STARGPT: A spatiotemporal transformer foundation model for geostationary satellite-based solar radiation nowcasting

Accurate solar radiation nowcasting (4 h) is crucial for the efficient operation and management of photovoltaic (PV) power grids. However, the dynamic evolution of cloud systems and complex atmospheric conditions introduce substantial uncertainty, making high-precision short-term forecasting challenging. Although geostationary satellite-based approaches have attracted increasing attention, their predictive performance still falls short of the stringent accuracy requirements of the PV power market. To address this challenge, this study proposes a novel spatiotemporal foundation model framework: Super-Resolution Time-space Augmented Radiation Generative Pre-trained Transformer (STARGPT), for high-resolution regional Global Horizontal Irradiance (GHI) nowcasting using satellite imagery. Built upon the GPT-2 architecture, STARGPT integrates an enhanced ResNet-based super-resolution module, enabling joint spatiotemporal feature learning and improving the spatial granularity and accuracy of satellite-derived radiation forecasts. In addition, an Exponential Moving Variance Adjustment (EMVA) model is introduced to fuse model forecasts with real-time ground observations, further enhancing ground-level radiation prediction. Extensive experiments are conducted using satellite-retrieved irradiance from Fengyun-4A, with model performance evaluated across four representative months. STARGPT is benchmarked against the optical flow method and UNet model, with validation performed at six radiation observation stations in Hubei Province, China. Results demonstrate that STARGPT consistently outperforms the benchmark methods in both regional and ground-level forecasts, particularly for challenging lead times of up to four hours. Furthermore, STARGPT operates independently of Numerical Weather Prediction (NWP) inputs and specific satellite configurations, providing strong generalization capability across different observational settings. The forecasts also exhibit consistently high temporal correlation with ground observations, highlighting the model’s reliability and practical value for PV power forecasting and power grid stability management