Residual classifier assisted robust optimization for resilience enhancement of power system against cyber attack

With the increasing intelligence and interconnectivity of facilities, the power system is facing increasingly complex cyber-physical security threats. Among them, the vulnerability of distance protection relays remains insufficiently addressed in prior research. To enhance the cyber-physical defense and resilience of transmission networks, a defender-attacker-defender framework is developed to optimize dynamic defense portfolios. The attacker is modeled to launch coordinated attacks on transmission lines and intrusion on distance protection relays. Although conventional column-and-constraint generation (C&CG) algorithms typically identify worst-case scenarios, they notably ignore historical attack-defense patterns. To overcome this limitation, a residual classifier is employed to identify critical scenario patterns. The Lagrange function is embedded in the training process to enhance solution feasibility, and a selective correction method based on the one-norm function is formulated to improve the quality of enumerated attack scenarios. These learning-assisted steps enable more efficient robust optimization under dynamic conditions. Finally, the algorithm is validated on the IEEE 24-bus system. The simulation test demonstrates significant performance gains of this algorithm, including 84.0 % and 82.5 % reductions in iterations and computation time compared to the baseline C&CG algorithm.