On The Efficacy of Physics-Informed Context-Based Anomaly Detection for Power Systems

The Automatic Generation Control (AGC), a fundamental frequency control system, is vulnerable to cyber-physical attacks. Coordinated false data injection attack, aiming to generate fake transient measurements, typically precedes unwarranted actions, inducing frequency excursion, leading to electromechanical swings between generators, blackouts, and costly equipment damage. Unlike other works that focus on point anomaly detection, this work focuses on contextual detection of stealthy cyber-attacks against AGC by utilizing prior information, which is essential for power system operation and situational awareness. More specifically, we depart from the traditional deep learning anomaly detection that is thoroughly driven by black-box detection; instead, we envision an approach based on physics-informed hybrid deep learning detection 'CLDPhy,’ which utilizes the combination of prior knowledge of physics and system metrics. Our method, to the extent of our knowledge, is the first context-based anomaly detection for stealthy cyber-physical attacks against the AGC system. We evaluate our approach on an industrial high-class PowerWorld simulated dataset - based on the IEEE 37-bus model. Our experiments observe a 36.4% improvement in accuracy for coordinated attack detection with contextual information, and our approach clearly demonstrates the superiority in comparison with other baselines.