Station-level demand prediction for bike-sharing systems planning with graph convolutional neural networks

Abstract

Accurately predicting bike-sharing demand at the station level is of paramount importance to facilitate station planning and enhance the efficiency of bike-sharing systems. In this study, we develop graph convolutional neural networks (GCNN) to predict station-level bike-sharing demand by modeling the spatial dependence of stations in two ways, namely trips and nearest neighbor, and compare the prediction performance with three machine learning models, including multiple linear regression, multi-layer perceptron (MLP) and random forest. The two GCNNs and three machine learning models were implemented and evaluated using a bike-sharing trip dataset in Zurich, Switzerland. The results show that the GCNN model based on the graph structure built by k nearest neighbor achieves the best prediction performance. The way of modeling spatial dependence of bike-sharing stations presents an influence on the prediction. This research is beneficial for decision-making in establishing new stations to support bike-sharing systems planning.