A Physics-Constrained Graph Generative Adversarial Network for Augmenting Scarce Voltage Violation Scenarios in Distribution Networks

The proliferation of distributed energy resources is creating significant voltage stability challenges for power distribution networks. However, the scarcity of real-world data on rare but critical voltage violation events hampers the development of robust AI-based risk assessment models. To address this data scarcity and imbalance, this paper proposes a novel Physics-Constrained Graph Generative Adversarial Network (PCG-GAN). The framework introduces a topology-aware Wasserstein critic that leverages Graph Convolutional Networks (GCNs) to learn the grid's structural dependencies. This is combined with a physics-informed loss function that enforces compliance with power flow equations. The generator architecture utilizes a conditional, perturbation-based mechanism to ensure stable training and targeted synthesis of specific voltage violation types. Validated on the IEEE 33-bus test system, the proposed PCG-GAN successfully generates high-fidelity, physically valid, and controllable risk scenarios, demonstrating its potential to enhance data-driven grid monitoring and security analysis.