Identificação de perda não técnica de energia elétrica

Abstract

Due to changes in the electrical systems of many countries and increasing competitiveness, electric utilities have had to seek to improve financial and technical performance for greater productivity, efficiency and profitability. One of the ways to improve and maximize available energy is by reducing fraud and theft involving the power system. This percentage of energy that is delivered but not billed due to bad consumer faith is characterized as non-technical loss of electricity. To solve this problem, distribution system operators carry out inspections of consumers. However, it is impossible to inspect all customers. A lower cost alternative is the application of machine learning algorithms such as Optimal Path Forest, Support Vector Machine, Decision Tree, Outliers Detection, and Self-Organization Map. The intuition is to get the probability that the consumer is committing fraud. Currently, studies for non-technical loss identification using different evaluation metrics and no comparisons with traditional supervised or unsupervised algorithms. This paper proposes an experimental analysis of the optimal path forest, SVM, supervised decision tree and unsupervised optimal path algorithms. Thus, the objective is to compare the algorithms to identify the probabilities of a consumer committing fraud in the electricity system, using the same metrics and database in order to understand the behavior of the algorithms. The experiments demonstrated that it is possible to create a database with synthetic fraud labels, and for these experiments SVM was superior to the Optimal Path Forest and Decision Tree algorithms.