Wavelet based feature extraction and multiple classifiers for electricity fraud detection

Electricity consumer dishonesty is a serious problem faced by all utilities. Finding efficient measurements for detecting fraudulent energy usage has been an active research area. The most effective way is to use intelligent/smart electronic meters that make fraudulent activities more difficult and easily detectable. In this paper the authors propose a new automatic feature analysis method using wavelet techniques and combining multiple classifiers to identify fraud in electricity distribution networks. Based on the assumption that meter-reading data present abnormalities when fraud events occur, the feature extraction scheme is carried out in both time and wavelet domains and the combination of multiple classifiers is applied through a cross identification and a voting scheme. Simulation results prove the proposed method to be effective in electricity fraud identification. For a relatively small amount of data, the classification accuracy reaches 78% on the training dataset and 70% on the testing dataset.