Spatiotemporal and land use factors in urban rail transit passenger flow prediction

Urban rail transit passenger flows exhibit pronounced spatiotemporal heterogeneity and are strongly influenced by complex patterns of land-use mixing, which challenges the robustness and interpretability of existing models. In this study a multimodule collaborative framework was proposed for passenger-flow prediction and causal factor analysis. A three-dimensional space–time–land-use feature engineering scheme was developed incorporating distance- and time-attenuation mechanisms to capture spatiotemporal dependence, and Shannon entropy to characterise the degree of land-use mixing, resulting in an 11-dimensional structured feature matrix. A spatiotemporal enhanced random forest model was designed with dual-objective optimisation to balance prediction accuracy and interpretability. Shapley Additive exPlanations values quantified factor contributions and causal forests verified causal relationships. Empirical validation using 30 days of data from 128 metropolis stations shows a global prediction R2 of 0.995 and stable performance across multiple spatial clustering scenarios, demonstrating high accuracy, robustness, and strong spatial adaptability.