The power flow calculation model based on deep learning can directly fit the mapping relationship between the initial value of the system power flow and the result of the power flow, and the calculation speed is extremely fast and the ill-posed power flow problem is not generated. However, the existing deep learning power flow calculation methods are mostly based on regression models, which can’t identify whether the power flow converges, resulting in false system power flow distribution still mapped to the input non-convergent power flow samples. To solve this problem, a power flow analysis method based on graph multi-task learning network is proposed, power flow analysis is performed on the input case combined with the physical characteristics of the power system. Finally, the proposed model is comprehensively simulated on the IEEE 14-node system, and 10 000 system samples containing different network topologies are generated. The simulation experiment verifies that the computational time of the proposed model is about a quarter of that of the Newton-Raphson method, and the accuracy of power flow convergence judgment reaches 98.81%, the calculation accuracy of power flow distribution calculation reaches 98.58%, and the effectiveness of the improvement in graph convolution and graph pooling is verified by comparative experimental ablation experiments.