Load Frequency Control with Deep Reinforcement Learning under Adversarial Attacks

Abstract

While deep reinforcement learning (DRL) ways are increasingly adopted in wind power systems, adversarial attacks to DRL may significantly degrade the control performance. This paper studies load frequency control (LFC) for single-area power systems under cyber-attacks based on robust DRL with state-space adversarial training. The cyber-attacks on LFC systems are modeled as malicious disturbances on the frequency measurement. The robustness of the control model is improved by adversarial training and stacked denoising auto-encoders (SDAE). Considering the continuity of the control action, this paper uses the continuous action search to correlate exploration noise over time. At last, the simulations are conducive to prove the effectiveness and feasibility of the proposed method.