Dynamic-parameter physics-informed neural networks for short-term photovoltaic power prediction: Integrating physics-informed and data driven

Abstract

In order to address the limitations of rigid physical constraints and sample imbalance in traditional hybrid prediction models, this paper proposes a novel short-term photovoltaic (PV) power prediction framework based on dynamic-parameter physical information neural network (DP-PINN). Based on Newton Raphson's optimized K-means++ (NBRO-Kmeans++) algorithm, the weather is classified into four types, and compared with K-means++, the silhouette coefficient is increased by 6.6–45.8 %. The Synthetic Minority Oversampling Technique (SMOTE) is used to dynamically balance minority samples, reducing RMSE by 50.5 % in this case. The physical equations are dynamically adjusted based on weather types, and the triple constraint loss function integrates data fitting, physical derivatives, and equation consistency, and dynamically adjusts the weights related to weather during the training process. The photoelectric conversion efficiency (η) and temperature coefficient (α) are learnable parameters optimized through backpropagation. The effectiveness of this method is verified through one-year operation data simulation of a 50 MW PV power station in China. Case analysis shows that under extreme weather conditions, RMSE is 50.8 % lower than CNN-LSTM, 34.08 % higher on sunny days compared to pure data-driven models, and 25.7 % lower on average RMSE compared to static parameter PINN (SP-PINN). This method provides a universal solution for predicting high volatility renewable energy with enhanced physical interpretability.