Abstract: Each year in the United States, weather-related power outages result in billions of dollars of restoration costs and economic losses. Utility companies, emergency management agencies, and other public and private entities affected by power outages attempt to anticipate and mitigate the effect of these outages by utilizing power outage prediction models. These models are typically developed for either a combination of weather events or specialized for specific weather events like tropical cyclones. Despite the fact that thunderstorms account for almost half of major power outage events, development of specialized models for thunderstorms is at an early stage. This study uses the random forest machine learning technique to develop specialized models for thunderstorm related power outage events. The models are trained using power outage data from 31 thunderstorm events along with 75 predictor variables that include forecast weather conditions and environmental variables that have been found to improve power outage prediction models in past research. Results showed modest model skill compared to baseline models. Variable importance measures showed that environmental variables had high importance and convective hazard probabilities issued by NOAA’s National Weather Service Storm Prediction Center (SPC) had low importance. This low importance of convective hazard probabilities potentially decreased model skill and we hypothesize that it is related to the spatial scale used in this study. Additionally, it is noted that the model has a tendency to underpredict outages in more intense thunderstorm events.
Authors: Stephen A. Shield, Steven M. Quiring, D. Brent McRoberts
Date: June 22, 2018