High-Fidelity Full-Sky Video Prediction for Photovoltaic Ramp Event Forecasting

A new paper from researchers Siyuan Wang and Fengqi You presents a dual-model AI framework that tackles one of solar energy's trickiest problems: predicting rapid power fluctuations caused by passing clouds. The system combines PhyDiffNet, which generates high-fidelity future sky video frames by modeling the slow but chaotic motion of clouds, with RaPVFormer, a ramp-aware PV output forecasting model. Together they can predict photovoltaic ramp events—sudden changes in solar generation—up to 16 minutes in advance at a one-minute temporal resolution. By capturing fine-grained spatiotemporal cloud patterns, the framework delivers state-of-the-art performance in both video quality and power output forecasting.

Quantitative evaluations show a 10% improvement in the Critical Success Index (CSI) for ramp detection compared to existing methods. The model also provides interpretability through attention visualization, highlighting which cloud occlusion regions most influence irradiance variability. This enables grid operators to understand why a forecast is issued, building trust in AI-driven decisions. The work demonstrates that multimodal sensing—combining full-sky imagery with time-series solar data—can significantly improve ultra-short-term solar forecasting. For power grids with high solar penetration, such accurate early warnings could reduce the need for spinning reserve capacity and support more reliable renewable energy integration.