Distributed photovoltaic power forecasting based on dynamic fusion of physical priors and numerical weather prediction

Abstract

Distributed photovoltaic power forecasting remains difficult because existing data-driven methods mainly learn statistical correlations from historical samples and often produce physically inconsistent outputs under rapidly changing weather. From a measurement perspective, this task requires reliable inference of an unobserved future power quantity from heterogeneous and uncertain inputs, including historical power measurements and numerical weather predictions. To address this issue, this paper proposes a forecasting framework that dynamically fuses physical priors and forecast meteorological information. Its novelty lies in treating the physical power envelope and baseline as interpretable measurement references rather than merely concatenating them with data-driven features. A physical power upper bound and a physical baseline power are first constructed from irradiance, temperature, wind speed, and humidity. A dual-branch TimeMixer encoder separately models the forecast meteorological and physical-prior sequences, while multi-scale cross attention and dynamic gating adaptively fuse weather evolution with physical power evolution. Residual correction around the physical baseline is further combined with physics-constrained losses to suppress over-bound and nighttime-positive predictions. Experiments on three distributed photovoltaic sites show that the proposed method achieves R^2 values of 0.874, 0.898, and 0.766 and RMSE values of 5.603, 9.854, and 6.423kW, respectively, with a FAR of 0 across all sites. Compared with the best-performing baseline at each site, it reduces RMSE by up to 35.8% under normal conditions and by up to 37.7% under complex weather; even in complex-weather scenarios, it maintains R^2 values of 0.679--0.801 and a FAR of 0. These results demonstrate improved accuracy, traceability to physically meaningful reference quantities, and robustness under challenging weather, with broader relevance to measurement-informed monitoring, grid operation, and energy management.