Dynamic multi-graph neural network for traffic flow prediction incorporating traffic accidents

Traffic flow forecasting is the foundation of intelligent transportation development and an important task in realizing intelligent transportation services. This task is challenging due to the complex spatiotemporal dependencies between road nodes and some other external factors. Most existing GCN-based methods usually use a single and fixed adjacency matrix to characterize the global spatiotemporal relationship of road networks, which limits the expressiveness of the model in different scenarios and ignores the dynamic nature of node relationships that change over time. In addition, sudden traffic accidents may also cause fluctuations in traffic flow in the short term, which may affect the accuracy of the model prediction. To address the above problems, this paper proposes a dynamic multi-graph neural network (DMGNN) incorporating traffic accidents for multi-step traffic flow prediction. First, to provide richer prior knowledge for the model, we construct multiple graphs to represent various contextual dependencies among nodes. Second, we designed a dynamic graph adjustment module to update the adjacency matrix used in each training step. Finally, we build a deep learning framework based on GAT and Bi-LSTM to focus on local fluctuations caused by traffic incidents and to extract sophisticated spatiotemporal correlations between data. We conducted extensive experiments on two real traffic datasets to evaluate the model, and the ablation experiments verified the effectiveness of each module. On the standard public dataset PEMSD3, compared to the optimal baseline model, our model improves the RMSE, MAE, and MAPE of the multi-step prediction by about 21%, 21%, and 22%, respectively.