Hourly RMSE histogram for 56 days of test data obtained by the universal model for 24-hour ahead ...... 48 4.8 Comparison between the seasonal and universal NARX models ...... 50 4.9 RMSE of the test days during summer season using recursive estimation for 24 hours ahead ...... 51 4.10 Seasonal comparison based on the hourly-averaged RMSE obtained by using seasonal and universal network models ...... 51 4.11 The RMSE ratio for the summer NARX model ...... 52 4.12 The RMSE ratio for the autumn NARX model ...... 53 4.13 The RMSE ratio for the winter NARX model ...... 53 4.14 The RMSE ratio for the spring NARX model ...... 53 4.15 Partial autocorrelation function (PACF) plot ...... 57 4.16 Autocorrelation function (ACF) plot ...... 57 4.17 Hourly RMSE for test days with higher errors ...... 58 4.18 Histogram of hourly RMSE for 30 test days using ARIMAX battery cell temperature estimation model ...... 58 4.19 NARX and ARIMAX Universal model comparison ...... 59 4.20 NARX and ARIMAX seasonal model comparison for summer ...... 59 4.21 NARX and ARIMAX seasonal model comparison for winter ...... 60 4.22 NARX and ARIMAX seasonal model comparison for spring ...... 60 4.23 NARX and ARIMAX seasonal model comparison for autumn ...... 61

5.1 Cost-beneﬁt of the entire PV-BESS plant ...... 64 5.2 Cost-beneﬁt of the PV plant without BESS ...... 65 LISTOF FIGURES xv

5.3 Daily cost associated with BESS operation in the plant ...... 70 5.4 Daily FiT beneﬁt with and without BESS in the plant ...... 71 5.5 Monthly demand charge with and without BESS integration in the plant ...... 71 5.6 Daily total cost for the plant with and without BESS ...... 72 List of Tables

3.1 Battery cell speciﬁcations ...... 20

4.1 Parameters for training the NARX networks ...... 47 4.2 Data selection and training outcomes from MATLAB Neural Net Time Series toolbox 47 4.3 Seasonal data selection for training and evaluation ...... 49 4.4 Summer data selection and training outcomes from MATLAB Neural Net Time Series toolbox ...... 49 4.5 Autumn data selection and training outcomes from MATLAB Neural Net Time Series toolbox ...... 49 4.6 Winter data selection and training outcomes from MATLAB Neural Net Time Series toolbox ...... 49 4.7 Spring data selection and training outcomes from MATLAB Neural Net Time Series toolbox ...... 50

5.1 Existing cooling system speciﬁcations ...... 67

xvi List of Abbreviations

PV Photovoltaic BESS Battery Energy Storage System SoC State of Charge DoD Depth of Discharge UQ The University of Queensland GSRF Gatton Solar Research Facility ANN Artiﬁcial Neural Network MLP Multilayer Perceptron NARX Nonlinear Autoregressive Network with Exogenous inputs TDL Tapped Delay Line RMSE Root Mean Square Error ARIMA Autoregressive Integrated Moving Average ARIMAX Autoregressive Integrated Moving Average with Exogenous inputs KPSS Kwiatkowski Phillips Schmidt Shin ACF Autocorrelation Function PACF Partial Autocorrelation ADF Augmented Dickey Fuller ARX Autoregressive Exogenous AR Autoregressive MA Moving-Average EMS Energy Management System EV Electric Vehicle HEV Hybrid Electric Vehicle PEV Plug-in Electric Vehicle PHEV Plug-in Hybrid Electric Vehicle

xvii xviii

OLTC On-Load Tap Changer SVR Static Voltage Regulator O&M Operation and Maintenance LCOE Levelized cost of electricity PLC Programmable logic Controller SCADA Supervisory Control and Data Acquisition CSS Central Supervisory System BMS Battery Management System RET Renewable Energy Target ToU Time-of-Use FiT Feed In Tariff EoL End of Life DUoS Distribution Use of System TUoS Transmission Use of System BoM Bureau of Meteorology 1 Introduction

The electricity produced from Renewable Energy Sources (RES) has increased significantly over the past couple of years in Australian power supply system [1–3]. The production of energy from RES is likely to increase further because of ongoing concerns about climate change. Major portion of the energy is provided by wind and solar energy compared to other RES, as the sources are considered to have the highest potential in Australia. Figure 1.1 shows the growth of different RES from 1991 to 2017. It is noticeable that solar generation increased significantly in recent years. It is reported in [3] that renewable energy reached to 16% of electricity generation in Australia in 2016-17. In the same year, the renewable energy increased by 6% with the help of RES namely solar, hydro, and wind. 31% of renewable electricity is contributed by wind generation in 2016-17 and wind generation is increased by 3% in the same year. Solar generation in Australia has been increasing strongly every year. 18% solar generation growth is experienced in 2016-17. This generation was accounted for 3% of total electricity generation in Australia. Large-scale PV installation growth remained strong at 47%. Solar PV system installation has been increasing significantly in residential and large-scale level with the help of Renewable Energy Target (RET) scheme. Moreover, With the advances in photovoltaic (PV) technology and cost reduction, solar generation has been increasing in recent years as a well-accepted renewable energy source. The PV installation growth in residential, commercial

1 2 and utility-scale sectors is reflected in Figure 1.2, where it is clearly indicated that large-scale PV plant has been increasing significantly. However, aforementioned renewable sources are not capable of providing smooth and uninterrupted continuous generation due to their nature of origin. Therefore, energy storage device [4–6] has become integral for ensuring secure power supply. Battery energy storage system (BESS) has become indispensable with the increasing rate of RES, more specifically the PV plants, all over the world to store its excess energy and utilise it effectively [7]. Interest in BESSs is increasing rapidly due to other factors such as their capability of smoothing unpredictable renewable energy generation. Nowadays, different types of battery such as lithium-ion, lead-acid, flow battery, salt-water battery are popular as energy storage for different applications [8]. Although, each type of battery has specific features that makes it suitable for certain applications, lithium-ion technology with high energy density is preferred in large-scale grid integrated PV plants, where fast charging/discharging operations with respect to variable PV generation are desired. Energy storage system contributes to reduce cost for PV plant and can provide ancillary services to the grid when it Among bioenergy sources, generation from bagasse during 2016–17 was is required. BESS accomplisheslower than such the featuresprevious withyear, itswith continuous more bagasse charging/discharging being used by the food operations. Although, a lot of advancementmanufacturing and cost sector reduction for other are stationery experienced energy in purposes. battery technology, degradation remainsFigure a big 3.6: concern Australian forelectricity this expensivegeneration from product. renewable sources

Source: Department of the Environment and Energy (2018) Australian Energy Statistics, Table O Figure 1.1: Australian electricity generation from renewable sources

3.3 Electricity generation[3] in calendar year 2017

Estimates of electricity generation were published in May 2018 for the 2017 calendar year, to improve the availability of up-to-date official data on total generation in Australia. Total electricity generation in Australia was estimated to be 259 terawatt hours in calendar year 2017, an increase of approximately 1 per cent compared with 2016 (Table 3.3).

Non-renewable sources contributed 220 terawatt hours (85 per cent) of total electricity generation in 2017, a decrease of 1.4 per cent compared with 2016. Coal continued to account for the majority of electricity generation, at 61 per cent of total generation in 2017. Gas-fired generation increased in 2017, to represent 21 per cent of total generation.

Renewable sources contributed 39 terawatt hours, or about 15 per cent of total electricity generation in 2017. The largest source of renewable generation was hydro (5 per cent of total generation), followed by wind (5 per cent) and small- scale solar (3 per cent). Renewable generation decreased by 7 per cent in 2017 from 42 terawatt hours in 2016, with a steep decline in hydro generation and slight drop in wind generation offsetting growth in other sources.

Generation varies quite a lot across Australia. In 2017, more than 70 per cent of electricity generation in Queensland, New South Wales and Victoria was coal fired (Figure 3.7). In Victoria, brown coal’s share was 77 per cent in 2017, down from 84 per cent in 2016. This is largely attributable to the closure of

Australian Energy Update 2018 24

APVI/UNSW Solar Trends Report for Solar Citizens funded by Lord Mayor’s Charitable Foundation

1.1 BACKGROUNDAND4000 MOTIVATION 3

3500

3000

2500

2000

Capacity (MW) 1500

1000

500

0 NSW VIC QLD WA SA TAS ACT NT

(a) Residential Comme rcial Utility

5000

4000

3000

Capacity (MW) 2000

1000

0 2007-01 2008-02 2009-02 2010-02 2011-02 2012-02 2013-02 2014-02 2015-02 2016-02 2017-02 2018-02

(b) Residential Comme rcial Utility

FigureFigure 1.2: 4: Residential,Residential, Commercial commercial and Large and-scale large-scale PV Capacity PV Installed capacity by State installed (a) September by states 2018, (b) (a) by Septembermonth 2018, (b) by month

[9]

1.1 Background and Motivation 7 | Page

Battery degradation remains a serious obstacle to its further acceptance in large-scale PV plant. Degradation occurs during charging and discharging as well as during the idle time [10–12]. Ex- treme temperature is another factor accelerating the battery degradation faster than it is expected [13]. Typically, the acceptable operating temperature of the Li-ion battery technology is between 25 ◦C to 30 ◦C [14, 15]. Any deviation from this range may adversely affect the battery life. According to [16–18], temperature rise accelerates the battery life degradation. They showed that the rate of change 1.2 OBJECTIVES 4 in battery degradation is significant for the upper bound temperature values. Although manufacturers of battery recommend maintaining the temperature in a range of 25 ◦C to 30 ◦C for extending battery life [19], temperature deviates from the range due to instantaneous battery activity of charging and discharging. In [20], it is shown that maximum battery cell temperature increases with the change of charge and discharge current. According to the extensive experimental analysis carried out in [21], every 10 ◦C increase in cell temperature would double the battery degradation. Depending on the temperature, different charge and discharge cycles combination lead to different capacity loss [22]. Moreover, accumulated heat in a utility-scale BESS, placed in a confined container, elevates the temperature inside battery container, which increases cooling cost by operating air conditioning units at a higher rate to maintain the temperature at an acceptable level [23]. This, in turn, escalates the cost of operation of the whole plant. Therefore, it is important to estimate battery cell temperature evolution in time with respect to a planned operation regime and ambient interactions to have a better evaluation of the incurred cost. In particular, accurate temperature estimation is important for an Energy Management System (EMS), which operates the entire plant including the BESS in an optimal way considering its degradation and cooling cost. Using a battery cell temperature estimation algorithm, an effective energy management strategy can be established with an accurate cell temperature model, which will be able to prevent unnecessary battery operation by taking different factors into account. As a result, the BESS lifetime can be extended, which positively affects economic benefits of the entire system. Moreover, a comprehensive cost analysis for the entire plant is able to render an insight of operation cost of BESS within PV plants. Analysing cost associated with each factor helps to identify the dominant cost factors, which can be considered by the EMS during operating the BESS. To summarize, there has been a continuous growth of BESS with the expansion of PV generation. Simultaneously, there are research gaps, which obstructs the acceptance of large-scale BESS in PV plant. Early degradation of the expensive BESS is a big concern for many owners. Therefore, a proper analysis of battery degradation with respect to its temperature and operation is required, along with the development of suitable models to reduce the problems. The models should be developed using theoretically sound methodologies and algorithms to enable its acceptance as a real-world solution.

1.2 Objectives

In order to thoroughly investigate and analyze thermal effects on battery degradation and develop a thermal model and cost formulation for the entire PV-BESS plant, the following objectives will be considered in this research. 1.3 THESIS OVERVIEW 5

• Investigating the impact of charge/discharge cycling of battery and ambient temperature on battery cell temperature and degradation based on battery temperature.

• Create a data-driven battery cell temperature estimation model for 24 hours ahead horizon using historical real-world data of effective parameters.

• Developing a cost function of battery degradation by adopting a linear degradation model.

• Developing a cost function of cooling mechanism for BESS storage container considering heat sources.

• Developing a cost formulation for the entire PV plant considering inﬂuential factors and con- ducting a comparison study of PV plant with and without BESS.

1.3 Thesis Overview

The current chapter provides an overall background to the research work. This chapter also renders a brief overview of BESS degradation and responsible causes. The motivations behind this research work are also highlighted. Finally, research objectives are underlined, which will be performed to accomplish the research work. Chapter 2 reviews the literature on topics of interest to this research study. Firstly, the literature related to the impact of temperature on battery degradation and its associated extra cost is reviewed in this chapter. Modelling-related literature such as battery thermal and degradation models are also reviewed in this chapter. The shortcomings of thermal modelling for a large-scale grid-tied PV-BESS plant are identified. Finally, the chapter provides an overview of operational cost of a PV plant with and without BESS. Chapter 3 explains the thermal behaviour of battery with respect to its operation and ambient temperature, and the impact of battery cell temperature on degradation. Dominant factors are considered in this investigation to identify the root cause of battery cell temperature. Real-world field data was considered for the entire analysis. Interesting outcomes were congregated and used for further research work. Chapter 4 explains the development of a data-driven battery cell temperature estimation model. Developed non-liner model is established and tested with field data. The simulation outcomes presented in this chapter confirms the robustness of the estimation model. In addition, a comparison study is also presented in this chapter to compare the performance of linear and non-linear data-driven approaches for battery cell temperature estimation. 1.3 THESIS OVERVIEW 6

Chapter 5 discusses cost of PV plant considering inﬂuential factors. Development of cost formulations for battery degradation and cooling system operation are explained in this chapter. Furthermore, adopted degradation model is discussed in this chapter. Finally, a comparison cost analysis between PV plant with and without BESS is carried out using actual PV-BESS plant data. Simulation results shown in the chapter offer a signiﬁcant amount of information, which can be utilised in developing an optimal BESS operational algorithm for EMS. Chapter 6 concludes the research work carried out with a summary and provides further direction of the research. 2 Literature Review

This chapter presents a comprehensive literature review on the relevant work to this thesis. Firstly, a review of thermal behaviour and characteristics of battery is addressed. Then, studies on thermal estimation model strategies are reviewed. Finally, relevant studies on cost and beneﬁt of using battery in different applications are reviewed.

2.1 Temperature Effects on BESS

Battery cell temperature has signiﬁcant impact on battery behaviour both on short and long term time scales. Therefore, it is important to control the operating temperature of battery considering its performance, degradation, and safety. It is shown in [24] that the performance of battery is affected by temperature in shorter time frame, while battery ageing mechanism reported in [17, 25–28] are indicating effects of temperature on battery on longer time scale. In addition, elevated temperature of battery causes a thermal runway reaction that might triggers a ﬁre and an explosion [29, 30]. Moreover, battery cell temperature changes with respect to its environment and operation, which affects the battery performance over time at different rates and consequently reduces its lifetime. The lifetime of a battery can be extended by knowing the root cause of degradation with respect to its

7 2.1 TEMPERATURE EFFECTS ON BESS 8 operation and surrounding factors. The two phenomenon causing battery degradation are Solid-Electrolyte Interphase (SEI) and loss of active materials. The degradation of lithium due to the SEI is the most usual and basic cause of battery capacity fade in lithium-ion batteries. Over hundreds of cycle can be reduced by the SEI formation. Elevated temperature is highly responsible for the SEI formation [31]. Matthew and Martin systematically established a theory of SEI growth and capacity degradation of battery in [32], where they asserted a temperature-dependent capacity loss because of SEI layer formation [33–35]. There is a good number of research conducted on the changes in SEI composition at elevated temperature [31, 36–38]. The growth of SEI progresses throughout cycling and idle situation, favoured by elevated temperature. Capacity fading due to loss of active materials under elevated temperature conditions are outlined in [39–47]. Lithium manganese dissolution at elevated temperature leads to active material losses. Authors have discussed two different manganese dissolution mechanism namely, manganese dissolution at low potential and acid dissolution. Both of the mechanisms are accelerated at higher temperature. Higher temperature facilitates the corrosion and lithium losses, which increases the capacity fade at higher rates [26, 48, 49]. Authors in [50–52] clearly indicated that the degradation process of a battery cell is highly influenced by temperature. In [52], degradation of battery is divided into actual capacity fading and capacity loss that occurs temporarily. Actual capacity fading indicates the permanent loss of cell due to the consumption of lithium-ion. This form of degradation rate is accelerated by high battery cell temperature. Temporary capacity loss indicates the temperature drop in a certain cycle and this is recoverable if temperature of battery returns to a certain level. Moreover, in [53], authors run an analysis of ambient temperature effects on battery cycle life. It is shown in the paper that the effect of ambient temperature on battery cycle life can be described by film growth on the electrode for lithium-ion battery. The cell oxidation leads to a film grown on the electrode, which non reversibly increases the internal resistance of lithium-ion batteries and finally causes a failure. Simulation results showed that more self discharge and capacity loss are experienced in presence of higher temperature during idle situation of battery compared to the ambient temperature. In addition, battery charging/discharging temperature and rate of State of Charge (SoC) have strong effects on battery capacity life degradation as shown in [54–59]. The high SoC rate in presence of elevated temperature creates side reactions at the positive electrode, which increases impedance in cells [60]. Therefore, reduction of active surface area is imminent. Furthermore, a good number of studies showed experimental analysis of calendar ageing with respect to temperature [26, 48, 61], where ageing was tested over different temperature, SoC and end of charge voltages. Battery cycle life tests are mostly performed in laboratory environment 2.1 TEMPERATURE EFFECTS ON BESS 9 where the positive results are achievable [62, 63] because of controlling facility for temperature and charge/discharge operation. In reality the batteries are under harsh operational condition. In [64], authors suggested that the lithium-ion battery for EV, they have used in the experiment could last for 2,000 cycles at 25 ◦C and 800 cycles at 55 ◦C. However, it is not practical to run this type of experiment when it comes to thousands of battery cell in a single system. Existing studies provide an insight of battery degradation with respect to temperature. However, most of the studies are carried out under controlled environment to examine the temperature distribution and its effect in different scenarios. Moreover, small-scale battery is used to study temperature behaviour and its effects on degradation. In addition to the degradation studies with respect to temperature, it is important to identify the root cause of temperature changes in a battery cell. In [65], authors presented the temperature behaviour of prismatic lithium-ion battery with respect to different discharge rates such as 1C, 2C, 3C and 4C, and various surrounding conditions such as temperature of the environment. Constant discharge current rate is used to identify the temperature contours. As per the results, it has been suggested that higher discharge rate accelerates the battery cell temperature. Authors in [66], analysed the thermal behaviour of a prismatic lithium-ion battery under constant discharge current rates of C/10, C/5, C/2, 1C, 2C, 3C, and 4C. In this study, surface temperature of battery is measured in different scenarios. However, in utility-scale BESS, discharge current rate varies with respect to generation of PV system and the load that is locally connected to PV plant. Authors in [67] examined temperature changes in battery by injecting different charge profiles. It is observed that higher charging rate increases the battery cell temperature significantly, even with the presence of active cooling system during the analysis. There is a huge difference between battery surface and internal temperature during charging operation at high current rate [68]. The temperature difference could be greater than 10 ◦C in high power application as indicated in [69]. As a result the degradation quantification with respect to battery external temperature will not be able to reflect the actual results. However, measuring internal temperature of battery cell for a large-scale BESS is impractical. Therefore, an easy approach is required for a large system where thousand of cells are installed. While a significant number of studies dealt with thermal management of electric vehicles (EVs), hybrid electric vehicles (HEVs) and plug-in hybrid electric vehicles (PHEVs), no research studies have investigated thermal issues in an upper range of temperature in large-scale BESS. Moreover, none of them are focusing on utility-scale BESS, where thousands of battery cells are installed in a container and operated based on PV generation and grid requirements. In this thesis, the impact of ambient temperature and charge/discharge activities on battery cell temperature and degradation 2.2 BATTERY CELL TEMPERATURE ESTIMATION METHODS 10 of BESS is investigated for a large-scale PV power plant in a tropical weather condition. Illustra- tive examples and analysis are provided to show how battery cell temperature varies with respect to charging/discharging current and ambient temperature. The illustrative cases are based on actual BESS operation, and therefore realistic results are rendered.

