Detecção de Perdas não Técnicas em Sistema de Distribuição Empregando a Transformada Wavelet

Abstract

Non-technical losses account for 7,5% of the energy injected into the Brazilian electricity grid. That losses cannot be fully passed on from distributors to consumers, which impacts on the profit of the companies, but also impacts in the responsible consumer, since part of this loss is prorated among regular users. The development of non-technical loss detection methodologies is important since, by discovering fraud, there is the possibility of revenue recovery and mitigation of other consequences. In this paper, two specific algorithmic solutions are proposed, capable of acting as a tool for extracting intrinsic characteristics of consumption patterns, which allow the detection of behavior associated with fraud. To develop both, a database with monthly consumption information in kilowatt-hour was created. The first technique makes use of discrete wavelet transform, so wavelet families were evaluated in order to define which one is the most suitable for implementing the technique. The Haar wavelet proved to be promising for the database in question, allowing a rate of correct answers in the classification of commercial customers around 77% and 86% accuracy in detecting fraud in the same customers. The second technique had a 79% hit rate for commercial customer ratings and a 78% fraud detection rate. As highlighted throughout the text, the complementarity of results for each of the techniques motivated the development of both.