Evaluating Data-Driven Detectors of Electricity Theft in Smart Grids

William Sanders, University of Illinois at Urbana-Champaign

Usage Details

Electricity theft is a billion-dollar problem faced by utilities around the world and current measures are ineffective against sophisticated theft attacks that compromise the integrity of smart meter communications. We are devising algorithms that detect such theft attacks, and that are based on mathematical techniques in statistics and machine learning. The goal is to detect and mitigate theft by identifying anomalies in consumption patterns of electricity consumers. We intend to use Blue Waters to evaluate the effectiveness of three algorithms that detect simulated anomalies. The best parameters for these algorithms need to be found using scanning techniques, and they need to account for a wide range of attack parameters that produce anomalies. In addition, the simulations need to evaluate the effectiveness with which the detectors are able to work for different types of consumption patterns. Our evaluation is based on a large dataset obtained from a real smart meter deployment.