Traffic flow prediction (TFP) plays a crucial role in intelligent transportation systems. Due to the need to collect data during the prediction process, TFP can face the risk of data leakage. Federated Graph Learning (FGL) is a novel method for protecting graph data privacy, but there are lower performance and Byzantine attacks challenges when applying FGL to TFP. In order to address these challenges, this paper proposes a distributed and secure traffic flow prediction (DS-TFP) framework based on federated graph learning, which includes three methods. Firstly, a New Spatial-Temporal Graph (NSTG) model is applied for the dynamic changes of temporal and spatial and simplifies the encoding procedures. Secondly, graph models are combined with federated method to ensure the protection of local data privacy. Thirdly, the resist Byzantine attacks method which utilizes clients dynamic cross-validation and a voting-based aggregation algorithm, is designed to enhance system security. Finally, extensive experiments are conducted on two datasets. The results show that compared with the baselines, the Fed-NSTG model improved accuracy by an average of 8.15%. The DS-TFP framework can maintain an average prediction accuracy of 87.57% even under the Byzantine attacks.