Public Bicycle Flow Forecasting using Spatial and Temporal Graph Neural Network

Public bicycle systems (PBSs) that connect end users’ houses to public mass transportation are typically viewed as the "last-mile" of public transportation. Because of the limited capacity of stations in a PBS, the PBS operator must dispatch bicycles between stations to ensure that there are always bicycles/spaces available for bicycle borrowing/returning. However, because the variances in flows among stations are large, bicycle dispatch is difficult without a precise flow forecasting approach. In this paper, we propose an innovative approach to forecast bicycle flow on the basis of a graph neural network (GNN). Instead of processing the temporal information using RNN, or 1D-CNN, our approach integrates both spatial and temporal information into graphs, and analyzes them using graph convolution. Our approach works well on NYCitibike open dataset in terms of prediction accuracy. From the experiment, our approach shows it capability in accurate forecasting of peak flows and self-adjustment while perceiving abnormal flows caused by sporadic situations.