LAUSR

Predicting hurricane power outages facilitates disaster response decision-making by electric power utilities as well as other organizations of critical importance to society. Predictive models can be built on the basis of statistical learning methods that use data from past hurricanes to capture the effects of climatological, geographical, and environmental variables on the power systems. When the dataset is largely zero-inflated, as power outage datasets often are, classical data mining methods that are based on a relatively balanced number of zeros and non-zeros may fail. General accuracy evaluation metrics also become misleading because they focus on the prevalent zero-valued responses in the dataset. We develop a new framework that operates in three stages by separating the prediction of whether or not power outages will occur from the number of customers without power. In the first stage, the zero-inflation problem is handled via a series of binary classifications. In the second stage, the severity of outages is predicted leveraging clustering techniques. In the final stage, regression models estimate the number of customers without power. We introduce a weighted accuracy metric and investigate its benefits over mean absolute error. We validate the models with data from hurricanes Dennis (2005), Ivan (2004), and Katrina (2005), and then predict power outages associated with hurricanes Matthew (2016) and Irma (2017) in the central Gulf region. The results demonstrate improvement over the traditional approaches in the context of power outage prediction.