Electricity Theft Detection Using Deep Neural Networks: A Machine Learning Approach for Automated Anomaly Identification in Power Distribution Networks

Abstract

Power theft is a major problem facing power distribution industries around the world and thus leading to a massive loss of revenue, grid instability and increased cost of tariffs to the legitimate users. Traditional methods of detection, which rely on the physical inspection of meters and the existence of fixed threshold hardening, have an inherent limitation of scalability, flexibility, and responsiveness. The present paper presents an automated electricity theft detection model, with a combination of a Random Forest classifier and a web-based inference engine, and suggests a further architecture that would include Deep Neural Networks (DNNs) to better identify pattern changes over time. Sequences of electricity consumption over multiple days are gathered, preprocessed in order to remove noise and missing data, and finally provided to each of the paradigms in the classes. The baseline of the Random Forests reaches high accuracy in the classification of labeled consumption profiles which include natural usage and artificial synthetically produced theft cases. The DNN extension uses several hidden layers that have ReLU activations on nonlinear and time-dependent consumption anomalies. The accuracy of experimental assessment is more than 96.00% in crossvalidated conditions. The trained model is deployed into the Flask web application, which allows real-time prediction of the field operators. Findings support the viability of data-based anomaly detection as a scalable alternative to traditional approaches, which is the obvious direction toward linking the isotope with the smart metering system and cloud-based systems of distribution management.