Microsoft AI predicts solar storm risks for US power grids

Microsoft Research has built an end-to-end machine learning pipeline to estimate localized geomagnetic risks across 66,935 substations in the continental United States. The system leverages solar-wind data from the L1 Lagrange point to forecast Auroral Electrojet (AE) and Disturbance Storm Time (Dst) indices. It then combines these forecasts with local geological conductivity, latitude, and grid-infrastructure data. A gradient-boosting model processes these inputs to calculate the rate of magnetic-field change, or dB/dt, which correlates with geomagnetically induced current (GIC) risks. The pipeline relies on public data sources, including NASA OMNI, Kyoto World Data Center, INTERMAGNET, the U.S. Geological Survey, and GridSFM-derived grid data. A team of 50 AI agents assisted in optimizing the pipeline's features and configurations.

During its 2020-2026 evaluation period, the system demonstrated high accuracy in predicting severe space-weather threats. The AE predictor achieved a root mean square error (RMSE) of 410.2 nT, while the Dst predictor achieved an RMSE of 7.2 nT, outperforming the traditional Burton equation on 62.2% of peak-activity hours. Integrating both predictors boosted severe storm detection by 1.2 percentage points. For GIC risks, the system successfully detected 76.5% of major events (at least 10 nT/min), 81.2% of severe events (at least 20 nT/min), and 64.1% of extreme events (at least 50 nT/min).

For grid operators, this tool changes how they prepare for solar storms, like the major geomagnetic event of May 2024. Instead of relying on broad, nationwide alerts, practitioners can receive location-specific risk estimates 30 to 60 minutes before an impact occurs. The pipeline performs inference for all 66,935 substations in approximately 333 milliseconds, enabling rapid real-time scenario analysis. This rapid warning allows utilities to proactively adjust reactive-power reserves or temporarily reconfigure transmission networks to protect critical infrastructure.