Anomaly Identification in Power Systems Using Dynamic State Estimation and Deep Learning

With the expansion of cyber assets in modern power systems, cyber vulnerabilities and the attack surface have significantly increased. This paper presents an anomaly identification scheme to ensure data validity against false data injection attacks (FDIA) by combining model-based and datadriven techniques, specifically dynamic state estimation (DSE) and deep learning. In the proposed method, DSE first detects the presence of anomalies. Next, a neural network (NN) model generates hypotheses about the possible root cause, which DSE then asserts or rejects the hypothesis. The asserted hypothesis provides the root cause of the anomaly. Once the root cause is identified, the scheme replaces compromised data with corrected values in real-time, enabling self-healing. Numerical experiments with various realistic cyber-attacks demonstrate the method's capability to identify and distinguish between types of anomalies. This study also evaluates different input feature sets for the NN model, including DSE residuals, sampled value measurements, and RMS measurements, highlighting the critical role of synchronized measurements in training NN models and enhancing their performance.