Short-term passenger flow prediction for multi-traffic modes: A Transformer and residual network based multi-task learning method

Abstract

Managing multiple traffic modes cooperatively is becoming increasingly important owing to the diversity of passenger demands. Short-term passenger flow predictions for multi-traffic modes can be applied to the management of the multi-traffic modes system. However, this is challenging because the spatiotemporal features of multi-traffic modes are complex. Moreover, the passenger flows of the multi-traffic modes differentiated and fluctuated significantly. To address these issues, this study proposes a multitask learning-based model, called Res-Transformer, for short-term inflow prediction of multi-traffic modes. The Res-Transformer consists of two parts: (1) modified Transformer layers comprising the Conv-Transformer layer and the multi-head attention mechanism, which helps extract the spatiotemporal features of multi-traffic modes, and (2) the structure of the residual network, which is utilized to obtain correlations among multi-traffic modes and prevent gradient vanishing and explosion. The proposed model was evaluated using two large-scale real-world datasets from Beijing, China. One was a traffic hub, and the other was a residential area. The results not only demonstrate the effectiveness and robustness of the Res-Transformer but also prove the benefits of considering multi-traffic modes jointly. This study provides critical insights into short-term inflow prediction of the multi-traffic modes system.