A physics-informed machine learning method for short-term solar radiation forecast with satellite-derived cloud optical depth

The variation of surface incident solar radiation (i.e., global horizontal irradiance, GHI) challenges the stable integration of solar energy, underscoring the necessity of accurate GHI forecasting. Although advances in machine learning have facilitated GHI forecasting using geostationary satellite observations, most methods remain purely data-driven, relying on historical GHI or its proxy variables (e.g., clearness index) as input, thereby lacking physical constraints and exhibiting poor interpretability. This study proposes a short-term (10-min – 4-h) GHI forecasting method with two steps. First, cloud optical depth (COD) and cloud mask from Geostationary Operational Environmental Satellite-16 (GOES-16) are input into deep learning models to forecast both variables, and the final COD forecasts are obtained by combining their predictions. Second, the forecasted COD, together with auxiliary variables (aerosol optical depth, water vapor, black-sky albedo, and solar zenith angle), is converted into GHI forecasts using an all-sky estimation model developed through radiative transfer simulations. Compared with purely data-driven methods, the proposed hybrid model explicitly accounts for multiple drivers of GHI, thereby enhancing interpretability and enforcing physical constraints. This leads to more reliable decision support, clearer attribution of influencing factors, and a stronger foundation for further model improvement. Within a 4-h lead time, the root mean squared error (RMSE) and mean bias error (MBE) of GHI forecasts ranged from 96.7–150.3 W/m2 and −6.9–0.2 W/m2, respectively. When the model trained on GOES-16 data mainly over the continental U.S. was directly applied to Himawari-8 data over East Asia, validation results indicated an RMSE of 137.0–159.2 W/m2 for GHI forecasts, demonstrating the model's potential for cross-regional applicability. Additionally, incorporating cloud mask predictions mitigates the overprediction of low COD (≤10), reducing MBE of GHI forecasts by 14.0–33.7 W/m2. This study proposes a partly interpretable GHI forecasting method that facilitates solar energy's large-scale and efficient utilization.