Abstract: The distribution network operates flexibly and the topology changes frequently. The existing fault section location method of the distribution network based on artificial intelligence algorithm has poor topological generalization. In this paper, the underlying model of fault section location in distribution network is proposed based on graph attention network(GAT). This model fully explores the fault characteristics based on the topology of the distribution network to improve its topological generalization. In addition, the conformal risk control(CRC) method is introduced to construct a reliable model of fault section location in distribution network, so that the prediction risk of the model is artificially controllable. The results of IEEE-33 system show that the fault section location model based on GAT and CRC has preferable accuracy, robustness and topological generalization. Moreover, the model has good performance under double fault and high resistance fault, and has certain generalization ability for the change of distribution network line parameters.