Outage prediction for extreme weather events based on machine learning
Lena Peter, Katarina Knezović, Muhammad Jawad, Neeraj Katiyar, Milos Subasic, Ashwin Shirsat · IET conference proceedings. · 2025
Extreme weather-related events, such as hurricanes or winter storms pose significant challenges to the reliable operation of the power distribution networks, resulting in prolonged power outages, substantial economic repercussions, and threats to human wellbeing. Electric utility companies strive to minimize the severity of outages caused by such events, thus paving the way to a smooth post-event restoration and ensuring uninterrupted service delivery. In that context, advanced predictive analytics can enhance the network's resilience since knowing the expected outages prior to the event would allow network operators to protect critical areas. In this paper, we propose a tree-based, probabilistic machine-learning approach for outage prediction prior to extreme weather events. The method predicts the proportion of customers without power for a specified service territory based on weather forecasts, such as wind speed, air pressure, and precipitation. The algorithm is flexible and can be used for different prediction horizons, from few hours up to several days ahead. We test and validate the algorithm based on selected evaluation metrics using historical data of extreme weather events (hurricanes, heavy rainfall, nor’easter, and snowstorms) and outages in different U.S. counties. The presented results show stable prediction performance with high accuracy across different regions and events of different intensities.