Km-scale solar irradiance ensemble forecasting with generative deep learning

Rapid variability and intermittency in surface shortwave downward irradiance (SWDR) impact photovoltaic (PV) generation, challenging power system stability. Yet, existing SWDR prediction methods often cannot simultaneously deliver high-resolution, wide-area, and well-calibrated probabilistic prediction, limiting their use for uncertainty-aware PV power prediction and operational decisions such as regional coordination and reserve scheduling. Here, we develop GenSolar, a generative deep learning model for global probabilistic SWDR prediction at ~5 km and 10-min resolution. By mapping large-scale atmospheric circulation information into fine-scale SWDR, GenSolar enables high-resolution ensemble prediction, providing essential inputs for PV power probabilistic estimation. Compared with a numerical forecasting model (the Global Forecast System, GFS), a deterministic deep learning baseline (U-Net), and a generative baseline (a conditional generative adversarial network, CGAN), GenSolar achieves ~15.8% reduction in root mean square error (RMSE) and ~46.5% reduction in continuous ranked probability score (CRPS) compared to GFS in the short-term forecasting range (1–72 h), while showing ~27.5% and ~25.4% CRPS reductions compared to U-Net and CGAN, respectively. Additionally, GenSolar exhibits performance gain over both baselines, even in mountainous regions with complex topography and tropical areas characterized by strong convection. GenSolar provides probabilistic SWDR ensembles with the potential to support downstream PV power estimation and power system operational planning in rapidly growing solar energy systems.