Research on traffic flow forecasting based on interactive dynamic meta-graph learning

Abstract

Traffic flow exhibits intricate dynamic spatial-temporal correlation. Aiming at the difficulty in capturing the long-term dependence and dynamic spatial correlations among concealed road nodes of in traffic flow, a new Interactive Dynamic forecasting model based on Meta-graph learning (Mega-ID) is proposed, which combines spatial-temporal transformer and interactive dynamic graph convolution (IDGCN) to optimize the Meta-graph module. Specifically, it optimizes a spatial-temporal meta-graph with memory and discrimination capabilities. The model introduces a Dynamic Graph Convolution (DGCN) embedded Interactive Learning structure, which simultaneously captures the hidden dynamic spatial correlations and long-term dependency of traffic flow. The experimental results demonstrate that the method proposed in this paper can capture the hidden dynamic spatial correlation and long-term dependence, leading to better forecasting performance compared to other baseline models.