Machine-Learning-Based Adaptive Anomaly Detection for Control Feedback Interferences in Solid-State Transformers

Solid-State Transformer (SST)-based Power Substations (SSPS) have emerged as a pivotal innovation for integrating distributed generation and energy storage systems within modern grid. However, SSTs' mixed-signal nature and network-dependent control make them vulnerable to evolving cyber-physical threats, which can disrupt real-time operations, especially as attack patterns continuously evolve, making static, batch-trained anomaly detection systems (ADS) ineffective. To address this, this paper proposes a Machine Learning (ML)-based Adaptive ADS (ML-A2D) designed to detect control feedback noise interference attacks that compromise the low-frequency closed-loop performance of SSTs. The proposed framework employs a semi-supervised online learning approach, enabling continuous adaptability to new anomalies while maintaining fine-grained, real-time detection. The system was evaluated in a realistic SST hardware testbed under practical and varying attack scenarios, demonstrating robust performance with detection accuracy exceeding 96%. With an effective detection time of 0.07 ms and an overall latency of less than 200 ms within a hierarchically controlled network of AC/AC converter modules, the proposed ML-A2D offers a scalable and reliable solution to enhance the resilience of SSTs in next-generation power systems.