A Hybrid Monotonic Neural Network Approach for Multi-Area Power System Load Frequency Control Against FGSM Attack

The integrity of data-driven load frequency control (LFC) in power system is increasingly threatened by adversarial attack. Addressing this concern, this paper introduces a novel hybrid approach that integrates adversarial reinforcement learning and monotonic neural network (ARL-HMNN) for LFC in multi-area power system. To holistically counter unforeseen uncertainties and to withstand the prevalent adversarial attack, the proposed ARL-HMNN approach builds a stable neural network structure with monotonicity constraints, and optimizes this neural network for LFC with adversarial training and deep deterministic policy gradient algorithm. By enforcing the deviation-command monotonicity constraints, the HMNN is enabled to satisfy Lyapunov stability conditions for LFC, which significantly enhances the stability and robustness of the power system. To further enhance the robustness of LFC, the classic fast gradient sign method (FGSM) adversarial attack is applied during the reinforcement learning training process. Through the integration of adversarial training, our method improves the system's resilience to FGSM attack under malicious threat from the communication network, while at the same time maintaining provable frequency stability. The superior performance of the developed approach is demonstrated by comparison to existing data-driven control methods on the IEEE 39-bus power system.