Abstract: The diversified mode of power system operation puts forward higher requirements for speed and topological generalization of weak branch identification. In this paper, graph deep learning and interpretation method are used to identify and analyze the weak branches. A weak branch identification model based on graph convolutional network via initial residual and identity mapping (GCNII) is constructed. The model can integrate feature based on topological relation, and evaluate the weakness of branches based on security situation of neighboring power grids. The interpretation method based on mutual information optimization is used to analyze the decision basis of the identification model and extract dominant factors of weak branches. The results of IEEE68 system and actual power grid example show that the identification model has preferable accuracy and topological generalization. The results of attribution analysis are consistent with the conclusions of traditional mechanism cognition, which can provide effective guidance for real-time warning and preventive control of cascading failure.