Performance analysis of lightgbm, xgboost, Random forest, and gradient boosting in Photovoltaic energy forecasting with Hyperparameter optimization

Accurate photovoltaic (PV) energy forecasting is crucial for effective grid integration and for predictive usage at residential and industrial levels, especially under increasing climatic variability. This study evaluates and compares four machine learning (ML) models, LightGBM, XGBoost, Random Forest, and Gradient Boosting, for hourly PV energy forecasting using real-time data from numerical weather model (NWM), PVGIS, and historical production data from operational PV plant in Southern Italy. Three hyperparameter strategies, namely default settings, Optuna optimization, and Grid Search, were tested. Results show that LightGBM achieved the best performance with Grid Search tuning, yielding an MAE of 2.85 kWh, RMSE of 5.45 kWh, and R2 of 0.71 over an 8-day forecasting horizon. Comparatively, XGBoost with Grid Search attained an MAE of 3.00 kWh, RMSE of 5.82 kWh, and R2 of 0.67. The findings highlight that hyperparameter tuning significantly improved forecast accuracy and provide actionable insights for selecting ML models and optimization techniques in PV management systems. Findings are specifically of interest for practitioners, researchers, and organizations associated with PV management and operations.