Predicting Thunderstorm-Induced Power Outages to Support Utility Restoration

Abstract

Strong thunderstorms have substantial impacts on power systems, posing risks and inconveniences due to power outages. Developing models predicting the outages before a storm is a high priority to support restoration planning. However, most power outage data are zero-inflated, which results in some challenges in predictive modeling such as bias and inaccuracy. Power outages are also stochastic and there always exists irreducible variability in outage predictions. The goal is to develop models to overcome the challenges caused by zero-inflation and to accurately estimate power outages in terms of probability distributions to better address inherent stochasticity and uncertainty in predictions. This paper proposes a novel approach integrating mixture models with resampling and cost-sensitive learning for predicting the probability distribution for the number of outages. Validating the models using power outage data, we demonstrate that our approach offers more accurate point and probabilistic predictions compared to traditional approaches, better supporting utility restoration planning.