Researchers are solving the challenges of spatial-temporal prediction by combining Federated Learning (FL) and graph models with respect to the constrain of privacy and security. However, there are still several issues left unattended: 1) Clients might not be able to access the server during inference phase; 2) The graph of clients designed manually in the server model may not reveal the proper relationship between clients. This paper proposes a new embeddings aggregation structured FL approach named node Masking and Multi-granularity Message passing-based Federated Graph Model (M3FGM) for the above issues. The server model of M3FGM employs a MaskNode layer to simulate the case of offline clients. We also redesign the decoder of the client model using a dual-sub-decoders structure so that each client model can use its local data to predict independently when offline. As for the second issue, A new GNN layer named Multi-Granularity Message Passing (MGMP) allows each client node to perceive global and local information.We conducted extensive experiments in two different scenarios on two real traffic datasets. Results show that the proposed model outperforms the baselines and variant models, achieves the best results in both scenarios.