Impact of Historical Storms in Machine Learning Outage Prediction Model

Accurate and timely prediction of storm-caused power outages is critical to helping utility companies restore power outages quickly and reduce impact to customers. This paper investigates the significance of the training sample size and representativeness of historical storms in machine learning (ML) outage prediction model (OPM) and introduced a novel method to enhance the predictive accuracy of OPM. In this study, we separately explored the influence of a random sample size selection for historical extratropical storms and thunderstorms on the performance of OPM. We found that OPM was more robust when trained with larger training sample sizes of historical storm events and that OPM with small training set didn't generalize well to storm events that fall outside the dynamic range of explanatory variables. Furthermore, we present an innovative method for training OPM by conditioning the training set with storms of different outage ranges, was explored for historical extratropical events. It demonstrated that OPM performed better when storms the training set in OPM was more representative to the outages.