Attention-Based Spatial-Temporal Fusion Networks for Traffic Flow Prediction

Jiaying Wang, Heng Yang, Jing Shan, Xiaoxu Song, Junyi Jiang

Abstract: Traffic flow prediction plays a crucial role in the development of intelligent transportation systems and smart cities. Recent advancements in deep network models have improved the accuracy of traffic flow prediction. However, existing methods often focus on the close proximity of distance and time, neglecting the dynamic characteristics of traffic flow influenced by various factors such as periodicity, tendency, functional similarity, etc. To address this issue and effectively capture the dynamic spatial-temporal correlation, we propose a novel approach called Attention-based Spatial-Temporal Fusion Networks (ASTFN) for modeling the fusion of dynamic temporal and spatial features. ASTFN utilizes a multi-layer encoder architecture and consists of three main modules: the traffic data embedding module, the spatial-temporal attention module, and spacial-temporal fusion module. To evaluate the performance of ASTFN, we conducted extensive experimental studies on four real-world traffic datasets. The results demonstrate the superiority of ASTFN over other state-of-the-art methods.