Forecasting space weather risks on power grids
Microsoft Rohan Kannan
Microsoft Research built a model that forecasts space-weather risk for 66,935 U.S. substations. It can flag likely trouble 30 to 60 minutes ahead, before a storm turns into grid damage.
Based on reporting by Microsoft, Rohan Kannan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft Research says it has built a machine-learning pipeline that turns solar-wind data into location-specific space-weather risk estimates for 66,935 substations in the continental United States. The idea is simple enough to say and hard enough to do: don’t just know a geomagnetic storm is coming, know where it’s likely to bite first.
The system ties together forecasted Auroral Electrojet and Disturbance Storm Time indices, solar-wind measurements from the L1 Lagrange point, local latitude, geology, ground conductivity, and grid-infrastructure data. It then runs those inputs through a gradient-boosting model to estimate dB/dt, the rate of magnetic-field change linked to geomagnetically induced currents. From there, it produces site-level risk estimates and rolls them up into a continental view.
There’s a practical reason for all that plumbing. During the May 2024 geomagnetic storm, utilities across North America were already bracing for impacts as auroras pushed far beyond their usual range. The source notes that storms like this can degrade GPS accuracy, affect satellite operations, and create current flows in transmission networks that raise operational risk. The point of the model is to give operators 30 to 60 minutes of warning before a specific risk shows up.
The pipeline is split into three stages. First, it forecasts AE and Dst from solar-wind observations and assembles geology and location features for each substation. Then it uses those inputs to estimate geomagnetic change. Finally, it converts that into location-specific risk and a continental assessment. Microsoft says a system of 50 AI agents helped explore features, validation strategies, and model settings, and that only public data sources were used, including NASA OMNI, NASA-aggregated Kyoto World Data Center data, INTERMAGNET, U.S. Geological Survey magnetometer observations, and GridSFM-derived grid data.
On the evaluation side, the AE predictor covered nearly the full observed range of activity from the 2020-2026 period. The Dst predictor outperformed the Burton equation on 62.2% of individual hours during the most active periods, and combining AE forecasts with the larger system improved severe-event detection by 1.2 percentage points. For the final risk stage, the model detected 76.5% of major events, 81.2% of severe events, and 64.1% of extreme events, with performance strongest at northern stations. Microsoft also says the full pipeline can produce estimates for all 66,935 substations in about 333 milliseconds.
The work isn’t claiming it’s grid-ready tomorrow. Microsoft says more validation with utilities and operational data would be needed before use in operations. But the direction is clear: space-weather forecasting is moving from a blunt warning to something much more local, which is exactly what a real grid would want.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI work: narrow, physical, and tied to a real failure mode instead of another demo about productivity. The useful trick here is not bigger hype, it’s smaller scope — a model that knows where the trouble is likely to matter. That’s the difference between science and slideware, and the grid has enough slideware already.
Read more about this at: Microsoft