FarSky Improves Intra-Hour Solar Forecasting with Generative AI

Researchers have developed FarSky, a novel generative forecasting framework designed to enhance intra-hour solar irradiance predictions. Accurate solar forecasting is crucial for the stable integration of photovoltaic (PV) power into electricity grids. Existing deep learning methods have improved accuracy but often struggle with deterministic predictions and anticipating sudden changes, known as ramp events. FarSky addresses these limitations by leveraging latent-space coupling to create task-aware representations from all-sky imager (ASI) data. The framework employs a multi-task autoencoder to learn a shared latent representation for both image reconstruction and irradiance estimation. A latent diffusion model then generates future latent states, conditioned on recent observations, from which solar irradiance forecasts are directly decoded. This generative approach inherently provides probabilistic forecasts through stochastic sampling. Evaluated against state-of-the-art methods using a multi-year ASI dataset, FarSky demonstrated superior overall deterministic and probabilistic forecasting performance, boosting forecast skill by up to 11 percentage points. Notably, it significantly improved ramp event detection, achieving F1-scores over 60%. These results highlight the potential of combining generative models with task-aware latent-space coupling for more reliable solar energy management.