Data-driven rebalancing support for station-based bike sharing: A hybrid centralized-distributed deployment in Chiba city

Station-based bike-sharing systems routinely experience spatiotemporal imbalances that lead to empty and full stations, thereby reducing travel opportunities. Although the literature offers many optimization-based rebalancing models, including dynamic formulations that re-optimize over a rolling horizon, live operations still often depend on field staff’s heuristics because rigid centrally prescribed routes are difficult to maintain under traffic uncertainty and rapidly changing station conditions. A further obstacle is that such methods must be operated continuously: the computational budget, real-time data pipelines, and tight central–field coordination they presuppose are difficult to sustain for many smaller or regional operators, for whom rebalancing must instead be absorbed into routine field work at minimal marginal cost. This study proposes a hybrid centralized–distributed rebalancing framework for HELLO CYCLING, a nationwide shared-mobility platform in Japan, and examines its deployment in the electric bike-sharing network of Chiba City. The proposed procedure computes time-of-day target inventory levels from historical demand and logged empty/full episodes, then prepares a station-level decision table that lists target inventory ranges and associated priorities. Field staff compare these targets with current station inventories and use the table as decision support when choosing which relocations to execute. The scheme adopts relatively simple rule-based calculations rather than large-scale optimization, keeping computational costs minimal and allowing straightforward integration into daily field operations. The framework is examined through a live two-week deployment and compared with a pre-intervention operational baseline and a randomized relocation baseline. The empirical analysis shows spatial and temporal heterogeneity in demand and indicates that many stations face exposure to both empty and full states, motivating intraday repositioning. During the deployment period, the proposed workflow shows higher avoided demand losses per relocated bicycle than both the randomized baseline and routine operations. The results suggest that the proposed support scheme can identify stations requiring near-term intervention and can also reveal stations that were more likely to be overlooked under knowledge- and experience-based operations. Although the present deployment relied on contractor-based execution, the same logic may also provide a basis for more distributed user-side rebalancing schemes in the future.