Abstract: To accurately predict traffic flow in large-scale road networks, a traffic flow prediction model is proposed based on Graph Transformer to capture the complex and dynamic spatiotemporal characteristics of traffic flows. Gate recurrent unit (GRU) module is adopted in the model to extract temporal features of historical traffic flow data in a road network. According to the connections among sensors distributed in a network, traffic graphs are then established with historical traffic flows as nodes and connections of sensors as edges. On this basis, spatiotemporal characteristics are captured using Graph Transformer-based deep learning technologies. To verify the effectiveness of the proposed model, comparisons with six baseline models using the PeMS highway dataset are conducted. Experiments show that the proposed prediction model results in the best performance.