Community power outage prediction modeling for the Eastern United States

In the United States, weather-related power outages cost the economy tens of billions annually, and there has been an upward trend in billion-dollar disasters over the last two decades. Thus, it is of growing importance to be able to predict outages and understand local resilience. However, many outage prediction models rely on utility infrastructure and outage data, which can be difficult to obtain when a study domain covers many utility territories. This study demonstrates two gradient-boosting machine-learning models driven by utility-agnostic non-proprietary data, eliminating the need for utility-specific data, and allowing individuals or communities to build and use such models for emergency planning or vulnerability analysis. Further, the framework is novel for its ability to incorporate data from various ecoregions, utilize infrastructure proxy data, and provide outage predictions for a breadth of storm types over a large and scalable domain. In this study, vegetation, land cover, energy infrastructure proxy, and weather data are used as model inputs to evaluate 15,872 events across 17 states in the Eastern U.S., where an event is defined as a unique combination of geographic county and storm episode ID. The model predicting all storm types except thunderstorms was validated using 10-fold cross-validation where folds were split chronologically, and demonstrates an r-squared value between predicted and actual outages of 0.61. Similarly, the thunderstorm-only model demonstrates an r-squared of 0.31. For future work, the addition of flooding data may be considered as the r-squared for the various-storm-type model increases to 0.77 when data from New York and New Jersey for Hurricane Sandy are removed. Additionally, the framework demonstrated here can be used to create a real-time outage prediction forecasting tool for storm events, and can be used to analyze resilience at a county resolution under future climate scenarios.