Research on fault sample generation and decision support system for converter stations combining gan and reinforcement learning

Abstract This paper addresses the critical challenges of data scarcity and decision optimization in High-Voltage Direct Current (HVDC) converter station fault management by integrating Generative Adversarial Networks (GANs) and Reinforcement Learning (RL). A Wasserstein GAN with gradient penalty (WGAN-GP) was employed to generate high-fidelity fault samples, achieving a low Fréchet Inception Distance (FID) of 15.2, effectively resolving data imbalance. Utilizing this augmented dataset, a Proximal Policy Optimization (PPO) decision model was trained, reaching a peak accuracy of 96.5%, representing a 4.8% improvement over models trained solely on real data. Experimental results demonstrate the system’s operational superiority, reducing average fault isolation time to 0.50 s—a 39% improvement over Deep Q-Networks (0.82s) and 61% over rule-based systems (1.28s). Furthermore, the proposed model exhibited exceptional robustness, maintaining 85.1% accuracy even under extreme noise conditions (30%), validating its potential for enhancing power grid reliability and safety.