Análise da aplicação de máquinas de vetores de suporte na detecção de perdas não-técnicas em sistemas de distribuição de energia elétrica.

Abstract

Non-technical losses represent a serious problem in electric power distribution sector. To combat them, concessionaires use on-site inspections which represent a high financial cost to utilities and are inefficient. Therefore, the selection of suspect consumers carries great importance, so that new technologies based on Data Mining and Machine Learning have emerged. In this regard, Support Vector Machines have proved to be a very efficient method, but has not been studied in a concrete way. The purpose of this study was to analyze the application of Support Vector Machines to detect non-technical losses in electric power distribution systems. A database containing consumer information from 9177 consumers, provided by a distributor from the state of Paraíba, Brazil, was used to carry out a series of tests that identified the best strategy for separating training and test sets to be used, the best time period, proportion of non-technical losses in database, and the best type of input. The best results were obtained with the use of 10 folders – cross validation, with 36 months of consumption and proportion of nontechnical losses in the database of 50%. The best type of input was consumption information, with inspections success rates of up to 76.6% and accuracy of detecting irregularities of 63.32%, which are much better than the methods currently used by energy utilities. The application of Support Vector Machines is thus a viable alternative for the detection of non-technical losses in power distribution systems.