Flow matching improves sub-hourly solar forecasting from weather models

⚡ AI Executive Summary

Researchers have developed a machine-learning framework to convert hourly weather forecasts into high-resolution, probabilistic 15-minute solar irradiance predictions. The approach combines physical constraints with generative modeling to capture rapid cloud-induced fluctuations and ramps that matter for grid operations. For solar-heavy grids, accurate sub-hourly forecasting is critical to reserve scheduling, battery dispatch, and voltage stability—yet operational weather models typically run at hourly resolution, masking the variability that challenges operators on high-penetration solar systems. This work bridges that gap by starting with a physically plausible high-resolution shape and then using probabilistic refinement to quantify forecast uncertainty. The framework shows significant error reduction compared to statistical and deep-learning baselines, better captures ramp events, and appears to work on new sites without retraining—important properties for real-world deployment across distributed solar portfolios.