Modeling Fraud in Residential Power Usage

Understanding patterns of power usage is fundamental to the security goals of automation in the energy sector. Security aspects include data corruption via cyber-attacks, device tampering, or bypassing meter readings in energy theft. Although the mechanisms of abuse vary, persistent corruption generates patterns that statistically deviate from the designs of normal power usage. In this paper, we study power usage using both unsupervised and supervised techniques. The clustering algorithm creates energy profiles corresponding to a threat level hierarchy. We define a test in a proper statistical context by clearly specifying the fraud model (allowing simulation) to permit the analysis of the specificity and sensitivity of the model. The variance across meter readings changes over months, and the Gamma distribution fits nicely as a statistical model. Sensitivity analysis shows that detection accuracy is generally above 70% and identifies the threat level accurately. As expected, the accuracy varies over months and is lower during the months of summer. Detecting fraud in power usage is a significant problem and is an active area of research. It is essential to use the proper statistical measures to compare tests using diverse techniques.