The Pulse
Microsoft Maps Space-Weather Risk Across 66,935 Substations
Microsoft Research describes a machine-learning pipeline that estimates space-weather risk 30 to 60 minutes ahead for 66,935 U.S. substations. In its evaluation, adding Dst forecasts alongside AE forecasts improved severe-event detection by
AI.info Team · Written with AI ·
“During my summer internship at Microsoft Research, I developed a machine learning system that forecasts space-weather risk across 66,935 substations in the continental United States.”
Rohan Kannan, intern, Microsoft Research
Microsoft Research published the system on September 30, 2026, describing a pipeline designed to estimate where geomagnetic activity could pose risks to U.S. power infrastructure. Its forecasts look 30 to 60 minutes ahead and cover 66,935 substations in the contiguous United States. The project aims to move beyond warnings that a solar storm is approaching and estimate which locations may face stronger ground-level effects.
From solar wind to substation risk
The pipeline starts with solar-wind measurements from the L1 Lagrange point, between Earth and the Sun. It uses those inputs to forecast two geomagnetic indices: the Auroral Electrojet, or AE, and Disturbance Storm Time, or Dst. The system then combines forecast values with each substation’s location, local geological conductivity and grid information.
A gradient-boosting model estimates the rate of change in Earth’s magnetic field, known as dB/dt, which is associated with geomagnetically induced current risk. Those estimates feed location-specific risk scores and a broader assessment across the continental grid. Microsoft says the pipeline uses public data, including NASA and Kyoto World Data Center records, magnetometer observations from INTERMAGNET and the U.S. Geological Survey, and grid data derived from GridSFM.
The location detail matters because the same geomagnetic disturbance can have different effects in different places. Microsoft’s account says local geology, latitude and transmission-line characteristics shape exposure; resistive bedrock can contribute to stronger induced currents. The system’s output is therefore a set of geographically differentiated estimates, rather than one alert applied uniformly across the country.
Detection rates vary with storm severity
Microsoft evaluated the system over a 2020–2026 period. In its reported tests, it detected 76.5% of major events, defined as magnetic-field change of at least 10 nanoteslas per minute; 81.2% of severe events, at 20 nT/min or more; and 64.1% of extreme events, at 50 nT/min or more. False alarms rose with storm severity, and detection performance varied by latitude, with the strongest rates at northern stations.
The risk stage has no widely deployed operational system for a direct industry comparison, Microsoft says, so the team compared it with simple linear regression. For its geomagnetic forecasts, the model outperformed the Burton equation during 62.2% of individual hours in the most active periods. Adding Dst forecasts to AE forecasts improved severe-event detection in the end-to-end system by 1.2 percentage points.
Microsoft also reports that producing estimates for all 66,935 substations took about 333 milliseconds during measured inference. That speed allows the team to evaluate many scenarios quickly, but it does not establish how the system would perform as a live utility warning service.
A research result, not an operating alert
The map Microsoft presents is a demonstration under a representative major-storm scenario, not a record of an active event. The company says the model could help grid operators and planners identify locations for closer review and consider targeted steps, such as adjusting reactive-power reserves or temporarily reconfiguring parts of a network. Those are potential uses, not actions the system has been shown to trigger in utility operations.
Microsoft says validation with utilities and operational data is still needed before the forecasts could be used in grid operations. The project’s next steps include extending the forecast horizon beyond 30 to 60 minutes, adapting the method to other regions and grid structures, and adding asset-specific data to estimate risk at the transformer level. For now, the published result is a substation-level research pipeline whose strongest practical claim is that it can turn space-weather forecasts into fast, location-specific estimates for further review.