Traffic flow is considered as a critical feature of intelligent transportation systems (ITS). Accurately forecasting future vehicular volumes is an effective means of mitigating traffic congestion. However, the nonlinear and complex traffic flow characteristics make the traditional approaches unable to achieve satisfactory prediction performance. Although existing methods based on deep learning models have improved the accuracy of traffic flow prediction, the spatio-temporal features of traffic flow data are still not fully explored. Moreover, existing methods pay little attention to the task of training models in a decentralized environment where data are distributed across multiple clients. To solve the problems mentioned above, we propose a novel network model called Graph Transformer Attention Network (GTAN) for traffic flow prediction, which can effectively extract traffic flow's temporal and spatial characteristics by considering all node locations' information in the traffic networks. Then, we propose a training strategy called Graph Federated Meta-learning (FedGM), solving the problem of topological heterogeneity by combining meta-learning and federated learning, to achieve an optimal initialization model which can quickly adapt to different traffic networks under low communication cost. Finally, the experimental results on a real data set show that the GTAN model has better prediction performance and faster meta-training speed. The model trained by FedGM can quickly adapt to different graph-structured data and achieve high accuracy.