Research on rebalancing of large-scale bike-sharing system driven by zonal heterogeneity and demand uncertainty

Abstract

Bike-sharing, as an open and intricate system, encompass a vast and diverse array of data, and are often affected by various time-varying and uncertain factors. Consequently, employing scientific and appropriate rebalancing in the age of big data is pivotal for the system’s sustained and healthy development. This paper takes the dockless bike-sharing system as the research object, considers regional heterogeneity and time-varying demand uncertainty, and proposes a rebalancing strategy that integrates initial inventory determination and remaining inventory uncertainty. Firstly, this paper considers using of Poisson distribution chi-square tests to assess the borrowing and returning behaviors within a zone, selects a unit time tailored to the zone’s unique circumstances to estimate the parameter rate, and independently establishes a non-stationary Markov chain for each zone to determine the initial expected inventory under dynamic zonal demand conditions. Using big data from bike-sharing operations in Nanjing for empirical validation, the results indicate that borrowing and returning behavior in most zones follows a Poisson distribution, and that zones with higher volumes of traffic have fewer bikes initially deployed. Secondly, we address the uncertainty of demand-driven residual inventory and spatial correlations, formulating a robust two-stage optimization model aimed at minimizing the worst-case scenario. We then transform this model into a computationally tractable form using polyhedral uncertainty sets. By analyzing the model structure and mathematical properties, we develop a column-and-constraint generation algorithm for customizing a two-stage robust optimization model based on residual inventory, and compare it with other traditional algorithms. The numerical experimental results show that the proposed model and algorithm have significant advantages in terms of solution accuracy and efficiency, and can play a role in real-world problems. Finally, the paper discusses the impact of various parameters in the model on the solution, yielding results consistent with our expectations.