A dual-channel parallel physical-statistical coupling model with dynamic fusion for day-ahead photovoltaic power forecasting

The integration of photovoltaic power generation at high penetration levels introduces significant intermittency and variability into modern power grids, making accurate day-ahead forecasting indispensable for grid stability and economic dispatch. To address the limitations of standalone physical or statistical models, this study proposes a novel dual-channel parallel physical-statistical coupling model featuring a dynamic fusion mechanism. The physical channel leverages an improved Sandia photovoltaic model to calculate theoretical power output, while the statistical channel employs an XGBoost regressor to correct systematic biases inherent in numerical weather prediction data. A gated recurrent unit network acts as a dynamic fusion gate, adaptively weighting the contributions of the two channels based on real-time meteorological context. Experimental validation on a dataset from a 10 MW utility-scale plant demonstrates the model's superiority. It achieves a root mean square error of 396.4 kW and a mean absolute error of 289.1 kW, representing a 39.2% reduction in RMSE compared to a standalone physical model. The dynamic fusion mechanism functions as intended, assigning greater weight to the physical channel under clear skies and shifting reliance to the statistical channel during complex weather conditions. This adaptive coupling provides a robust and reliable forecasting tool for grid scheduling.