GCN for Identifying Critical Branches Based on Unlabeled Power Outage Data

This letter proposes graph convolutional neural networks (GCN) with self-supervised learning to identify the critical branches comprehensively in power systems. Firstly, we construct a two-layer graph data model comprehensively considering the propagation properties and topological properties. Secondly, we model a GCN consisting of two convolution layers and a linear layer to extract the features of the proposed data model. Furthermore, we adopt a self-supervised learning method based on different training strategies to train unlabeled data to identify the critical branches. The simulation of the IEEE 39-bus system verifies the effectiveness of proposed method.