Abstract: In the power system distribution network, accurate and rapid location of the fault section is the prerequisite for isolating and restoring the normal power supply status of the fault area. This paper proposes a distribution network fault section location and restoration reconstruction technology based on graph neural network. First, the distribution network is transformed into a multivariate data set model that integrates node measurement parameter information in the time dimension and node topology information in the spatial dimension. Secondly, a graph neural network model based on a combination of attention aggregation function and gating mechanism update function is constructed. Finally, a distribution network fault experimental model is built in MATLAB/Simulink using the IEEE33 node standard system. The simulation results show that the model's accuracy under normal conditions is 98.92%, which is higher than the 95.85% and 97.63% of the RNN and CNN models. Under conditions of high interference and data loss rates, the accuracy is still better than the RNN and CNN models, with an accuracy of 88.45%.