User-Behaviour-Based Dynamic Clustering Optimisation Algorithm for True Demand Prediction of Shared Bikes

Dockless bike-sharing systems are increasingly important for sustainable urban mobility, yet they frequently suffer from spatial misallocation between supply and demand. Conventional demand forecasting relies primarily on historical trip records and therefore systematically underestimates true demand when users abandon bike-finding attempts. To address this limitation, we propose a user-behaviour-based dynamic clustering optimisation algorithm that integrates observed riding behaviour with latent unmet demand through the concept of Golden Distance—a district-adaptive service-radius threshold estimated heuristically from operational logs of successful rides and inferred unmet-demand events, serving as a behaviourally motivated proxy for aggregate walking tolerance. Building on a prior AIoT-enabled demand-prediction framework, the method first applies HDBSCAN density-based clustering to discover intrinsic demand topology, then selectively refines only those clusters that violate Golden Distance coverage constraints via an elongation-aware adaptive k-medoids formulation. District-level true demand is predicted using an XGBoost regression model and subsequently downscaled to the cluster level based on historical activity shares. Experiments on one full year of operational data from three Hong Kong districts (Tseung Kwan O, Sha Tin, and Tuen Mun) show that the proposed pipeline produces Golden-Distance-compliant service zones at a substantially finer operational resolution than the DBSCAN baseline: 86.5% to 91.6% of clustered demand points lie within the Golden Distance district of their assigned medoid, at cluster coverages of 68.0% to 78.5%, against 60.0% to 71.1% at coverages of 49.7% to 88.1% for the baseline. A resolution-fair evaluation shows that the cluster-level RMSE advantage reported previously largely reflects the finer reporting unit rather than better prediction: the same fixed district-level forecast spans a five- to seven-fold range of RMSE when scored on progressively coarser spatial units, and at a matched clustering resolution the DBSCAN baseline equals or exceeds the proposed pipeline on cluster-level metrics. Accuracy is therefore evaluated with scale-free metrics, on which the proposed method is not the more accurate of the two, and the contribution of this work is positioned on operationally deployable, behaviourally constrained spatial zoning rather than on per-cluster forecasting accuracy.