Dynamic Graph Convolutional Recurrent Networks Based on Progressive Learning

Traffic flow prediction is the core challenge of transportation, and its key lies in effectively capturing the spatio-temporal dynamic dependencies. Aiming at the deficiencies of existing methods in modeling global temporal relations and dynamic spatial heterogeneity, this paper proposes a dynamic graph convolutional recurrent network (DGCRN) based on interactive progressive learning. First, the interactive progressive learning module (IPL) is designed to segment the input sequences through a tree structure, synchronize the extraction of spatiotemporal features using the interactive learning of parity subsequences, and adaptively capture the dynamic associations among nodes by combining with the dynamic graph convolutional recursive module (DGCRM). Secondly, a spatio-temporal embedding generator (STEG) is constructed to fuse temporal and spatial embedding to generate dynamic graph structures. Experiments validate the effectiveness of DGCRN on the PEMS04 and PEMS08 datasets with MAE, RMSE and MAPE metrics improved by 2.20%, 0.84% and 0.16% (PEMS04) and 5.78%, 3.33% and 0.35% (PEMS08), respectively, compared with the optimal baseline model. The visualization results show that the model can accurately capture the periodicity and trend of the traffic flow and exhibits superior prediction robustness, especially during peak periods. This study provides a new idea for dynamic spatio-temporal dependence modeling, significantly improving long-duration traffic prediction accuracy and reliability.