Unsupervised Approach for Electricity Theft Detection Combining Recurrent Neural Networks and Rule-Based Policy

The Advanced Metering Infrastructure (AMI) is a bi-directional communication system between utilities and consumers and is an integral part of smart energy grid development. AMI offers several benefits such as automated collection of energy measurements, outage management, connection and disconnection requests, and system status tracking. However, given the increased dependency on connectivity and data-driven decisions, AMI and smart energy systems have become vulnerable to multiple security threats. In this paper, we consider electricity theft, which is a well-known data-falsification cyber attack on AMI. In this attack, smart meters are compromised to report falsified energy consumption data. We propose an electricity theft attack detection approach using deep machine learning models, specifically a long-short term memory recurrent neural network (LSTM-RNN) model. The proposed approach combines an LSTM-RNN deep learning model and rule-based policy for attack detection. Unlike prior work that mainly considers supervised machine learning models, we propose an unsupervised learning approach that is more practical in real-world environments. We compare the performance of both approaches and show the effectiveness of the proposed attack detection method under different attack models. Various tests are conducted to evaluate the attack detection performance using a publicly available dataset containing energy consumption data for over 5,000 customers. We show that the proposed model achieves high attack detection accuracy and low false alarm rates.