A model forecasts grid space-weather risk

A Microsoft Research post dated September 30 describes work by intern Rohan Kannan: a machine-learning system that forecasts space-weather risk for 66,935 substations in the continental United States. It estimates how solar activity can affect Earth’s magnetic field and, from there, the power grid. The inputs are solar-wind observations, forecasts of the Auroral Electrojet (AE) and Disturbance Storm Time (Dst) indices, physics-informed constraints, local geological conductivity, and grid data. The output is a risk estimate for each location, 30 to 60 minutes ahead.

The forecast is hard because several systems are coupled. The solar wind changes quickly, its interaction with the magnetosphere is irregular, and the ground effect depends on local conditions. Resistive bedrock can see stronger geomagnetically induced currents than more conductive geology. Line orientation, latitude, and the power system itself also change how exposed a given asset is.

The pipeline has three stages. First, solar-wind measurements from the L1 Lagrange point forecast the AE and Dst indices, and geological conductivity and location features are assembled for each substation. Second, a gradient-boosting model combines those forecasts with the location inputs to estimate dB/dt, the rate of magnetic-field change tied to induced-current risk. Third, the predictions become location-specific risks and a continental assessment. About 50 AI agents helped explore features, validation strategies, and model configurations. Only public data were used, including NASA OMNI, NASA-aggregated Kyoto World Data Center data, INTERMAGNET and U.S. Geological Survey magnetometers, and grid data derived from GridSFM.

The AE forecast targets rare, high-intensity geomagnetic activity that drives infrastructure risk. Over the 2020–2026 evaluation period, the forecasts spanned nearly the full observed range of AE. Root-mean-square error was 410.2 nT, lower than the empirical and solar-wind-only baselines listed in the post. Dst root-mean-square error was 7.2 nT. During the most geomagnetically active hours of that period, the machine-learning model beat the Burton equation on 62.2 percent of individual hours, and its prediction range was wider than Burton-style approaches. Combined with the AE forecasts, severe-event detection in the end-to-end system improved by 1.2 percentage points.

The induced-current stage had no widely deployed operational system to use as a direct industry benchmark, so the model was compared with simple linear regression. Detection was 76.5 percent for major events (at least 10 nT per minute), 81.2 percent for severe events (at least 20), and 64.1 percent for extreme events (at least 50). False-alarm rates rose with storm severity. The author treats that as a tradeoff between missed events and cautious alerts, and does not give the false-alarm percentages. Detection varied by latitude and was highest at northern stations, where geomagnetic activity is strongest.

The last stage turns the forecast into a local dB/dt using each substation’s latitude and geology, rather than one alert for the whole United States. The map in the post is a demonstration under a representative major-storm scenario, not a record of a live operational event. In a measured run, the pipeline produced estimates for all 66,935 substations in about 333 milliseconds.

The author writes that earlier, location-specific information could help utilities prioritize engineering review and consider steps such as adjusting reactive-power reserves or temporarily reconfiguring part of the network. Further validation with utilities and operational data would be needed before the system could be used in grid operations. Later directions include extending the lead time past the current 30 to 60 minutes, adapting to regions with different geology and grids, connecting forecasts to existing human decisions before any higher level of automation, and moving from substations to individual transformers.