SkyEvo: Generative probabilistic solar forecasting with sky image evolution

Accurate short-term solar irradiance forecasting is essential for the reliable operation of photovoltaic (PV) systems. However, probabilistic methods often estimate output uncertainty without explicitly representing the multiple possible evolutions of future cloud fields, which limits their ability to characterize rapidly changing solar conditions. This study presents SkyEvo, a two-stage probabilistic forecasting framework that integrates stochastic sky-image generation with multi-horizon irradiance prediction. In the first stage, a Generative Sky Image Evolution (GSE) module based on variational latent dynamics generates multiple future sky-image trajectories from historical observations, thereby representing uncertainty in future visual conditions. In the second stage, each generated trajectory is combined with irradiance or PV-power measurements through a multimodal forecasting module, producing an empirical predictive distribution from 15 stochastic samples. SkyEvo is evaluated over 10–60 min forecasting horizons on two datasets: the Folsom dataset for global horizontal irradiance forecasting and the Stanford dataset for PV power forecasting. On the Folsom dataset, SkyEvo maintains PICP above 89% across all forecasting horizons, achieves a 10-min nRMSE of 14.51%, and provides stable interval widths and quantile calibration. The GSE module also achieves lower MSE and higher PSNR and UQI than VanillaVAE and a lightweight diffusion-based video forecasting baseline. Under direct Folsom-to-Stanford transfer without fine-tuning, SkyEvo achieves the highest PICP at five of the six forecasting horizons and the lowest CRPS at all horizons, while maintaining competitive prediction-interval sharpness. These results demonstrate that stochastic future sky evolution can provide forecast-relevant visual scenarios and support reliable sample-based probabilistic forecasting across different datasets and forecasting targets.