Power Theft Detection in Residential Distribution Grids using Dirichlet Process Mixture Modelling

In developing nations like India, distribution companies annually face a loss of billions of dollars due to electricity theft. Conventional manual inspection of power theft can be very time-consuming and ineffective. Hence, advanced methods need to be developed to accurately detect electricity theft. The technique proposed in this work is based upon irregular-scenario classification, feature extraction and clustering to identify suspicious customers. The proposed methodology makes use of Bayesian Non-Parametric Dirichlet Process Mixture Modelling (BNP-DPMM) which is an advanced clustering algorithm and eliminates the problem of having a fixed number of clusters. A rigorous analysis is carried out and the proposed technique successfully group irregular and regular customers in very distinct clusters. For verification, another analysis on actual Smart meter data is also carried out and the technique is able to again successfully categorize different types of customers in distinct clusters.