Dynamic Graph Convolutional Network With Spatiotemporal and Multiscale Feature Fusion for Multistep Traffic Flow Prediction

Abstract

Accurate multistep traffic flow prediction is crucial for intelligent transportation systems (ITS). Existing graph convolutional networks (GCNs) still have limitations. They struggle to capture the complex dynamic relationships between adjacent nodes in road networks, to model long-term dependencies, and fully mine cross-layer feature expressions. These limitations make long-term predictions challenging. To address these challenges, we propose a GCN based on dynamic spatiotemporal and multiscale feature fusion (DGCN-STMSF). The model adopts a shared and independent stacked architecture, taking into account cross-layer feature sharing and personalized modeling, this architecture enables better characterization of the complex dynamic relationships between nodes. We introduce an exponential smoothing attention mechanism (ESA). It embeds the exponential decay mechanism into graph-structured temporal modeling. This improves the unique feature representation between layers and enhances temporal stability and trend perception. Furthermore, a fine-grained correlation enhancement module (FGCE) is proposed, featuring a dual-layer mechanism design, combined with spatial location enhancement learning (SLEL) and cross-spatial integration learning strategy (CSIL), which are used to characterize the instantaneous changes of adjacent nodes and the interactive effects between different road sections, respectively, and dynamically model complex local and global spatial dependencies. Extensive experiments on four real datasets show that the DGCN-STMSF model demonstrates superior effectiveness and performance. On various datasets, mean absolute error (MAE), root mean square error, and mean absolute percentage error are reduced by up to 2.14%, 5.34%, and 3.12%, respectively, demonstrating good generalization ability and application value.