GST-Net: A GIS-based Hybrid Prediction Model for Shared Bike Traffic Flow

To achieve reasonable placement of shared bikes, it is necessary to make accurate traffic predictions at bike-sharing stations. Different from single station's shared bike traffic flow prediction, we investigate the problem based on bike traffic flows among multiple stations. In this paper, we propose a GIS-based hybrid prediction model (GST-Net) which uses Gaussian Mixture Model (GMM) clustering and Deep Spatial-Temporal 3D Convolutional Neural Networks (ST-3DNet) to predict shared bike traffic flow. Firstly, we use GMM clustering to cluster multiple shared bike stations and extract the shared bike traffic flows among multiple stations to combine with the geographic information. Then, we merge bike traffic flows information between clusters and geographic information, and use ST-3DNet to predict the shared bike traffic flow. Finally, we conduct benchmark tests on New York's shared bicycle dataset. The experiment shows that our proposed GST-Net is superior to baseline, and the error rate is reduced by more than 15% compared with other models.