Predicting Power Outages from Ice Storms Using Machine Learning Models with SHAP Interpretability

Ice storms pose significant challenges, causing extensive damage and widespread power outages. Accurate prediction of outages during such events and their impact on distribution infrastructure is critical for mitigation efforts. This study presents a machine learning-based approach to predict the likelihood of distribution line power outages due to ice storms. Utilizing a comprehensive Ice Storm dataset from the cities of Hope and Mayville of North Dakota, containing meteorological and power outage information, features such as temperature, dew point, humidity, wind gust, wind speed, wind direction and precipitation metrics are used in the dataset. Various machine learning (ML) algorithms, including Random Forest Classifier (RFC), Logistic Regression Classifier (LRC), Support Vector Machine Classifier (SVC), and Gradient Boosting Classifier (GBC), are used to develop predictive models. The preliminary results indicate that RFC has the highest precision of 1.000. Feature importance using Shapley additive explanations (SHAP) analysis highlights wind gust, and wind direction as critical variables. This work can aid utility companies and emergency management agencies, enabling proactive measures to reduce the impact of ice storms.