Abstract
The growing worldwide penetration of solar energy into power grids has increased the need for accurate forecasting to reduce supply–demand imbalances and grid congestion. Physics-based models perform well for average conditions. However, edge cases such as very low or high irradiance are challenging for traditional physics-based models, which often underperform under these conditions. Data-driven models offer improved performance in these conditions but typically lack interpretability and insight into the underlying physical mechanisms. To address these limitations, this article presents a combined modeling framework that integrates physics-based principles with machine learning inspired by Physics-Informed Neural Networks (PINNs). The main novelty of this work is a machine learning architecture for solar energy power prediction that explicitly isolates angle of incidence losses and temperature losses in the form of the respective parameters entering a PV power equation as a strict output constraint, which improves interpretability and generalizability, while maintaining high forecasting precision. The methodology is demonstrated using only commonly available commercial system data and validated on both a small-scale research system and a commercial floating solar installation. The results demonstrate significantly improved accuracy over basic physics-based models (0.018 and 0.018 versus 0.023 and 0.029 normalized mean absolute error), particularly for non-standard conditions, and isolate key loss mechanisms.