Power Outage Model Uses 5 Data Streams to Predict at Census Tract Level

A new paper from researchers at (presumably) a U.S. university proposes a powerful two-stage hurdle model that predicts both the occurrence and severity of power outages at the census-tract level. The framework fuses five distinct data sources: 15-minute utility outage feeds, OpenMeteo hourly weather records, American Community Survey socioeconomic data, CDC social vulnerability indices, and GIS-derived vegetation coverage. By combining these streams, the model can attribute outage risks to specific demographic and environmental factors, offering a sensitivity analysis that helps identify which communities are most vulnerable.

The model was validated using a high-resolution dataset covering 290 census tracts in the Detroit metro area over 14 months, with a remarkable 15-minute temporal resolution. This allows it to capture dynamic changes during extreme events like storms or heatwaves. The work has been accepted for presentation at the 58th North American Power Symposium (NAPS). For utilities and emergency managers, this means a tool to pinpoint hyper-local outage risk hours in advance, potentially saving lives and reducing economic losses.