DRL²FC: An Attack-Resilient Load Frequency Control Based on Deep Reinforcement Learning

Modern power grids utilize Information and Communication Technologies (ICT) for a more efficient system monitoring and control. However, the extensive digitalization exposes electrical systems to numerous cyber risks. Load Frequency Control (LFC) is a power system automation that is heavily threatened by cyberattacks, particularly False Data Injection Attacks (FDIAs). FDIAs are a special type of cyberattacks that can make LFC unstable by deliberately tampering its measurements and control signals. To mitigate the impact of FDIAs on LFC, a Deep Reinforcement Learning (DRL) approach, called DRL<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>FC, is proposed in this paper. The presented methodology replaces the conventional integral controller of LFC with double Q-learning neural networks that have been trained to stabilize the system both after load fluctuations and under FDIAs. The introduced controller learns the desired attack-resilient control policy by interacting with simulated power system environments that incorporate the LFC dynamics under cyberattacks. The experimental results confirm the effectiveness of the proposed defense mechanism, its scalability to large power systems, its robustness against noisy measurements and its superiority over other similar works.