Machine Learning Applications to Ice-Storm Power Outage Forecasting for Distribution System Resilience Enhancement

This study introduces a logistic regression-based machine learning (ML) classifier that can predict day-ahead electric power grid outages caused by ice storms. Data utilized in this study incorporate historical outage time and severity, extreme weather conditions, and physical grid locations and operating criteria. The proposed ML method applies the Majority Under-Sampling and Minority Over-sampling (MUMO) algorithm, which enhances the the forecasting accuracy of ice storm outages. The proposed MUMO algorithm is further combined with different feature selection methods, like Chi-Square and Minimum Redundancy Maximum Relevance (MRMR), and the ML model is trained with specific features which are highly correlated with outage status of each distribution feeder. We consider a practical utility case in our studies and our forecast results for ice storm outages are compared with those of Under Sampling (US) technique. The proposed ML model for ice storm outages, which could forecast feeder level day-ahead outages, would offer power system operators with a strategic advantage to execute precise and effective resilience measures for reducing outage occurrences and durations.