Efficient bicycle-sharing rebalancing via demand distribution-based integer programming model and neural lagrangian...

The development of the artificial intelligence of things (AIoT) has transformed bicycle sharing systems into large-scale, spatially distributed networks of physical resources. While this transformation has strengthened their prominence as an effective solution for alleviating urban traffic congestion and reducing carbon emissions, uncertain user demand and uneven bicycle distribution often lead to supply-demand imbalances at stations. In practice, bicycle rebalancing strategy, which transfers bicycles among stations, is used to satisfy supply-demand balance. Integer Programming (IP) optimization formulation is widely used to address the redeployment of bicycles. However, existing IP rebalancing methods suffer from two major limitations: the use of historical average demand models that overlook fluctuations, and high computational complexity that prevents real-time application. To address these limitations, this paper proposes demand distribution-based integer programming with neural-Lagrangian acceleration (DD-IP-NLA) framework, where DD-IP captures demand fluctuations via demand distribution vectors, while NLA predicts near-optimal Lagrangian multipliers using a neural network. For large-scale systems, DBSCAN clustering with a distance-demand similarity metric further reduces computational complexity. Extensive experiments on synthetic and real datasets demonstrate that DD-IP-NLA can better satisfy user demand, improve profitability, and scale well for large-scale bicycle sharing systems in real time.