Quickly and accurately locating the fault section is crucial for the safe operation of a distribution network. Traditional fault location methods have low fault tolerance and take a long time, and most deep learning algorithms have insufficient generalization to topological changes. Thus a fault section location method for a distribution network based on a one-dimensional convolutional neural network (1D-CNN) and a graph neural network (GNN) is proposed. This method combines the original information of the distribution network with deep learning algorithms such as GNN for modeling. First, an attention-based spatial-temporal graph convolutional network (ASTCGN) is used to extract fault features from telemetry data at different temporal and spatial scales. The graph attention network (GAT) is used to fuse multi-source remote signaling data. Then, 1D-CNN is used to adjust the feature output dimension to realize the mapping of node features to fault branches. Finally, the fault section location results are output by adding a fully-connected network (FCN). A 10 kV neutral ungrounded distribution network system for simulation and testing is built based on the Matlab/Simulink platform. It is verified that the proposed method possesses superior positioning performance, can be flexibly applied to various low, medium, and high resistance grounding fault scenarios, and has strong generalizability to system topology changes and good robustness to incomplete fault data.