A weather model now forecasts clouds accurately for 12 hours using satellite data.

Predicting cloud cover matters for forecasting temperature, radiation, and solar-power operations. Short-range methods track observed clouds well at first, but fail as clouds form, dissolve, and change shape. Traditional physics models capture long-term atmospheric shifts, but they often miss the real cloud layout seen by satellites at the start. To solve this, researchers built CloudCast v2, a machine-learning model that predicts cloud cover up to 12 hours ahead. The model first learns how clouds move and change using historical European weather data. Then, a generative technique called conditional flow matchingconditional flow matchingA generative machine-learning method that transforms random noise into forecasts based on conditions. turns random noise into cloud forecasts, guided by current satellite observations and physics model inputs. CloudCast v2 cuts mean absolute error by 10% compared to CloudCast v1, proving observation-guided machine learning can reliably forecast beyond the usual 1-3-hour window.

Why it matters

This helps solar power operators plan better and improves weather predictions for agriculture and energy grids. It extends accurate cloud forecasts from a few hours to half a day.

Questions answered