Forecasting space weather risks on power grids

A Microsoft Research blog describes an end-to-end system that turns forecast-time solar-wind information into location-specific risk scores for 66,935 continental U.S. substations. It combines AE and Dst index forecasts with latitude, geology, and ground conductivity, aiming to warn operators 30 to 60 minutes before a specific risk appears.

In evaluation, the pipeline detected nearly 80% of major space-weather events. The May 2024 geomagnetic storm — which stretched auroras and disrupted GPS/satellite operations affecting farms — is the motivating backdrop for why grid operators need place-specific timing, not just storm headlines.

We’re filing this under critical-infrastructure AI that doesn’t need a hype cycle.

Geomagnetic storms are rare until they aren’t; induced currents don’t care about your chatbot roadmap. Substation-level risk with half an hour of lead time is useful if utilities can act on it. Physics-informed features plus local geology beat a generic storm alert.

I’m watching whether this leaves internship-blog status for operational deployment with grid operators. Nearly 80% detection still means missed majors — operators need calibrated uncertainty, not a single score.

We’ll celebrate ML that protects transformers. This is the kind of applied research that makes “AI for science” concrete.

Microsoft Research blog on space-weather risk forecasting for U.S. power grids.