Evaluation of classifiers for non-technical loss identification in electric power systems

Abstract

Recent works have proposed different machine learning classifiers for non-technical loss identification. However, it is still not clear what is the level of suitability of different classifiers and there is a lack of comparison between the proposed approaches in bibliography. Another problem is there are only few works that presents real field inspections results, so, it is not possible to evaluate the reliability of classifiers. To fill up these gaps is a valuable contribution given by the present paper. In the study, 23 different classifiers were evaluated regarding its performance, runtime and reliability. The classifiers are derivate from the 10 most used algorithms in bibliography plus two ensemble algorithms that have yet not been used for non-technical loss identification. The F1-score was utilized as performance parameter and was obtained from a cross-validation process using a dataset of 261,489 consumers from a Brazilian power utility. The classifiers were also used to identify non-technical loss in new consumers and about 1,400 real field inspection were executed. Furthermore, different approaches were also evaluated. One of them consists in perform a preliminary data clustering before the classification process and another is a voting criterion, combing many classifiers’ results. From the obtained results it was possible to conclude that classifiers based on ensemble methods are the most suitable for non-technical loss identification. The Gradient Boosted Three presented a F1-score of 0.45 and the Rotation Forest presented a precision of 66.50% in real field inspections, deviating only 6.86% of the simulated result.