A two-stage Machine Learning based approach for Power Outage Prediction during High-Wind events

Extreme weather events, such as hurricanes and typhoons, are expected to increase both in frequency and severity in the next decades as a symptom of climate change. Such events pose significant challenges for electrical utilities, often resulting in severe power outages and high operational costs. Power outage forecasts can help utilities to prepare better for such events, by localizing critical areas and plan post-event restoration more effectively. Predictive analytics based on machine learning (ML) can help to get an understanding of the disaster extent prior to the event, and support utilities in crew planning. In this paper, we propose a two-stage machine learning approach for power outage prediction during high-wind events: the first stage comprises a so-called "online event detector" which identifies the start of high-wind events in real-time. The second stage comprises a ML-based model that is tailored to power outage prediction during high-wind events. The method was evaluated on multiple high-wind events in the U.S. and demonstrated strong predictive performance across regions.