Modeling Dynamic Traffic Flow as Visibility Graphs: A Network-Scale Prediction Framework for Lane-Level Traffic Flow Based on LPR Data

Emerging applications in real-time traffic management put forward urgent requirements for lane-level traffic flow prediction. Limited by extremely unstable traffic volumes and heterogeneous spatiotemporal dependencies in urban road networks, network-scale prediction for lane-level traffic flow is still a critical challenge. This study models the dynamic characteristics of lane-level traffic flow as complex networks and proposes a deep learning framework for network-scale prediction. Relying on the visibility graph, we transform the temporal dependence learning task into spatial correlation mining on temporal complex networks. For spatial dependency extraction in urban traffic flows, we establish three topological graphs from traffic, statistical, and semantic perspectives to investigate the static and dynamic correlations. Then, a network-scale traffic volumes prediction model, i.e., spatiotemporal multigraph gated network (STMGG), is proposed to learn spatiotemporal correlations on visibility graphs and spatial topological graphs. This model designs an attention-based gated mechanism to incorporate global features from multigraphs. Additionally, a Seq2Seq structure is integrated to enhance multistep prediction stability. We employ two license plate recognition (LPR) datasets as case studies, and STMGG expresses superiorities over various advanced deep learning models. Meanwhile, an ablation experiment is conducted to evaluate its components, and numerical tests further reveal its impressive inductive learning capability.