Microsoft Research’s Experimental AI Forecasts Space-Weather Grid Risk 30–60 Minutes Ahead

Power-grid operators could get a more targeted warning of space-weather trouble: which substations face elevated risk, rather than just whether a storm is approaching. In a September 30 research post, Microsoft Research intern Rohan Kannan described an experimental machine-learning pipeline that estimates risk at 66,935 substations across the continental United States, 30–60 minutes ahead of potential impact. Microsoft says further validation with utilities and operational data is needed before the system can be used in grid operations.

Extreme space weather can induce currents in transmission networks, damage equipment and increase operating risk. But a storm’s effects are not uniform. Latitude, local geology and transmission-line orientation influence exposure. Regions with more resistive bedrock can experience stronger geomagnetically induced currents than places with more conductive geology. That makes location a necessary part of the forecast, not simply a detail added to a national warning.

The pipeline starts with solar-wind measurements collected at the L1 Lagrange point. It uses those measurements to forecast two indicators of geomagnetic activity: the Auroral Electrojet, or AE, index and the Disturbance Storm Time, or Dst, index. Those forecasts are combined with each substation’s location and geological conductivity, alongside grid-infrastructure data.

A gradient-boosting model then estimates how quickly the magnetic field will change at each location. That quantity, written as dB/dt, is associated with the risk of storm-induced currents. The final stage converts the predictions into local risk estimates and a continental assessment. The output is therefore an estimate of exposure—not a direct measurement of current flowing through a particular transformer.

The project used only public data, including NASA solar-wind and geomagnetic records, INTERMAGNET and U.S. Geological Survey magnetometer observations, and GridSFM-derived grid data. Its AE predictor was evaluated over 2020–2026 and produced forecasts spanning nearly the full observed range of activity, according to Microsoft, while outperforming several empirical solar-wind-based approaches.

Microsoft’s reported detection rates varied substantially by severity. The categories were defined by magnetic-field changes of at least 10, 20 and 50 nanoteslas per minute, respectively—not by observed equipment damage. Detection was highest for severe events and lowest for extreme ones. Northern stations also had the highest detection rates, where geomagnetic activity was strongest.

76.5%Major events

Detection rate for events at or above 10 nanoteslas per minute.

81.2%Severe events

Detection rate for events at or above 20 nanoteslas per minute.

64.1%Extreme events

Detection rate for events at or above 50 nanoteslas per minute.

False-alarm rates increased with storm severity, reflecting a tradeoff between missed events and cautious alerts. For the grid-risk stage, the researchers compared the model with simple linear regression. Microsoft says there is no equivalent widely deployed operational system that provides a direct industry benchmark for this calculation.

The Dst predictor had a separate comparison: it outperformed the Burton equation on 62.2% of individual hours during peak geomagnetic activity. Adding Dst forecasts to the AE-based pipeline improved severe-event detection by 1.2 percentage points. That is a measured contribution from a second storm signal, rather than evidence that every component delivers the same gain.

During measured inference, the pipeline generated estimates for all 66,935 substations in approximately 333 milliseconds. Its continental map illustrates how the same storm scenario can produce different local risk levels. The published map is a demonstration under a representative major-storm scenario, not a record of a live operational event.

Microsoft describes the potential use as helping utilities prioritize engineering review and consider targeted protective actions, such as adjusting reactive-power reserves or temporarily reconfiguring parts of the network. Its proposed next steps include testing forecasts within existing decision-making workflows before considering higher levels of automation. Longer forecast horizons and transformer-level estimates are also future research directions, not demonstrated capabilities of the current pipeline.