A Comprehensive Framework for Dynamic Bike Rebalancing in a Large Bike Sharing Network

Bike sharing is a vital component of a modern multi-modal transportation system. However, its implementation can lead to bike supply-demand imbalance due to fluctuating spatial and temporal demands. This study proposes a comprehensive framework to develop optimal dynamic bike rebalancing strategies in a large bike sharing network. It consists of three components, including a station-level pick-up/drop-off prediction model, station clustering model, and capacitated location-routing optimization model. For the first component, we propose a powerful deep learning model called graph convolution neural network model (GCNN) with data-driven graph filter (DDGF), which can automatically learn the hidden spatial-temporal correlations among stations to provide more accurate predictions; for the second component, we apply a graph clustering algorithm labeled the Community Detection algorithm to cluster stations that locate geographically close to each other and have a small net demand gap; last, a capacitated location-routing problem (CLRP) is solved to deal with the combination of two types of decision variables: the locations of bike distribution centers and the design of distribution routes for each cluster.