AI-DRIVEN ELECTRICITY THEFT DETECTION IN SMART GRIDS: A CONVOLUTIONAL LSTM APPROACH WITH DEEP REINFORCEMENT LEARNING

Abstract

The integration of smart grids into modern energy infrastructures has enabled notable advances in efficiency, monitoring and automation. However, it also presents criti- cal challenges and risks, including electricity theft, cyberattack susceptibility, and data integrity threats. As traditional rule-based and statistical methods often fail to adapt to evolving fraudulent activities, intelligent artificial intelligence (AI)-driven solutions are required. I propose a convolutional long short-term memory model (ConvLSTM)- based model for anomaly detection that captures spatial and temporal correlations in electricity consumption data, effectively identifying anomalous behaviors indicative of electricity consumption fraud. A key challenge in achieving smart grid security is the susceptibility of AI models to adversarial attacks, in which an intelligent attacker al- ters consumption patterns to evade detection. Anomaly detection models such as the proposed ConvLSTM have been tested and evaluated against machine-generated adver- sarial attacks, such as the Fast Gradient Sign Method (FGSM) and Carlini & Wagnar (C&W). However, these types of attacks are designed mainly to prevent anomaly de- tection without considering a possible reduction in the reported energy consumption, thus overlooking the problem of energy theft. Furthermore, the lack of generalization of adversarial attacks evaluated to other models remains a concern. In fact, conven- tional anomaly detection methods do not detect new adversarial attacks generated by artificial intelligence. Thus, To assess and enhance the involved model resiliency, I include a deep reinforcement learning (DRL)-based attack framework for an agent to learn to generate adversarial perturbations that bypass detection. Experimental results demonstrate that the proposed ConvLSTM substantially outperforms traditional ma- chine learning models in detecting electricity theft. However, DRL-generated adversar- ial attacks reduce the model’s accuracy, highlighting potential security vulnerabilities. After defensive training, the detection model exhibits improved resilience, maintain- ing high detection rates even under adversarial attacks. Overall, this research aims to advance the security and contribute to the ongoing research efforts focused on enhanc- ing the resilience, reliability, and stability of the smart grid infrastructure through an adaptive AI-based anomaly detection framework.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering