Census Tract-Level Power Outage Prediction and Sensitivity Analysis During Extreme Events

Abstract

This paper develops a two-stage hurdle model for predicting power outage occurrence and severity at the census-tract level. The proposed framework is then used to assess the sensitivity of power outage to socioeconomic, demographic, and environmental factors during extreme events. Five heterogeneous data streams are integrated at the census tract level: 15-minute customer outage data, OpenMeteo hourly weather records, American Community Survey (ACS) socioeconomic indicators, Centers for Disease Control (CDC) social vulnerability indices (SVI), and Geographic Information System (GIS) derived vegetation coverage. The proposed framework is validated using a high-resolution power outage dataset covering 290 census tracts in the Detroit area over a period exceeding 14 months, with a temporal resolution of 15 minutes.