Microsoft AI Forecasts Space Weather Risk Across 66,935 U.S. Substations
Microsoft Research's AI forecasts geomagnetic risk for 66,935 U.S. substations 30 to 60 minutes ahead, detecting 76.5% of major events in 2020 to 2026 tests.
Summary
Microsoft Research intern Rohan Kannan built a machine learning pipeline that uses L1 solar wind measurements, forecast Auroral Electrojet and Disturbance Storm Time indices, physics constraints, latitude, geology, ground conductivity and grid data to estimate dB/dt, a proxy for geomagnetically induced current risk, at 66,935 substations in the continental United States. It provides location-specific warnings 30 to 60 minutes ahead and aggregates them into a continental assessment. The work follows the May 2024 geomagnetic storm, which extended auroras, prompted utility preparations, degraded GPS and satellite operations, and affected farming in North America and Europe.
The three-stage system forecasts AE and Dst, assembles location and conductivity features, then uses gradient boosting to estimate and map risk. Fifty AI agents explored features, validation and configurations using public NASA OMNI, NASA-aggregated Kyoto World Data Center, INTERMAGNET, U.S. Geological Survey and GridSFM-derived data. In 2020 to 2026 tests, AE achieved 410.2 nT RMSE, covered nearly the observed range and beat empirical and solar-wind-only baselines. Dst achieved 7.2 nT RMSE, beat the Burton equation during 62.2% of peak-activity hours and raised severe-event detection 1.2 percentage points. Against linear regression, detection reached 76.5% for major events at 10 nT/min or above, 81.2% for severe events at 20 or above, and 64.1% for extreme events at 50 or above. False alarms rose with severity, while northern stations performed best.
All substations were scored in about 333 milliseconds. Targeted warnings could support engineering reviews, reactive-power adjustments and temporary network reconfiguration, but utility and operational-data validation is required because no widely deployed direct benchmark exists. Next steps include longer forecasts using temporal transformers, international expansion, operational integration and transformer-level estimates. Combining the pipeline with Microsoft Research's GridSFM could link hazard forecasts, grid topology and AC power-flow analysis.
Positives
- 30 to 60 minutes of advance warning could give grid operators time to review assets and take targeted protective action.
- 81.2% of severe events at 20 nT/min or above were detected during the 2020 to 2026 evaluation.
- 62.2% of peak-activity hours saw the Dst model outperform the Burton equation.
- 333 milliseconds was sufficient to estimate risk across all 66,935 continental U.S. substations.
- 1.2 percentage points were added to severe-event detection when Dst forecasts were combined with AE forecasts.
- Location-specific modeling accounts for latitude and geological conductivity instead of issuing one nationwide alert.
Risks & concerns
- 64.1% detection for extreme events at 50 nT/min or above leaves more than one-third undetected.
- False-alarm rates increased with storm severity, creating a trade-off between cautious alerts and operational disruption.
- 30 to 60 minutes remains a short forecast window, prompting plans to test temporal-transformer models.
- No widely deployed operational benchmark exists for direct comparison of the final risk calculations.
- Utility and operational-data validation is still required before the system can support live grid operations.
- Continental U.S. coverage and substation-level estimates do not yet provide international or transformer-specific risk.