Conditional Graph Diffusion Model for Synthetic Feeder Fault Data under Extreme Weather

Extreme weather events pose serious challenges to the safe operation of power distribution systems. However, the scarcity of fault data under extreme conditions hinders the effectiveness of data-driven prediction models. To address this issue, this paper proposes a Conditional Graph Diffusion Model (CGDM) for generating realistic feeder fault samples under extreme weather scenarios. The model integrates graph neural networks (GNNs) with a conditional diffusion framework, enabling structure-aware and physically consistent data generation. Specifically, the model integrates node-level features through a GNN-based noise estimation network, which is trained to recover extreme-weather fault characteristics by predicting and removing the noise added during the diffusion process. Experimental results on a real-world distribution system dataset from southern China demonstrate that CGDM significantly outperforms baseline models in terms of distributional accuracy, sample quality, and diversity. Case studies further validate that the proposed model preserves key variable characteristics and produces physically meaningful scenarios.