A Generative Approach for Surface Solar Irradiance Nowcasting

Accurate short-term forecasting of surface solar irradiance (SSI) is essential for renewable energy integration and trading considerations. In operations, it enables flexibility mechanisms, provides a hedge against rapid weather transitions and overall facilitates decision-making for intraday arbitrage. The ability to anticipate rapid ramp events in particular allows for the proactive management of renewable assets, maximizing capture prices and minimizing imbalance settlements. 

To this end, we present a new probabilistic framework for SSI forecasting over the contiguous United States (CONUS), developed within the NVIDIA Earth-2 platform. The framework builds upon Stormscope, NVIDIA's latest generative model for short-term Geostationary Operational Environmental Satellite (GOES) imagery forecasting, which serves as its core component. Stormscope predicts the spatio-temporal evolution of cloud fields, producing probabilistic satellite imagery sequences that capture atmospheric variability at high temporal resolution across eight spectral bands.

On top of this forecasting backbone, we apply a diagnostic diffusion model to estimate surface solar irradiance from GOES imagery using the National Solar Radiation Database (NSRDB) as reference data. This diagnostic model converts predicted satellite imagery into uncertainty-aware irradiance fields. Real-time inference is performed through Earth2Studio, providing continuous processing of live GOES data streams suitable for operational deployment. 

We evaluate the system’s performance against the High-Resolution Rapid Refresh SSI forecasts, demonstrating improved skill in capturing rapid irradiance fluctuations and cloud-driven variability at short lead times. The integration of Stormscope and the diagnostic diffusion model represents a significant expansion of TotalEnergies' global weather forecast capabilities, bridging the gap between real-time and medium-range weather forecast. This work advances the reliability of solar resource prediction and contributes to improving the profitability of renewable asset portfolios in increasingly volatile merchant markets.