2.2 Battery Cell Temperature Estimation Methods

In order to reduce the battery degradation and its associated costs, it is important to know the battery cell temperature in advance for a given charge/discharge profile. Such a model can be used by the EMS of a large-scale BESS to operate the storage system in an optimal way using the estimated temperature values. In this way, early battery degradation can be reduced by considering the cost and benefit of battery operation. A good number of research has been published to date to address this issue and proposed different approaches for estimating battery cell temperature. However, most of the works conducted are based on the battery thermal management of EVs, HEVs and PHEVs, where they proposed general models of battery thermal behaviour [70–76] or specialized models for lower range of temperature [23, 77]. Battery heat generation or cell temperature evolution is a complex phenomenon, which requires an in-depth knowledge of electrochemical reaction with respect to time and temperature, and distribution of current within a large size battery. A good number of literature is considered in thermal and electrochemical coupled or decoupled to establish a thermal model [23, 70, 71, 73]. To calculate the heat generation of battery, a fully coupled model of thermal-electrochemical utilises newly generated parameters for current and future from the model, so that the prediction can be made for the temperature distribution [78]. Decoupled model utilises empirical equations using experimental data. To capture temperature distribution within battery cell, energy balance is included in many one-dimensional electrochemical model as an extension. Gu and Wang in [23, 78] have shown a thermal electrochemical coupled modelling approach which considered multi-scale physics in lithium-ion battery. This includes electrochemical kinetics, solid-phase lithium transport, and ion transport in electrolyte, charge conservation and thermal energy conservation. Although the above works delivered an expected outcomes, this is very complex approach to follow and not a feasible way to estimate cell temperature for large-scale BESS. A local heat generation method shown in [79–81] gives an acceptable accuracy than a lumped model [82–85]. Thermal energy equation and multi-phase microscopic electrochemical models is coupled in this paper by considering temperature-dependent physico-chemical properties. However, microscopic electrochemical model is too complicated and validation requires experiment which is 2.2 BATTERY CELL TEMPERATURE ESTIMATION METHODS 11 time consuming. Moreover, an equivalent circuit with RC elements is used to capture thermal behaviour using current-voltage performance in real-time with low computational effort and versatile simulation of transient voltage profile of lithium-ion secondary battery employing internal equivalent electric circuit [86]. Due to computation complexity and time, it is difficult to describe the current-voltage character- istic of lithium-ion battery. In [87], a two-stage approximation of radially distributed thermal model is used to determine charging strategy. RC model and two-stage thermal model is combined in the paper. Battery surface and internal temperature are considered in this study and the variables related to this model have been determined under a laboratory condition. Moreover, A three-dimensional model has been proposed to determine the thermal behaviour of lithium-ion battery during discharge events [83]. This model includes battery core region, contact layer and battery case without any sim- plification. To enhance the accuracy at boundaries, this study considered convection and radiation simultaneously. However, these studies require complex computational efforts and huge number of parameters based on laboratory analyses. For estimating battery internal temperature in real-time, a thermoelectric coupled model is proposed [88]. The model is developed to observe the relationship between battery thermal and electrical behaviours. Then, estimation of internal temperature is carried out by using an extended Kalman filter. Ohmic heating (thermal generation) is considered in [88, 89] to reflect the effect of large load current. This study assumed that battery shell and internal temperature have uniform thermal char- acteristic and uniform heat distribution. Although heat transfer includes heat convection, conduction and radiation, heat conduction is assumed to be the only heat transfer between battery internal and shell, and between battery shell and ambient. Average battery temperature and heat generation rate is represented by an equivalent circuit model in [90]. The technique estimates heat generation and average temperature in real time by using only surface temperature and cooling rate measurement. However, only average temperature considered in the literature reduces the accuracy of estimation. To determine the thermal behaviour of battery, the first law of thermodynamics has been used in [91]. This study focused on heat transfer to the coolant in EVs by means of cooling plates. It is also assumed that inlet temperature of the air is equal to the ambient. Heat generation of battery cell is not taken into account by the authors. In [88], an online estimation technique is developed for determining thermal state of a pouch type lithium-ion battery. Lumped thermal model and state observer are utilised to describe the internal temperature and heat generation rate of a battery cell. They have used state space model using the states of battery surface temperature and average temperature as the state vector. Computer simulation results based on similar cell components and lumped capacitance method, are used only to identify 2.2 BATTERY CELL TEMPERATURE ESTIMATION METHODS 12 the parameters of model. A potter filtering algorithm is used in [92] to determine cylindrical battery core temperature by combining potters filter and Maybecks adaptive estimation strategy. It neglects axial temperature variation while considering lumped thermal model to estimate the core temperature from skin temperature readings. Only electrical model is integrated without the impact of convection and conduction of heat transfer to ambient. Therefore, actual temperature results from the analysis cannot be made. In [93, 94] a finite element method is used to observe thermal behaviour of lithium-ion battery during charging events. Current density distributions in a lithium-ion battery are predicted as a function of charging time during constant current-constant voltage charging pattern by using finite element method. To predict the temperature distribution in the battery, heat generation rates as a function of charging time are calculated. Although, good prediction was made with charging operation, discharging events are ignored in the study. Furthermore, experimental approaches has also been considered in understanding the thermal behaviours of a battery in different conditions. The behaviour of a prismatic lithium-ion battery temperature distribution has been shown based on the laboratory experiment at different discharge current rates and different boundary or ambient temperature in [95–97]. It is identified from the results that larger discharging current at higher ambient temperature generates more temperature at the end of discharge cycle compared to lower discharge rates with lower temperature. In [97], an investigation of electrical and thermal performance of lithium-ion battery at different discharge current is performed. Using MATLAB Simulink, a battery thermal model has been created and was validated with the experimental data of temperature, voltage, heat generation and internal resistance at different discharge rate. Temperature non-uniformity in the heat generation on a surface of battery was observed using an IR camera shown in Figure 2.1. To determine the heat generation in a battery cell, this study considered two primary sources (i) entropy changes due to electrochemical reactions and (ii) Ohmic heating for discharging current. The results show that higher discharge rate will cause higher surface temperature distribution. Above studies only show temperature behaviour of battery for discharging events. Moreover, adopting this single cell experimental approach for a large-scale BESS is quite impractical and it is not capable of proving estimated values for future time steps. The motivation of this research comes from a stationary application, BESS installed in a PV plant, where arrangement of battery and its operation is completely different from the existing applications discussed in the above studies. Existing literature focused on EVs and PHEVs, where operation of battery does not fluctuate with respect to weather. Moreover, BESS in a PV plant highly contributes the grid by its continuous operation. Furthermore, thousands of battery cells are installed in a 2.3 BATTERY DEGRADATION AND OPERATIONAL COST 13

Figure 2.1: Thermal images at the start (a), intermediate (b) and at the end (c) of discharge during discharge of 4.5C [65] single system and stored in a conﬁned container, indicating different thermal scenario from the small- size battery inside a EVs or PHEVs. However, Only single battery cell is considered for estimating temperature with respect to its operation in most of the studies. The existing battery cell temperature estimation methods require intensive computational efforts and complex thermal modelling. In addition, analytical methods require measurement of time-varying internal parameters of the BESS periodically. Therefore, available methods are unable to be integrated in EMS, where real time data is required for estimating battery cell temperature in advance. An easy practically feasible solution is required, which can contribute EMS signiﬁcantly without following complex and expensive solutions.

2.3 Battery Degradation and Operational Cost

Analysing total PV plant cost with respect to BESS operation is essential to identify the dominant cost factors. In other words, it is important to know the behaviour of cost factors prior to establishing an optimal operational algorithm. Degradation and its associated cost is one of the major concerns 2.3 BATTERY DEGRADATION AND OPERATIONAL COST 14 along with the other operational cost factors in a PV plant. The integral part of finding the cost of degradation is to adopt an accurate degradation model, which will require low computational efforts and easy to implement in EMS. There have been a con- siderable number of research studies conducted on the battery degradation model to predict capacity fade of lithium-ion batteries [33, 71, 81, 98–101]. Theoretical approach, focusing on the degradation mechanism of lithium-ion and active materials in a battery is established in [57, 100, 102–105]. These degradation models render a comprehensive explanation about degradation based on battery usages in different situations. However, because of their incapability of reflecting the practical data from real-world battery operations, they are not well accepted in estimation application. On the other hand, empirical models are capable of rendering suitable and straight-forward solution to meet the requirements of estimation applications. However, this type of model has some certain limitations due to limited available experimental data, and also models are not applicable in irregular operations, which is very normal in PV-BESS plant. Furthermore, most of the models are nonlinear which cannot be used in optimal EMS with linear optimisation. A developed model by Bloom [26] was adopted by Wang [16], to develop a prediction model for lithium-iron-phosphate batteries. In this model, a large number of experimental data is used to create prediction model. The ageing mechanisms of this particular battery are evaluated using destructive physical and non-destructive electrochemical analysis. In this literature, the experimental parameters namely time, Depth of Discharge (DoD), discharge rate, and the temperature were considered and their effects were outlined. However, fitting parameters and coefficient values differed from one type of battery to another type of battery. The model is derived and investigated based on lithium iron phosphate battery, which is widely used in automotive and space industries for its chemical and thermal stability and low cost, however, it is different from the battery used in this thesis for large- scale PV plant in terms of application and type of battery. Adopting this model for quantifying the degradation of battery is not a solution because the values observed in the experiment varies with the types of battery. In order to identify the total cost of PV-BESS plant, cost factors related to battery operation such as charging/discharging, demand charge, cooling system operation are important to be considered besides most prominent cost factor occurs for degradation. A number of research studies developed cost functions for different applications. A multi-objective cost function is developed in [89] to optimize battery operation based on thermal effects of working temperature and electrical power used for thermal management. It is determined from the results that deviation from safe temperature range promotes battery ageing. In other words, the thermal management strategy with respect to cost is less attractive when more electrical power is 2.3 BATTERY DEGRADATION AND OPERATIONAL COST 15 used for actuators namely, cooler valve, fan, pump, chiller, compressor, and circuit valve. Authors in [106] considered different cost functions including- time-to-charge, energy loss, and temperature rise index to find the daily cost of PHEV. The current and voltage profile are also discussed in this study by considering the cost function as a weighted sum of time-to-charge, energy loss, the square of the difference of the SoC from the final desired SoC, and a temperature rise index. Authors in [15] used energy cost for heating the battery in a cold temperature and the degradation cost due to battery operation in cold environment in a thermal management system The energy cost is calculated based on the cost of electricity generation from fuel, i.e., 0.4USD/kWh. Although, above studies show some cost estimation methods in terms of temperature variation, no research to date considered large-scale BESS in the higher range of temperature. Moreover, BESS container arrangement is completely different from small-scale battery storage in EVs or PHEVs. Therefore, different cost function formulation and technical constraints need to be considered. In [107], a cost-benefit analysis is carried out in order to design a battery size more accurately. Multiple factors such as annual cost of battery and benefit from on-load tap changer (OLTC)/Static Voltage Regulator (SVR) work stress, benefit from peak-shaving operation, and load shifting, are considered for cost analysis. However, BESS operational related cost such as grid charging activities of BESS, degradation of battery with respect to its operation and cooling cost for maintaining the storage temperature, are not in consideration for the cost analysis. Moreover, number of studies analysed cost and benefits of BESS integrated in presence of RESs. Sardi et al. in [108] deliver a cost and benefit analysis of grid integrated BESS with solar PV generation. A number of benefits and cost factors such as load shaving, charging during off-peak period, operation and maintenance (O&M) or operating and maintenance, are identified for BESS in the study and the strategy proposed by the authors helps to find the optimal position of BESS, which increases the net present value. However, the analysis cannot be used for calculating daily cost-benefit of BESS in PV plant. As a result, daily optimal operation cannot be made by EMS. Furthermore, an optimization model is proposed by Han et al. [109] for grid-connected micro- grids with solar PV and BESS, where a cost-benefit analysis is carried out considering initial investment of PV and BESS, O&M cost and equipment replacement costs. The literature only focuses on the discharging operation of battery to estimate operational cost for BESS. In [110] authors considered a number of operational costs such as maintenance, facilities, insurance, management, self-discharge losses, control systems and charging operation by purchasing energy from the grid. A degradation cost function for BESS is also proposed in the research, where annual degradation of the energy capacity of the BESS is estimated based on calendar ageing mechanism ignoring operational related 2.4 SUMMARY 16 degradation. Cardoso in [111] considered linear battery ageing model to quantify degradation of battery in terms of calendar and cyclic ageing mechanisms. Moreover, residential PV-BESS is explained in [112] with a consideration of battery degradation cost and energy savings characteristics of the system. Calendar and cycle are utilised in this study as the two different capacity degradation mechanisms. However, cycle degradation is given importance to establish a long-term optimal operational model. In this analysis, two different factors such as number of cycle of charging/discharging operation and level of DoD after discharging operation are considered. Optimal operational algorithm is established considering the cost and benefit of PV-BESS system, where PV generated power to meet the residential demand, PV and grid energy to charge the BESS, and purchased electricity from the grid to meet the residential demand are considered. However, cost analysis study was carried out considering small-scale PV and BESS energy sources. Levelized Cost Of Electricity (LCOE) is used for cost related evaluation of residential PV-BESS [113]. Furthermore, the LCOE includes all cost related factors over the system’s lifetime. Capital cost, cost of operation and maintenance, and initial investment are considered in this literature to evaluate the operational strategies. However, these approaches are not capable of providing cost estimation on a daily basis. Daily basis BESS operation cost considering important factors is crucial for optimal operation. One of the important cost factors, cooling cost is analyzed by [114], where a data center is considered in the formulation of the cost associated with heat sources inside the data center. The load on the cooling equipment is calculated based on power consumption by installed hardware in the data center. However, no such research considered cooling cost for a BESS storage room in a PV plant. Although above studies offered different cost estimation methods in terms of battery degradation and maintenance, no research to date considered cost analysis for a large-scale BESS considering temperature-oriented factors. This is important in terms of thermal behaviour because the BESS container arrangement is completely different from small-scale battery used in EVs or PHEVs. Moreover, operation of BESS in PV plant is completely different from a small-scale battery. It highly depends on the weather and the load supported by the plant. Furthermore, daily optimal operation cannot be made by EMS without the knowledge of cost-benefit of BESS considering effective parameters on a daily basis.

2.4 Summary

There are numerous research papers on different aspects of storage technologies such as battery thermal and degradation modelling, and cost analysis of battery operation. However, several significant factors (thermal issues for , battery operation with respect to PV generation, temperature oriented 2.4 SUMMARY 17 costs) are not covered in these studies in order to represent a practical grid-tied BESS in PV plant. A new thermal model needs to be created for large-scale BESS, which is completely different from small size battery used in EVs and PHEVs. The existing literature dealt with complex models to estimate battery cell temperature, which are not practical to be used in EMS. Therefore, a computationally tractable battery cell temperature model will be developed in this thesis considering BESS operation and surrounding factors for large-scale BESS within a PV plant. The proposed model will be suitable for day-ahead estimation for given charge/discharge profile which will let EMS to operate BESS in an optimum way. Moreover, the cost analysis of PV-BESS plant will be developed where all BESS related cost and benefit will be dominantly formulated extensively. For this thesis, data from an actual PV-BESS plant will be considered in order to study the impact of temperature on BESS and develop a solution to the issues. The study will assist real-world scenarios and set-up a platform for future studies. Both industry and academic will be able to use this work for the development of battery storage energy usages in renewable energy sector. 3 Battery Cell Temperature and Degradation Study

3.1 Introduction

1 Battery cell temperature is one of the major factors in battery degradation. Excessively high battery cell temperature due to its charge/discharge operations with high magnitude of current, and surrounding temperature accelerates battery cell degradation. In order to study BESS degradation and its corresponding correlated parameters, the operation of a BESS in a large-scale PV plant is investigated. It is useful to apprehend battery cell temperature behaviour in a large-scale BESS working within a renewable energy plant, which consists of thousands of battery cells in a conﬁned room. The University of Queensland (UQ) has such facility that provides the opportunity for extensive investigation to determine the temperature behaviour of battery in different weather and operational scenarios. The University of Queensland Gatton Solar Research Facility (UQ GSRF) located at the

1This chapter has materials from the following reference published by the MPhil candidate. 1. Md Mehedi Hasan, S. Ali Pourmousavi, Feifei Bai and Tapan K. Saha. The impact of temperature on battery degradation for large-scale BESS in PV plant. (IEEE AUPEC 2017).

18 3.2 PV INTEGRATED BESS UNDER STUDY 19

University of Queensland (UQ) Gatton campus, Australia, consists of a grid-tied 3.275MWp PV plant with Lithium-Polymer battery storage system. The PV plant with BESS has been in operation for a few years now and provides high resolution data for battery cell temperature evaluation, among other things. The chapter is organized as follows: Section 3.2 explains the system under study which covers the operation rules and cooling mechanisms for the battery storage room. Section 3.3 presents the data storage and selection for the study. Section 3.4 delineates the analysis of battery temperature behaviour based on selected data. Section 3.5 presents the battery degradation with respect to temperature. Finally, the chapter is concluded with a summary in Section 3.6.

3.2 PV Integrated BESS under Study

In this study, UQ GSRF is considered for the entire work. The plant is connected to the 11kV local distribution network. It is a large research facility, funded by the Australian Federal Government, which contains state-of-the-art technologies, such as different PV tracking systems, Li-Polymer battery storage, and smart metering and control systems. In the following subsections, a complete overview of the BESS and data acquisition system within the plant is presented.

3.2.1 Physical Layout

A 600 kW/760 kWh Li-Polymer BESS is installed in the UQ GSRF. Figure 3.1 shows the structure of the plant. The schematic diagram shows BESS connection to the PV plant, UQ Gatton campus substation and local utility grid. The BESS is partitioned into two 300 kW/380 kWh battery bank. Each bank is interfaced to the grid using two 300 kVA bi-directional inverter with 415 V, 3-phase AC output. The inverters are connected to the campus substation through a 1 MVA step-up transformer, as shown in Figure 3.1. The voltage of each bank varies between 576 V and 748 V DC based on the battery SoC and internal resistances. In addition, the BESS is capable to source/sink reactive power at ±0.9 power factor. The total BESS capacity and power are divided into two battery banks. Each battery bank has four racks of battery modules in parallel and each of the racks comprises with ten series of battery modules. Each module has two parallel strings of 18 battery cells in series. Battery cell speciﬁcation details are shown in Table 3.1. The battery arrangements in the storage room are shown in Figure 3.2. To render critical information in every level of battery namely, module, rack and bank level, Battery management System (BMS) is integrated in each level. 3.2 PV INTEGRATED BESS UNDER STUDY 20

Figure 3.1: Single line diagram of Gatton PV plant with BESS [115]

Table 3.1: Battery cell speciﬁcations Description Speciﬁcation

Battery Cell Type Lithium Polymer Cell Capacity 75 Ah Cell Voltage 2.7 V to 4.1 V, average 3.7 V Maximum Continuous Discharging Current 2C(150A) at 23 ± 3 ◦C Maximum Continuous Charging Current 5C(375A) at 23 ±3 ◦C Peak Discharging Current 8C(150A) Cycle-Life 4000 Cycles at 80% Depth of Discharge (DoD), 1C(Charge)/1C(Discharge) Charging Temperature 10 to 35 ◦C Discharging Temperature -10 to 55 ◦C

3.2.2 Operational Rules

As a part of the plant, the BESS system is operated by the central supervisory system (CSS) through a Programmable Logic Controller (PLC) and SCADA data communication and acquisition. PLC issues commands for BESS operation considering PV plant output, campus load demand and other relevant parameters for network. The operational commands are decided by the CSS based on the collected measurements and processed information of the campus load and PV generation. The control system for the BESS operation is established in a programmable manner to use the whole range of capabilities 3.2 PV INTEGRATED BESS UNDER STUDY 21

Figure 3.2: Battery arrangement and bank level management system for testing of new BESS operation strategies. Currently, there are seven modes of operation for the BESS such as PV charging, off-peak charging, and export to the grid. Multiple operation modes can be enabled at the same time. However, only one mode of operation at each moment drives the BESS based on a pre-deﬁned priority list. The modes of operation, their priority and timing are explained as below: Delta Solar: Discharging of BESS occurs in this option when there are ﬂuctuations in solar power. It helps to smooth output generation by PV plant. The battery is discharged when the PV generation changes at 10 kW/s rate, and the discharge rate decreases with the reduction of solar power change slowly. Maximum Supply: The mode is to ensure the level of imported energy is controlled by discharging BESS to reduce the cost of the campus due to demand charge.The operation takes place when power supply without BESS exceed the maximum import level. Peak Swap: Another cost-effective mode is BESS discharging operation during the energy peak period. The two battery banks have their own operational time by maintaining 50% of discharge limit as below:

• Battery Bank 1 discharges between 7pm and 10pm

• Battery Bank 2 discharges between 8pm and 11pm

Export Limit: The BESS charges when supply level without BESS drop below 100kW. Charge rate increases at a maximum rate of 10kW/s and the quantity ramps back to zero at a rate of 10kW/s. Solar Charge: One of the important operations is to charge both battery banks with PV generated power. The battery banks are charged according to the time and conditions given below:

• Charging operation is activated between 11am and 7pm 3.2 PV INTEGRATED BESS UNDER STUDY 22

• It only charges when solar generated PV power exceeds 800kW

Off-Peak Charge: Another charging operation is required to ensure the BESS is charged adequately in the absence of excess PV power. The off-peak charge operations take place as per below time and conditions.

