Integrating Spatial Heterogeneity and Nonlinear Dynamics for Interpretable Dockless Bike-Sharing Demand Prediction

Dockless bicycle-sharing systems (DBSS) have emerged as critical solutions for first/last-mile connectivity in urban transportation networks. While demand forecasting is fundamental to optimising system deployment and improving service efficiency, existing models often fail to account for the spatial heterogeneity of urban functional zones. In this study, we utilise social network-based check-in data to characterise the functional composition of urban areas, offering a dynamic and data-driven perspective on interpreting travel demand patterns. This study proposes a methodological framework integrating demographic, land-use and built environment variables to model DBSS demand patterns. The geographic detector method is employed to quantify the contributions of multidimensional factors, enhancing variable selection efficiency and improving model robustness. To address spatial heterogeneity and nonlinear dynamics, a high-precision hybrid model, GWRBoost, is proposed by combining the spatial heterogeneity-capturing capability of geographically weighted regression (GWR) with the nonlinear learning strength of extreme gradient boosting (XGBoost). Benchmark testing against conventional approaches demonstrates enhanced predictive accuracy and improved spatial explicability. The findings provide practical guidance for urban transportation planning and the optimisation of shared mobility systems.