| Abstract: |
Modern power distribution networks have seen the widespread deployment of smart meters, resulting in an unprecedented level of operational visibility and also significant cyber security vulnerabilities. The smart grid infrastructure hinges on advanced metering infrastructure (AMI) that constantly gathers, transmits and processes high-frequency consumption data meaning there is an immense attack surface vulnerable to energy theft, false-data injection, denial-of-service attacks and even meter tampering. Machine Learning (ML) and artificial intelligence (AI)-based techniques have proven to be the best-suited paradigms for detecting anomalies in smart meter data as they can model complex, non-linear consumption patterns while adapting to changes in threat factors without any explicit rule-based programming. In this reviews paper, we provide the most comprehensive meta-analysis of literature for AI enabled anomaly detection frameworks that have deployed within smart grid cybersecurity contexts. This work synthesizes results from thirty peer-reviewed studies published from 2015 to 2024, analyzing the most effective deep learning architectures, ensemble methods, federated learning approaches and hybrid AI models applied to smart meter anomaly detection. The review categorizes four major threats energy theft, false data injection attacks (FDIA), meter malfunction, and communication layer intrusions and analyzes existing detection algorithms in terms of their performance for each threat. We characterise them on several important performance metrics, including detection accuracy, false positive rates, computational overhead and scalability. The analysis indicates that deep learning models, especially long short-term memory (LSTM) networks and convolutional neural networks (CNN), perform better than traditional machine-learning techniques on controlled benchmarks with over 97% detection accuracy achieved, whereas federated AI and privacy-preserving AI architectures are emergi |
