Machine Learning Modelling for Predicting Roadside Forest Risk on Distribution Powerlines
Abstract
Roadside tree failures are accountable for nearly 90% of storm-related outages in the forested Northeastern US. Vegetation management is considered as an indispensable intervention mechanism to reduce vegetation risk on utility infrastructure. Optimal implementation of tree trimming practices and grid hardening mechanisms largely center on accurate information on physical structure and health condition of roadside vegetation. The goal of this study is to use machine learning techniques to understand how local environmental variables, vegetation management practices, and infrastructure settings affect tree-related power outages at a finer spatial scale. Using the Random Forest (RF) algorithm, we developed and compared five candidate vegetation risk models (VRM). Each model consisted of different combinations of variables representing vegetation, soil and terrain, tree trimming, and infrastructure information at device exposure zone level. Accuracy metrics suggested that the candidate model consisting of all variables as the best performing VRM with an AUC-ROC of 0.8320. We observed a significant (p<0.05) reduction of accuracies when candidate models lack vegetation and infrastructure related variables. Based on the best performing model (i.e., all environmental variables included), LiDAR-derived proximity pixel variables, primary overhead length, median height, and wire properties (covered or not) reported as the highest importance variables. Soil and terrain variables exhibited a marginal contribution to the overall model performance. The findings of this study could guide utilities in prioritizing treatment areas for conducting tree trimming and hazard tree removal programs. Optimal targeting of vegetation management programs would greatly benefit both broad-scale ecological and economical productivity by reducing risk to infrastructure and improving roadside tree health.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMGC42T0953W