Detection of anomalous electricity consumption by customers in a Russian region based on robust spectral clustering

Abstract

This paper presents a robust method for detecting anomalous electricity consumption using spectral clustering, tailored for real-world smart meter data from a Russian region. The study analyzes hourly readings from 1745 customers in the North Caucasus, specifically addressing challenges like significant data gaps and zero values. Our three-stage method involves constructing aggregated consumption profiles (daily, weekly, monthly), filtering out customers with negligible consumption, and applying spectral clustering with t-SNE visualization. This approach effectively categorizes customers into three distinct groups: those with stable load profiles, anomalous consumption, and insignificant consumption. The highest clustering quality was achieved using the filtered monthly consumption profile, with a Davies-Bouldin Index of 0.9741 and a Silhouette Coefficient of 0.4535. The results confirm the method's effectiveness for monitoring and analyzing electricity consumption. The developed universal method is applicable to any energy consumption dataset. Furthermore, we provide an open dataset of 27-month hourly consumption data from customers to support further research in electricity theft detection. The proposed method serves as an efficient preliminary screening tool to identify suspicious customers for subsequent expert verification.