Análise da contribuição de atributos derivados do histórico de consumo para a detecção de perdas não técnicas.

Abstract

Non-technical losses are mainly caused by the electricity irregular consumption due to fraud or theft. Their reduction is one of the main objectives of electricity distribution companies. Currently, companies have mainly used in loco inspections to identify irregular customers. However, these inspections are often associated with high costs and low effectiveness. Then, many companies have resorted to data mining techniques in order to increase assertiveness in the selection of irregular customers for inspections, using cadastral information such as class, supply voltage, type of connection and, mainly, historical consumption data. In this work, the contribution of attributes derived from the consumption electric energy history in non-technical losses detection using data mining techniques is analyzed. Therefore, new attributes are created from the consumption data, using seasonality characteristics, statistical information, monthly consumption variations, fall rates and consumption information in the frequency domain. In order to define the best attributes considering the original attributes and the attributes created subsequently, the attribute selection methods Correlation Based Feature Selection and Relief are used. Afterwards, the multilayer perceptron artificial neural networks algorithm is applied to classify the database clients between regular and irregular using the selected attributes. From the results, it was verified that the new attributes addition contributed to increase the artificial neural networks assertiveness, providing approximately a 10 percentage point gain, which can represent significant savings on the money spent by concessionaires with not assertive inspections. Therefore, it can be emphasized that the attributes analysis presented in this work can be used to reduce costs associated with non-technical losses detection by improving assertiveness in the potential irregular client’s identification.