Residual Physics-Informed Neural Networks for Robust Photovoltaic Cell Temperature Modeling Across Seasonal and Data-Scarce Conditions

Abstract

Accurate photovoltaic (PV) cell temperature prediction is essential for improving power forecasting, efficiency estimation, and long-term reliability analysis of photovoltaic systems. However, conventional empirical thermal models often lack adaptability under varying environmental conditions, whereas purely data-driven methods may exhibit poor generalization and physically inconsistent predictions. This study proposes a Residual Physics-Informed Neural Network (Residual-PINN) that integrates a simplified lumped thermal energy balance model with residual learning to enhance both predictive accuracy and physical consistency. The physics-based model first estimates the dominant thermal behavior, while the neural network learns only the residual temperature component associated with nonlinear environmental effects. The proposed framework was comprehensively evaluated through random-split, cross-season, extreme-irradiance, physics-consistency and low-data experiments. Experimental results demonstrate that the proposed Residual-PINN achieves the best overall performance with a MAE of 1.82°C, an RMSE of 2.61°C, and an R² value of 0.95, outperforming all benchmark models. Compared with the conventional multilayer perceptron, the proposed approach reduces RMSE by approximately 20% while decreasing the nighttime prediction error to 0.36°C and limiting the monotonicity violation ratio to 0.9%. Furthermore, the model maintains superior robustness under seasonal distribution shifts and low-data conditions, achieving an RMSE of 2.95°C when trained using only 20% of the available data. These results demonstrate that integrating simplified thermal physics with residual learning provides a robust, interpretable, and data-efficient framework for photovoltaic temperature prediction.