Bridging the spatiotemporal scale gap: A physics-informed stacking ensemble for probabilistic photovoltaic power forecasting

To mitigate phase lags in photovoltaic (PV) power forecasting associated with numerical weather prediction scale mismatch, this study develops a physics-informed probabilistic Stacking framework. A theoretical power baseline anchors the diurnal cycle, while an empirical irradiance-gradient proxy serves as a non-anticipative surrogate for unresolved cloud-advection effects. Under the ERA5-Land reanalysis setting, the framework achieves a root mean square error (RMSE) of 2.293 MW on a 26.0 MW station. Operational diagnostics using European Center for Medium-Range Weather Forecasts day-ahead forecasts yield an RMSE of 2.472 MW and a low time-delay index of 0.14 steps, indicating robust phase tracking under forecast-boundary uncertainty. The Stacking architecture stabilizes ensemble variance and provides a low-noise foundation for probabilistic forecasting. By combining quantile regression with median-centered calibration and strict physical bounding, the framework achieves 80.75% coverage for the 80% interval with a prediction interval normalized average width of 0.1685. Dual-station evaluation further supports cross-site reproducibility under local retraining, showing that the physics-augmented feature system adapts across the evaluated temperate-monsoon and hot-desert settings while data-driven weights remain site-specific.