Physics-informed deep learning for data-efficient and robust photovoltaic power forecasting

Photovoltaic (PV) power forecasting is challenged by its inherent variability. Pure data-driven models struggle with generalization under data scarcity and complex weather conditions. In this paper, we introduce a physics-informed deep learning hybrid model (PIDL-HM) that systematically generates physically grounded input features (e.g., plane-of-array irradiance and module temperature) for a convolutional neural network–long short-term memory (CNN–LSTM), establishing a principled integration framework beyond simple ensemble methods. Rigorously validated across multiple PV power plants in China and Australia for 15-min, 4-h, and 24-h forecasting, our approach demonstrates superior performance, with a reduction of up to 8.93% in root mean square error compared to a purely data-driven baseline. Crucially, the model shows remarkable data efficiency, maintaining high accuracy with only three months of training data, and exceptional robustness, providing a 6.87% improvement in performance under strong cross-seasonal data distribution shifts. This work provides a reliable and data-efficient forecasting solution, establishing the PIDL-HM as a foundational element for next-generation forecasting systems.