SOLAR POWER SYSTEM PRODUCTION FORECASTING ON MICROGRID SYSTEM USING PHYSICS-INFORMED MACHINE LEARNING METHOD

• Main Author: Hanadi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81490
• Institution: Institut Teknologi Bandung

The uncertainty and stochastic nature of solar energy pose challenges in maintaining power quality and system stability in microgrids. Solar power generation is influenced by complex external factors such as solar radiation, ambient temperature, and sun position. Therefore, accurate prediction of solar power generation is crucial for effective and efficient microgrid management. Data-driven models like Machine Learning can handle high data complexity and make highly accurate predictions when trained with large datasets. However, these models may produce unrealistic predictions due to inconsistencies with the underlying physical principles of the phenomena. On the other hand, physicsbased models provide accurate predictions with minimal validation data requirements but are complex, time-consuming, and susceptible to uncertainties in physical parameters. To overcome these limitations, this study proposes a Physics-informed Machine <p align="justify">Learning method that integrates physical knowledge into the Machine Learning model. This method aims to reduce prediction errors caused by fluctuations in physical parameters and ensure that predictions are consistent with relevant physical principles. The proposed approach uses a modified XGBoost algorithm that incorporates physical constraints into its learning process by combining productivity formulas with the mean-square error (MSE) loss function. A hyperparameter is introduced to balance the productivity formula and MSE to create this physics-guided adaptive loss function. The model's performance is evaluated using several metrics, emphasizing the alignment between estimated and actual solar energy production values. The data used in this study includes irradiance, ambient temperature, humidity, wind speed, solar panel power production, solar path, and microgrid operation modes. The results show that the PC-XGBoost model provides more accurate and robust predictions compared to the pure data-driven model. The PC-XGBoost model demonstrates a lower mean absolute error (MAE) of 1.05% compared to 1.24% for the pure data-driven model, and an improvement in the coefficient of determination (R²) from 0.90 to 0.92. This study makes significant contributions to the development of solar energy production prediction in microgrid systems, ultimately aiming for better alignment between estimates and actual values, supporting more efficient and stable microgrid management.<p align="justify">