• Battery Bank 1 is activated to charge from 12am to 6am

• Battery Bank 2 charging operation activates from 1am to 7am

• Rate of charge is at 50% of CSS deﬁned charge limit

• The operation continues until the battery charge reaches to 60%

Trickle Charge: The mode is to charge the BESS when the charge level fall below minimum. The operational rules for this plant are:

• It is activated when the SoC of the BESS is below 5%

• The operation stops when the SoC reaches above 20%

• The rate of charge is at 50kW

Both battery banks in the BESS are operated following the above operational rules for maximum beneﬁt out of large-scale storage system in the PV plant.

3.2.3 Cooling Mechanisms

In order to keep the battery cell and battery room temperature within an acceptable range, a cooling system is used in the battery container. Battery module and cell ventilation is achieved through a) passive cooling mechanism including air vent holes along the side of the battery casing and spacing of cells, which ensures an even temperature distribution between the cells, and b) active cooling consists of rack fans and air-conditioning unit with 7.7 kW rated cooling capacity. Three fans are placed at the top of each battery bank to draw air up through vents in the front panel by passing through the modules and out at the top. The fans start operating when cell temperature reaches 29 ◦C. The consumption of each rack fan is 44.64W, which is powered by an external +24V DC system. Battery modules are spaced inside the rack with gaps to allow airﬂow. The air conditioning unit is set at the room temperature of 23 ◦C, which indicates that the evaporator of the air conditioning unit starts its operation along with fan when the room temperature reaches above 23 ◦C. The fan installed inside the air conditioning unit solely operates when room temperature is below or equal to 23 ◦C. 3.3 DATA FOR ANALYSIS 23

The room temperature is measured and recorded by the thermostat installed with the air conditioning unit. More energy consumption by the cooling system occurs due to higher room temperature.

Figure 3.3: Cooling system arrangement for BESS

Figure 3.3 shows the cooling system inside the conﬁned container to keep the battery room temperature at an acceptable level. It is noticeable that cooling system is installed in all level of the container to conﬁrm proper heat dissipation generated from battery.

3.3 Data for Analysis

As a dedicated research facility, UQ GSRF has sophisticated data logging and collection system with utmost importance for understanding the system behaviour. This section explains the data management system for the entire plant and selection criteria that has been used for preliminary investigation of battery cell temperature behaviour with respect to most inﬂuential factors.

3.3.1 BESS Monitoring and Data Logging System

A comprehensive set of data containing electrical, control, and physical parameters of BESS, operational parameters of the PV system and the network measurement are monitored and logged. In order to investigate energy storage system operation, a sophisticated BMS for each battery bank is installed, which can provide critical information such as cell temperature, voltage, current, power etc.. About 1390 points within the BESS is monitored through measurement and data logging system, which gives detailed insights into the battery operation on the cell and module levels. CSS PLC receives all the data from each responsible PLC within the plant, e.g., import/export energy from/to the grid, BESS, capacitor bank, onsite diesel generator, and ﬁve solar array generation through a sophisticated SCADA system. All the CSS collected data for the plant is remotely accessible through a Wonder- ware Historian system interfaced. A signiﬁcant number of variables including weather parameters, DC and AC side parameters of different inverters, are measured and stored in Wanderware Historian 3.3 DATA FOR ANALYSIS 24 system. The plant data logging, including BESS, is performed using a Delta Mode operation. Al- though data is sampled at a 1-second rate, the Delta Mode operation is adopted to reduce data storage capacity requirement. In this procedure, a measurement is recorded when the value is different from the previous measurement by a certain pre-set threshold [116]. In this thesis, 1-minute sampled data has been considered for all modelling, investigations and validation studies.

3.3.2 Data Selection for Battery Cell Temperature Thermal Behaviour Inves- tigation

In order to analyse thermal behaviour of battery cell temperature with respect to its operations and ambient temperature, a good number of data is collected from UQ GSRF. 1-minute sampling rate data of a year from 26th March 2016 to 25th March 2017 has been considered for this investigation. More than 100 incidents of charging and discharging have been selected from 365 days of data to observe the charge/discharge and ambient temperature effects on the battery cell temperature. In order to eliminate the impact of ambient temperature, only incidents with battery cell temperature above the ambient temperature for the whole period of charge/discharge are selected for this investigation. In addition, it is made sure that the ambient temperature is not increasing during the period of charge/discharge to further neglect the ambient temperature impact. Finally, an investigation has been carried out by selecting a good number of days with higher ambient temperature to observe its influence on battery cell temperature. Moreover, the impact of charging/discharging operations during high ambient temperature is also examined. The plan is to evaluate the impact of charging/discharging current, rate of temperature change, and charge on the battery temperature. Using a nonlinear relationship between temperature and battery degradation, an accumulated extra degradation is estimated for the charging and discharging incidents above 30 ◦C. In an attempt to find a relationship between effective parameters on the battery temperature, several parameters are defined that will be investigated on the data, as follows:

• Temperature Rising Slope ( ◦C/min): the slope of the straight line going through the minimum and maximum temperature values during an event.

• Peak Temperature ( ◦C): the maximum temperature occurred during an event.

• Absolute Temperature Change ( ◦C): the difference between the maximum and minimum temperature during an event.

• Peak Current (A): the peak value of the current during the event under study. 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 25

• Charge (C): total charge of a charging or discharging event in hourly rate.

• Temperature Rising Delay (min): the difference between the time at the beginning of an incident and the time in which battery temperature starts to increase. This parameter is important for battery thermal modelling.

The initial study revealed a similar behaviour among different battery cells and modules in terms of temperature and charge/discharge current. Therefore, the average cell temperature of a single battery module is considered in the analysis representing the overall BESS. Moreover, due to different behaviour and performance of battery during charging and discharging modes, their analyses is carried out separately. For the degradation analysis, however, both charging and discharging modes are combined.

3.4 Analysis of Battery Cell Temperature Behaviour

A comprehensive investigation study of battery cell temperature behavior is carried out in this section using the selected data described in 3.3.2. Most influential factors for battery cell temperature upsurge are considered for investigation and finally a battery degradation model has been used to draw a temperature-dependent degradation rate for various battery cell temperature regimes. In the following sections 3.4.1 and 3.4.2, ambient temperature effects have been excluded from the analysis to observe only charging and discharging current influences on battery cell temperature.

3.4.1 Discharging Operation of BESS

One of the primary operations of the BESS is to discharge to smooth PV output during any sporadic changes in solar irradiance and export energy to the grid during peak-time periods. Operating in these occasions, battery cell temperature rises signiﬁcantly based on the peak discharge current, and duration of the discharge event. In this study, 50 discharging events have been selected to observe battery temperature behavior. As an example, a discharging event from 18:00 to 23:00 on 25/03/2017 has been shown in Figure 3.4. It is noticeable that the battery cell temperature increases with the discharging current while the ambient temperature is below battery cell temperature. It indicates that the ambient temperature has no effect on the rising cell temperature in this occasion. The cell temperature starts rising at 19:08 with 4 minutes delay from discharge starting time (19:04). It then ramps up to the peak temperature at 27.25 ◦C, steadily. It can be seen from Figure 3.4 that the temperature starts to fall gradually when the discharging current decreases. In this case, the rate of 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 26 temperature change (from point B to point C) is lower than the increasing rate. The reason is twofold: reduction in discharge current rate and cooling system operation.

Figure 3.4: Current and ambient temperature effect on battery cell temperature during a typical discharging event

Figure 3.5 shows a meaningful relationship between peak current and battery cell temperature rising slope. Most of the discharging events demonstrate that the temperature rising slope increases with increase in peak discharging current. The events with higher current (110A-125A) veriﬁes the same hypothesis as the events with lower discharge current (42A-65A). Due to inadequate data in between the highest and lowest range, it is difﬁcult to draw a conclusion for the entire range of peak current. In Figure 3.6, the peak temperature values are plotted against the peak current where a strong association can be realized. It appears that the peak temperature behavior can be described by the peak discharge current for the lower and upper range of peak current. In addition, the cell temperature increases linearly with upsurge of the peak discharge current with a linear function. The impact of charge on the temperature rising slope is shown in Figure 3.7. As analysis shows, two different clusters can be recognized. While the charge of the event is below 80 coulomb (C), the temperature follows a linear relationship with a small slope. When the value of the charge reaches around 80C and beyond, it is noticeable that temperature rising slope jumps upward. It proves that the higher charge accelerates battery temperature rising slope. As battery temperature increases gradually after starting an event, cell temperature takes time to reach its peak value due to heat convection delay, cooling system operation, and the battery physical layout in the container. The average temperature rising delay is 5.5 minutes, where standard deviation is 5 minutes. Couple of incidents shown in Figure 3.8, have 9 minutes delay. The average time delay 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 27

0.1000 0.0900

0.0800 º C/min) º 0.0700 0.0600 R² = 0.9231 0.0500 0.0400 0.0300 0.0200 0.0100 Temperature Rising Slope ( Slope Rising Temperature 0.0000 0 20 40 60 80 100 120 140 Peak Current (A)

Figure 3.5: Peak current and temperature rising slope: A linear relationship for discharging events

31.0 R² = 0.8924 30.5 30.0 29.5

º C) º 29.0 28.5 28.0 27.5 27.0 26.5

Peak Temperature ( Temperature Peak 26.0 25.5 0 20 40 60 80 100 120 140 Peak Current (A)

Figure 3.6: Peak current and peak temperature: A linear relationship for discharging events

0.1000 0.0900

0.0800 º C/min) º 0.0700 0.0600 R² = 0.6475 0.0500 High Charge States 0.0400 Low Charge States 0.0300 0.0200 0.0100 R² = 0.1981 Temperature Rising Slope ( Slope Rising Temperature 0.0000 0 20 40 60 80 100 120 Charge (C)

Figure 3.7: Temperature rising slope vs charge of discharging events shows a reduction to 5 minutes without these incidents. The different ambient temperature and/or the condition of the cooling system at the beginning of the discharging event must have caused the 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 28 differences.

10.00 9.00 8.00 7.00 6.00 5.00 4.00 3.00 2.00

Temperature Rising Delay (min) Delay Rising Temperature 1.00 0.00 1 6 11 16 21 26 31 36 41 46 Number of Event

Figure 3.8: Temperature rising delay for each discharging event

5.0

4.5 C) ° 4.0 3.5 3.0 2.5 2.0 1.5 1.0

AbsoluteTemperature( Change 0.5 0.0 1 6 11 16 21 26 31 36 41 46 Number of Event

Figure 3.9: Absolute temperature change during each discharging event

Although temperature rising slope differs from one event to another as per in Figure 3.9, most of the events yield only 1 ◦C increase on average. As per results shown in Figure 3.9, six discharging periods are yielding signiﬁcantly higher changes in temperature. The root-cause of the unusual behavior is possibly high discharge current and comparatively higher cell temperature prior to discharging period together with possible malfunction in the cooling system.

3.4.2 Charging Operation of BESS

Like discharging operation, charging is an indispensable operation in large-scale PV plant where excess solar energy is stored for later utilization. Through the year from April 2016 to March 2017, the BESS was in different operational mode for research purposes. A total number of 61 charging 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 29 events from one-year available data have been considered in this study for analysis. Charging event on 11th February, 20:00 to 12th February 08:00 in 2017 is depicted in Figure 3.10 to show a relationship between cell temperature and the charging current. The cell temperature is higher than the ambient temperature for the period of charging. More importantly, the ambient temperature decreases over the same period of time, which does not correlate with the cell temperature increase. Cell temperature increases from the beginning of charging period. Then, it starts to decrease from peak point, B, to post charging minimum point, C, because of the reduction in the magnitude of the charging current.

Figure 3.10: Current and ambient temperature effect on battery cell temperature during a typical charging event

0.06000

0.05000

ºC/min) R² = 0.3128 0.04000

0.03000

0.02000

0.01000 Temperature Rising Slope ( Slope Rising Temperature 0.00000 0.00 20.00 40.00 60.00 80.00 100.00 Peak Current (A)

Figure 3.11: Relationship between peak current and temperature rising slope for charging events

Figure 3.11 and Figure 3.12 show relationships between temperature rate of a change and peak 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 30 current, and peak temperature and peak current, respectively for each charging event. Both indicate that overall time series only have approximately 30% correlation between temperature changes and peak current. The R-squared value is not large enough to show a strong correlation. Therefore, it is not possible to explain cell temperature changes with these two predictors individually using a linear correlation. To ﬁnd the reasons for this phenomenon, a couple of hypotheses can be drawn:

• Although the ambient temperature never exceeds the battery temperature over the period of the charge event, its changes can have impact on the battery cell temperature. For instance, the ambient temperature in Figure 3.4 is slightly decreasing while the one in Figure 3.10 is decreasing quickly. The absolute change in the ambient temperature is more signiﬁcant in the charging event shown in Figure 3.10, which can contribute in reducing the temperature rising slope. Therefore, a combined thermal model of the battery, cooling system, and container accounting for the ambient temperature is necessary to model the external factors more effectively.

• In this study, a linear relationship between the battery temperature and effective parameters is investigated. However, the relationship might be of higher order and nonlinear with a combination of multiple effective parameters.

• As it was mentioned earlier, only charging/discharging events are selected for analysis where battery temperature is always higher than the ambient temperature. Therefore, a small sample of events could meet the requirements for this study. For a strong statistical analysis, a larger sample size is required.

50 45 40 C) R² = 0.2624 º 35 30 25 20 15

Peak Temperature ( Temperature Peak 10 5 0 0.00 20.00 40.00 60.00 80.00 100.00 Peak Current (A)

Figure 3.12: Relationship between peak current and peak temperature for charging events

Although the temperature rising delay is more than 14 minutes during couple of charging events, as shown in Figure 3.13, the average delay is around 6.4 minutes without these events. The average 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 31

20.00 18.00 16.00 14.00 12.00 10.00 8.00 6.00 4.00

Temperature Rising Delay (min) Delay Rising Temperature 2.00 0.00 1 6 11 16 21 26 31 36 41 46 51 56 61 Number of Event

Figure 3.13: Temperature rising delay for every charging incident rising time will rise to around 7 minutes with the incidents with higher delay. On average, an absolute temperature change is 2.5 ◦C for the charging event. It is about 60% more compared to the average value of discharging events. Only one charging event in Figure 3.14 has less change (0.5 ◦C) in battery cell temperature.

4.50

4.00

C) º 3.50

3.00

2.50

2.00

1.50

1.00

Absolute Temperature Change ( 0.50

0.00 1 6 11 16 21 26 31 36 41 46 51 56 61 Number of Event

Figure 3.14: Absolute temperature change during each charging event

According to Figure 3.15, the relationship between the rate of temperature changes and charge is negligible as the R-squared value is very small. As explained earlier, it could be linked to the effect of ambient temperature, cooling system operation, and small sampling size.

3.4.3 Ambient Temperature Effects during BESS Operations

Although, lithium battery performs well at higher temperature, prolong exposure to the high temperature increases cell temperature. Higher cell temperature is responsible for accelerating battery life 3.4 ANALYSIS OF BATTERY CELL TEMPERATURE BEHAVIOUR 32

0.06000

0.05000 ºC/min) 0.04000

0.03000 R² = 0.0439

0.02000

0.01000 Temperature Rising Slope ( Slope Rising Temperature 0.00000 0.00 20.00 40.00 60.00 80.00 100.00 120.00 Charge (C)

Figure 3.15: Temperature rising slope vs charge of charging events degradation. Higher ambient temperature will hamper a charging operation by increasing battery cell temperature. The ambient temperature will have a strong negative impact on increasing battery cell temperature in a confined BESS container. Figure 3.16 shows that battery cell temperature increases with the changes in ambient temperature in absence of BESS operation for first couple of days. It can be seen that every time the ambient temperature starts increasing, the battery cell temperature is also increasing with a delay and similar patterns is followed when the ambient temperature decreases during night. On 1st January 2017, the battery cell temperature increased to 40 ◦C with the elevating ambient temperature. The cell temperature of battery start decreasing after 2nd January 2018 and it is noticeable that the overall ambient temperature during the same time is also decreasing. The battery cell temperature decreased to below 30 ◦C from 40 ◦C with the changing ambient temperature from the day range of 22 ◦C - 27 ◦C. In order to reduce the impact of charging and discharging current on the battery cell temperature mentioned in earlier sections, the analysis has been carried out for the events when there is no operations of BESS. Figure 3.16 demonstrates that battery cell temperature is increasing with the elevated ambient temperature while battery is completely idle. More specifically, there is no BESS operations from 27th December to 29th December, however, battery cell temperature increases to 29 ◦C from 24 ◦C. It also suggests that the cooling system operations such as battery rack fans and air conditioning unit were not capable enough to reduce the battery cell temperature at an acceptable level by reducing external effects. Therefore, accumulated heat inside the battery room, accelerated battery cell temperature. Moreover, it is noticeable that battery cell temperature increasing rate upsurges on 29th December when BESS under operation. It is also observed that battery cell temperature is above ambient temperature during the BESS operations, which accelerates heat losses through convection, conduction, and radiation. Therefore, both charging/discharging and ambient temperature are 3.5 TEMPERATURE-DEPENDENT BATTERY DEGRADATION STUDY 33 responsible for battery cell temperature behaviour. The investigations suggest a strong relationship between battery cell temperature and ambient temperature, which has to be considered in an effective temperature estimation algorithm.

Figure 3.16: Ambient temperature effects on battery cell temperature

3.5 Temperature-Dependent Battery Degradation Study

So far, it has been shown that charge/discharge events and ambient temperature increases battery temperature beyond the safe range. Quantification of the extra degradation imposed on the battery because of excessive temperature is discussed in this subsection. To do so, the modified Zhurkov model is used for battery degradation analysis from [117]. It estimates the impact of battery temperature on its degradation compared to a reference temperature, as follows: T−Tnom Tfact× Tnabs DegradationRATE = e Ta (3.1) where, Tfact = 0.0693 is the coefficient of temperature in thermal ageing model; T is the actual battery ◦ ◦ temperature in C; Tnom is the reference battery temperature (25 C in this study); Tnabs is the reference battery temperature in Kelvin (K) and is calculated by Tnabs=Tnom+273; and Ta is the battery absolute temperature in Kelvin (K). To quantify the extra battery degradation caused by excessive temperature, Eq. (3.1) has been used in two following scenarios. In Scenario I, the extra degradation rate is calculated for the actual measured temperature. In Scenario II, however, it is assumed that the battery temperature never exceeds 30 ◦C, which is an ideal case [15, 89]. By comparing two Scenarios, the 3.5 TEMPERATURE-DEPENDENT BATTERY DEGRADATION STUDY 34 extra degradation imposed on the battery by the charge/discharge events and the resultant excessive heat, can be quantified.

Figure 3.17: Average extra degradation rates for the actual temperature in each charge/discharge event

Figure 3.17 and Figure 3.18 illustrate the extra degradation rate because of higher temperature during charging and discharging events. It is observed from a couple of incidents illustrated in Fig- ure 3.18 that same charging or discharging event causes extra degradation for temperature above 30 ◦C. Therefore, battery degradation rate would be signiﬁcantly lower if the temperature can be kept at maximum allowable range, i.e., 30 ◦C. The degradation rate jumps to 104% from 52.43% for only 3.9 ◦C excess temperature. The extra degradation can be avoided by regulating charge/discharge current.

Figure 3.18: Comparison of degradation rates between Scenario I and Scenario II 3.6 SUMMARY 35

3.6 Summary

This chapter of thesis made an effort to render preliminary investigations on the battery cell temperature behavior and the battery degradation caused by excessive temperature for different charge/discharge regimes and ambient temperature in a large-scale PV plant. Actual BESS field data has been used to investigate temperature behavior based on charging and discharging current and ambient temperature. Both states, i.e., charging and discharging, show that the battery temperature starts increasing due to high current. While there is a strong correlation between the battery temperature and the temperature rising slope, peak current, and charge in discharging incidents, no particular linear correlation has been identified for the charging events. In addition, it is determined that continuous higher ambient temperature does have effects on the cell temperature increasing rate. In addition, battery cell temperature changes with the variations in ambient temperature. Finally, the estimation of degradation rate shows that, significant number of events in a year are responsible for extra battery degradation and the rate exponentially increases with the higher temperature. The next Chapter will further extend the study in this chapter to develop a model to estimate battery cell temperature accurately. In addition, development of two different models for 24 hours ahead battery cell temperature estimation will be discussed. 4 Battery Cell Temperature Modelling

4.1 Introduction

1 Battery technology is playing a vital role in the power system for safe and secure operation. However, ageing of battery is one of the major concerns for its well acceptance in the power system. Elevated battery cell temperature is one of the highly responsible factors for degradation of battery. Moreover, battery cell temperature is directly responsible for increasing temperature of a conﬁned storage room, which escalates extra cooling cost. Therefore, it is important to estimate battery cell temperature evolution over time with respect to a planned operation regime and ambient interactions to have a better evaluation of the incurred cost. In this chapter of thesis, an unprecedented data-driven approach is developed for estimating battery cell temperature considering effective parameters. In addition to the non-linear approach, linear data-driven approach is also developed for establishing a

1This chapter has materials from the following references submitted/published by the MPhil candidate. 2. Md Mehedi Hasan, Ali Pourmousavi, Tapan K. Saha, A data-driven approach to estimate battery cell temperature using NARX neural network model. (Submitted to Applied Energy). 3. Md Mehedi Hasan, S. Ali Pourmousavi and Tapan K. Saha, Battery Cell Temperature Estimation Model and Cost Analysis of a Grid-Connected PV-BESS Plant. (IEEE ISGT Asia 2019)

36 4.1 INTRODUCTION 37 comparison analysis. Both of the models are developed for 24 hours ahead battery cell temperature estimation. The rest of the chapter is organized as follows: Subsection 4.1.1 and 4.1.2 explain about existing estimation approaches and signiﬁcance of the proposed method, respectively. Section 4.2 explains the proposed methodology. Section 4.3 presents the system under study, Section 4.4 explains a wide range of results and discussions, and Section 4.5 shows a comparison study between linear and non- linear models with a good number of results and discussions. Finally, the chapter is concluded in Section 4.6 with a summary of the model.

