Behavioral Diagnosis on Individual Electricity Consumption: Formulation Using a Neural Network Based on Adaptive Resonance Theory

This research aims to study the daily consumption behavior of individual customers connected to the electricity distribution network and, extending it to longer periods, seek evidence of fraud, classified as non-technical losses. It should be noted that current Brazilian legislation authorizes distribution companies to pass non-technical losses on to electricity tariffs, consequently increasing the tariff for consumers who comply with their contractual obligations. In contrast to this practice, this research aims to develop a system for studying consumer behavior collaboratively and in complement to existing techniques, thereby mitigating or eliminating these losses. To achieve this objective, we propose the development of an inference system based on ANNs from the adaptive resonance theory (ART) family of [1] and Grossberg [2, 3]. Specifically, a Fuzzy-ART network, known for its ability to learn reliably and in real time, was employed. The customer consumption data used to develop this detection system comes from real customers of the Commission for Energy Regulation (CER) of Ireland, utilizing data from only one year to extract different consumption patterns across various seasons. Each sample, or input vector, corresponds to a customer's daily consumption in 30-minute intervals, allowing for the capture of information about the customer at different times of the day. Given the difficulty of obtaining real data, seven types of fraud were generated to represent, as closely as possible, the various types of fraudsters that might be encountered in real life. To avoid biasing the model due to the typical predominance of benign data, the database was balanced, consisting of 3,500 days of benign customer data and 3,500 days of fraudulent customer data.