Adap-STWT: Real-Time Traffic Flow Prediction Based on Multi-Scale Graph Adaptive Fusion and Spatio-Temporal Wavelet Transformer

Accurate traffic prediction is crucial for managing and making decisions in urban traffic systems. Recent studies have shown outstanding performance in traffic prediction by using methods that employ graph adjacency matrices predefined by heuristic rules or learn graph structures through trainable parameters. These studies tend to focus on interactions between nodes based on macroscopic traffic node attributes, overlooking the profound impact of individual behaviors and decisions of microscopic traffic participants on the transportation system. This valuable microscopic information has not been utilized to guide graph structure learning. In this paper, we proposed a spatio-temporal model with the capability of multi-scale adaptive graph structure learning, named Adap-STWT. Initially, an optimal graph structure learning module is constructed for dynamic traffic systems, encompassing multiple scales from microscopic traffic participants to macroscopic node attributes. Subsequently, in conjunction with this optimal graph structure, an optimized spatio-temporal prediction module is designed to facilitate multi-step traffic flow prediction. Ultimately, iterative optimization of the two modules is achieved through an alternating training approach. Notably, owing to the limited availability of large-scale public traffic datasets containing trajectory information, evaluations are conducted on two relatively small public datasets. The results underscore the potential of graph structure learning based on microscopic trajectories for traffic flow prediction. The source code is publicly available athttps://github.com/caisj11/Adap-STWT.git