A model forecasts risk for 66,935 substations up to an hour before a solar storm hits

Microsoft Research has built a machine learning pipeline that turns solar-wind data into local risk estimates for 66,935 substations across the continental United States. For grid operators, the useful product is a 30-to-60-minute warning tied to specific locations, not the mere presence of AI.

Solar wind becomes a substation risk map in 333 milliseconds

The system first uses solar-wind measurements from the L1 Lagrange point to forecast the AE and Dst indices. A gradient-boosting model then combines those forecasts with latitude, geology, ground conductivity and transmission-grid data. It estimates the rate of magnetic-field change associated with the risk of geomagnetically induced currents.

The pipeline uses only public data and produces estimates for all 66,935 substations in about 333 milliseconds. The team also used 50 AI agents during development to explore features, validation strategies and model configurations. Those agents are not part of the operational forecasting path.

Local geology makes the warning more useful than a national colour code

A geomagnetic storm does not affect every location equally. Resistive bedrock can amplify ground currents, while latitude and transmission-line orientation change the exposure of individual assets. A local estimate could tell an operator where to review reactive-power reserves or temporarily reconfigure part of the network.

That is a more useful product layer than one warning for an entire continent. The model shortens the path from knowing that a storm is approaching to deciding which substations deserve attention within the next hour.

Detection falls for extreme events as false alarms increase

Across the 2020-to-2026 evaluation period, the system detected 76.5% of major events at a threshold of at least 10 nT/min, 81.2% of severe events at 20 nT/min and 64.1% of extreme events at 50 nT/min. False-alarm rates also increased with storm severity. Performance was strongest at northern stations, where geomagnetic activity is greatest.

There is no comparable, widely deployed operational system for the final risk calculation, so the authors used simple linear regression as the baseline. Microsoft explicitly says further validation with utilities and operational data is required before the system can support grid operations.

A live shift with grid operators will decide its value

The proposed next steps include longer forecast horizons, expansion beyond the United States, integration into operational decisions and a move from substation-level to transformer-level risk. Each step requires data that the current public datasets do not contain.

The most important signal will be a pilot alongside working dispatchers. It must show whether 30 to 60 minutes is enough for safe action and whether false alarms remain low enough that operators do not ignore the one warning that matters.

Lilith's verdict

A solar storm does not read a press release. A dispatcher needs half an hour, a map of specific substations and an alarm that does not scream whenever the sky twitches.