Detection of commercial losses in electric power distribution systems using data mining techniques

Non-technical or commercial losses are one of the main challenges faced by electricity distribution utilities, especially in developing countries. Frauds and energy theft, such as illegal tampering with meters and clandestine connections, are primarily responsible for commercial losses, as well as billing errors and faulty or broken meters. The most used way to identify such losses is the inspection of the consumer units. This procedure requires considerable allocation of financial resources, making it necessary to pre-select customers with unusual consumption behavior to optimize the detection of non-technical losses. In this paper, the clustering techniques K-Means and K-Medoids was used to find the groups of clients that have suspect energy consumption. Those methods were chosen because the unsupervised tools are not widely used in the commercial losses problem, besides being useful when the information about the results of the previous inspections is not available. The results showed that both techniques presented a performance similar to supervised methods reported in literature. However, there is a need to carefully define the input data and the number of clusters. The methods could be integrated with other techniques and more analyses should be done considering different unsupervised methods.