A public transit passenger flow prediction method based on mode decomposition and deep learning under station classification

To address insufficient accuracy in public transit station passenger flow prediction, this study proposes a hybrid deep learning approach based on station classification using EMD-GCN-BiGRU. The method incorporates Dynamic Time Warping distance measurement with K-means clustering to classify stations based on temporal flow characteristics; applies Empirical Mode Decomposition to mitigate noise interference in original flow data; and constructs a deep learning framework integrating Graph Convolutinal Networks and Bidirectional Gated Recurrent Units to effectively capture spatial topological relationships between stations and temporal dependency features for category-specific predictions. Validation using Beijing bus card data demonstrates that transit stations can be classified into four categories based on weekday and non-weekday flow patterns: commuter-dominated, low-flow steady-state, multi-functional mixed, and evening outbound-dominated. Compared to baseline models, the proposed method improves R² by 0.01-0.31, reduces MAE by 0.12%-11.77%, and reduces RMSE by 1.97%-12.43%. Prediction accuracy is significantly enhanced after station classification compared to non-classified predictions.