LTiT: A Deep Learning Model for Subway Section Passenger Flow Prediction Based on LSTM-TSSA-iTransformer

Area of Science:

• Urban transportation systems
• Predictive modeling
• Data science

Background:

• Subway passenger flow prediction is critical for urban transit efficiency.
• Current methods often focus on overall passenger numbers or origin-destination demand.
• Predicting passenger flow within specific subway sections offers better insights for management.

Purpose of the Study:

• To develop an advanced model for subway section passenger flow prediction.
• To improve the accuracy and robustness of short-term passenger flow forecasting.
• To support better resource allocation and operational management in subway systems.

Main Methods:

• Proposed a novel subway section passenger flow prediction model named LTiT (LSTM-TSSA-iTransformer).
• Utilized an iTransformer encoder architecture combined with Long Short-Term Memory (LSTM) for temporal feature extraction.
• Incorporated Token Statistics Self-Attention (TSSA) to adaptively weight critical temporal information.

Main Results:

• The LTiT model demonstrated superior performance compared to baseline models (SARIMA, BP, LightGBM, LSTM).
• Achieved higher accuracy based on evaluation metrics including R², MAE, MSE, and MAPE.
• Showcased robustness under various noise conditions and optimized input parameters.

Conclusions:

• The proposed LTiT model offers a significant advancement in subway section passenger flow prediction.
• This method provides a more direct reflection of passenger dynamics on line segments.
• The findings support enhanced management and resource allocation in urban public transportation.