An intern's AI pipeline warns of grid-hitting solar storms up to an hour early

Microsoft Research says a machine learning pipeline can give utilities a half hour to an hour of notice ahead of a geomagnetic storm, pointing to the substations across the continental United States that face the most risk. In its own evaluation it caught nearly 80% of major space-weather events, but the company says it still needs validation with utilities.

Three stages, one forecast

The system, described in a company blog post by Rohan Kannan, who is listed as an intern at Microsoft Research, works in three steps.

First, it uses measurements from the L1 Lagrange point, a spot between Earth and the Sun where spacecraft watch the solar wind, to forecast two standard geomagnetic indices.

Second, a gradient-boosting model combines those forecasts with features of each substation, including latitude, geology and conductivity. It estimates how fast the local magnetic field is changing, which is linked to geomagnetically induced currents in power lines.

Third, the local estimates are rolled up into a continental view. According to Microsoft, one run scores all 66,935 substations in about 333 milliseconds.

The numbers, and their limits

By Microsoft's account, the system catches 76.5% of major events, 81.2% of severe ones and 64.1% of extreme ones, with severity defined by how quickly the magnetic field changes. The extreme tier is the worst performer, and the company notes that false alarms rise as storms get stronger.

Performance also varies by place. Northern stations show the highest detection rates.

The post says the training data came from public sources, including NASA OMNI, the Kyoto World Data Center, INTERMAGNET and U.S. Geological Survey magnetometers.

Fifty agents in the loop

One detail stands out. Microsoft says 50 AI agents were used to explore features, validation approaches and configurations across the pipeline. The humans set the problem. The agents did much of the searching.

Microsoft says the system needs further validation with utilities before it can be used operationally, and that the 30 to 60 minute warning window could be extended. Nobody has yet shown that grid operators would act on its alerts, or how many false alarms they would tolerate before ignoring it.