Abstract: False data injection attack(FDIA) can evade traditional bad data detectors and pose challenges to the stable operation of smart grids. Therefore, an FDIA detection method based on an improved graph attention network(GATv2) model is proposed. Firstly, based on the structure of the power system and the characteristics of FDIA, the required dataset is constructed for the model. Then, graph data is established based on the topology and operational information of the power system. A detection model is designed based on GATv2 to extract spatial features of power grid diagram data and detect injected false data attacks. Finally, the effectiveness of the GATv2 model is validated on IEEE 14 node and IEEE 118 node systems, and simulation results show that the detection performance of the GATv2 model is superior to other models, with a detection accuracy of over 98%, and it has good robustness in different attack node numbers and attack intensities.