PV Forecasting with Constrained Ensemble Learning and Context-Aware Stacking

We propose a hybrid ensemble learning architecture for short-term photovoltaic (PV) power forecasting. The system integrates a physics-based model, a Bi-LSTM network, and an XGBoost regressor within a two-layer prediction pipeline. A constrained Ridge regression assigns robust weights to the base models, while a context-aware stacking layer dynamically refines predictions based on temporal and meteorological interaction features. Post-processing filters ensure physical consistency and enhance resilience to anomalies. Evaluated on real-world PV datasets, the proposed architecture consistently outperforms individual models and baseline ensembles, demonstrating superior accuracy, robustness, and readiness for deployment in intelligent energy management platforms.