Data-Driven Detection Of Low-Current Arcing Events In Power Distribution Systems

Wildfires caused by electric equipment have become a challenge for electricity distribution operators and utilities in vulnerable regions, as witnessed by the recent catastrophic cases. Part of the challenge in preventing such events is lack of effective ways for monitoring equipment condition that may produce arcing and sparks. In the meantime, high-resolution, high-fidelity sensor measurements can be used to detect unique signatures of equipment malfunction and anomalies such as arcing faults that can potentially cause outages and wildfires. However, even with high-speed measurement data, low-current arcing events are notoriously difficult to detect due to their short bursts of duration and low amplitudes. In this paper, we propose a combination of unsupervised and supervised classification framework to detect anomalous events such as voltage regulation, overcurrent, fuse open etc., along with more instantaneous low-amplitude current arcing events, using phasor measurements. While the difficulty the proposed supervised learning algorithm evaluates the probability of detected low-current arcing events based on an existing library of labeled arcing signatures.