Electric theft detection using unsupervised machine learning-based matrix profile and K-means clustering technique

Abstract

Electric theft is the major issue faced by utility companies in different countries as it causes significant revenue losses and affects the power grid reliability. This paper presents a novel electric theft detection framework based on an unsupervised machine learning technique employing matrix profile and K-means clustering algorithm. The proposed framework is based on three stages to identify the fraudster consumers in a conventional electric consumption meter dataset acquired from Pakistan's power distribution company. Initially, the missing and inconsistent observations are filtered out from the acquired dataset. After that, the matrix profile from each consumer’s consumption profile is computed to identify the irregular and sudden changes present in them. Later, the K-means clustering algorithm is used on the datasets divided based on their computed matrix profile values in order to label each consumer into “Healthy” and Theft.” The developed framework is compared against the latest state of art machine learning algorithms and statistical-based outlier detection methods. The proposed technique achieved an accuracy of 93% and a detection rate of 91%, which is greater than all compared models.