A Bayesian Prediction Framework of Weather Based Power Line Damages in the Northeast

This study aims to evaluate the predictability of damages to overhead power distribution lines from severe weather events in the New England area. During storms, trees and branches can come down and interact with power lines that results in significant interruptions to electricity distribution, causing major interruptions to residents and monetary losses to the utility company. In Connecticut, a densely forested state, severe winds and precipitation (in the form of rain and snow) from storms are key weather factors that challenge the power grid infrastructure vulnerability. Evaluating the local predictability of these impacts may aid local power utilities with crew allocation and preparedness during an event. A probabilistic approach to damage prediction caused by trees subjected to severe weather is being investigated in the region. This study specifically, explores the feasibility of applying Bayesian inversion technique to weather parameters by developing a damage decision tree composed of various meteorological and static parameters, like wind gust, precipitation (rain and snow accumulation and rates), high canopy forest density and tree trimming history for the power distribution lines. The resulting decision tree can be used as a Bayesian inversion database to predict the probability distribution of damages given a storm forecast. The Bayesian database is based on a historical data source provided by The Connecticut Light & Power Company (Connecticut's primary power utility) containing geographical information of trouble spots caused by thunderstorm and winter/snow-storm events; power line specifications and trimming history; and high-resolution model analysis of those storms. The analysis is based on a 2-sqkm model grid cropped over the state of Connecticut comprising a database of 3,307 pixels per storm. Each storm pixel is flagged to contain power line damages or no-damages. A total of 50 storm simulations is used to build the database. Pairs of probability density functions of damaged and non-damaged grid points are used to indicate a threshold value for each parameter. The several sets of combined threshold values are used as conditional probabilities. Using Bayesian inversion, the conditional probabilities are then related to the probability of damage occurrence per square area in a forecasted storm event.