FarSky AI model boosts solar forecast skill by 11 points

Solar forecasting is critical for grid stability as photovoltaic penetration grows, but current deep learning methods often produce deterministic predictions and struggle with sudden cloud-driven ramp events. FarSky, proposed by Yann Fabel and colleagues from DLR, RWTH Aachen, and CIEMAT, addresses this with a generative framework that captures uncertainty while maintaining accuracy. The architecture first trains a multi-task autoencoder to embed sky images into a latent space that jointly supports image reconstruction and irradiance estimation, forcing the representation to retain physically relevant cloud information. Then, a latent diffusion model samples future latent states conditioned on recent observations, yielding probabilistic irradiance forecasts through stochastic inference. This design decouples representation learning from forecasting, enabling task-aware features without sacrificing generative flexibility.

Evaluated on a multi-year all-sky imager dataset from Plataforma Solar de Almería in Spain, FarSky was benchmarked against persistence baselines, end-to-end deep learning models, and other generative approaches across two independent test sets. It achieved the best overall deterministic and probabilistic skill, improving forecast skill by up to 11 percentage points over existing methods. Notably, FarSky substantially enhanced ramp event detection, reaching F1-scores above 60%, a significant improvement for anticipating rapid irradiance changes that threaten grid operations. These results suggest that combining generative models with task-aware latent-space coupling is a promising path for high-resolution solar forecasting, potentially enabling more reliable integration of solar power into modern electricity grids.