Granular Spatio-Temporal Grid-based Machine Learning for Forced Outage Prediction in Electrical Networks

Abstract

We developed a Granular Grid-based Outage Prediction Model, GG-OPM, that utilizes outage dependencies between electric grid substations and related feeders, benefiting from more granular weather data to improve outage prediction in the distribution network. The model incorporates a spatial and temporal module and tunes the hyperparameters to allow selecting components that optimize the task performance. Sixteen levels of spatio-temporal granularities were tested, using a large real-life dataset from a major Texas utility, which consisted of six years of historical outage events. The results obtained indicate an improvement of up to 50 times in F1 score over an earlier developed baseline for the most challenging prediction task, and an overall improvement across all levels of temporal and spatial granularity, providing solid grounds for utilizing the method in real-life outage prediction scenarios.