A Passenger Flow Prediction Method Based on a Multi‐Feature Fusion Hybrid Model

ABSTRACT

To address the limitations of traditional methods in modelling spatial interactions between metro stations and the incompleteness of feature learning, this paper proposes a combined passenger flow prediction model based on multi‐feature fusion, named MFGAT‐attention‐Mamba. The model integrates multiple correlated features such as weather, air quality and temporal proximity, holiday, weekly cycle feature and so on. A multi‐branch structure is designed to strengthen the association between spatial and temporal features of stations. A dual‐layer graph attention network (GAT) is employed to capture dependencies from three types of correlation graphs: a graph based on passenger OD travel patterns, a graph based on spatial distances between stations and a graph based on passenger flow similarity. These fused spatial features are then input into an attention mechanism and parallel Mamba blocks to extract temporal dependencies. The attention mechanism enhances the modelling of temporal patterns, whereas the Mamba module specialises in deep temporal sequence learning. Experiments conducted on five representative hub stations with the highest passenger flow​ in the Hangzhou metro dataset demonstrate that compared to baseline models (ARIMA, SVR, ASTGCN, STSGCN, GAT‐BiLSTM‐transformer), the proposed MFGAM model achieves higher prediction accuracy and lower prediction error across multiple time granularities. The results highlight the effectiveness of the proposed approach for short‐term passenger flow forecasting at major hub stations.