Global Attention-Based Dynamic Multi-Graph Convolutional Recurrent Network for Traffic Flow Forecasting

Accurate traffic flow forecasting is a challenging task in intelligent transportation system. With traffic flow forecasting being formulated as a spatio-temporal graph modelling problem, graph convolution network (GCN) is increasingly used in recent research. However, most approaches employ a single predefined or adaptive graph for convolution, which cannot adequately represent complicated dependencies inherent in real-world traffic flow data. And they are limited in learning relationships between long-distance time steps. To address these concerns, we propose a global attention-based dynamic multi-graph convolutional recurrent network (GA-DMGCRN). Specifically, we design a dynamic multi-graph convolution module based on dynamic graph learning network that generates graphs by adjusting to time-varying input data throughout the training and testing phases, allowing for the effective extraction of dynamic spatial and semantic dependencies. To capture temporal features, we propose the dynamic multi-graph convolution recurrent unit, and multihead ProbSparse self-attention with linear biases is developed to model global temporal dependencies. The proposed GA-DMGCRN is evaluated on three real traffic datasets. Compared with the baseline models, our model achieves an average improvement of 1.97%, 3.11%, and 2.01% under MAE, RMSE, and MAPE metrics, which can provide real-world value by improving traffic efficiency, mitigating congestion, and optimizing route planning.