4.1.1 Thermal Behaviour Estimation Approaches

A number of research have been conducted on the battery thermal management of EVs and HEVs, where they proposed general models of battery thermal behaviour in [70–76] or specialized models for cold weather or lower range of temperature [23, 77]. However, there is no such research work to investigate thermal behaviour of a utility-scale BESS. In this thesis, thermal behaviour of the battery with respect to charge/discharge activities and ambient weather conditions are considered. It is noticeable that in most of the studies, thermoelectric, battery physical model and thermal-conductivity based approaches have been followed to analyse thermal behaviour of lithium-ion batteries [23, 65, 118]. These approaches require laboratory-based experiments for a single battery module, considering various factors such as battery ambient temperature, age, SoC and other operational conditions. In [119, 120], complicated thermo-dynamical models are developed for sealed battery packs in EVs and HEVs in contact with a cooling agent. While the models are accurate, the application of such models could become overwhelmingly complicated in the utility-scale BESS due to the large number of battery modules. Moreover, some of these models typically need several internal parameters to be measured continuously [78], which might become an expensive and unscalable solution for the utility-scale application. In addition, detailed thermal modelling of the entire utility-scale BESS is time consuming as it requires complex thermal modelling and intensive computational efforts that is not feasible for the energy management application. In other words, modelling the complex nature of thermal interactions between thousands of battery cells with each other and the larger surrounding area, and active/passive cooling mechanisms through thermodynamics models will be a computationally expensive approach.

4.1.2 Signiﬁcance of Data-Driven Approach

Based on the discussion in Subsection 4.1.1, a physical-thermodynamic model of utility-scale BESS is not feasible nor necessary. Therefore, a physics-free data-driven modelling is preferred in 4.1 INTRODUCTION 38 this study. The approach only requires input and output data of the system to identify its behaviour with respect to different input variations. Another advantage of data-driven models over analytical methods is that they can cope with the time-varying internal parameters of the BESS by receiving new data and re-training. For instance, higher internal resistance of the battery cell is inevitable when the cell ages, which contributes to higher cell temperature by excessive heat generation during charging/discharging activities. For analytical methods to work accurately, internal resistance should be measured periodically by interrupting battery operation and spending money on measurement tools and personnel. With data-driven method, however, new measurement of external parameters and re- training the model can modify it according to the new changes in the battery cell operation. It can be done continuously without requiring expensive procedures. Therefore, the overall accuracy of the model will improve substantially with this low-cost solution over the lifetime of the BESS. Artificial intelligence methods are widely used in various research fields to model the unknown or complicated physical systems based on measured input and output parameters, such as fuzzy system, nonparametric regression, wavelets, neural network, or a combination of them which is known as hybrid system [121–123]. These methods are extensively used to model non-linear black-box systems in different applications [124–126]. The studies conducted with these methods indicate that any complex nature of dynamics and nonlinearity can be modelled effectively. Therefore, these methods can be considered for modelling battery thermal dynamics in this study using available input and output data of the battery operation. In this procedure, historical data of the input and output parameters are used to train a model considering the most influential independent variables. In this study, Nonlinear Autoregressive Network with Exogenous inputs (NARX) is proposed to estimate lithium-polymer battery cell temperature. It is shown that NARX is more effective in terms of learning the long temporal dependencies than the other complicated methods such as wavelets, nonparametric regression, and fuzzy models [127]. Instead of developing a model for a single cell in a controlled experimental setup, the proposed model is created from the operation of thousands of cells affecting each other in a large-scale BESS. The main goal is to create a thermal model of the battery cells within a container equipped with passive and active cooling mechanisms, which can be utilised in the EMS. Also, the non-linear nature of the phenomenon (i.e., the cell temperature variations due to charge/discharge regime and thermal interactions among cells with outside and cooling mechanisms) can be effectively captured by the NARX model. As a black-box modelling approach, NARX does not need a physical model of the underlying system and the thermal interactions among different mediums and effective parameters [128]. The model can be created by battery operation and ambient temperature historical values measured in a utility- scale PV-BESS plant. In this study, the best structure of the NARX model (i.e., number of neurons 4.2 DATA-DRIVEN BATTERY CELL TEMPERATURE MODELLING 39 and feedback delay) is selected by sensitivity analysis. Minute-by-minute field data, collected from The UQ GSRF, is used in this study by dividing data into training and test sets for the NARX model development and evaluation, respectively. In addition, in an attempt to improve accuracy of the model and to evaluate seasonal impact on the cell temperature modelling, seasonal NARX models are also developed using training dataset and evaluation of each model is done using different set of seasonal data (i.e., test dataset) for each season. The results obtained from the model considered entire-year data are compared with the ones obtained from seasonal models, where the latter outperformed the former. Finally, a linear data-driven model has also been established to show a comparison between non-linear and linear models.

4.2 Data-Driven Battery Cell Temperature modelling

The NARX approach based on ANN is proposed for battery cell temperature modelling considering exogenous parameters and feedback signal. The NARX model is based on the linear autoregressive model, which is commonly used in time-series modelling. The model accepts exogenous parameters as well as a feedback from the output with a certain delay. In this thesis, exogenous parameters are the battery current and ambient temperature. Battery current (i.e., charge/discharge and idle regime) has direct inﬂuence on the battery cell temperature because of the ohmic and non-ohmic internal losses [74]. Also, the ambient temperature changes the temperature inside the container, which essentially changes cell temperature with some delay. Battery cell temperature, similar to many other natural phenomena, does not change like a step function. As it was shown in [129], battery cell temperature starts increasing gradually after a charge or discharge incident due to heat convection delay, physical layout of batteries in the container, and cooling system operation. Therefore, limited number of samples from the past can be helpful to estimate future output with higher accuracy, which requires a feedback from output in the model. The NARX model can be mathematically represented as,

y(t) = f (y(t − 1),y(t − 2),..,y(t − ny),u(t − 1),u(t − 2),..,y(t − nu)) (4.1) where, y is the output or dependent variable; u refers to the input parameters; t denotes to the discrete time-step; and nu and ny represents the number of time-step delays for input and output parameters, respectively. Non-linear mapping function, f (·), is generally unknown and it is represented by using a standard multilayer perceptron (MLP) network, a dominant structure of learning any type of continuous non-linear mapping, and proper training algorithms [130, 131]. To train an MLP feed-forward ANN, backpropagation algorithm can be used as a supervised learning technique [132]. 4.2 DATA-DRIVEN BATTERY CELL TEMPERATURE MODELLING 40

Similar to typical ANN structure, the NARX network consists of three layers, namely input, hidden and output layers. Figure 4.1 shows block diagram of an NARX model. Each of the layers consists of speciﬁc number of neurons. In this study, a three-layer feed-forward network with a sigmoid transfer function and a linear function in the hidden and output layers, respectively, is used for the approximation of the function, f (·). The estimated output is used as an input parameter to estimate the battery cell temperature in the future time steps. The weights and biases, shown in Fig- ure 4.1, are important to optimise the network performance. Tuning the values of the weights and biases occurs during training of the NARX model. If there are any difference between estimated and actual output values, weights and biases will be adjusted during training process using gradient descent-based (Levenberg-Marquardt in this study) algorithm to minimize the errors [133, 134]. The network training process will continue with the same input and output values until the error reaches to an acceptable limit. Estimated output data is then fed back to work as input with other exogenous input parameters to estimate the battery cell temperature during testing stage. Tapped delay line (TDL), shown in Figure 4.1, is an embedded memory in feed-forward network, which stores present and previous time series as per the deﬁned delay.

Hidden Layer Input Layer Output Layer

Exogenous TDL weight weight Inputs Sigmoid Linear Estimated + + Function Function Output TDL weight bias

bias

Figure 4.1: NARX structural model based on MLP neural network

Two different architectures of NARX are used for training and ﬁnally estimating battery cell temperature, which are series-parallel and parallel architectures [135]. Series-parallel architecture, shown in Figure 4.2 and Eq. (4.2), is used during training stage, where actual battery cell temperature is available from the past observations. The future values are estimated from the present and past values of the input variables, u(t)...u(t − nu) and true battery cell temperature from the past, i.e., y(t −n),...,y(t −ny). The advantage of this model is that the trained network has a purely feed-forward architecture and static backpropagation can be used for training. The series-parallel architecture can 4.2 DATA-DRIVEN BATTERY CELL TEMPERATURE MODELLING 41 be represented as:

yˆ(t)Training = fˆ(y(t − 1),y(t − 2),..,y(t − ny),u(t),u(t − 1),u(t − 2),..,u(t − nu)) (4.2)

u(t), ,u(t-nu) Inputs (Current and TDL Ambient Estimated output Feedforward (Battery Cell Temperature) Temperature Network y(̂ t) Actual Output TDL Training (Battery Cell Temperature)

y(t-1), , y(t-ny)

Figure 4.2: Series-parallel architecture for NARX during training where,y ˆ(t)Training is the Series-parallel mode output. However, the series-parallel architecture is not capable of estimating multi-step ahead of battery cell temperature. Therefore, parallel architecture, shown in Figure 4.3 and Eq. (4.3), is used to perform future estimation for the test data and real-world simulation. The estimation is performed by taking the past and present values of the input parameters, i.e., u(t),...,u(t − nu) and previously estimated battery cell temperature values, yˆ(t)Testing,...,y ˆ(t − ny)Testing during testing stage, the results of which will be shown and discussed in Section 4.4. The parallel NARX architecture can be shown mathematically by:

yˆ(t)Testing = fˆ(yˆ(t − 1)Testing,yˆ(t − 2)Testing,..,yˆ(t − ny)Testing,u(t),u(t − 1),u(t − 2),..,u(t − nu)) (4.3) where,y ˆ(t)Testing is the parallel mode output.

u(t), ,u(t-nu) Inputs (Current and TDL Estimated output Feedforward Ambient (Battery Cell Temperature) Network Temperature y(̂ t)Testing TDL

y(̂ t-1)Testing , , y(̂ t-ny)Testing

Figure 4.3: Parallel architecture for NARX during testing and actual simulation

The number of neurons in the input and output layers of the NARX model are equal to the number of independent (i.e., current and ambient temperature) and dependent variables (i.e., battery cell 4.2 DATA-DRIVEN BATTERY CELL TEMPERATURE MODELLING 42 temperature), respectively. Only one hidden layer is considered in the NARX model as per universal- ity theorem [136], where neural network with single hidden layer is able to approximate continuous functions with desired accuracy most of the time. Selecting the number of neurons in the hidden layer is important to have the most efficient and accurate network. Employing few or too many neurons in the hidden layer will result in underfitting or overfitting, respectively. Underfitting occurs when small number of neurons are unable to learn the patterns in the training signal adequately. On the other hand, overfitting happens when the trained model rely on the training dataset too much, which might provide unacceptable results for different test dataset. In addition to the importance of the number of neurons in the hidden layer, the time delay (i.e., the number of samples from the past) of the exogenous input and feedback signals are important in the performance and accuracy of the model. A sensitivity analysis has been proposed in this thesis to select the best number of neurons in the hidden layer and time delay in the feedback signal [137]. In this analysis, different time delays (10, 20, 30, 40, 50 and 60 minutes) and number of neurons (2, 4, 6, 8, 10, 12, 14, 16, 18, 20) have been used to train the neural network and then tested the model with completely different set of data. Data is nor- malised for training and model evaluation purposes throughout this chapter, where maximum value of input and output parameters in the historical data are considered as the base value to normalise the data. Normalising the original data can eliminate negative influence of large numbers and is also apt to improve the convergence rate of the training process [138]. Three standard measures are used in this study to quantify the accuracy of the NARX model in estimating battery cell temperature. Root mean squared error (RMSE), correlation coefficient (R) and 2 2 adjusted R (RAd justed) given in Eq. (4.4), Eq. (4.5) and Eq. (4.6), respectively, are used to determine the trained network outcomes in training, testing and validation stages. The best trained NARX model 2 is indicated by the RMSE value close to 0, R and RAd justed value close to 1.

v u N u u (T − T )2 u∑ ai mi t i=1 RMSE = (4.4) N

N (T − T ) × (T − T ) ∑ ai a mi m i=1 R = v (4.5) u N N u (T − T )2 × (T − T )2 t∑ ai a ∑ mi m i=1 i=1

(1 − R2)(N − 1) R2 = 1 − (4.6) Ad justed N − k − 1 where, in Eq. (4.4), Tai , Tmi and N represent actual battery cell temperature, estimated battery cell 4.3 SYSTEM UNDER STUDY 43 temperature using NARX model and the total number of samples, respectively. In Eq. (4.5), T a denotes mean value of actual temperature for battery cell. Similarly, T m represents the average of estimated temperature. In Eq. (4.6), k represents the number of variables in the model. Since network training involves solving an optimisation problem, the values of biases and weights might be different in multiple runs of the training process with the same input and output values [139]. This also can affect the performance of the network when it is used for different test dataset because different NARX network can give altered results for the same set of test samples. To resolve this issue, neural network is trained several times for the better accuracy. In this model, one year of minute-by-minute data has been used to train an NARX network for estimating battery cell temperature. One year of data contains significant number of charging/discharging and ambient temperature incidents, which covers many possible charging/discharging and ambient temperature combinations. In this study, two types of models are created. In the first type, the data from the entire training set is used to develop a single NARX model for the entire year (called universal NARX model). In the second type, training data is partitioned in four different seasons and a specialised NARX model is developed for each season (called seasonal NARX model). In this case, apart from whole-year modelling structure, seasonality has been given priority since ambient temperature is highly dependent on the seasonal variations with different characteristics. Moreover, PV generation behaves differently in various seasons because it dominantly depends on the solar irradi- ation and ambient temperature. As solar radiation varies in seasons, the charging and discharging activities of the BESS varies accordingly. Therefore, more fluctuations in charging and discharging may occur during winter compared to summer or vice versa. As will be shown in the simulation results, the seasonal NARX model outperforms the universal model considerably.

4.3 System Under Study

In this modelling, similar UQ-GSRF is considered for operational data. Figure 4.4 shows a genral overview of the battery conﬁguration in the plant. Please refer to Section 3.2.1, [129] and [116] for more details. In this study, 1-minute sampled data has been considered for training and testing the model, which are further explained in the next section. 4.4 SIMULATION STUDY 44

Connected to Campus Substation 11 kV

1 MVA Transformer 0.04 kV

Bi-Directional Inverter

Battery Bank 1 Battery Bank 2

Figure 4.4: Schematic diagram of the BESS conﬁguration in the PV plant

4.4 Simulation Study

In this study, two independent variables, i.e., battery current and ambient temperature, have been considered to create and test battery cell temperature estimation model. The following data selection and analysis are carried out in the simulation studies for training and assessing the NARX networks: • For training purposes, 1-minute data of around 12 months from 1st April 2016 to 31st March 2017 have been selected, which consists of 516,960 samples (359 days excluding 6 days used for one of the seasonal testing purposes) overall. Entire year is covered in the data to represent seasonal differences in the real-world condition. A universal NARX model is created for battery cell temperature estimation for the entire year. Another 90 days equivalent data from 1st April 2017 to 7th July 2017 is used for validation and testing the trained model during training process. • Seasonality analysis has been carried out by developing separate NARX model for each season (i.e., seasonal NARX model). Similar procedure as universal model has been followed for training each seasonal model. Multiple days were selected from each season as for assessing performance of the respective seasonal model. • For further assessment of universal network model, 56 days of data from 10th July 2017 to 31st October 2017 and test days (6 days x 4 seasons) used for seasonal model assessment are considered. The impact of time delay on the performance of the model has also been analysed for feedback signal 4.4 SIMULATION STUDY 45 by testing the network with test days and NARX with feedback configuration, as shown in Figure 4.3. • NARX training has been repeated for different number of neurons (i.e., 2, 4, 6, 8, 10, 12, 14, 16, 18 and 20 minutes) and time delay (i.e., 10, 20, 30, 50 and 60 minutes), separately, as a sensitivity analysis to reveal their impact on the accuracy of the model and to find the best number of neurons and time delay for the rest of the simulation studies. The assessment of each model with different time delay was carried out by arbitrarily selected test data. • A comparison study between the universal and seasonal NARX models has been carried out in the simulation study to determine the best NARX model. MATLAB/Neural Net Time Series toolbox is used in this study for training and testing NARX models. During training process in MATLAB Neural Net Time Series toolbox, battery cell temperature is estimated for 1 minute ahead. In reality, however, we need estimated cell temperature for longer time horizons. For instance, a day-ahead EMS of a microgrid plans the system operation for 24 hours ahead. In this case, it is needed to estimate battery cell temperature for any charge/discharge profile in the next 24 hours. To estimate temperature in multi-hour ahead, the trained model is recursively used, where the estimated cell temperature from previous timestep (1 minute before) is used as an input parameter in the feedback loop. For instance, if 24 hours ahead estimation is required for given charge/discharge and ambient temperature profiles, the estimation will be repeated 1440 times to cover the entire period.

4.4.1 Sensitivity analysis on the number of neurons and time delay

To find the appropriate number of neurons, different NARX model is trained with different number of neurons in the hidden layer, while input and output samples were remained the same. The number of neurons in the hidden layer is increased from 2 to 20 with an interval of two neurons to determine the optimal network model. Each model was run for 20 times for a specific number of neurons to reduce the impact of initialisation and optimisation of the weights and biases on the results. Average estimation error on the test dataset is used to determine the best number of neurons. As per Figure 4.5, the average value of actual error is higher when the number of neurons is at the lower and higher ranges. 10 neurons resulted in only 0.59 ◦C average absolute error, indicating the best performance among different cases. Therefore, NARX model with 10 neurons in the hidden layer is used in rest of the chapter. In order to find the best time delay of the feedback signal for the parallel network shown in Figure 4.3, different NARX models are trained for different time delays (i.e., 10, 20, 30, 40, 50 and 60 minutes). In each case, the number of neurons in the hidden layer was 10 and the performance 4.4 SIMULATION STUDY 46

2.50

2.00

C) 1.50 º

ERROR ( ERROR 1.00

0.50

0.00 2 4 6 8 10 12 14 16 18 20 NUMBER OF NEURONS IN HIDDEN LAYER

Figure 4.5: Average error of different NARX model with different number of neurons in the hidden layer

10 mins 20 mins 30 mins 40 mins 50 mins 60 mins 1.4

1.2

1 C)

º 0.8

0.6 RMSE ( RMSE

0.4

0.2

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

Figure 4.6: Sensitivity analysis results on the cell temperature feedback delay of the NARX model is assessed by using the test dataset. 56 days of data is randomly selected with 24 hours horizon for assessing the network performance. Hourly average RMSE values are calculated from minute-by-minute data of 56 days of test data, which are shown in Figure 4.6. From the ﬁgure, it is clear that 20 minutes feedback delay yields the best performance on the test dataset. 10 and 30 minutes time delay also shows reliable performance. However, creating a model with 20 minutes delay is comparatively more efﬁcient and accurate. For the rest of the chapter, 10 neurons is considered in the hidden layer and cell temperature feedback of 20 minutes will be used. Please note that the sensitivity analyses are done for 1 minute ahead estimation.

4.4.2 The universal NARX model

This section presents the universal NARX model training and its performance estimation with completely different set of test data. Table 4.1 and Table 4.2 show the parameters for network training 4.4 SIMULATION STUDY 47 and selection of data with outcomes during training process, respectively. 449 days (excluding 6 days used for evaluating seasonality in Section 4.4.3) of the data have been used for training, validation and testing. 359 days (80%) or a complete year of data has only been used to train the NARX model. 90 days (20%) of data for validation and testing stages (45 days each) are randomly chosen from the remaining data to validate the model during training stage and evaluate the trained model during testing stage. Since the data selection affects the performance of the NARX model, the training process has been repeated 10 times to minimise the impact of the training process and to choose the best NARX model. It has been observed that although different combinations provide different results, 2 the performance of the model (i.e., RMSE, R and RAd justed) remains almost the same. Validation dataset is used to minimize overﬁtting during training stage by MATLAB toolbox. This data set does not adjust the weights of the network model but weights are checked for accuracy in the model estimation [140]. Testing dataset is completely unseen set of samples that is used by the Neural Net Time Series toolbox at the end of training to provide an unbiased assessment of the trained model. A good performance of a model on the test dataset is the key to a successful modelling.

