Machine-Learning-based Anomaly Detection System for Cyber-Physical Attack on a Solid-State Transformer

With the evolving characteristics and dynamics of the grid, Solid-State Transformer (SST)-based Power Substations (SSPS) are being proposed as a new solution to effectively incorporate distributed generation and energy storage systems. However, these power converter-based solutions are vulnerable to cyber and physical threats due to their mixed signal characteristics (analog and digital data processing) and communication requirements, leading to severe consequences if the attacks remain undetected and unmitigated. To identify potential threats, this paper proposes a real-time supervised machine learning-based anomaly detection system (ML-ADS) for detecting Electromagnetic Side-channel Noise Injection (EM-SNI) attacks on SSTs. The proposed system can detect anomalies with fine granularity, meeting real-time latency requirements for effective mitigations. The model was deployed and optimized for evaluation into a realistic hardware-in-the-loop (HIL) SST testbed environment. Our experimental evaluation demonstrates a promising performance with high accuracy (more than 99%) and low false-negative rates (less than 1%), with feasible latency (≈200ms).