A Robust Deep Q-Network Based Attack Detection Approach in Power Systems

Abstract

To achieve safe and reliable operation of a power system, accurate and timely attack detection is required. The proposed detection policy can be applied to small noise attacks or attacks by adversarial signals. To improve the robustness of the DRL policy, a robust Deep Q-Network (DQN) is designed to defend against attack perturbations in the state observations of a power system in this paper. Accordingly, we formulate the attack detection as a change point detection problem in which the detection delay and accuracy are optimized. A robust policy regularizer is included in the DQN to allow a defender to learn a policy that can efficiently detect an attack. A new metric can be modeled to evaluate the robustness of the proposed algorithm. Numerical simulations on the IEEE 14-bus system verify the effectiveness of the proposed robust DQN.