Hybrid photovoltaic power forecasting by combining physical and machine learning models

Hybrid models integrating physical and machine learning (ML) techniques demonstrate strong potential for photovoltaic (PV) power forecasting. This manuscript provides a comprehensive analysis of hybrid models for deterministic and probabilistic day-ahead PV forecasting. Two hybrid approaches are implemented: physics-informed ML, which relies primarily on the ML model, and physics-guided models, primarily driven by the physical approach. Different physical and ML models, from classical to more detailed formulations, are used to study whether increasing physical and ML complexity affects the final performance of hybrid models. Results indicate that the ML model is the most influential factor in final forecasting accuracy, with nonlinear (especially ensemble) methods achieving the best results. In contrast, hybrid models’ performance is weakly sensitive to physical model refinements and to the hybridization approach, which affects the final metrics by less than 1%. To validate the results, data from PV systems in Brazil and the Netherlands are used, representing different climate conditions and installation types. Comparative tests confirm that the synergy between physical and ML improves accuracy and robustness, offering a reliable solution for PV forecasting.