Understanding and Predicting the Short-Term Passenger Flow of Station-Free Shared Bikes: A Spatiotemporal Deep Learning Approach

Abstract

Accurate passenger flow prediction of shared bikes provides useful information for operators to optimize the repository of bikes in different regions. However, it is very challenging as the usage of shared bikes is affected by many complex factors. In this article, we propose an end-to-end deep learning architecture, termed spatial-temporal fusion network (STFNet), to forecast short-term passenger flow in the new station-free bike-sharing system. The architecture utilizes the different neural network structures jointly to capture the complex nonlinear relationships of spatiotemporal dependencies and external factors from multiple data sources. Furthermore, the attention mechanism is introduced to improve the model’s interpretability and prediction ability. Based on two real-world data sets collected from Chinese cities, Beijing and Shenzhen, detailed spatiotemporal usage patterns of shared bikes are analyzed. Ablation studies are performed to test the effectiveness of different components on the whole framework. Experiment results show that the proposed STFNet can effectively capture the spatiotemporal correlations, and the predictions outperform state-of-art baselines in different predicting horizons.