What if rebalancing fleets could adapt? A two-stage stochastic model for dynamic bike redistribution
提出了一个集成战术车队部署和运营车辆再平衡的优化框架，通过动态再平衡组和机器学习启发式算法，帮助共享单车系统管理者在需求不确定下高效决策，减少未满足需求。
• Presents an optimization framework that integrates tactical fleet deployment and operational vehicle rebalancing, providing unified decision support for bike-sharing systems. • Introduces dynamic rebalancing groups that adjust to changing demand in each period, improving flexibility compared to static grouping strategies. • Incorporates supervised machine learning-based models into a two-stage stochastic framework to approximate second-stage costs and reduce computational time. • Employs data-driven demand forecasting to create realistic rebalancing scenarios for rentals and returns. • Offers managers practical decision support for bike-sharing systems. Bike-sharing systems are an important mode of transportation, enabling individuals to rent bikes for short trips and return them to any station throughout the city. However, the dynamic nature of user arrivals at each station leads to imbalances between bike supply and demand, resulting in unsatisfied users. An essential challenge lies in efficiently deploying and scheduling rebalancing vehicles for bike redistribution, as these decisions have a considerable effect on the efficiency of the system. To tackle this challenge, we propose a dynamic rebalancing model that integrates tactical and operational decisions within a single optimization framework. Unlike approaches that treat these decisions separately, our model captures the interaction between the two: in the first stage, it determines how many vehicles should be deployed over the planning horizon (tactical decision), and in the second stage, it assigns stations to dynamic rebalancing groups and allocates vehicles to these groups in response to demand realizations (operational decisions). To address the computational challenge, we propose two approaches: an Improved Integer L-shaped decomposition algorithm and a heuristic that combines machine learning with an early stopping criterion to estimate the second-stage cost function. Moreover, we generate forecasts of rental and return demand and incorporate them into the optimization model to enhance decision-making under demand uncertainty. Our numerical results show that the proposed heuristic is highly effective in minimizing the unsatisfied demand while reducing the computational costs efficiently.