On predicting clouds: AI, satellites, and the future of solar grid management

There is a particular kind of cloud that grid operators have learned to dread.

Not the storm clouds of a frontal system, those you can see coming. The trouble comes from things like fair-weather cumulus (Cumulus humilis), the small popcorn clouds that form on warm summer afternoons over nothing at all. Pleasant to look at. Genuinely hard to predict.

As a cumulus shadow drifts across a utility solar plant, production can drop by 100 megawatts and recover just minutes later. The event is fast and isolated, yet entirely real. It also happens to be invisible to the forecasting tools that most operators rely on.

Traditional numerical weather prediction runs too infrequently and resolves at too coarse a spatial scale to track individual cloud structures. The forecaster can’t see the clouds. The operator can’t predict the ramp. And so they carry expensive reserves they may not need or curtail solar they could have used. Both outcomes cost money that doesn’t have to be spent.

A new service from Xweather is built specifically to change that.

Forecast on a satellite clock

The Xweather solar irradiance and cloud-cover nowcasting service runs 24/7 in production, refreshing every few minutes and timed to geostationary satellite scans around the globe. Forecasts extend three hours into the future and are delivered as gridded irradiance fields and as point forecasts for individual assets, sized to plug directly into the scheduling, market-bidding, and energy-management workflows operators already run.

The service is deployed across multiple continents — the Americas, Europe, and Australia — using GOES in the western hemisphere and equivalent geostationary satellites elsewhere. This builds on the Xweather Solar Model 3, which produces the most accurate GHI-derived production estimates available for current solar generation. The nowcasting service takes that foundation into a harder problem: predicting how the sky above a plant will look not now, but in ninety minutes.

A combination of transformer and state-of-the-art AI diffusion models has been trained to evolve natural cloud fields that are combined with surface observations to produce new levels of forecast skill.

Where forecast accuracy becomes dollars

The new AI nowcasts deliver a 30% reduction in error over previous forecasts at lead times under 30 minutes.

The improvement widens as lead time grows, because AI nowcasting is strongest where traditional machine learning forecasts degrade fastest: predicting the arrival, density, and coverage of cloud structures the satellite is already tracking upstream of the plant.

For a grid operator or solar asset owner, that accuracy translates directly into cost savings. When the next thirty minutes are predictable, the operating reserve needed to cover sudden solar drops can be sized smaller. The avoided cost of carrying that capacity returns to the system.

For solar independent power producers bidding into intraday and real-time markets, tighter forecasts reduce imbalance settlements and improve realized prices on every cloudy day. For battery storage co-located with solar, accurate ramp prediction changes the economics of dispatch — batteries charged when they should be, discharged when they’re needed, rather than reacting after the fact. 

Utility solar is now one of the dominant sources of new generation capacity globally. The variability that comes with it isn’t a political or engineering failure — it’s a physics problem, and physics problems respond to better measurements and forecasting. Nowcasting won’t make clouds stop moving. But it can tell you, with meaningful accuracy, where they’re going and when they’ll arrive. At grid scale, that’s worth a great deal.