A Quasi-Binomial Regression Model for Hurricane-Induced Power Outages during Early Warning

Abstract

Hurricanes can cause devastating damage to overhead distribution lines leading to large power outages in electric grids. Power outage prediction models can help utilities to plan for an expedited power recovery by identifying the extent of power disruptions before the arrival of a hurricane. These models often use multiple input parameters, including early warning forecasts of hurricane characteristics, environmental data, power system details, and demographic information. We propose a quasi-binomial regression model to advance power outage models and overcome their existing limitations, such as unbounded outage predictions, limited extrapolation, and high uncertainties at low and high winds. This paper shows that the quasi-binomial model allows us to better capture the mechanics of power system failures due to hurricanes. We fitted our model to power outage data across 2,322 cities for four historical hurricanes: Harvey (2017), Michael (2018), Isaias (2020), and Ida (2021). We validated our model for the outages in Florida during Hurricane Ian (2022). The quasi-binomial model outperformed existing random forest and negative binomial regression models with a 7% error versus 50% and 76%, respectively. To demonstrate the quasi-binomial model’s good performance more comprehensively, we also tested a new beta regression model for outages. We show the quasi-binomial model had a smaller cross-validation root-mean squared error of 0.23 compared with 0.28 for the beta model. Finally, we show that our model also captures that grids with more redundant components can be more resilient to hurricane-caused outages. Thus, our proposed quasi-binomial model advances the state of the art for power outage predictions.