RESEARCH · RESEARCH · #1290
Microsoft Research ML pipeline forecasts geomagnetic risk for 66,935 U.S. substations
Microsoft Research developed an end-to-end machine-learning pipeline that combines L1 solar-wind observations, AE and Dst forecasts, local geology and grid data to produce location-specific dB/dt and geomagnetically induced current (GIC) risk estimates for 66,935 continental U.S. substations with 30–60 minute lead time. In a 2020–2026 evaluation the system detected ~76.5% of major events (≥10 nT/min), ~81.2% of severe events (≥20 nT/min) and ~64.1% of extreme events (≥50 nT/min), outperformed some empirical baselines (including a Burton-equation comparison for Dst) and was developed using only public datasets.
KEY POINTS
- Microsoft Research developed an end-to-end machine-learning pipeline that combines L1 solar-wind observations, AE and Dst forecasts, local geology and grid data to produce location-specific dB/dt and geomagnetically induced current (GIC) risk estimates for 66,935 continental U.S.
- substations with 30–60 minute lead time.
- In a 2020–2026 evaluation the system detected ~76.5% of major events (≥10 nT/min), ~81.2% of severe events (≥20 nT/min) and ~64.1% of extreme events (≥50 nT/min), outperformed some empirical baselines (including a Burton-equation comparison for Dst) and was developed using only public datasets.
WHY IT MATTERS
Location-specific short-term forecasts of geomagnetic activity help grid operators target interventions to reduce damage from GICs and increase electrical-system resilience.