The effect of weather forecast uncertainty on machine-learning outage prediction modeling

Abstract:

Weather-related power outages affect power reliability that electric utilities need to address using efficient restoration strategies. Predicting weather-related power outages in lead-times up to 5 days could help utility companies with their planning of crew and equipment allocation for faster and cost-efficient power restoration. Based on 273 historical outage events that occurred since 2016 for the Northeastern region of the United States, including Connecticut, Massachusetts and New Hampshire service territories of Eversource Energy and United Illuminating, we create a regional machine-learning outage prediction model (OPM) that exhibits mean absolute percentage error (MAPE) of 38%. The predictors for the model are weather (wind speed, gust, precipitation, etc.), land cover, tree canopy, vegetation, and utility infrastructure variables. Uncertainty of weather forecasts is a key component of the uncertainties in OPM-based outage forecasts. In this study, we investigate how the uncertainties of weather forecasting propagate to the uncertainties of outage prediction for lead-times ranging between 1 and 5 days, based on 200 weather-based outage events in the Northeastern United States. The study exhibits that the uncertainties of outage prediction and those of weather forecasting show decreasing trends from 5 days to 1 day lead-times; we show the complexity of interactions between weather characteristics and outage prediction, and nonlinear relationships between uncertainties of weather forecasting parameters and outages; at lead-time of four days NCRMSE of outage prediction magnifies more than any weather parameter.