A Hoeffding D Statistic Approach for Detecting Electricity Theft

Electricity theft is a major financial loss for utility companies worldwide, costing them billions of dollars each year. Traditional methods of detecting theft either require large labelled datasets or can be expensive to implement, requiring the installation of expensive sensors. Our approach addresses these limitations by recognizing that the loss due to power theft and fraud is more intricately associated to the energy consumption measurements of the dishonest customers rather than that of normal consumers. Building on this understanding, we have framed the issue of power theft detection as a time- series correlation analysis problem. In our work, we have used Hoeffding's D method as a measure of association between ETL and power consumption of consumers. Hoeffding's D operates on a non-parametric basis, implying that it does not rely on any specific assumptions about the probability distribution inherent within the dataset. Hoeffding's D measure exhibits reduced sensitivity to outliers and serves as a more potent tool for detecting associations that are nonlinear and non-monotonic in nature. To validate our method, we conducted a series of experiments on real-world energy consumption data, encompassing a range of attack scenarios. The outcomes of these experiments demonstrate that our proposed technique notably enhances the accuracy of identifying deceitful customers when compared to alternative state-of-the-art methods based on correlation analysis.