In an article published September 30, 2026, Microsoft Research describes a machine learning system for forecasting space weather risks to power grids. It produces location-specific estimates for 66,935 substations across the continental United States. The stated forecast window is 30–60 minutes ahead of potential impact.

From solar wind to a local estimate

The pipeline starts with solar-wind measurements from the L1 Lagrange point. It uses those measurements to forecast the Auroral Electrojet and Disturbance Storm Time, or Dst, indices. A gradient-boosting model then combines the forecasts with location-specific inputs to estimate dB/dt: the rate at which the magnetic field changes, associated here with the risk of geomagnetically induced currents.

These are distinct steps in the reported system. The index forecasts provide inputs; the dB/dt estimate adds a location-specific view of risk. An estimate for a substation should therefore be read as a model output, rather than as a measurement of current at that substation or a prediction of a particular piece of equipment failing. The reported 30–60 minutes describes how far ahead the system estimates potential impact, not how long the model takes to run.

Microsoft Research says the project used public sources including NASA OMNI, NASA-aggregated Kyoto World Data Center data, magnetometer observations from INTERMAGNET and the U.S. Geological Survey, and GridSFM-derived grid data. It also reports that a system of 50 AI agents helped explore features, validation strategies and model configurations. That description places the agents in the development process; the published forecasting pipeline is described through its inputs and models.

What the reported evaluation measures

During the most geomagnetically active periods in the 2020–2026 evaluation period, the machine learning Dst predictor outperformed the Burton equation on 62.2% of individual hours. The scope matters: this is a comparison of Dst prediction during the specified active periods. It is not a measured rate of successful grid interventions, nor a percentage of substations protected from an outcome.

The article reports detection rates at three dB/dt thresholds: 76.5% for major events at 10 nT/min or higher, 81.2% for severe events at 20 nT/min or higher, and 64.1% for extreme events at 50 nT/min or higher. Each percentage belongs to its named threshold. The figures show how the reported detector performed under those event definitions; they do not describe every possible operating condition with one universal accuracy figure.

Measured inference took approximately 333 milliseconds to generate estimates for all 66,935 substations. This result concerns computation time for the set of estimates. It should be kept separate from the 30–60-minute forecast horizon, which concerns the time before potential impact. Both numbers help characterize the research system, but answer different questions about it.

What remains to be established

Microsoft Research says further validation with utilities and operational data is needed before the system could be used in grid operations. That qualification is central to interpreting the results. The article presents a research pipeline, reported evaluation figures and a measured inference time; it does not establish an operational deployment or show that the forecasts have prevented equipment damage or outages.

For a reader assessing the announcement, the useful distinction is between a forecast of a risk indicator and a demonstrated operational result. The location-specific estimates may make the reported output relevant to individual substations, and the stated horizon describes when those estimates would be available relative to potential impact. Whether the estimates support decisions in actual grid operations is the question the article leaves for validation with utilities and operational data.