Physics-Informed Deep Learning for Solar Irradiance Forecasting Using a Constrained LSTM and Hybrid Physical–Machine Learning Models

Abstract

Accurate forecasting of Rs (global solar irradiance) is essential for the reliable integration of photovoltaic energy into modern power systems. This work proposes a physics-informed deep learning framework that combines clear-sky solar radiation modeling with data-driven approaches to improve short-term solar irradiance prediction. A hybrid physical-machine learning model based on XGBoost integrates theoretical clear-sky global horizontal irradiance (GHI) with meteorological variables, while a physicsconstrained Long Short-Term Memory (LSTM) network learns the residual component between measured irradiance and physical estimates. Temporal cross-validation, statistical significance tests, SHAP-based interpretability analysis, and robustness experiments across multiple years and solar stations are conducted. Results demonstrate that the proposed hybrid and constrained LSTM models significantly outperform conventional machine learning and deep learning baselines, achieving higher accuracy, better generalization, and enhanced physical consistency. The proposed framework is suitable for real-world deployment in smart grids and renewable energy management systems.