Traffic Flow Prediction Based on Dynamic Time Slot Graph Convolution

Flow prediction, a critical component of intelligent transportation systems, is essential for travel planning and traffic control. However, existing methods often struggle with inflexible sharing patterns and difficulty capturing dynamic global temporal dependencies. To address these issues, this paper proposes a traffic flow prediction model based on dynamic time-gap graph convolution (DTGCN). The DTGCN model achieves parameter sharing and cross-layer independence through independent and shared modules, enabling the utilization of distinct patterns between layers while capturing stable patterns across layers. Additionally, the paper introduces a novel method for constructing a dynamic time slot graph by viewing historical time slots as nodes, effectively modeling the ever-changing temporal interactions. Lastly, a new temporal convolution module is designed to capture flexible global temporal dependencies. Experimental results on two widely used traffic network datasets, METR-LA and PEMS-BAY, demonstrate the effectiveness of the proposed model.