Dockless Shared-Bike Demand Prediction with Temporal Convolutional Networks

Dockless shared-bikes systems are being widely developed in many countries and have been studied extensively within the transport domain. This study aims to develop a dynamic travel demand prediction model for dockless shared-bikes by using the deep learning method. Firstly, a geographical algorithm was utilized to create the traffic analysis zones (TAZ), considering the spatial correlation of sharing bicycle trips. Analyses on cycling trips were conducted to show the imbalance of spatial and temporal distribution. Additionally, the temporal convolutional networks (TCN), which constantly performs well on a vast range of tasks, were then applied to predict the demand over the TAZs. The statistical model and other machine learning models were developed to validate the TCN model. The results confirmed the TCN model's advantages beyond baseline approaches. Finally, the models predicted the gap between trip production and attraction at each TAZ, which provides useful strategies for rebalancing the dockless bicycles system.