Real-time reposition management of bike-sharing systems: a synchronous predict-then-optimize approach

As a convenient and low-carbon transport service to address the “last mile” problem, bike-sharing systems (BSSs) have been rapidly developed worldwide. However, the salient spatiotemporal imbalance between demand and supply has led to the bike-sharing repositioning problem (BSRP), aiming to reposition bikes from surplus stations to insufficient stations efficiently in BSS. This paper proposes a synchronous prediction then instantaneous optimization (SPtIO) approach, which consists of a multi-task multi-gate mixture of topology adaptive graph convolutional networks (3M−TAGCN) station relocation demand prediction model and a transformer policy-based reinforcement learning (TPRL) bike-sharing repositioning model. The 3M−TAGCN model makes synchronous and dynamic predictions for the real-time relocation demands of all stations by learning features and relationships from historical inflow and outflow spatiotemporal data. Leveraging the advantage of “offline training + online optimizing”, the TPRL model instantaneously figures out the dynamic BSRP based on predicted relocation demands. The policy of the TPRL model consists of a transformer network and a mask method, and the training process incorporates the policy gradient algorithm. Experiments on the New York Citi Bike dataset demonstrate that the 3M−TAGCN prediction model outperforms other baseline models in various scenarios. The TPRL bike-sharing repositioning model effectively determines near-optimal repositioning schemes. Evident results have shown significant improvements in the proposed SPtIO approach over the service quality and repositioning efficiency of BSSs.