ADST: Forecasting Metro Flow Using Attention-Based Deep Spatial-Temporal Networks with Multi-Task Learning.

Area of Science:

• Artificial Intelligence
• Urban Planning
• Transportation Engineering

Background:

• Passenger flow prediction is vital for traffic management and public safety.
• Existing models often overlook complex temporal correlations and spatial similarities in urban metro systems.
• Accurate prediction requires understanding intricate spatiotemporal dynamics.

Purpose of the Study:

• To propose an attention-based deep spatiotemporal network (ADST-Net) for citywide passenger flow prediction.
• To address the challenge of complex non-linear spatiotemporal correlations in urban metro data.
• To improve the accuracy and generalization of future passenger flow forecasts.

Main Methods:

• ADST-Net utilizes three channels to model recent, daily-periodic, and weekly-periodic spatiotemporal correlations.
• Employs residual networks, rectified blocks with attention mechanisms, and multi-scale convolutions for feature extraction.
• Incorporates external factors (e.g., weather) via an embedding mechanism and multi-task learning for enhanced generalization.

Main Results:

• ADST-Net successfully captures both steady trends and sudden changes in passenger flow.
• Demonstrates significant improvement and superior performance over state-of-the-art baseline methods.
• Experimental results on real-world datasets validate the model's effectiveness.

Conclusions:

• The proposed ADST-Net effectively models unique urban metro spatiotemporal correlations.
• The attention mechanism and multi-task learning contribute to robust and accurate passenger flow prediction.
• ADST-Net offers a promising solution for intelligent urban traffic management and public safety.