Dynamic Ensemble Learning with Explainability for Photovoltaic Power Prediction

Photovoltaic (PV) power prediction is essential for integrating solar energy into the power grid and optimizing energy supply and demand management. However, accurate prediction is challenging due to the variability and intermittency of solar radiation, influenced by diverse weather conditions. This study enhances PV power prediction accuracy by evaluating various ensemble machine-learning models: CatBoost, Random Forest (RF), Gradient Boosting (GB), and XGBoost. These models are chosen for their capacity to model complex nonlinear patterns and enhance predictive accuracy through ensemble aggregation of weak learners. SHapley Additive exPlanations (SHAP) are employed to identify critical variables affecting PV power prediction, confirming solar radiation as the most significant factor. To account for time dependence and focus on essential variables, reduced dynamic models are developed. These models omit non-essential variables and incorporate lagged solar irradiation and PV power data, effectively capturing temporal dependencies and improving accuracy. Evaluation using real-world data from five PV systems in Brisbane, Australia, demonstrates that reduced dynamic models consistently outperform their static counterparts. Among the dynamic models, Gradient Boosting achieves the highest average R$^{2}$ of 0.979, followed closely by CatBoost and Random Forest at 0.977, while XGBoost achieves 0.975. In contrast, static models such as XGBoost, Random Forest, and Gradient Boosting yield lower average R$^{2}$ values of 0.9668, 0.9658, and 0.9452, respectively. These results underscore the critical influence of solar radiation on PV system performance, highlighting the effectiveness of dynamic modeling approaches in enhancing prediction precision and supporting informed decision-making in energy management strategies.