Table 4.1: Parameters for training the NARX networks Parameter Description

Problem Nonlinear auto-regressive time-series network with external input (NARX) Training Algorithm Levenberg-Marquardt Input charging/discharging Current and Ambient Temperature Output Battery Cell Temperature Data Division Training (80%), Validation (10%) and Testing (10%) Hidden layer size 10 neurons Feedback Delay 20 minutes

Table 4.2: Data selection and training outcomes from MATLAB Neural Net Time Series toolbox Data Type Number of Sample Sample Percentage Regression (R) Adjusted R2 RMSE (◦C)

Training 516960 80% 0.989 0.978 0.452 Validation 64800 10% 0.987 0.974 0.551 Testing 64800 10% 0.99 0.98 0.415

The performance of the universal model in comparison with the actual data is tabulated in Ta- ble 4.2 for training, validation, and testing processes during training in the MATLAB Neural Net Time Series toolbox, where one time step ahead estimation is carried out. The RMSE, coefﬁcient of determination, i.e., R, and Adjusted R2 are calculated using Eq. (4.4), Eq. (4.5) and Eq. (4.6), respectively. There is a close agreement between actual battery cell temperature values with those estimated 4.4 SIMULATION STUDY 48

2 by the network. According to Table 4.2, the RAd justed values for training, validation and testing are very close to 1, which indicates a good model fit. Moreover, it indicates that the model is not overfit- ted by providing a true goodness of fit by yielding similar performance for test dataset. In addition, the RMSE is 0.454 ◦C, 0.55 ◦C and 0.42 ◦C for training, validation and testing, respectively, which is very accurate for many applications such as EMS. As per both assessment criteria of the trained network, it can be stated that the universal model is able to provide a good estimation of the battery cell temperature for many combination of input parameters. The universal NARX model is further validated by testing with completely different set of data. In this analysis, 56 days of data (outside of training, testing and validation dataset) have been considered to estimate battery cell temperature in 24 hours ahead (minute-by-minute samples) using the trained model recursively. Figure 4.7 shows the hourly- averaged RMSE histogram of the 56 days of estimation, where 96.1% of incidents are experiencing RMSE in below 1 ◦C. Only 3.9% of incidents are yielding more than 1 ◦C. Majority of incidents are happening in between 0.17 ◦C and 0.22 ◦C RMSE range. Also, it is clear from the histogram that estimated RMSE never exceeded 1.95 ◦C for the 56 test days. It indicates a relatively robust performance of the NARX model, which can be effective in determining battery cell temperature with a high accuracy in advance.

Figure 4.7: Hourly RMSE histogram for 56 days of test data obtained by the universal model for 24-hour ahead

4.4.3 Seasonal NARX Model

In order to evaluate the seasonal effects, separate NARX models are trained for each season in this subsection, and their performance compared with the universal model. The battery cell temperature estimation model for each season have been assessed using completely different set of data extracted 4.4 SIMULATION STUDY 49 from each season, as given in Table 4.3.

Table 4.3: Seasonal data selection for training and evaluation Season Months Year Days for training Test Days for evaluation

Spring September, October, November 2017 85 6 Days Summer December, January, February 2016-2017 84 6 Days Autumn March, April, May 2017 86 6 Days Winter June, July, August 2017 86 6 Days

Similar NARX conﬁguration with 10 neurons in the hidden layer and feedback time delay of 20 minutes are used in this analysis. Seasonal test days are selected randomly for assessing the trained model. In order to make fair comparison, the same six days were taken out from the training data set in the universal model.

Table 4.4: Summer data selection and training outcomes from MATLAB Neural Net Time Series toolbox Data Type Number of Sample Sample Percentage Regression (R) Adjusted R2 RMSE (◦C)

Training 84672 70% 0.989 0.978 0.479 Validation 18144 15% 0.986 0.972 0.593 Testing 18144 15% 0.992 0.984 0.342

Table 4.5: Autumn data selection and training outcomes from MATLAB Neural Net Time Series toolbox Data Type Number of Sample Sample Percentage Regression (R) Adjusted R2 RMSE (◦C)

Training 86688 70% 0.991 0.982 0.286 Validation 18576 15% 0.993 0.986 0.276 Testing 18576 15% 0.99 0.98 0.31

Table 4.6: Winter data selection and training outcomes from MATLAB Neural Net Time Series toolbox Data Type Number of Sample Sample Percentage Regression (R) Adjusted R2 RMSE (◦C)

Training 86688 70% 0.995 0.99 0.33 Validation 18576 15% 0.992 0.984 0.337 Testing 18576 15% 0.992 0.984 0.336

Table 4.4 to Table 4.7 show the number of samples and outcomes during training, validation and testing stages for each seasonal NARX model from MATLAB Neural Net Time Series toolbox. Adjusted R2 values for each season are close to 1 in all cases, which indicates an outstanding ﬁt to 4.4 SIMULATION STUDY 50

Table 4.7: Spring data selection and training outcomes from MATLAB Neural Net Time Series toolbox Data Type Number of Sample Sample Percentage Regression (R) Adjusted R2 RMSE (◦C)

Training 85680 70% 0.991 0.982 0.343 Validation 18360 15% 0.983 0.966 0.702 Testing 18360 15% 0.991 0.982 0.358 the model by drawing a close relationship between estimated and actual battery cell temperature for all four seasons. The RMSE value has also been considered to evaluate the model. According to the results reported in Table 4.4 to Table 4.7, the RMSE values are below 0.71 ◦C. Most importantly the RMSE values during testing process are below 0.5 ◦C in all seasons, which shows a close agreement between estimated and actual battery cell temperature. Figure 4.8 depicts a comparison between the seasonal and universal models during training and testing processes as well as evaluation with completely different set of test days. The hourly RMSE values for selected test data set during model training for summer, autumn, winter and spring are 0.475 ◦C, 0.382 ◦C, 0.405 ◦C and 0.425 ◦C, respectively, whereas seasonal RMSE values for the similar seasons are 0.342 ◦C, 0.31 ◦C, 0.336 ◦C and 0.358 ◦C, respectively. It indicates an 18% to 39% improvement in the accuracy of the estimated values with seasonal model compared to the universal model. Evaluation of the trained model with test days shows better outcomes by seasonal network model as well. The RMSE values of the universal network model for the test days during summer, autumn, winter and spring are 0.541 ◦C, 0.403 ◦C, 0.412 ◦C and 0.489 ◦C, respectively, where seasonal models yield 0.332 ◦C, 0.202 ◦C, 0.314 ◦C and 0.322 ◦C for summer, autumn, winter and spring, respectively. This comparison demonstrates that seasonal networks yield 23% to 50% less RMSE values than the universal network, indicating better accuracy by the seasonal network model.

Training Testing Test Days Universal Network

0.552

0.541

0.489

0.480

0.475

0.442

0.425

0.422

0.412

0.405

0.403

0.401

0.382

0.358

0.343

0.342

)

0.336

0.332

0.330

0.322

0.314

0.310

ºC

0.286

RMSE ( RMSE 0.202

SUMMER AUTUMN WINTER SPRING

Figure 4.8: Comparison between the seasonal and universal NARX models 4.4 SIMULATION STUDY 51

Figure 4.9 shows the performances of the seasonal NARX models for the test data on an hourly (i.e., RMSE hourly) basis. The battery cell temperatures are estimated recursively for 24 hours ahead, as explained at the beginning of Section 4.4. The average RMSE values are calculated for every hour. The values are shown for summer only in this figure due to similar performance experienced for the rest of the seasonal models. It can be observed from the figure that the RMSE values are lower than 1 ◦C, except couple of hourly incidents on day 3, and the differences between days are minor. The RMSE outcomes during hour 1 to 3 and hour 10 to 16 (where battery operates rarely) of each day are comparatively better than the RMSE values during other hours. However, the differences are insignificant, which pointing out to a robust model performance throughout the whole day.

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 1.8 1.6 1.4

) 1.2 C

° 1 0.8

RMSE ( RMSE 0.6 0.4 0.2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 HOUR

Figure 4.9: RMSE of the test days during summer season using recursive estimation for 24 hours ahead

Summer Autumn Winter Spring Universal network 0.8

0.7

0.6

0.5

C) ° 0.4

RMSE( 0.3

0.2

0.1

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

Figure 4.10: Seasonal comparison based on the hourly-averaged RMSE obtained by using seasonal and universal network models

Figure 4.10 shows the average RMSE comparison between seasonal and universal models. In this case, the average RMSE values of the test days for each season using seasonal NARX model and 24 days (4 Seasons × 6 Days) using the universal NARX model have been considered for each 4.4 SIMULATION STUDY 52 hourly RMSE. It can be seen that summer and spring seasons are yielding maximum of 0.66 ◦C on average using the seasonal network model. On the other hand, autumn, winter and spring have RMSE values of 0.46 ◦C, 0.43 ◦C and 0.42 ◦C, respectively. Although couple of hours during summer show more RMSE than autumn, winter and spring, the daily average RMSE value differences between seasons are not signiﬁcant, which suggests a better performance of the seasonal models in general. In comparison with the universal model, it is clear that the seasonal NARX models are yielding at least 0.1 ◦C to 0.3 ◦C less RMSE during most of the hours of day for the same test days compared to the universal model. In order to investigate further on the performance of the seasonal and universal NARX models, the RMSE ratio has been calculated. Figure 4.11 to Figure 4.14 show comparison between performance of the seasonal and universal NARX models using Eq. (4.7) to determine quantitative comparison between the two models. It helps to determine which model is providing better results with less RMSE value.

RMSE of the Seasonal Model RMSE Ratio = (4.7) RMSE of the Universal Model When the RMSE ratio is lower than 1 (red dashed line in the ﬁgures), it represents a better performance of seasonal network model and above 1 always indicates better performance for the universal model. In this comparison, the same six days of test data has been used for both types of models and hourly estimated values up to 24 hours has been considered to compare the results.

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 4

3.5

2.5

1.5 RMSE RATIO RMSE 1

0.5

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

Figure 4.11: The RMSE ratio for the summer NARX model

As per Figure 4.11, the RMSE ratio is below 1 in most of the hourly incidents during summer test days, which denotes a better performance by the seasonal model. Figure 4.12 shows almost similar outcome during autumn, where the hourly ratios are less than 1 most of the time, except a couple of instances. Similar conclusion can be drawn for winter in Figure 4.13. While the ratios are closer 4.4 SIMULATION STUDY 53

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 3.5

2.5

1.5 RMSE RATIO RMSE 1

0.5

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

Figure 4.12: The RMSE ratio for the autumn NARX model

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 3

2.5

1.5

RMSE RATIO RMSE 1

0.5

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

Figure 4.13: The RMSE ratio for the winter NARX model

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 3

2.5

1.5

RMSE RATIO RMSE 1

0.5

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

Figure 4.14: The RMSE ratio for the spring NARX model to 1 compared to summer, most of the cases are still below 1, indicating a better performance of the seasonal model compared to the universal model over the selected test days. During spring season, shown in Figure 4.14, the RMSE ratio for couple of incidents from different test days are bigger than 1. 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 54

However, more than 80% of the time, ratios are far below 1, which demonstrates a better performance of the seasonal network model than universal model for this season too. In few instances, hourly ratios are more than 2 in every season, which indicates the universal model is performing more than two times better than the seasonal model. However, 74.3%, 77.8%, 78.5%, 80.1% of the hours are showing better results for the seasonal model in summer, autumn, winter and spring, respectively. Overall, it can be concluded that training the neural network with seasonal data provides a better results than the network trained with the universal data. Although there are some variations between seasons, on average, all four seasons are performing equally well and more accurate results can be obtained compared to the universal model.

4.5 Comparison between linear and non-linear model

A comparison study has been introduced between non-linear and linear models for battery cell temperature estimation. The non-linear NARX model is compared with another data-driven estimation model called Autoregressive Integrated Moving Average with eXogenous inputs (ARIMAX), which is a linear model. It is a powerful tool to outline the dynamics of a time series. It is capable of considering effective exogenous parameters for estimating battery cell temperature for longer horizon. EMS can be easily solved using linear models instead of using non-linear models. Although, non-linearity of battery thermal behaviour cannot be modelled by this approach, linear model performance is investigated in this thesis to see if the outcome is comparable with nonlinear model. There is a number of applications where linear model are utilised for prediction or estimation purposes [141–143]. In this section of thesis, the ARIMAX model is developed with similar the same training and test data sets are employed for training and testing the established model.

4.5.1 Battery Cell Temperature Estimation with ARIMAX model

The ARIMAX concept is proposed over ARIMA as a multivariate method that can include independent variables, which are important in the battery cell temperature model. In ARIMA, only dependent variable is regressed on its own lagged values (i.e., AR terms), error values generated in previous time steps by the model (i.e., MA terms) and the number of nonseasonal differences needed for stationarity of time-series data. In addition to ARIMA, ARIMAX model takes the impact of exogenous variables into account. In this study, the charging/discharging current of the battery and ambient temperature are considered as the external variables to estimate dependent variable, i.e., battery cell temperature 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 55

[129]. ARIMAX model can be mathematically represented as,

yˆt = βxt + φ1yt-1 + ... + φpyt-p − θ1zt-1 − ··· − θqzt-q + zt (4.8) wherey ˆt represents dependent variable based on differencing of time series data; yt is the dependent variable denoting actual time series; xt represents exogenous inputs at time t and β is the coefficient of exogenous inputs; zt represents the forecast errors; and φp and θq are the estimated coefficients of the respective variables; and t denotes the time-step of the series. The proposed model is classified as ARIMAX(p, d, q) model. There are three main parameters, namely p, d and q, to be set to determine the model. p is the number of Auto-regressive (AR) terms, d is the number of nonseasonal differences needed for converting the non-stationary time-series to a stationary one, and q denotes the number of lagged forecast errors. The identification of the model is performed in three steps. First step is to analyse the trend of time series to determine whether transformation is needed. In order to check the stationarity of the time series, statistical analysis will be carried out. Two tests are used in this study to check the stationarity of the time series:

• Augmented Dickey Fuller (ADF) unit root testing is chosen to examines the null hypothesis of a unit root [144].

• Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test is utilised to determine the stationarity of a time series around a mean or a linear trend [145].

Both tests collectively indicate the stationarity of a time series when existence of a unit root is rejected by ADF test and the mean/trend-stationarity of the time series is not rejected by KPSS test. Second step is to find the best order of the AR model for estimating the battery cell temperature. An AR model predicts the dependent variable (i.e., battery cell temperature), where specific lagged values of the dependent variables are used as predictor variables. Partial Autocorrelation Function (PACF) is used to identify the order of the autorcorrelation or AR(p) model [146]. The other component of ARIMAX is the Moving Average (MA) with the order of q. In a time series model, a MA term is a past error multiplied by a coefficient. Autocorrelation Function (ACF) is used to determine the order (q) of MA model [146]. These methods will be used in Section 4.5.1.1 for modelling using actual data. Third step is to build the model using training dataset after identifying p, d and q order. The coefficients of the model are identified during the training process. The trained model is then tested with a separate set of independent data (called test data), which has not being used during the training 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 56 process, to evaluate the performance of the model. In order to quantify the accuracy of the battery cell temperature prediction, RMSE, given in Eq. (4.9), is used as a standard measure.

s N 2 ∑ (Ta − Tm ) RMSE = i=1 i i (4.9) N where Tai , Tmi and N represent time series of actual battery cell temperature, predicted battery cell temperature using ARIMAX model, and the total number of samples, respectively.

4.5.1.1 Simulation Study For ARIMAX model and comparison study

The following data selection and analysis are carried out in the simulation studies for training and assessing the ARIMAX model and comparison study: • For training the ARIMAX model, 12 months equivalent data with 1-minute sampling rate from 1st April 2016 to 31st March 2017 have been selected, which consists of 516,960 samples (359 days) overall. Another 90 days equivalent data from 1st April 2017 to 7th July 2017 is used for validation and testing the trained model during training process. • For further assessment of ARIMAX prediction model, 30 days of data from 10th July 2017 to 31st October 2017 have been used. • 56 days of data from 10th July 2017 to 31th October 2017 have been used for comparison study between non-linear model, NARX and linear model, ARIMAX. • Seasonal ARIMAX model is also developed to draw a comparison between ARIMAX and NARX seasonal models. Seasonal data is used for training the ARIMAX model. 6 days of data is extracted from corresponding months of each season for testing purposes. Similar test days are also used for NARX model to show comparison outcomes. For training and testing the ARIMAX prediction model, ’R’ time series and forecasting package is used in this study. During training process, 1 minute ahead battery cell temperature predicted is carried out.

4.5.1.2 ARIMAX Estimation Model Outcomes

Model order has been identified firstly before training the time series data and estimating the battery cell temperature. As mentioned earlier, PACF, ADF-KPSS and ACF tests are considered to find the best order of p, d and q, respectively. Stationarity of the time series signal is assessed by using ADF and KPSS tests. The p-value is 1.55e-302, which is far below 0.05 (for 95% confidence interval). It gives the indication of rejecting null hypothesis not having a unit root in time series. In addition to that, the KPSS test fails to reject 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 57 the null hypothesis. p-value is <0.01 and test critical values are 0.347, 0.462 and 0.744 for 10%, 5% and 1%, respectively. The test critical values are more than the test statistic value, 0.0733. Therefore, it can be concluded that the time-series data is a stationary one. Therefore, the order of I(d) will be 0 or I(0). ACF for AvgTemperature_B1_R1 Figure 4.15shows 1 the PACF values for different lags of data. An approximate test that a given 0.5 +- 1.96/T^0.5 partial correlation 0 is zero (at a 95% significance level) is given by comparing the sample partial -0.5 √ autocorrelations-1 against the critical region with upper and lower limits given by ±1.96/ T, where T 0 10 20 30 40 50 60 is the number of samples of the time-series. There is a cut-off experienced after the lag 26. It suggests lag an AR model with order 26, i.e. AR(26).PACF for AvgTemperature_B1_R1

1 0.5 +- 1.96/T^0.5 0 PACF -0.5 -1 0 10 20 30 40 50 60 Lag

Figure 4.15: Partial autocorrelation function (PACF) plot

Figure 4.16 shows the autocorrelation values of different lags. It shows that there is no cut-off where the peak of the autocorrelation value for lag falls below the critical region. The value remains almost the same even after observation of 60 lags. ACF for AvgTemperature_B1_R1 1 0.5 0 ACF -0.5 +- 1.96/T^0.5 -1 0 10 20 30 40 50 60

Lag

PACF for AvgTemperature_B1_R1 Figure 4.16: Autocorrelation function (ACF) plot 1 0.5 +- 1.96/T^0.5 0 From the above-0.5 analysis with training data, it is determined that the order of model is ARIMAX -1 th (26, 0, 0). This means 0 that the 10 best model 20 should have30 26 order 40 of AR 50 model with 60 the component of 0 order MA and the non-seasonality time-series data.lag In order to verify the model, 30 days equivalent of test data is selected. The good performance of the model with the completely different set of data is the key to a successful modelling. Figure 4.17 shows the hourly outcomes of 10 test days with the ARIMAX model. The days with higher RMSE value are shown in the ﬁgure. It can be observed that except two days (Day 4 and Day 6), all the test days hourly RMSE value is equal or below 3◦C. 65% of hourly RMSE value is below 2◦C, while only 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 58

14.58% of hourly RMSE value is above 2.5◦C. This indicates a good performance by the ARIMAX model in determining battery cell temperature with high accuracy in advance.

DAY 1 DAY 2 DAY 3 DAY 4 DAY 5 DAY 6 DAY 7 DAY 8 DAY 9 DAY 10 4.5 4 3.5

) 3 ºC 2.5 2

RMSE ( RMSE 1.5 1 0.5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 HOUR

Figure 4.17: Hourly RMSE for test days with higher errors

Hourly-averaged RMSE histogram of the 30 days of estimation is shown in Figure 4.18. It shows that majority of the hourly estimation (63.61%) yields equal or below 1◦C of RMSE value. 14.86% of hourly estimation renders RMSE value of above 2◦C, which suggests a strong agreement between actual and ARIMAX estimated battery cell temperature.

