Traffic flow forecasting based on dynamic graph convolution and triplet attention

Hong Zhang, Min Yi, Xijun Zhang

Abstract: Due to the complexity and dynamics of traffic flow, a new traffic flow prediction model (TS-DGCRN) is proposed to capture the dynamic spatio-temporal correlation and solve the computationally time-consuming problem. The model uses a new dynamic graph generator (STE generator) to establish a dynamic graph structure that does not depend on predefined adjacency matrices and a dynamic graph convolutional recursive module to simultaneously capture the dynamic spatio-temporal features of traffic flow. Meanwhile, the triplet attention mechanism is designed to capture the cross-dimensional features of traffic flow spatial dimension and channel dimension, which emphasizes the importance of information interaction of global feature dimension and further enhances the global spatio-temporal representation capability. In addition, to improve computational efficiency, ScConv is introduced to reduce spatial redundancy in the convolutional layer and shrink computational costs. Experimental results on four datasets demonstrate that the TS-DGCRN model outperforms most baseline models regarding predictive performance.