基于GAT-Attention-Mamba组合模型的轨道交通短时客流预测OA

Traditional methods for short-term passenger flow prediction in urban rail tran-sit often rely on a single feature or a single spatial correlation graph.To address these lim-itations,multiple associated features such as weather,air quality,and temporal proximity were incorporated.A short-term passenger flow prediction method based on the GAT-At-tention-Mamba hybrid model was proposed.A multi-branch structure was employed to enhance the correlation between temporal and spatial features of stations.A two-layer Graph Attention Network(GAT)was utilized to capture the dependencies from three types of relational graphs:Origin-Destination based passenger travel graph,spatial dis-tance-based station graph,and passenger flow similarity-based station graph.The fused relational features were subsequently fed into the Attention mechanism and parallel Mam-ba blocks to capture temporal dependencies.The Attention module strengthens the repre-sentation of temporal features,while the Mamba module focuses on the deep modeling of sequential patterns.Experimental results show that,compared with baseline models,the proposed method achieves a 16.3%～19.7%reduction in MAE and a 22.6%～32.9%re-duction in RMSE under 10～30 min forecasting horizons,validating its effectiveness in multi-feature fusion and spatiotemporal dynamic modeling.