Figure 4.18: Histogram of hourly RMSE for 30 test days using ARIMAX battery cell temperature estimation model

4.5.1.3 Comparison Outcomes

In the comparison study, both the seasonal and universal models using ARIMAX and NARX approaches, are trained and tested with the same data. Universal Model Comparison: The comparison between NARX and ARIMAX is shown in Figure 4.19, where, the hourly average values of RMSE from 56 days of similar test days are used. It is noticeable that universal NARX model is always yielding less RMSE value compare to ARIMAX model. At the beginning of the day 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 59 the difference reached to maximum of 1.48 ◦C. However, rest of the hourly difference is much lower and stays in between 0.2 ◦C to 0.7 ◦C. Overall, the NARX model outcomes is signiﬁcantly better than ARIMAX model in estimating battery cell temperature using the universal network.

NARX UNIVERSAL NETWORK ARIMAX UNIVERSAL NETWORK 2.5

2 )

C 1.5 °

1 RMSE RMSE (

0.5

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

Figure 4.19: NARX and ARIMAX Universal model comparison

NARX ARIMAX 1.8 1.6 1.4

1.2 C)

° 1 0.8

RMSE ( RMSE 0.6 0.4 0.2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 HOUR

Figure 4.20: NARX and ARIMAX seasonal model comparison for summer

Seasonal Model Comparison: Figure 4.20 to Figure 4.23 show the comparison between two different models season by season. Similar test days are used in each seasonal comparison and hourly average values are shown in each ﬁgure. It is noticeable from results that seasonal NARX model in every season has less RMSE value than seasonal ARIMAX model. When it comes to the hourly comparison, ARIMAX models are always yielding higher RMSE values than NARX models, except 4th and 5th hours during summer season. Error differences between two models are comparatively less during summer. It shows maximum of 1.44◦C more RMSE by ARIMAX than NARX. On the other hand, the RMSE differences 4.5 COMPARISONBETWEENLINEARANDNON-LINEARMODEL 60

NARX ARIMAX 5 4.5 4 3.5

C) 3 ° 2.5 2 RMSE ( RMSE 1.5 1 0.5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 HOUR

Figure 4.21: NARX and ARIMAX seasonal model comparison for winter between ARIMAX and NARX is slightly higher and on average the value is 2.24 ◦C, where ARIMAX is expressing 0.56 ◦C, 1.65 ◦C and 1.9 ◦C more RMSE values than NARX models during summer, autumn and winter, respectively. It is perceptible that the RMSE value differences are at higher range during the ﬁrst couple of hours in winter and autumn. It shows maximum of 4 ◦C more RMSE values by ARIMAX than NARX model in these seasons.

NARX ARIMAX 4 3.5 3

2.5

C) ° 2

1.5 RMSE ( RMSE 1 0.5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 HOUR

Figure 4.22: NARX and ARIMAX seasonal model comparison for spring

It can be concluded from the results discussed in this section that both universal and seasonal NARX models are indicating signiﬁcantly better outcomes compared to the linear model, ARIMAX. It proves the existence and prominence of nonlinearity in the utility-scale BESS behaviour, which cannot be modelled using a linear model, such as ARIMAX. 4.6 SUMMARY 61

NARX ARIMAX 5 4.5 4 3.5

C) 3 ° 2.5

2 RMSE ( RMSE 1.5 1 0.5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 HOUR

Figure 4.23: NARX and ARIMAX seasonal model comparison for autumn

4.6 Summary

In this chapter, a model based on ANN is proposed to estimate battery cell temperature. The proposed method does not require to consider any complicated battery dynamics and thermo-dynamical model of the battery container. Only two dominant input variables for battery cell temperature estimation, namely charging/discharging current and ambient temperature, are shown to be sufficient to estimate cell temperature. The proposed model is able to estimate battery cell temperature accurately using a relatively large historical data set. Time-series NARX model has been chosen due to non-linear behaviour and time dependencies of the battery cell temperature. Moreover, seasonal impact on the modelling is investigated by creating different NARX models for each season. Based on the simulation outcomes, the universal model yielded a small RMSE value, where the RMSE was below 1 ◦C most of the times, which is quite acceptable. Compared to the universal model, the seasonal NARX models provided more accurate results. In addition, proposed recursive estimation with NARX network shows good estimation results for an entire day. The comparison study indicates that non- linear model has better agreement in estimating battery cell temperature than ARIMAX model, which is a linear model. However, as a linear approach, it is providing effective way to find the battery cell temperature in advance, which can be beneficial for EMS with linear optimisation. The next chapter will analyse the cost and benefit of using BESS in a PV plant and propose a comprehensive cost function for the entire PV-BESS plant. Moreover, a cost comparison study between PV plant with and without BESS will be shown. 5 Development of Cost Function

5.1 Introduction

1 In order to operate a PV-BESS plant in an optimal way, the cost and benefit associated with grid-tied PV and BESS should be accounted within an EMS. In other words, a comprehensive cost assessment considering influential factors is able to render an opportunity to develop an optimal EMS for the entire PV plant. While utilization of storage device at a maximum possible rate seems reasonal, the economic benefits of the operation regime should exceed the battery degradation caused by that operation regime. While a number of research studies described energy management and optimization frameworks of various types of energy storage system in [23], to the best of our knowledge, a comprehensive cost function for a utility-scale BESS, including influential factors such as battery capacity degradation, cooling system operation and other operational costs related to BESS operations, has not been reported in the literature.

1This chapter has materials from the following reference published by the MPhil candidate. 3. Md Mehedi Hasan, S. Ali Pourmousavi and Tapan K. Saha, Battery Cell Temperature Estimation Model and Cost Analysis of a Grid-Connected PV-BESS Plant. (IEEE ISGT Asia 2019)

62 5.2 COST FUNCTION DEVELOPMENT 63

A comprehensive cost function, including battery cycling and calendar degradation costs, battery degradation cost because of cycling and calender life, cooling cost, energy cost associated with charging/discharging of BESS and PV generated energy, is developed for a grid-tied PV-BESS plant. Using the cost function, a comparison study between PV plant with and without BESS has been carried out in this study to identify the effectiveness of BESS integration in the plant. Through such study, the dominant factors can be created and cost-effective energy management system can be introduced considering balanced BESS operation for more beneﬁt out of this expensive investment. The primary structure of the chapter is as follows. First, the different cost factors are discussed for a PV plant with and without BESS in Section 5.2. Then, each factor is explained and relevant mathematical formulations are derived in the same section. Finally, simulation study is carried out with ﬁeld data and results are analysed in Section 5.3. The chapter is concluded with a summary in Section 5.4.

5.2 Cost Function Development

In order to find the total cost of PV plant, dominant cost-benefit factors with and without BESS in the plant are identified and separated cost functions are developed for each case. The dominant factors for both scenarios, such as battery degradation cost due to cycling and calendar life, cost of cooling system, and the cost associated with charging/discharging operation of BESS have taken into account to develop a complete cost function for a PV-BESS plant. Each cost factor taken in consideration in this study is explained in detail and formulation of each cost factor is delivered step by step.

5.2.1 PV plant cost with and without BESS

In this section, the plant operation costs with and without battery will be formulated step-by-step. The cost functions are developed based on UQ GSRF [116]. The primary objectives of the plant are to fulﬁl campus electricity requirements and reduce demand charge. More details about the plant and operation rules are given in [116] and in chapter 3. A cost comparison of the PV plant with and without energy storage system is established focusing on the operational rules of BESS in the plant, as explained in Chapter 3.1.

5.2.1.1 Total cost function with BESS

Although BESS is primarily integrated to the PV plant to store excess PV energy and utilise it effectively, its continuous operation in the plant incurs cost. The dominant cost factors for operating 5.2 COST FUNCTION DEVELOPMENT 64 a BESS in the PV plant are cost associated with battery degradation, cooling system operation, and importing power from the grid to charge the BESS.

DailyDeg

Figure 5.1: Cost-beneﬁt of the entire PV-BESS plant

Figure 5.1 shows the daily cost-benefit associated with BESS operation in the plant. Continuous battery operation with charging and discharging mode and idle situation are responsible for battery degradation. Although, a significant benefit is experienced by discharging operation of the BESS during peak hours, extra cost incurs through battery charging by the grid and avoiding selling excess PV generation to the grid at FiT price (yellow area in Figure 5.1). As per the plant’s operational rules discussed in Subsection 3.2.2, the BESS charging operation only takes place during off-peak time, when the price is relatively lower. In addition, BESS will be charged by the excess PV power in the middle of the day, if available. Therefore, this is considered as a cost in a PV plant operation with

BESS. The total PV plant cost with BESS, i.e., CostWB, on a daily basis is calculated using Eq. (5.1), where the cost associated with aforementioned dominant factors are considered.

CostWB = CostDailyDeg +CostDTC +CostDI +CostCFiT (5.1) where CostDailyDeg refers to the total daily cost associated with degradation; CostDTC and CostDI represent cost of cooling mechanism operation and imported energy from the grid to charge the BESS, respectively. CostCFiT is the cost of charging the BESS with excess PV energy, which could be exported to the grid at FiT.

5.2.1.2 Total Cost of PV Plant Without BESS

PV plant without BESS is only able to export energy to the grid at FiT (shown in Figure 5.2). Due to unavailability of PV generation or energy storage system during evening peak, the plant is unable to reduce the peak demand and consequently the demand charge cost will be higher. It is the cost of a 5.2 COST FUNCTION DEVELOPMENT 65

Figure 5.2: Cost-beneﬁt of the PV plant without BESS single monthly peak demand that should be paid alongside energy cost. It is calculated as the average of measured instantaneous power in pre-deﬁned intervals. Therefore, extra cost occurs for the PV plant without BESS. The total daily cost for the PV plant without BESS, i.e., CostWOB, is estimated using Eq. (5.2).

CostWOB = CostDE +CostDC − FiTBeneﬁt (5.2) where, CostDE and CostDC are the cost of energy and demand charge during peak hours, respectively.

5.2.2 Battery Degradation Cost

Degradation of a battery is a continuous process. Couple of factors affect degradation at the same time [147]. A simpliﬁed linear cyclic and calendar ageing models are considered in this study to quantify the degradation of the BESS. Battery undergoes charge and discharge operations as well as idle situation throughout its lifetime. Therefore, both processes continuously drive battery ageing. In this section, the two degradation processes, namely cycling and calendar ageings, are formulated considering operational conditions of the BESS. Finally, a cost function is developed to quantify cost of degradation on a daily basis. The hourly degradation cost is calculated using Eq. (5.3), where hourly operational condition of BESS is considered to quantify the degradation of battery.

CostDEG = CostBESS × (DEGCyclic + DEGCalendar) (5.3) where CostDEG presents the total degradation cost for one day or 24 hours; DEGCyclic and DEGCalendar represent the daily degradation of BESS because of cyclic and calendar ageing mechanisms, respectively. The total cost of BESS is calculated considering energy and power costs, shown in Eq. (5.4).

CostBESS = CkWh × Ecap(kWh) +CkW × Pcap(kW) (5.4) where the BESS rated power and energy capacity are represented by Ecap(kW) and Pcap(kW), respectively, and prices per kWh and kW are denoted by CkWh and CkW. 5.2 COST FUNCTION DEVELOPMENT 66

5.2.2.1 Cyclic Ageing

Degradation of battery for cyclic ageing mechanism is quantiﬁed with respect to the energy throughput for a certain time interval, and the entire energy throughput of BESS throughout its lifetime, given in Eq. (5.5). ETh DEGCyclic = (5.5) BLT where ETh and BLT are representing energy throughput for a deﬁned time interval and total BESS energy throughput until reaching to its End of Life (EoL), respectively. Charging and discharging activities of BESS are observed using Eq. (5.6) for each time-interval to quantify the energy throughput.

t t ETh = ηCH ∑ Pj,CH × hour + ηDCH ∑ | Pj,DCH | ×hour (5.6) j=1 j=1 where ETh is the total energy throughput; ηCH and ηDCH represent the efﬁciency of charging and discharging operation, respectively; and Pj,CH and Pj,DCH are the real power in kW for charging and discharging operations, respectively. Total energy throughput of the BESS is calculated using Eq. (5.7), which considers EoL, total rated cycle number at the rated DoD, and the rated energy capacity of BESS.

NR ηCH EoL − 1 BLT = 2 × DoDR × × Ecap × ∑ (1 + × j) (5.7) ηDCH j=0 NR where BLT is the total kWh throughput of the battery over its lifetime; DoDR represents the rated

DoD for BESS; Ecap is the actual energy capacity of the BESS at the beginning of its operation in the

ﬁeld; and NR is the total rated battery cycle number; EoL refers the end of battery life in p.u (0.65 N in utility-scale energy storage application). ∑ R (1 + EoL−1 × j) in the above equation, models the j=0 NR linear degradation of battery capacity until it reaches to the EoL [148].

5.2.2.2 Calendar Ageing

Calendar ageing is estimated by counting number of idle hours experienced by the BESS and the total calendar life of the battery, as given in Eq. (5.8).

TIdle DEGCalendar = (5.8) TLife where TIdle and TLife are the time in which battery is idle within a speciﬁc period of time and the entire calendar life of the BESS, respectively. Hourly idle time during a day and a total battery calendar life is calculated using Eq. (5.9) and Eq. (5.10), respectively. t TIdle = ∑ Sj × hour (5.9) j=1 5.2 COST FUNCTION DEVELOPMENT 67 where Sj is the On/OFF status of the battery. TIdle represents the aggregated idle time of the battery until time t. The status of the battery is counted and accumulated for a certain period of time as per requirement. The battery is considered OFF when Pj,CH = 0 and Pj,DCH = 0 at a given time. Otherwise, battery is under operation, i.e., ON.

TLife = 365 × 24 × TRATED (5.10) where TRATED is the rated calendar ageing of the battery in year. Hourly quantiﬁed degradation achieved from Eq. (5.5) and Eq. (5.8) is calculated on a daily basis by observing degradation for 24 24 hours. The total degradation cost on a daily basis can be calculated as: CostDailyDeg(t) = ∑ j=1 CostDEG, j.

5.2.3 Cooling Cost

Apart from passive cooling mechanism to ensure even temperature distribution, an active cooling consists of rack fans and air-conditioning unit with 7.7 kW cooling capacity is dedicated to keep the conﬁned storage room temperature at an acceptable range as explained in Subsection 3.2.3. The rack fan and air-conditioning unit start operating when the battery cell temperature is equal or above 29 ◦C and the room temperature reaches to 23 ◦C and above, respectively. 85% of power is consumed by compressor unit compared to the condenser-evaporator fans. Table 5.1 shows the details of each cooling system units.

Table 5.1: Existing cooling system speciﬁcations Parameters Description

Air-conditioning Unit Cooling Capacity 7.7 kW

Energy Efﬁciency Ratio(EER) 2.8

Air-conditioning Unit Power Consumption 2.3 kW(Compressor), 0.43 kW(Fans)

Rack Fans Power Consumption 0.36 kW

Operation of a dedicated cooling system unit is controlled by certain rules given in Eq. (5.11), which shows the type of operation taking place for different thermal conditions. Minute-by-minute temperature values are considered to calculate total energy consumption and associated cost by the cooling components.  C , T < 23◦C  F R   ◦ CCOM +CF , TR ≥ 23 C CCool(T) = (5.11) C +C +C , (T ≥ 29◦C) ∧ (T ≥ 23◦C)  COM F RF C R   ◦ ◦ CF +CRF , (TC ≥ 29 C) ∧ (TR < 23 C) 5.2 COST FUNCTION DEVELOPMENT 68 where, CCool represents the total cooling cost at temperature T; TR and TC refers to battery room temperature and average battery cell temperature, respectively; CCOM, CF and CRF are the cost function of each cooling component, namely compressor, fan of air-conditioning unit, and rack fans on top of each battery bank. In order to calculate total daily energy cost, ﬁrst hourly energy cost is calculated for each cost t=24 component. Then, daily cooling cost is calculated using CostDTC = ∑n=1 CCool(n), where CostDTC refers to the total daily energy cost associated with cooling system operation in the BESS container. Due to different energy tariff during a day, shown in Eq. (5.12), hourly cost is calculated using hourly energy consumption, (Energy(kWh) = PRATED(kW)×hour) by different cooling system units.

PRATED values for each cooling component are shown in Table. 5.1. Energy consumption cost is

CostEnergy($) = Energy(kWh) × τ(t)($).  6.545 /kWh, t ∈ [11PM,12AM),(12AM,7AM] τ(t) = (5.12) ¢ 11.78 /kWh, t ∈ [7AM,11PM] ¢ where, τ(t) represents the local Time-of-Use (TOU) tariff over a day.

5.2.4 Cost of imported energy from the grid and demand charge

Imported energy for BESS charging operation takes place during night when the grid energy cost is lower. The imported energy is calculated using simple equation, EC(kWh) = PC(kW) × Hour(h), where PC represents the power measured at the point of connection to the grid to charge the BESS. Eq. (5.13) is utilised to calculate the daily cost associated with the grid imported energy for charging the BESS. 5 CostDI = ∑ EC(t) × τ(t) (5.13) n=1 where CostDI is the total daily cost of imported energy from the grid. Minute-by-minute imported st th grid energy is represented by EC. The charging operation of BESS takes place in between 1 and 5 hours of each day in this PV plant. BESS is playing a vital role by reducing peak load by its discharging operation. In the absence of BESS, the PV plant will need to buy the energy from the grid at peak-period price. The cost associated with the excess energy is calculated using Eq. (5.14).

23 CostDE = ∑ ED(t) × τ(t) (5.14) n=18 where, CostDE represents the total daily cost due to discharging operation of BESS to the campus substation during peak demand. The discharging operation of BESS for peak shaving takes place in 5.3 SIMULATION STUDY 69 between 18th and 23th hours of each day in this PV plant, when the grid energy cost is more and peak demand more prone to happen. Hourly energy by discharging operation is represented by ED. Similar ToU tariff mentioned in (5.12) is used to calculate the cost.

5.2.5 Demand Charge

The demand charge is calculated based on the maximum apparent power (kVA). kVA is calculated for each 30 minutes period using, kVA = p(kW)2 + (kVAr)2. The maximum kVA demand is estimated for each billing period (30 days). Therefore, recorded peak value is used for the entire month’s demand charge calculation shown in Eq. (5.15).

CostMDC = kVAMaximum × (σDUoS + σTUoS) (5.15) where, CostMDC represents the demand charge for monthly billing period. σDUoS and σTUoS de- note distribution use of system (DUoS) and transmission use of system (TUoS) charges (σDUoS =

$12.49/kVA/month and σTUoS = $1.169/kVA/month) per month, respectively.

As demand charge is calculated based on 30-day period, daily demand charge cost, i.e., CostDC, is calculated by multiplying cost of demand charge per kVA per month by the peak monthly power demand.

5.2.6 FiT Beneﬁt

The beneﬁt obtained by the FiT scheme is calculated based on the PV generated energy that is exported to the grid. The cost of charging the BESS with PV energy, CostCFit and exported PV energy,

FiTBeneﬁt are calculated using the Eqs. (5.16) and (5.17), respectively:

CostCFiT(t) = PCPV(t) × hour × ξFiT , ξFiT = 2.356 /kWh (5.16) ¢

FiTBenefit(t) = PEPV(t) × hour × ξFiT (5.17) where FiTBenefit and CostCFiT is the total benefit by selling PV generated energy to the grid and the cost of charging battery instead of exporting to the grid, respectively. PCPV and PEPV are the PV generated power for charging and exporting to the grid, respectively; and ξFiT refers to price of FiT.

5.3 Simulation Study

In this study, minute-by-minute ﬁeld data consisting of a large number of charging/discharging and ambient temperature and its corresponding battery cell temperature incident, is collected from UQ 5.3 SIMULATION STUDY 70

GSRF. It covers a signiﬁcant number of plausible combinations of charging/discharging and ambient temperature operations. The following data is selected and analyses are carried out in the simulation studies for evaluating the cost of the system with/without battery:

• 13 months data, sampled every 1 minute, from 1st April 2016 to 30th September 2017 have been selected for this purpose. Also, 13 days of data different seasons are selected to assess seasonality impact on the PV plant.

• 1 day with the highest peak demand from each billing period is selected for demand charge calculation for the entire billing period.

5.3.1 Cost Analysis

The developed cost functions are used to outline the cost and beneﬁt of using BESS in a grid- connected PV plant. Total operational cost of the plant and comparison study between PV plant with and without BESS for 13 selected days from equivalent number of months are shown in Figure 5.3 to Figure 5.6.

