Development of a Comprehensive Outage Prediction Model for Utilities

Under a changing climate, New York is experiencing an increase in the frequency of extreme weather events, many of which directly produce widespread disruptions in the power grid. The upward trend in extreme weather is leading to an increase in the frequency of power outages, resulting in growing economic losses to utilities, businesses (including the especially vulnerable digital economy), and residential customers, and is an increasing burden to utilities and municipalities. Individual power outages caused by extreme weather can result in utility costs exceeding $1 million per day, and local and regional economic costs exceeding tens of millions of dollars due to productivity losses, lost inventory, and transportation disruption. In New York State, the top ten power outages have all been weather-related, with power losses for these events affecting upwards of hundreds of thousands of customers.

Here, we present a comprehensive, state-scale, all-storm-type Outage Prediction Model (OPM) framework that enables forecasting of electrical power outages induced by different types of extreme weather including nor’easters, tropical cyclones, winter storms, high wind events, and thunderstorms. The OPM is a synthesis of advanced machine learning (ML) algorithms developed at the University of Connecticut (UConn) and the high resolution (1 km horizontal) Wind Extremes Forecast System (WEFS) developed by the State University of New York, University at Albany (UAlbany), and MESO, Inc. WEFS incorporates the Weather Research and Forecasting (WRF) modeling system optimized for forecasting extreme winds. The UConn OPM uses weather data, land cover, vegetation (e.g., leaf area index) and electric distribution network infrastructure data to predict power outages using ML techniques. Input weather data is obtained from the WRF simulations (analysis and forecasts) in the form of summary weather variables (e.g., maximum and sustained winds, duration variables, e.g. hours of sustained winds above established thresholds, etc.) derived from output data of the WRF simulations. Infrastructure data describing assets in the overhead distribution network are provided by the electric utility. Environmental data are fed into the ML models, trained with a historical database of weather events and validated using a leave-one-storm-out cross-validation. The output of the machine learning models is optimized based on historical performance and displayed as outage maps available to utility operators and in-house meteorologists. The new OPM is trained on unique observations from the 127 station New York State Mesonet (NYSM), with validation leveraging measurements from wind LiDARs and microwave radiometers at 17 sites comprising the profiler network. The system began real-time experimental operation in mid-2024. Here we present preliminary results on the OPM performance