Machine learning evaluation of storm-related transmission outage factors and risk

Abstract

This paper develops several machine learning models to evaluate the significance and correlation between variables related to vegetation, weather, infrastructure, the physical environment and storm-related outages in transmission networks, and acts as a proof of concept for machine learning to improve outage risk forecasting. The analysis covers the Eversource Energy transmission network territories in Connecticut, Western Massachusetts and New Hampshire. Random forest models are trained on 44 storms between May 2015 and December 2020, with 111 sustained outages incurred in total. Variable importance is performed to determine and remove insignificant variables, and for those remaining variables the correlations with transmission outages are plotted using partial dependence plots. The risk forecasting results more than doubled the number of transmission lines with outages classified in the top 20% of risk-score predictions for the best machine learning model relative to a baseline risk sorting algorithm, from 25 to 51. The variable importance and partial dependence plots demonstrate infrastructure related variables are significant predictors of transmission outage risk, with a magnitude of influence on outage levels on the same scale as significant vegetation and weather-related predictors.