Abstract
Modern power grids depend on the Advanced Metering Infrastructure (AMI) for consumption monitoring, energy management and billing. However, AMIs are vulnerable to electricity theft cyber attacks due to addition of communication layer. Electricity theft is one of the major Non-Technical Losses (NTLs) in the electricity distribution systems that has become a global concern, recently. Although the machine learning techniques are widely used for Electricity Theft Detection (ETD) in literature, some significant challenges need to be address. (i) The consumption data is usually unlabeled, there should be proper method to label the data. (ii) The fair consumers significantly outnumber the fraudulent consumers, which negatively impacts the performance of classification algorithm. (iii) The performance of classifier must be validated using proper performance evaluation measures. In this paper, an enhanced ETD model is proposed that is an optimized classifier Differential Evaluation Random Under Sampling Boosting (DE-RUSBoost) is used for classification. Proposed classifier DE-RUSBoost is optimized using a meta-heuristic optimization algorithm named Differential Evaluation (DE). The proposed method is evaluated on a real-world dataset, i.e., State Grid Corporation of China (SGCC) datasets. DE-RUSBoost achieves the highest accuracy of 96% and low false detection rate of 0.004. The proposed method outperforms its counterparts in terms of accuracy and false detection rate.