Probabilistic Restoration Modeling of Wide-Area Power Outage

The timely restoration of electricity services following extreme weather events is crucial to meet customer energy resilience as well as for the economic and national security of the United States. Electricity restoration plans are needed to monitor multi-state power restoration operations, undertake resource planning, and analyze system vulnerabilities. However, these plans are proprietary to utility companies and not readily available to first responders and decision-makers. The purpose of the Restoration of Power Outage from Wide-area Severe Weather Disruptions (RePOWERD) project was to (i) determine which type of model - empirical, statistical, or probabilistic-most accurately predicts restoration times for distribution-level power outages caused by Category 2 or higher hurricanes, and (ii) identify the impact on restoration times of various predictor variables, such as power outage impact (i.e., customers impacted), storm characteristics, land-use patterns, and baseline customer density at county-service-area resolution. Seven models were developed for hurricanes that made landfalls from 2017 - 2022 along the Southeast region of the United States (Irma, Michael, Harvey, Laura, and Zeta). Comparing methods for predicting the time to restore power to 95% of impacted customers for these hurricanes revealed that: 1) outage magnitude (i.e., initial number of customers experiencing outages and their spatial distributions) is the strongest predictor of recovery time; 2) the performance of the log-linear regression model was similar to more complex, less interpretable models (e.g., accelerated failure time); and 3) the final log-linear regression model achieved strong overall performance, but it struggled with certain hurricanes (overall adjusted R2 of 0.6730, with a minimum of 0.4006 for Harvey and maximum of 0.8636 for Zeta). Using the log-linear regression model to forecast restoration time is viable, as all input data are publicly available prior to or at storm onset; however, the model reliability would benefit from expanding the scope of predictors and training data.