Distributed Photovoltaic Power Prediction Model Based on Temporal Classification-Driven Structural Switching

In distributed photovoltaic (PV) power prediction, frequent local weather disturbances lead to severe fluctuations in PV output and a decline in prediction accuracy. Traditional single-structure prediction models struggle to adapt to both stable and disturbed scenarios, suffering from high structural rigidity and poor adaptability to varying conditions. To address this, this paper proposes a distributed PV power prediction model based on temporal classification-driven structural switching. The method first constructs a disturbance recognition module, training an advanced temporal classification model with numerical weather prediction (NWP) time-series features to determine the disturbance level of input meteorological sequences. Then, based on the recognition results, the model dynamically selects different structural paths and fuses multi-path prediction results through disturbance-aware weighting, achieving structurally adaptive modeling. This mechanism effectively enhances the model’s responsiveness and robustness under strong disturbance weather conditions. Experimental results on a real multi-site NWP–PV dataset show that the proposed model achieves significantly lower prediction errors in highly disturbed weather compared to traditional static-structure models, and outperforms multiple benchmark methods in overall RMSE and MAE metrics, validating its effectiveness.