BikeNet: Accurate Bike Demand Prediction Using Graph Neural Networks for Station Rebalancing

Abstract

Bike sharing systems are widely operated in many cities as a green transportation means to solve the last mile problem and to reduce traffic congestion. One of the key challenges in operating high quality bike sharing systems is rebalancing bike stations from being full or empty. To this end, operators usually need to foresee the bike demands and schedule trucks to reposition bikes among stations. However, an accurate prediction of city-wide bike demands is not trivial due to the spatial correlation and temporal dependency of user mobility dynamics. Moreover, finding an optimal station rebalancing strategy from potentially enormous candidates is challenging given resource optimization objectives. In this work, we propose a two-phase framework to accurately predict city-wide bike demands and effectively rebalance bikes stations leveraging state-of-the-art deep learning techniques. First, we build a spatiotemporal graph neural network (ST-GNN) to model and predict city-wide bike demands, simultaneously capturing the spatial correlation by Graph Convolutional Networks (GCN) and the temporal dependency by Gated Recurrent Units (GRU). Then, we formulate the truck-based station rebalancing problem as an optimization problem with transportation cost objectives, and effectively solve the problem with Integer Linear Programming (ILP) algorithm. Experiments on real-world datasets from New York City validate the performance of the proposed framework, reducing 13% of prediction error and 5% of transportation cost compared with the baseline methods.