HYBRID PV SYSTEM PRODUCTION MODELLING USING PHYSICS-INFORMED XGBOOST

• Main Author: Manik Kottama, Hanif
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85721
• Institution: Institut Teknologi Bandung

The utilization of Solar Power Plants (PLTS) as a renewable energy source is increasingly prevalent in Indonesia, aligned with the government's commitment to achieving a 23% share of new and renewable energy. However, power fluctuations caused by weather dependence pose a significant challenge in PLTS operations. To address this, an accurate power prediction model is necessary, particularly for hybrid PLTS systems integrated with energy storage systems. This study develops a power production prediction model for PLTS using the Physics-Informed eXtreme Gradient Boosting (XGBoost) method, combining machine learning with physical knowledge. The model considers changes in operating modes (grid-tie and islanded) as well as the impact of load on the PLTS system. Feature Interaction Constraint (FIC) is used to separate the features of grid-tie and islanded operations, while physical constraints are incorporated through the use of a Custom Loss Function (CLF). Three model variations were created: Model 1 using XGBoost without load and battery features; Model 2 using FIC, CLF, and adding load and battery features. From Models 1 and 2, optimal FIC and CLF configurations were obtained. Subsequently, Model 3 was developed using the previous research's scheme, predicting one day ahead with one month of training data to allow for comparison. Model 2 achieved a reduction in RMSE by 53,02% and MAE by 56,02%. Additionally, RMSE and MAE reductions were observed for each operating mode: 7.76% and 64.97% for grid-tie, and 37,94% and 41,80% for islanded mode. The implementation of Model 3 further reduced RMSE and MAE by 55,71% and 58,26%, respectively. Based on these evaluation results, the inclusion of load, battery, FIC, and CLF features enhances the model's prediction capability and enables it to differentiate between operating modes. Keywords: PLTS, Physics-Informed, Machine Learning, XGBoost, Operating Mode.