In order to comprehensively capture the Spatio-temporal characteristics of the traffic road network, analyze the complexity and variability of road conditions, and achieve accurate prediction of road congestion and emergencies, a Spatio-temporal graph attention neural network model is proposed. By modeling the road network as a series of time-varying graphs, the Graph Attention Network (GAT) mechanism was used to focus on the spatial characteristics of key road network graphs and capture the dynamic full-graph spatial features. Gated Recurrent Neural Network (GRU) was used to fully capture the temporal correlation of adjacent road network graphs and reduce model redundancy. The results of the experiments using the PEMSD dataset show that the proposed method outperforms the existing methods in terms of accuracy and recall compared with the baselines. It is concluded that the proposed model further improves the prediction accuracy of traffic anomalies.