Microsoft Research AI predicts solar storm risks to US power grids

Microsoft Research has developed a machine learning system designed to predict the impact of solar storms on electrical infrastructure. The system provides location-specific risk estimates for 66,935 substations across the continental United States, offering a warning window of 30 to 60 minutes before potential damage occurs.

Unlike general space weather alerts, this model focuses on geomagnetically induced currents that can saturate transformers and cause long-term equipment failure. By combining solar-wind data, Auroral Electrojet (AE) and Disturbance Storm Time (Dst) forecasts with local geological conductivity and grid data, the pipeline aims to identify exactly where the grid is most vulnerable.

During its evaluation period, the system detected nearly 80% of major space-weather events. The model's detection rates vary by intensity: it achieved a 76.5% detection rate for events of 10 nT/min or more, and 81.2% for severe events at 20 nT/min. The system is calibrated to prioritize early warnings to prevent critical infrastructure damage.