Deep Learning-Based Detection of Electricity Theft Cyber Attacks in Solar PV Distributed Generation

Abstract

In the past, electricity thefts were committed through physical means like tapping a line or bypassing the energy meter. However, the advent of smart meters has added another possible means of committing electricity theft that is through cyber-attacks. Existing research in this area focuses on detection of cyber-attacks that are aimed at reducing electricity bills by sending lower consumption readings to the utilities. This thesis describes artificial intelligence-based methods to identify cyber-attacks in Solar Photovoltaics (PV) distributed generation smart meters installed in houses that generate solar power for self-consumption as well as for sending excess power to the grid in exchange for incentives. In this work, we propose Deep Learning models: Feed Forward, Gated Recurrent Unit (GRU) and Convolutional Neural Network - Gated Recurrent Unit (CNN-GRU) to detect electricity theft cyber-attacks aimed at falsifying the generated energy readings for unlawful gains. A unique deep learning-based detector that trains on multiple datasets is also introduced herein thesis. It is found that such a detector presents a higher predictive performance. Hyperparametric-tuning of the models using cross-validated random-search for enhanced performance is also carried out in this thesis.