Multi-Scale Spatial-Temporal Graph Convolutional Networks for Traffic Speed Prediction
YIN Xueyan1, QIAN Youcheng2*
1. School of Artificial Intelligence and Information Technology, Nanjing University of Chinese Medicine, Nanjing 210023, China; 2. School of Statistics and Data Science, Nanjing Audit University, Nanjing 211815, China
Abstract: Traffic speed prediction is challenging due to its complex dynamic spatial-temporal characteristics. To deeply explore the latent information in traffic data, this paper proposes a multi-scale spatial-temporal graph convolutional network for traffic speed prediction. The method integrates modules that capture different temporal-scale features, as well as dynamic temporal and spatial features, into a unified neural network framework. Specifically, the model first employs multi-scale temporal partitioning to learn traffic patterns at different time scales, including short-term, medium-term, and long-term scales. Then, it combines adaptive graph convolution and dilated causal convolution to capture dynamic spatiotemporal dependencies at various scales. Next, a parameter matrix-based approach is used to fuse the outputs from different scales, followed by stacked activation functions and linear layers to generate the final prediction. Experimental results on multiple real-world datasets demonstrate that the proposed method improves prediction accuracy. Additionally, ablation studies validate the effectiveness of each module.
