Weather-aware Feeder Fault Prediction Based on Graph Learning with Synthetic Rare Event Generation

The increasing frequency and intensity of extreme weather events have significantly challenged the reliability of power distribution systems. Accurate prediction of weather-related feeder fault states under such conditions is hindered by two fundamental issues: (1) the mismatch between the Euclidean assumptions of traditional machine learning and the inherently graph-structured nature of distribution feeders, and (2) the severe class imbalance caused by the rarity of weather-related fault data. To overcome these challenges, this paper proposes a novel framework that integrates a graph convolutional network (GCN) with a conditional generative adversarial network (cGAN). The proposed GCN-cGAN framework captures topological dependencies of feeders while synthesizing realistic rare fault events under extreme weather scenarios. A graph attention-based model is further developed for feeder fault prediction, leveraging both real and synthetic data samples. Extensive experiments on real distribution topology and meteorological data show consistent gains under multiple evaluation scenarios, including strong wind, heavy rainfall, and their combination, with marked robustness to class imbalance. Under the present prediction setting, compared with conventional and state-of-the-art baselines, Recall improves by about 28 percent and F1-score by about 21 percent.