Publication

Abstract : Accurate prediction of photovoltaic generation, wind generation, and electric demand is crucial for grid reliability, economic dispatch, and the integration of renewable resources across different operational timeframes. This study proposes a forecasting framework that evaluates multiple machine learning models, including XGBoost, Random Forest, NGBoost, CatBoost, SVR, and KNN, with LightGBM identified as the most effective learner when applied to 5 -minute data aggregated into daily intervals. To enhance transparency, Explainable AI techniques are integrated: SHAP quantifies global and local feature contributions, highlighting the influence of lagged signals, while LIME provides case-specific explanations for individual forecasts. The methodology employs chronological data splits with train-only scaling to preserve the integrity of time-series learning and uses R2 and RMSE metrics for evaluation. Results indicate that forecasting performance varies across datasets, underscoring the need for customized models. Comparative analysis confirms that LightGBM achieves the best trade-off between accuracy, stability, and interpretability. By combining robust forecasting with explainable insights, the proposed framework supports AI-enhanced energy management, enabling situational awareness, informed decision-making, and scalable deployment in renewable-rich power systems.