HD-Net: A hybrid dynamic spatio-temporal network for traffic flow prediction

Abstract

Accurately predicting traffic flow is crucial for intelligent transportation systems (ITS). In recent years, many deep learning-based prediction models have been widely applied in traffic flow prediction, and various spatio-temporal networks have been proposed. However, most of the existing models follow a general technical route to extract the spatio-temporal features, which lack the capacity of extracting the important historical information with the high spatial and temporal correlations dynamically and deeply. How to develop a well-performance traffic flow prediction model for a complex transportation network is still facing some challenges. In this paper, a hybrid dynamic spatio-temporal network (HD-Net) for traffic flow prediction is proposed. In HD-Net, the authors first extract the dynamic spatio-temporal features using dynamic graph convolution and bidirectional gate recurrent uni (BiGRU). Subsequently, the authors extract the important features with high spatial and temporal correlations from the obtained dynamic spatio-temporal features using an auto-correlation mechanism from a local perspective, and self-attention mechanism from a global perspective, respectively. Extensive experiments have been conducted on two real-world traffic datasets. The experimental results demonstrate that the proposed HD-Net outperforms the baselines in the field of capturing the dynamic and important spatio-temporal features with high correlations.