160 100 84.08 140 85.68 80.84 81.24 82.79 80 120 77.93 77.32 76.52 73.73 100 62.52 63.82 62.82 60 54.04

80 % 40

COST ($) COST 60 40 20 20 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 DAY BESS Grid-Charging Cost Cooling Cost BESS Degradation Cost Degradation Cost Percentage (%)

Figure 5.3: Daily cost associated with BESS operation in the plant

Figure 5.3 presents the daily cost of each dominant factor for operating the BESS. It is clear from the figure that the cooling cost has less impact on daily total cost compared to the degradation and grid-charging cost. Cooling cost is only 5.6% of of the total cost on average, whereas the grid charging cost and degradation cost are responsible for 20.3% and 74.1% of the total daily cost on average, respectively. There are five days shown in the figure when the degradation cost yields more than 80% of the total daily cost associated with the BESS in the plant. FiT benefit by the PV plant is shown in Figure 5.4, where a comparison is made between the plant with and without BESS. The PV plant is able to export more energy to the grid in the absence of 5.3 SIMULATION STUDY 71

160 12 140 10.12 9.90 10 120 8 100 6.58 80 6 60 4 FITBENEFIT ($) 3.40 3.52 3.40 40 2.93 2.96 1.92 2

20 1.28 1.30 1.14 FIT BENEFIT WITHOUT BESS ($) FIT WITHOUT BENEFIT BESS 0 0.03 0 1 2 3 4 5 6 7 8 9 10 11 12 13 DAY FiT With BESS FiT Without BESS FiT Benefit Without BESS

Figure 5.4: Daily FiT benefit with and without BESS in the plant the BESS since storage system store PV energy to charge itself for future use. However, the BESS store this cheap energy and fulfil the load demand partially during peak hours when energy price is higher and monthly demand charge prone to be experienced during this time. Only three days from the figure yield more than $5 benefit in the absence of the BESS in the plant. 76.9% of the time, the benefit without BESS is less than $4. It suggests that the opportunity loss of the PV plant due to battery charging instead of exporting at FiT rate, is marginal most of the time. Figure 5.5 shows the monthly demand charges in 13 months with and without BESS. 80% of the time, the demand charge for each billing period without BESS is around $30,000, where demand charge is between $50,000 and $60,000 for the rest of the time. The number signifies the importance of demand charge management in each billing period. It is noticeable that the BESS is providing utmost efforts in reducing the charge during each billing period. The time the demand charge is reduced by more than 10% with the help of BESS in 46.1% of the time. Except the 0.48% cost reduction in 4th month, demand charge is reduced by more than 5% (>$1,600) in more than 90% of the time with the help of BESS in the plant.

70000 14 12.78 60000 11.97 11.88 13.15 12 11.11 50000 10.37 10 40000 8 7.28 6.92 % 30000 6.04 6.40 6 5.31 5.64 20000 4

10000 2 DEMAND CHARGE ($) CHARGE DEMAND 0 0.48 0 1 2 3 4 5 6 7 8 9 10 11 12 13 MONTH Monthly Demand Charge Without BESS Monthly Demand Charge With BESS Demand Charge benefit(%)

Figure 5.5: Monthly demand charge with and without BESS integration in the plant

Total daily plant cost are compared in Figure 5.6 with and without BESS for 13 different test 5.4 SUMMARY 72

6000 3.50 2.86 3.00 5000 2.35 2.50 4000 1.83 1.98 2.00 1.47 1.50

3000 1.03 1.00 % 0.62 0.76 0.34 0.50

2000 0.11 0.21 0.13 0.00 DAILY COST ($) COST DAILY 1000 -0.50 -0.89 -1.00 0 -1.50 1 2 3 4 5 6 7 8 9 10 11 12 13 DAY Total Cost With BESS Total Cost Without BESS Benefit /Loss (%)

Figure 5.6: Daily total cost for the plant with and without BESS days in 13 months. It can be seen that only three days are yielding more than $90 benefit with BESS and the rest of the days are showing profit of less than $50. On day four, there is about 1% loss that encountered in the plant with BESS in comparison to the PV plant operation cost without BESS. However, more than 90% of times, there is higher profit in running BESS within the PV plant. As per above analyses, it is clear that BESS is playing a vital role by reducing the demand charge from the monthly bill. However, when it comes to the total cost between PV plant with and without BESS (after subsidizing the cost of BESS from benefit), the profit with BESS is not significant most of the time. The most influential cost factor, namely battery degradation cost, is responsible for reduction in the BESS benefit. Therefore, it is important to establish an optimal operation algorithm for BESS to reduce the battery degradation. In addition, operating BESS during peak demand with higher magnitude will increase the benefit.

5.4 Summary

In this chapter, a cost-benefit analysis is carried out considering battery operational cost in the PV plant. The analysis helps to outline the operational cost associated with each factor in the overall PV- BESS operation cost. In addition, analysing the benefit from BESS renders the opportunity to develop an optimal operation algorithm of BESS in a PV plant to achieve the maximum benefits of the device. Cost functions are developed to carry out a comparison study between PV plant with and without BESS. The developed cost functions are analysed with a real PV-BESS plant data set, where a large number of BESS operational incidents were considered. The cost-benefit analysis shows an untimely battery capacity loss, which reduces the benefit from BESS operation in the plant. It also proves that the BESS integration into the PV plant increases the overall net benefits (up to 2.86%), which can be improved further by alleviating temperature-dependent costs and operating BESS in a way so that more benefit can be achieved without degrading battery excessively, which is responsible for 80% of 5.4 SUMMARY 73 the total cost. This established cost-benefit analysis can be utilised for setting up an optimal battery operation algorithm in the EMS to obtain more benefit out of BESS in the PV plant and to prolong the battery life. 6 Conclusion and Future Work

6.1 Conclusion

This research work investigated the thermal behaviour of battery with respect to its operation in a PV plant and developed a new data-driven approach to estimate battery thermal behaviour for a given charge/discharge proﬁle, in advance, which is necessary for an EMS to operate the BESS optimally in a cost-effective way. The study considered real-world scenarios to develop the models so that the solutions can be used in different applications. The contribution of this thesis has been described brieﬂy in the following points:

• A comprehensive thermal behaviour investigation of a grid-tied BESS within a PV plant was conducted. Excessive high temperature is an important factor for battery power and capacity degradation. Every charge/discharge activity along with high ambient temperature escalates cell temperature, which results in higher degradation. Therefore, considering the impact of charge/discharge activities on battery temperature and consequently degradation rate is an indispensable step in establishing an optimal operation strategy for batteries. One year of operational data from a utility-scale PV plant with BESS is used for this investigation. The results

74 6.1 CONCLUSION 75

of investigation provided detailed insights of battery cell temperature behaviour. The insights from the analyses are utilised in establishing a battery thermal model.

• Primary objective of this research work was to develop a thermal model for battery cell temperature for 24-hour horizon. It is well-known that the battery cell temperature is a key parameter in battery life degradation, safety and dynamic performance, thus accurate battery cell temperature estimation can play a significant role in ensuring BESS optimal operation considering its degradation. In order to estimate battery cell temperature as accurate as possible, a non-linear data-driven model based on NARX neural network was proposed considering strongly correlated independent variables (i.e., charging-discharging current and ambient temperature). Due to different temperature and weather characteristics in each season, seasonal NARX model was also developed and compared with the universal model. The proposed model was evaluated using field data collected from the UQ GSRF plant. The simulation results show a high accuracy of the model compared to the measured data for both seasonal and universal models without considering the complexity of the large-scale battery and container thermal dynamics interactions. The simulation results show that the estimated temperature yield RMSE below 1◦C in different conditions more than 95% of the time, which confirms the validity and accuracy of the model. Moreover, seasonal models show a better performance with 18% to 50% less RMSE on average (for 1 hour to 24 hours ahead estimation) compared to the universal model.

• The next contribution of this research work was to conduct a comparative study between linear and non-linear approaches for modelling thermal behaviour of battery. The developed data- driven linear ARIMAX model was trained and tested with exactly the same data used for training and testing of the non-linear NARX model for battery cell temperature estimation. As per the simulation results, the NARX model outperformed ARIMAX model in every comparison criterion. However, as a linear model approach, overall performance by ARIMAX was acceptable for some applications, such as EMS with linear optimisation.

• The final segment of this thesis rendered a cost-benefit analysis of BESS in the PV plant. Dom- inant cost factors were considered to outline daily operational cost and benefit of using BESS in the plant. Individual cost functions were developed for each cost and benefit factors. The proposed cost functions were analysed with real field data from UQ GSRF, where a significant number of BESS operational incidents were available. In addition, comparison study proved a marginal benefit of integrating battery storage into a PV plant and determined influential factors to consider for optimal battery operation systems. The analysis could be used in establishing 6.2 FUTURE WORK 76

an optimal BESS operational algorithm in EMS to achieve higher beneﬁts.

In summary, the research conducted in this thesis has addressed the thermal issues of utility-scale BESS, which is highly responsible for battery degradation. Thermal behaviour model and cost-benefit analysis for BESS in the PV plant together are offering a solution, which can be implemented in future research for alleviating degradation of battery and obtaining more benefit out of this expensive device. As the renewable energy plants (i.e., solar PV and wind power plants) are expected to be integrated with more large-scale BESS in the future, this work will be very beneficial and can be implemented in any PV-BESS and wind generation plant. As both developed linear and non-linear models are providing good estimation outcomes, any model can be integrated with respect to EMS’s compatibility. Moreover, extensive cost analysis and battery thermal model can play a vital role in developing a control strategy for optimal operation of BESS. The thesis will be beneficial for both academia and practicing engineers in the industry to evaluate the BESS integration in the PV or wind generation plant.

6.2 Future Work

Future research will focus on the following points:

• Battery thermal behaviour highly depends on the type of battery. For other storage technologies, perhaps temperature behaviour with respect to its operation is different from the lithium polymer battery which was used in this study. Therefore, thermal behaviour investigation of battery should be done for different types of battery technologies installed in a renewable energy plant.

• The cost analysis should be extended by considering more cost factors such as the cost associated with installation and maintenance of energy storage system. In addition, the price of battery is continuously dropping to more affordable price ranges, therefore the operational strategy of BESS in a PV plant might be changed to store more energy from PV generation and utilise it during peak period.

• As temperature is one of the dominant factors for accelerating battery degradation, it is indispensable to develop a degradation model based on the battery dynamic temperature. The proposed battery cell temperature model in this study can be used in developing temperature- dependent battery degradation model.

• A BESS operational algorithm should be developed for optimal battery operation considering the operational cost-beneﬁt of BESS in a PV plant. In the process of establishing optimal battery 6.2 FUTURE WORK 77

operation, proposed thermal model should be utilised to predict the battery cell temperature for 24-hour ahead to quantify the highly responsible cost factors namely degradation. In this case, Bureau of Meteorology (BOM) prediction data can be considered for understanding the weather condition and operational pattern of BESS in a plant with respect to weather. References

[1] D. of Environment and Energy. Australian energy update 2016. Tech. rep., Australian Gov- ernement (2016).

[2] D. of Environment and Energy. Australian energy update 2017. Tech. rep., Australian Gov- ernement (2017).

[3] D. of Environment and Energy. Australian energy update 2018. Tech. rep., Australian Gov- ernement (2018).

[4] E. C. Foundation. Roadmap 2050; a practical guide to a prosperous, low-carbon europe (2010).

[5] D. Heide, L. Von Bremen, M. Greiner, C. Hoffmann, M. Speckmann, and S. Boﬁnger. Seasonal optimal mix of wind and solar power in a future, highly renewable europe. Renewable Energy 35(11), 2483 (2010).

[6] M. G. Rasmussen, G. B. Andresen, and M. Greiner. Storage and balancing synergies in a fully or highly renewable pan-european power system. Energy Policy 51, 642 (2012).

[7] E. K. Hart, E. D. Stoutenburg, and M. Z. Jacobson. The potential of intermittent renewables to meet electric power demand: current methods and emerging analytical techniques. Proceed- ings of the IEEE 100(2), 322 (2012).

[8] K. Divya and J. Østergaard. Battery energy storage technology for power systemsan overview. Electric power systems research 79(4), 511 (2009).

[9] A. B. T. B. N. H. Mike Roberts, Jessie Copper and R. Hu. Solar trends report for solar citizens. Tech. rep., UNSW and Australian PV Institute (2018).

78 REFERENCES 79

[10] L. Liao, P. Zuo, Y. Ma, X. Chen, Y. An, Y. Gao, and G. Yin. Effects of temperature on charge/discharge behaviors of lifepo4 cathode for li-ion batteries. Electrochimica Acta 60, 269 (2012).

[11] M. Dubarry, C. Truchot, B. Y. Liaw, K. Gering, S. Sazhin, D. Jamison, and C. Michelbacher. Evaluation of commercial lithium-ion cells based on composite positive electrode for plug-in hybrid electric vehicle applications. part ii. degradation mechanism under 2 c cycle aging. Journal of Power Sources 196(23), 10336 (2011).

[12] F. Leng, C. M. Tan, and M. Pecht. Effect of temperature on the aging rate of li ion battery operating above room temperature. Scientiﬁc reports 5, 12967 (2015).

[13] T. Yuksel, S. Litster, V. Viswanathan, and J. J. Michalek. Plug-in hybrid electric vehicle lifepo4 battery life implications of thermal management, driving conditions, and regional climate. Journal of Power Sources 338, 49 (2017).

[14]J.L opez´ Sanz, C. Ocampo-Martinez, J. A. Alvarez´ Florez,´ J. M. Moreno Eguilaz, and R. Ruiz Mansilla. Thermal management in plug-in hybrid electric vehicles: A real-time non- linear model predictive control implementation. IEEE transactions on vehicular technology 66(9), 7751 (2017).

[15] J. Jaguemont, L. Boulon, Y. Dube,´ and F. Martel. Thermal management of a hybrid electric vehicle in cold weather. IEEE Transactions on Energy Conversion 31(3), 1110 (2016).

[16] J. Wang, P. Liu, J. Hicks-Garner, E. Sherman, S. Soukiazian, M. Verbrugge, H. Tataria, J. Musser, and P. Finamore. Cycle-life model for graphite-lifepo4 cells. Journal of Power Sources 196(8), 3942 (2011).

[17] M. Broussely, S. Herreyre, P. Biensan, P. Kasztejna, K. Nechev, and R. Staniewicz. Aging mechanism in li ion cells and calendar life predictions. Journal of Power Sources 97, 13 (2001).

[18] E. Thomas, I. Bloom, J. Christophersen, and V. Battaglia. Statistical methodology for predicting the life of lithium-ion cells via accelerated degradation testing. Journal of Power Sources 184(1), 312 (2008).

[19] E. Karden, S. Ploumen, B. Fricke, T. Miller, and K. Snyder. Energy storage devices for future hybrid electric vehicles. Journal of Power Sources 168(1), 2 (2007). REFERENCES 80

[20] T. Liu, L. Cheng, Z. PAn, and Q. Sun. Cycle life prediction of lithium-ion cells under complex temperature proﬁles. Eksploatacja i Niezawodnos´c´ 18(1) (2016).

[21] O. Gross and S. Clark. Optimizing electric vehicle battery life through battery thermal management. SAE International Journal of Engines 4(1), 1928 (2011).

[22] J. Groot, M. Swierczynski, A. I. Stan, and S. K. Kær. On the complex ageing characteristics of high-power lifepo4/graphite battery cells cycled with high charge and discharge currents. Journal of Power Sources 286, 475 (2015).

[23] Q. Wang, B. Jiang, B. Li, and Y. Yan. A critical review of thermal management models and solutions of lithium-ion batteries for the development of pure electric vehicles. Renewable and Sustainable Energy Reviews 64, 106 (2016).

[24] A. Jossen. Fundamentals of battery dynamics. Journal of power sources 154(2), 530 (2006).

[25] M. Ecker, J. B. Gerschler, J. Vogel, S. Kabitz,¨ F. Hust, P. Dechent, and D. U. Sauer. Devel- opment of a lifetime prediction model for lithium-ion batteries based on extended accelerated aging test data. Journal of Power Sources 215, 248 (2012).

[26] I. Bloom, B. Cole, J. Sohn, S. Jones, E. Polzin, V. Battaglia, G. Henriksen, C. Motloch, R. Richardson, T. Unkelhaeuser, et al. An accelerated calendar and cycle life study of li-ion cells. Journal of Power Sources 101(2), 238 (2001).

[27] I. Bloom, S. A. Jones, E. G. Polzin, V. S. Battaglia, G. L. Henriksen, C. G. Motloch, R. B. Wright, R. G. Jungst, H. L. Case, and D. H. Doughty. Mechanisms of impedance rise in high- power, lithium-ion cells. Journal of power sources 111(1), 152 (2002).

[28] A. Eddahech, O. Briat, E. Woirgard, and J.-M. Vinassa. Remaining useful life prediction of lithium batteries in calendar ageing for automotive applications. Microelectronics Reliability 52(9-10), 2438 (2012).

[29] X. Feng, M. Fang, X. He, M. Ouyang, L. Lu, H. Wang, and M. Zhang. Thermal runaway features of large format prismatic lithium ion battery using extended volume accelerating rate calorimetry. Journal of power sources 255, 294 (2014).

[30] N. Williard, W. He, C. Hendricks, and M. Pecht. Lessons learned from the 787 dreamliner issue on lithium-ion battery reliability. Energies 6(9), 4682 (2013). REFERENCES 81

[31] G. Pistoia, A. Antonini, R. Rosati, and D. Zane. Storage characteristics of cathodes for li-ion batteries. Electrochimica Acta 41(17), 2683 (1996).

[32] M. B. Pinson and M. Z. Bazant. Theory of sei formation in rechargeable batteries: capacity fade, accelerated aging and lifetime prediction. Journal of the Electrochemical Society 160(2), A243 (2013).

[33] P. Ramadass, B. Haran, P. M. Gomadam, R. White, and B. N. Popov. Development of ﬁrst principles capacity fade model for li-ion cells. Journal of the Electrochemical Society 151(2), A196 (2004).

[34] A. Smith, J. C. Burns, X. Zhao, D. Xiong, and J. Dahn. A high precision coulometry study of the sei growth in li/graphite cells. Journal of The Electrochemical Society 158(5), A447 (2011).

[35] H. J. Ploehn, P. Ramadass, and R. E. White. Solvent diffusion model for aging of lithium-ion battery cells. Journal of The Electrochemical Society 151(3), A456 (2004).

[36] A. M. Andersson, K. Edstrom,¨ and J. O. Thomas. Characterisation of the ambient and elevated temperature performance of a graphite electrode. Journal of Power Sources 81, 8 (1999).

[37] G. Amatucci, A. Du Pasquier, A. Blyr, T. Zheng, and J.-M. Tarascon. The elevated temperature performance of the limn2o4/c system: failure and solutions. Electrochimica Acta 45(1-2), 255 (1999).

[38] P. Ramadass, B. Haran, R. White, and B. N. Popov. Capacity fade of sony 18650 cells cycled at elevated temperatures: Part i. cycling performance. Journal of power sources 112(2), 606 (2002).

[39] A. Du Pasquier, A. Blyr, A. Cressent, C. Lenain, G. Amatucci, and J. Tarascon. An update on the high temperature ageing mechanism in limn2o4-based li-ion cells. Journal of power sources 81, 54 (1999).

[40] T. Aoshima, K. Okahara, C. Kiyohara, and K. Shizuka. Mechanisms of manganese spinels dissolution and capacity fade at high temperature. Journal of Power Sources 97, 377 (2001).

[41] D. Aurbach, M. Levi, K. Gamulski, B. Markovsky, G. Salitra, E. Levi, U. Heider, L. Heider, and R. Oesten. Capacity fading of lixmn2o4 spinel electrodes studied by xrd and electroanalytical techniques. Journal of Power Sources 81, 472 (1999). REFERENCES 82

[42] H. Yamane, T. Inoue, M. Fujita, and M. Sano. A causal study of the capacity fading of li1. 01mn1. 99o4 cathode at 80 c, and the suppressing substances of its fading. Journal of power sources 99(1-2), 60 (2001).

[43] H. Oka, S. Kasahara, T. Okada, E. Iwata, M. Okada, T. Shoji, H. Ohki, and T. Okuda. Structural analysis of lithium-excess lithium manganate cathode materials by 7li magic-angle spinning nuclear magnetic resonance spectroscopy. Solid State Ionics 144(1-2), 19 (2001).

[44] S. Komaba, N. Kumagai, T. Sasaki, and Y. Miki. Manganese dissolution from lithium doped li-mn-o spinel cathode materials into electrolyte solution. Electrochemistry 69(10), 784 (2001).

[45] E. Iwata, K.-i. Takahashi, K. Maeda, and T. Mouri. Capacity failure on cycling or storage of lithium-ion batteries with li–mn–o ternary phases having spinel-framework structure and its possible solution. Journal of power sources 81, 430 (1999).

[46] J. Cho, G. B. Kim, H. S. Lim, C.-S. Kim, and S.-I. Yoo. Improvement of structural stability of limn2 o 4 cathode material on 55 c cycling by sol-gel coating of licoo2. Electrochemical and Solid-State Letters 2(12), 607 (1999).

[47] P. Arora, R. E. White, and M. Doyle. Capacity fade mechanisms and side reactions in lithium- ion batteries. Journal of the Electrochemical Society 145(10), 3647 (1998).

[48] R. B. Wright, C. G. Motloch, J. R. Belt, J. P. Christophersen, C. D. Ho, R. A. Richardson, I. Bloom, S. Jones, V. S. Battaglia, G. Henriksen, et al. Calendar-and cycle-life studies of advanced technology development program generation 1 lithium-ion batteries. Journal of power sources 110(2), 445 (2002).

