Fault Classification in Distribution Networks Using Graph Neural Networks and Discrete Stockwell Transform

Deep learning models have shown great potential for fault location and classification tasks in distribution systems. Emerging multi-scale data sources such as waveform measurement units, synchro-phasors, smart meters, weather data, and other information from electricity markets provide a rich source of information that can be leveraged for various applications in power system operations. However, the lack of sufficient training datasets poses a significant challenge in effectively training these models to achieve high accuracy and generalization. This paper proposes a novel architecture for the generation of fault datasets from the distribution system and the fault location tasks. It then develops a fault classification solution using graph convolutional neural network that can capture topology information, combined with discrete orthogonal Stockwell's transform for extracting frequency-domain features from time-series data. Pre-processing of the bus voltage angle and magnitude from strategic locations was used. The method was tested to predict faulted buses in the Cigre distribution test system. Comparisons were made with various standard classifiers. The results show the usefulness of the data generation solution for developing fault location models and the benefit of integrating topology data into the classifier with a graph convolutional neural network technique, as it outperforms other techniques in most scenarios.