Smart grids (SGs) are becoming increasingly complex with the integration of communication, protection, and automation technologies. However, this digital transformation has introduced new vulnerabilities, especially false data injection attacks (FDIAs) and Denial of Service (DoS) attacks. FDIAs can subtly corrupt measurement data, misleading operators without triggering traditional bad data detection (BDD) methods in state estimation (SE), while DoS attacks disrupt the availability of sensor data, affecting grid observability. This paper presents a deep learning-based framework for detecting and localizing FDIAs, including under DoS conditions. A hybrid CNN, Transformer, and BiLSTM model captures spatial, global, and temporal correlations to forecast measurements and detect anomalies using a threshold-based approach. For further detection and localization, a Multi-layer Perceptron (MLP) model maps forecast errors to the compromised sensor locations, effectively complementing or replacing BDD methods. Unlike conventional SE, the approach is fully data-driven and does not require knowledge of grid topology. Experimental evaluation on IEEE 14–bus and 118–bus systems demonstrates strong performance for the FDIA condition, including precision of 0.9985, recall of 0.9980, and row-wise accuracy (RACC) of 0.9670 under simultaneous FDIA and DoS conditions. Furthermore, the proposed method outperforms existing machine learning models, showcasing its potential for real-time cybersecurity and situational awareness in modern SGs.
Shrestha et al. (Mon,) studied this question.