Microsoft Research ML pipeline forecasts space-weather risk for 66,935 US substations

A machine learning pipeline developed at Microsoft Research forecasts space-weather risk for 66,935 substations in the continental United States, providing grid operators with 30 to 60 minutes of advance warning before potential impacts. The system combines solar-wind observations, forecasts of the Auroral Electrojet (AE) and Disturbance Storm Time (Dst) indices, and local geological data to estimate geomagnetically induced currents (GICs).

End-to-end prediction architecture

The pipeline operates in three stages to address the complexity of coupled space-weather systems. First, solar-wind measurements from the L1 Lagrange point generate forecasts for the AE and Dst indices, while geological conductivity and location features are assembled for each substation. Second, a gradient-boosting model combines these inputs to estimate dB/dt, the rate of magnetic-field change associated with GIC risk. Finally, predictions are converted into location-specific risk estimates and aggregated into a continental assessment. A system of 50 AI agents assisted in exploring features, validation strategies, and model configurations.

Performance and evaluation

During the 2020-2026 evaluation period, the model demonstrated significant improvements over baseline approaches. The AE predictor achieved a 410.2 nT RMSE, capturing extreme-event activity better than empirical baselines. The Dst predictor reached a 7.2 nT RMSE, outperforming the Burton equation on 62.2% of peak-activity hours and improving severe-event detection by 1.2 percentage points when integrated into the final system.

The GIC risk stage was evaluated against simple linear regression, as no direct industry benchmark exists. The system achieved the following detection rates:

• 76.5% for major events (≥10 nT/min)
• 81.2% for severe events (≥20 nT/min)
• 64.1% for extreme events (≥50 nT/min)

False-alarm rates increased with storm severity, reflecting a trade-off between missed events and cautious alerts. Performance was highest at northern stations where geomagnetic activity is strongest.

Operational implications

The pipeline generates estimates for all 66,935 substations in approximately 333 milliseconds, allowing for rapid scenario evaluation. This speed enables utilities to move from broad space-weather warnings to targeted views of which locations warrant closer analysis. Potential applications include adjusting reactive-power reserves or temporarily reconfiguring network parts. Future work aims to extend forecast horizons beyond the current 30-60-minute window using temporal-transformer approaches and to scale the system internationally.