Dual-temporal inflow-outflow dependency modeling for short-term metro outflow prediction.

Area of Science:

• Deep Learning
• Transportation Science
• Urban Mobility Analytics

Background:

• Current deep learning models struggle with metro passenger flow prediction due to inadequate modeling of inflow-outflow dependencies.
• Predefined station correlation graphs limit flexibility and representational capacity in existing approaches.

Purpose of the Study:

• To propose a novel dual-temporal inflow-outflow dependency model (DTIOD) for accurate short-term metro passenger flow prediction.
• To address limitations in modeling inflow-outflow dependencies and spatial correlations.

Main Methods:

• Decomposing inflow influence into short-term and long-term temporal components.
• Employing an asymmetric feature extraction scheme and a dual-branch cross-attention mechanism for implicit spatial correlation learning.
• Incorporating sample-level origin-destination (OD) matrices as attention biases.

Main Results:

• DTIOD achieved significant reductions in RMSE (10.75%), MAE (11.60%), and WMAPE (6.84%) compared to baseline models.
• The model demonstrated practical applicability with efficient training times (under 70 seconds).

Conclusions:

• DTIOD offers a superior balance between predictive accuracy and computational efficiency for metro passenger flow forecasting.
• The model's ability to implicitly learn spatial correlations and model temporal dependencies enhances prediction performance.