Accurate photovoltaic power prediction via temperature correction with physics-informed neural networks

Abstract

Photovoltaic (PV) power generation, an essential part of renewable energy, is affected by both irradiance and module temperature. Accurately predicting PV module temperature and power output is essential for optimizing system operations and management. This paper proposes a PV module temperature prediction model based on physics-informed neural networks (PINN). The model uses an ordinary differential equation (ODE) to simulate the energy exchange between the PV module and its environment, accurately predicting the module's temperature. The temperature features generated by the PINN are then integrated with a long-short term cross attention mechanism (LSCAM) as part of the input for PV power prediction. This fusion of mechanism data-driven features enables precise forecasting of PV power generation. Experimental validation on a test set from a PV site in Zhejiang Province, China, demonstrates high R-squared values for both temperature prediction (0.9808, 0.9602, 0.9806, 0.9811) and power prediction (0.9880, 0.9720, 0.9829, 0.9872) across different seasons. The results show that the model significantly improves the prediction accuracy and enhances generalization, offering strong support for the future intelligent control and optimization of PV systems.