Channel Information Exchange-Induced Spatiotemporal Graph Convolutional Network

Accurately predicting traffic flow is a highly crucial task essential for providing functional services to urban road networks. Urban traffic is a complex and constantly evolving system, influenced not only by factors within individual regions but also by interactions among different regions within the entire city network. The majority of current traffic flow prediction methods rely on static geographic information, thus ignoring the cross regional flow of traffic within cities. To tackle this issue, this paper proposes a channel information exchange-induced spatiotemporal graph convolutional network (CIE-STGCN). This network constructs both a static adjacency matrix constructed from geographic information and a dynamic adjacency matrix based on adaptive parameter learning for nodes. The static and dynamic spatiotemporal graph convolutional networks operate on separate channels to extract features. Additionally, a channel information exchange module based on channel attention and gating mechanisms is designed to achieve global complementarity of static and dynamic features in traffic flow data. Validation using multiple real-world traffic flow datasets demonstrates the efficacy of the proposed model in reliably predicting traffic flow.