[49] K. Amine, C. Chen, J. Liu, M. Hammond, A. Jansen, D. Dees, I. Bloom, D. Vissers, and G. Henriksen. Factors responsible for impedance rise in high power lithium ion batteries. Journal of power sources 97, 684 (2001).

[50] J. P. Fellner, G. J. Loeber, and S. S. Sandhu. Testing of lithium-ion 18650 cells and character- izing/predicting cell performance. Journal of Power Sources 81, 867 (1999).

[51] C. Gui. Inﬂuence of temperature on lifepo4 lithium-ion power battery. Battery Bimonthly 41(2), 88 (2011).

[52] L. Lam and P. Bauer. Practical capacity fading model for li-ion battery cells in electric vehicles. IEEE transactions on power electronics 28(12), 5910 (2013). REFERENCES 83

[53] K. Qian, C. Zhou, Y. Yuan, and M. Allan. Temperature effect on electric vehicle battery cycle life in vehicle-to-grid applications. In CICED 2010 Proceedings, pp. 1–6 (IEEE, 2010).

[54] K. Smith, T. Markel, G.-H. Kim, and A. Pesaran. Design of electric drive vehicle batteries for long life and low cost. In Accelerated Stress Testing and Reliability (ASTR), IEE Workshop on, pp. 6–8 (2010).

[55] S. Tong, M. P. Klein, and J. W. Park. Comprehensive battery equivalent circuit based model for battery management application. In ASME 2013 Dynamic Systems and Control Conference, pp. V001T05A005–V001T05A005 (American Society of Mechanical Engineers, 2013).

[56] A. Barre,´ B. Deguilhem, S. Grolleau, M. Gerard,´ F. Suard, and D. Riu. A review on lithium- ion battery ageing mechanisms and estimations for automotive applications. Journal of Power Sources 241, 680 (2013).

[57] J. Vetter, P. Novak,´ M. R. Wagner, C. Veit, K.-C. Moller,¨ J. Besenhard, M. Winter, M. Wohlfahrt-Mehrens, C. Vogler, and A. Hammouche. Ageing mechanisms in lithium-ion batteries. Journal of power sources 147(1-2), 269 (2005).

[58] M. Koller, T. Borsche, A. Ulbig, and G. Andersson. Deﬁning a degradation cost function for optimal control of a battery energy storage system. In 2013 IEEE Grenoble Conference, pp. 1–6 (IEEE, 2013).

[59] C. Zhou, K. Qian, M. Allan, and W. Zhou. Modeling of the cost of ev battery wear due to v2g application in power systems. IEEE Transactions on Energy Conversion 26(4), 1041 (2011).

[60] M. Broussely, P. Biensan, F. Bonhomme, P. Blanchard, S. Herreyre, K. Nechev, and R. Staniewicz. Main aging mechanisms in li ion batteries. Journal of power sources 146(1-2), 90 (2005).

[61] R. P. Ramasamy, R. E. White, and B. N. Popov. Calendar life performance of pouch lithium-ion cells. Journal of Power Sources 141(2), 298 (2005).

[62] B. A. Johnson and R. E. White. Characterization of commercially available lithium-ion batteries. Journal of power sources 70(1), 48 (1998).

[63] W.-h. Huang, Y.-s. Yan, C.-r. Wan, C.-y. JIANG, and J.-r. YING. Effect of anisole as an additive in electrolyte on the cycle performance of lithium-ion rechargeable battery. Chinese Journal of Power Sources 25(2; ISSU 140), 91 (2001). REFERENCES 84

[64] J.-M. Tarascon, A. Gozdz, C. Schmutz, F. Shokoohi, and P. Warren. Performance of bellcore’s plastic rechargeable li-ion batteries. Solid State Ionics 86, 49 (1996).

[65] S. Panchal, I. Dincer, M. Agelin-Chaab, R. Fraser, and M. Fowler. Experimental temperature distributions in a prismatic lithium-ion battery at varying conditions. International Communi- cations in Heat and Mass Transfer 71, 35 (2016).

[66] S. Panchal, I. Dincer, M. Agelin-Chaab, R. Fraser, and M. Fowler. Experimental temperature distributions in a prismatic lithium-ion battery at varying conditions. International Communi- cations in Heat and Mass Transfer 71, 35 (2016).

[67] A. Hoke, A. Brissette, K. Smith, A. Pratt, and D. Maksimovic. Accounting for lithium-ion battery degradation in electric vehicle charging optimization. IEEE Journal of Emerging and Selected Topics in Power Electronics 2(3), 691 (2014).

[68] C. Forgez, D. V. Do, G. Friedrich, M. Morcrette, and C. Delacourt. Thermal modeling of a cylindrical lifepo4/graphite lithium-ion battery. Journal of Power Sources 195(9), 2961 (2010).

[69] R. R. Richardson, P. T. Ireland, and D. A. Howey. Battery internal temperature estimation by combined impedance and surface temperature measurement. Journal of Power Sources 265, 254 (2014).

[70] M. Doyle, T. F. Fuller, and J. Newman. Modeling of galvanostatic charge and discharge of the lithium/polymer/insertion cell. Journal of the Electrochemical Society 140(6), 1526 (1993).

[71] T. F. Fuller, M. Doyle, and J. Newman. Simulation and optimization of the dual lithium ion insertion cell. Journal of the Electrochemical Society 141(1), 1 (1994).

[72] M. Doyle, J. Newman, A. S. Gozdz, C. N. Schmutz, and J.-M. Tarascon. Comparison of modeling predictions with experimental data from plastic lithium ion cells. Journal of the Electrochemical Society 143(6), 1890 (1996).

[73] J. Newman and K. E. Thomas-Alyea. Electrochemical systems (John Wiley & Sons, 2012).

[74] T. F. Fuller, T. Bandhauer, and S. Garimella. Electrochemical-thermal modeling and microscale phase change for passive internal thermal management of lithium ion batteries. Tech. rep., Sandia National Laboratories (2012).

[75] D. Bernardi, E. Pawlikowski, and J. Newman. A general energy balance for battery systems. Journal of the electrochemical society 132(1), 5 (1985). REFERENCES 85

[76] H. Beelen, L. Raijmakers, M. Donkers, P. Notten, and H. Bergveld. A comparison and accuracy analysis of impedance-based temperature estimation methods for li-ion batteries. Applied Energy 175, 128 (2016).

[77] T. M. Bandhauer, S. Garimella, and T. F. Fuller. A critical review of thermal issues in lithium- ion batteries. Journal of the Electrochemical Society 158(3), R1 (2011).

[78] W. Gu and C. Wang. Thermal-electrochemical modeling of battery systems. Journal of The Electrochemical Society 147(8), 2910 (2000).

[79] V. Srinivasan and C. Wang. Analysis of electrochemical and thermal behavior of li-ion cells. Journal of The Electrochemical Society 150(1), A98 (2003).

[80] L. Song and J. W. Evans. Electrochemical-thermal model of lithium polymer batteries. Journal of the Electrochemical Society 147(6), 2086 (2000).

[81] P. M. Gomadam, J. W. Weidner, R. A. Dougal, and R. E. White. Mathematical modeling of lithium-ion and nickel battery systems. Journal of power sources 110(2), 267 (2002).

[82] C. R. Pals and J. Newman. Thermal modeling of the lithium/polymer battery ii. temperature proﬁles in a cell stack. Journal of the Electrochemical Society 142(10), 3282 (1995).

[83] S. Chen, C. Wan, and Y. Wang. Thermal analysis of lithium-ion batteries. Journal of power sources 140(1), 111 (2005).

[84] W. Fang, O. J. Kwon, and C.-Y. Wang. Electrochemical–thermal modeling of automotive li- ion batteries and experimental validation using a three-electrode cell. International journal of energy research 34(2), 107 (2010).

[85] E. Prada, D. Di Domenico, Y. Creff, J. Bernard, V. Sauvant-Moynot, and F. Huet. Simpliﬁed electrochemical and thermal model of lifepo4-graphite li-ion batteries for fast charge applications. Journal of The Electrochemical Society 159(9), A1508 (2012).

[86] L. Zhang, X. Hu, Z. Wang, F. Sun, and D. G. Dorrell. Experimental impedance investigation of an ultracapacitor at different conditions for electric vehicle applications. Journal of Power Sources 287, 129 (2015).

[87] K. Liu, K. Li, Z. Yang, C. Zhang, and J. Deng. Battery optimal charging strategy based on a coupled thermoelectric model. In 2016 IEEE Congress on Evolutionary Computation (CEC), pp. 5084–5091 (IEEE, 2016). REFERENCES 86

[88] H. Kim, S. Kim, T. Kim, C. Hu, and B. D. Youn. Online thermal state estimation of high power lithium-ion battery. In 2015 IEEE Conference on Prognostics and Health Management (PHM), pp. 1–6 (IEEE, 2015).

[89] J. Lopez-Sanz, C. Ocampo-Martinez, J. Alvarez-Florez, M. Moreno-Eguilaz, R. Ruiz- Mansilla, J. Kalmus, M. Graeber,¨ and G. Lux. Nonlinear model predictive control for thermal management in plug-in hybrid electric vehicles. IEEE Transactions on Vehicular Technology 66(5), 3632 (2017).

[90] T. Wang, K. Tseng, J. Zhao, and Z. Wei. Thermal investigation of lithium-ion battery module with different cell arrangement structures and forced air-cooling strategies. Applied energy 134, 229 (2014).

[91] H. Park. A design of air ﬂow conﬁguration for cooling lithium ion battery in hybrid electric vehicles. Journal of power sources 239, 30 (2013).

[92] I. Arasaratnam, J. Tjong, R. Ahmed, M. El-Sayed, and S. Habibi. Adaptive temperature monitoring for battery thermal management. In 2013 IEEE Transportation Electriﬁcation Confer- ence and Expo (ITEC), pp. 1–6 (IEEE, 2013).

[93] S. Basu, K. S. Hariharan, S. M. Kolake, T. Song, D. K. Sohn, and T. Yeo. Coupled electrochemical thermal modelling of a novel li-ion battery pack thermal management system. Applied Energy 181, 1 (2016).

[94] C. Zhang, J. Jiang, Y. Gao, W. Zhang, Q. Liu, and X. Hu. Charging optimization in lithium-ion batteries based on temperature rise and charge time. Applied energy 194, 569 (2017).

[95] S. Panchal, I. Dincer, M. Agelin-Chaab, R. Fraser, and M. Fowler. Experimental temperature distributions in a prismatic lithium-ion battery at varying conditions. International Communi- cations in Heat and Mass Transfer 71, 35 (2016).

[96] S. Panchal, I. Dincer, M. Agelin-Chaab, R. Fraser, and M. Fowler. Experimental and theoretical investigation of temperature distributions in a prismatic lithium-ion battery. International Journal of Thermal Sciences 99, 204 (2016).

[97] S. Panchal, I. Dincer, M. Agelin-Chaab, R. Fraser, and M. Fowler. Experimental and simulated temperature variations in a lifepo4-20 ah battery during discharge process. Applied Energy 180, 504 (2016). REFERENCES 87

[98] J. Christensen and J. Newman. Stress generation and fracture in lithium insertion materials. Journal of Solid State Electrochemistry 10(5), 293 (2006).

[99] C. Fellner and J. Newman. High-power batteries for use in hybrid vehicles. Journal of Power Sources 85(2), 229 (2000).

[100] G. Ning and B. N. Popov. Cycle life modeling of lithium-ion batteries. Journal of The Electro- chemical Society 151(10), A1584 (2004).

[101] B. Y. Liaw, R. G. Jungst, G. Nagasubramanian, H. L. Case, and D. H. Doughty. Modeling capacity fade in lithium-ion cells. Journal of power sources 140(1), 157 (2005).

[102] G. Ning, R. E. White, and B. N. Popov. A generalized cycle life model of rechargeable li-ion batteries. Electrochimica acta 51(10), 2012 (2006).

[103] M. Safari, M. Morcrette, A. Teyssot, and C. Delacourt. Multimodal physics-based aging model for life prediction of li-ion batteries. Journal of The Electrochemical Society 156(3), A145 (2009).

[104] R. Spotnitz. Simulation of capacity fade in lithium-ion batteries. Journal of Power Sources 113(1), 72 (2003).

[105] Q. Zhang and R. E. White. Capacity fade analysis of a lithium ion cell. Journal of Power Sources 179(2), 793 (2008).

[106] A. Abdollahi, X. Han, N. Raghunathan, B. Pattipati, B. Balasingam, K. Pattipati, Y. Bar- Shalom, and B. Card. Optimal charging for general equivalent electrical battery model, and battery life management. Journal of Energy Storage 9, 47 (2017).

[107] Y. Yang. Optimization of battery energy storage systems for pv grid integration based on sizing strategy (2014).

[108] J. Sardi, N. Mithulananthan, M. Gallagher, and D. Q. Hung. Multiple community energy storage planning in distribution networks using a cost-beneﬁt analysis. Applied energy 190, 453 (2017).

[109] X. Han, H. Zhang, X. Yu, and L. Wang. Economic evaluation of grid-connected micro-grid system with photovoltaic and energy storage under different investment and ﬁnancing models. Applied energy 184, 103 (2016). REFERENCES 88

[110] A. S. Sidhu, M. G. Pollitt, and K. L. Anaya. A social cost beneﬁt analysis of grid-scale electrical energy storage projects: A case study. Applied energy 212, 881 (2018).

[111] G. Cardoso, T. Brouhard, N. DeForest, D. Wang, M. Heleno, and L. Kotzur. Battery aging in multi-energy microgrid design using mixed integer linear programming. Applied Energy 231, 1059 (2018).

[112] A. Yoshida, T. Sato, Y. Amano, and K. Ito. Impact of electric battery degradation on cost-and energy-saving characteristics of a residential photovoltaic system. Energy and Buildings 124, 265 (2016).

[113] G. Angenendt, S. Zurmuhlen,¨ H. Axelsen, and D. U. Sauer. Comparison of different operation strategies for pv battery home storage systems including forecast-based operation strategies. Applied energy 229, 884 (2018).

[114] C. D. Patel and A. J. Shah. Cost model for planning, development and operation of a data center (2005).

[115] S. Pourmousavi and T. K. Saha. Evaluation of the battery operation in ramp-rate control mode within a pv plant: A case study. Solar Energy 166, 242 (2018).

[116] M. Alam, R. Yan, T. Saha, A. Chidurala, and D. Eghbal. Learning from a 3.275 mw utility scale pv plant project. In CIGRe´ (2016).

[117] A. Millner. Modeling lithium ion battery degradation in electric vehicles. In Innovative Tech- nologies for an Efﬁcient and Reliable Electricity Supply (CITRES), 2010 IEEE Conference on, pp. 349–356 (IEEE, 2010).

[118] L. Saw, Y. Ye, and A. Tay. Electro-thermal analysis and integration issues of lithium ion battery for electric vehicles. Applied energy 131, 97 (2014).

[119] U. S. Kim, J. Yi, C. B. Shin, T. Han, and S. Park. Modelling the thermal behaviour of a lithium-ion battery during charge. Journal of Power Sources 196(11), 5115 (2011).

[120] L. Fan, J. Khodadadi, and A. Pesaran. A parametric study on thermal management of an air-cooled lithium-ion battery module for plug-in hybrid electric vehicles. Journal of Power Sources 238, 301 (2013).

[121] S. S. Haykin, S. S. Haykin, S. S. Haykin, and S. S. Haykin. Neural networks and learning machines, vol. 3 (Pearson Upper Saddle River, NJ, USA:, 2009). REFERENCES 89

[122] G. Zhang, B. E. Patuwo, and M. Y. Hu. Forecasting with artiﬁcial neural networks:: The state of the art. International journal of forecasting 14(1), 35 (1998).

[123] M. Castilla, J. Alvarez,´ M. Ortega, and M. Arahal. Neural network and polynomial approxi- mated thermal comfort models for hvac systems. Building and Environment 59, 107 (2013).

[124] I. A. Basheer and M. Hajmeer. Artiﬁcial neural networks: fundamentals, computing, design, and application. Journal of microbiological methods 43(1), 3 (2000).

[125] S. Chen and S. Billings. Neural networks for nonlinear dynamic system modelling and identi- ﬁcation. International journal of control 56(2), 319 (1992).

[126] A. Mellit, S. Kalogirou, S. Shaari, H. Salhi, and A. H. Arab. Methodology for predicting sequences of mean monthly clearness index and daily solar radiation data in remote areas: Application for sizing a stand-alone pv system. Renewable Energy 33(7), 1570 (2008).

[127] L. Ljung. Black-box models from input-output measurements. In Instrumentation and Mea- surement Technology Conference, 2001. IMTC 2001. Proceedings of the 18th IEEE, vol. 1, pp. 138–146 (IEEE, 2001).

[128] E. Pisoni, M. Farina, C. Carnevale, and L. Piroddi. Forecasting peak air pollution levels using narx models. Engineering Applications of Artiﬁcial Intelligence 22(4-5), 593 (2009).

[129] M. M. Hasan, S. A. Pourmousavi, F. Bai, and T. K. Saha. The impact of temperature on battery degradation for large-scale bess in pv plant. In Universities Power Engineering Conference (AUPEC), 2017 Australasian, pp. 1–6 (IEEE, 2017).

[130] A. A. Ferreira, T. B. Ludermir, and R. R. de Aquino. Comparing recurrent networks for time- series forecasting. In Neural Networks (IJCNN), The 2012 International Joint Conference on, pp. 1–8 (IEEE, 2012).

[131] J. Buitrago and S. Asfour. Short-term forecasting of electric loads using nonlinear autoregressive artiﬁcial neural networks with exogenous vector inputs. Energies 10(1), 40 (2017).

[132] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal representations by error propagation. Tech. rep., California Univ San Diego La Jolla Inst for Cognitive Science (1985).

[133] S. Chen, S. Billings, and P. Grant. Non-linear system identiﬁcation using neural networks. International journal of control 51(6), 1191 (1990). REFERENCES 90

[134] S. Mohanty, P. K. Patra, and S. S. Sahoo. Prediction of global solar radiation using nonlinear auto regressive network with exogenous inputs (narx). In Systems Conference (NSC), 2015 39th National, pp. 1–6 (IEEE, 2015).

[135] H. Xie, H. Tang, and Y.-H. Liao. Time series prediction based on narx neural networks: An advanced approach. In Machine Learning and Cybernetics, 2009 International Conference on, vol. 3, pp. 1275–1279 (IEEE, 2009).

[136] K. Hornik. Approximation capabilities of multilayer feedforward networks. Neural networks 4(2), 251 (1991).

[137] B. Kermanshahi. Recurrent neural network for forecasting next 10 years loads of nine japanese utilities. Neurocomputing 23(1-3), 125 (1998).

[138] E. Ogasawara, L. C. Martinez, D. De Oliveira, G. Zimbrao,˜ G. L. Pappa, and M. Mattoso. Adaptive normalization: A novel data normalization approach for non-stationary time series. In Neural Networks (IJCNN), The 2010 International Joint Conference on, pp. 1–8 (IEEE, 2010).

[139] D. McLean, Z. Bandar, and J. OShea. The evolution of a feedforward neural network trained under backpropagation. In Artiﬁcial Neural Nets and Genetic Algorithms, pp. 518–522 (Springer, 1998).

[140] X. Glorot and Y. Bengio. Understanding the difﬁculty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artiﬁcial intelligence and statistics, pp. 249–256 (2010).

[141] H.-T. Yang, C.-M. Huang, and C.-L. Huang. Identiﬁcation of armax model for short term load forecasting: An evolutionary programming approach. In Power Industry Computer Application Conference, 1995. Conference Proceedings., 1995 IEEE, pp. 325–330 (IEEE, 1995).

[142] G. P. Zhang. Time series forecasting using a hybrid arima and neural network model. Neuro- computing 50, 159 (2003).

[143] J. H. Cochrane. Time series for macroeconomics and ﬁnance. Manuscript, University of Chicago (2005).

[144] W. A. Fuller. Introduction to statistical time series, vol. 428 (John Wiley & Sons, 2009). REFERENCES 91

[145] N. Korolko, Z. Sahinoglu, and D. Nikovski. Modeling and forecasting self-similar power load due to ev fast chargers. IEEE Transactions on Smart Grid 7(3), 1620 (2016).

[146] G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung. Time series analysis: forecasting and control (John Wiley & Sons, 2015).

[147] R. Hausbrand, G. Cherkashinin, H. Ehrenberg, M. Groting,¨ K. Albe, C. Hess, and W. Jaegermann. Fundamental degradation mechanisms of layered oxide li-ion battery cathode materials: Methodology, insights and novel approaches. Materials Science and Engineering: B 192, 3 (2015).

[148] S. A. P. Kani, P. Wild, and T. K. Saha. Improving predictability of renewable generation through optimal battery sizing. IEEE Transactions on Sustainable Energy (2018).