물리 기반 기계 학습 및 딥러닝 모델을 이용한 에너지 발전량 및 저장 손실 예측 연구 = Prediction of Energy Generation and Energy Storage Loss by Physics driven Machine and Deep Learning Models

다국어 초록 (Multilingual Abstract)

Fast growing of solar power technology in recent years is vastly dependent on location, climate and weather conditions, which hinders the proper Photovoltaic (PV) power generation and its distribution. Managing the PV power generation by smart management like machine learning, deep learning etc. is crucial to maintain photovoltaic system and its operation. This work investigates the PV power generation forecasting in the vertical bifacial silicon solar panels using the deep learning approach like long short-term memory (LSTM) and LSTM with multi-head attention models. Raw dataset of power generation was collected from four vertical solar panels (such as two monofacial and two bifacial), which were placed in South-North (S-N) and East-West (E-W) directions. The solar panels are situated at the New & Renewable Energy Materials Development Centre (NewREC), Buan-gun, Jeonbuk, Republic of Korea. Data for the vertical PV power system was collected from March 2022 to February 2023. Prior to performing the forecasting, several variables, including insolation, temperature, wind speed, and humidity, were used as input, and the feature importance was checked using the Pearson correlation coefficient. The collected dataset was processed and evaluated for PV power generation forecasting using LSTM and LSTM models with multi-head attention. As compared with monofacial solar panels, bifacial solar panels showed excellent accuracy in terms of PV power generation forecasting. Vertical bifacial solar system in S-N direction expressed the highest accuracy of PV power generation forecasting having low MAE = 0.0102 and RMSE = 0.01781 by LSTM with multi-head attention. The incorporation of multi-head attentions in the LSTM model may be helpful in improving the precision of PV power generation forecasting in vertical solar systems.
In the second part, the major obstacle to the widespread adoption of lithium-ion batteries (LIBs) in electric vehicles (EVs) is degradation, which reduces battery life and operational efficiency. The degradation process in LIBs may cause the loss of initial lithium, damage the active material, and elevate internal resistance, thereby reducing the overall capacity and performance. To manage these degradation processes, the proper and fast tools are required for better battery health and performance. By adopting machine/deep learning algorithms, the capacity fade and battery life (which typically reaches more than 50% of End-of-life (EoL)) can be accurately predicted to manage future battery health. In this work, the capacity-based predictions of capacity fade and 50% EoL were performed by several machine and deep learning models. The cell degradation dataset used in this work was collected from a public dataset of ‘Toyota’, in which 124 commercial LIBs with the nominal capacity of 1.1 Ah and a nominal voltage of 3.3 V were cycled to failure under fast-charging conditions. The estimation of capacity fade was carried out by 4 models, including Linear regression (LR), Light Gradient Boosting Machine (LightGBM), Neural Network (NN) and Physics informed Neural Network (PINN), wherein the internal resistance (IR), temperature, and cycle number were used as input features. Among all models, the PINN model showed the best accuracy with the MAE=0.0242, RMSE= 0.0371, and R2= 0.99987 for capacity fade estimation. The best PINN model was well-visualized with actual data, as observed, with a capacity fade of ~20% at 876 cycles. Moreover, Arrhenius equation was used to investigate the temperature effect on capacity fade. 3 different temperatures, such as 31°C, 45°C, and 60°C, were selected, and it was found that a proportional increase in capacity fade occurred with the increase in temperature. At the end, for long-term battery forecasting, the best PINN model showed that battery life severely decreased as the temperature increased from 31 to 60°C. At 31°C, the 50% EoL was recorded at 1169 cycles, whereas 50% EoL dropped significantly to 732 cycles at 45°C and 463 cycles at 60°C. Thus, the deep learning model (PINN) demonstrated the best performing model to estimate and forecast capacity fade and 50% EoL of the battery.

목차 (Table of Contents)

• Chapter 1 1
• Background 1
• 1.1. Introduction 2
• 1.1.1. Energy Generation- Solar Panel 8
• 1.1.2. Energy Stroage- Lithium-ion Battery(LiB) 12
• 1.1.2.1. Capacity Fade 15
• 1.1.2.2. End-of-Life(EoL) 17
• 1.1.2.3. Temperature Effect on Capacity Fade 20
• Chapter 2 24
• Methodology : PV Power Generation Forecasting in vertical PV panels by Deep Learning Models 24
• 2.1. Descrition of Methodology for PV Power Generation Forecasting 25
• 2.1.1. Pre-process Data 25
• 2.1.2. Building Model and Training 25
• 2.1.3. Forecasting and validation 26
• 2.2. Data Collection from installed Vertical Solar Panels 27
• 2.3. Data Description 32
• 2.4. Feature Selection by pearson Correlation coefficient(PCC) 41
• 2.5. Data Spitting 44
• 2.6. Models. 46
• 2.6.1. Long- Short Term Memory (LSTM) 46
• 2.6.2. LSTM with multi-head attention. 49
• 2.7. Evaluation Metrics 52
• 2.8. Result and Discussion. 54
• Chapter 3 62
• Capacity fade estimation and 50% End-of-Life management in battery by Machine learning and Deep Learning algorithms 62
• 3.1. Methodology 63
• 3.2. Experimental Dataset Descrition 63
• 3.3. Capacity Fade estimation Methodology 68
• 3.3.1. Capacity fade estimation. 70
• 3.3.2. Feature Selection by PCC 72
• 3.3.3. Models 72
• 3.3.3.1. LightGBM. 72
• 3.3.3.2. Physics informed Neural Network (PINN). 77
• 3.3.4. Result and Discussion. 81
• 3.4. End-of-Life (EoL) Forecasting 92
• 3.4.1. Methodology of Temperature Effect using Arrhenius Equation 95
• 3.4.1.1. Arrhenius Equation.. 98
• 3.4.1.2. Data Aggregation and Decay Constant Determination 99
• 3.4.1.3. Arrhenius Linearization and Parameter Extraction.. 101
• 3.4.1.4. Result and Discussion.. 103
• Chapter 4 107
• Conclusion 108
• References 110
• 요약 (국문초록) 128
• Acknowledgments 129
• List of Publications 131
• List of attended Conferences/Symposiums 132