Dynamic Spatial-Temporal Memory Augmentation Network for Traffic Prediction.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Smart city development relies heavily on accurate traffic flow prediction.
• Existing models struggle with complex spatio-temporal dynamics, hierarchical temporal features, and spatial heterogeneity.
• Effective traffic management requires advanced models to capture these nuances.

Purpose of the Study:

• To introduce a novel model, DSTMAN, for enhanced traffic flow prediction.
• To address limitations in capturing dynamic spatio-temporal contexts and spatial heterogeneity.
• To improve the accuracy and efficiency of traffic flow forecasting.

Main Methods:

• Developed three spatial-temporal embeddings to capture dynamic contexts and encode time/space characteristics.
• Integrated embeddings into a multi-scale block for hierarchical spatio-temporal dependency extraction.
• Introduced a meta-memory node bank for adaptive neighborhood graphs and secondary memory mechanism to learn spatial heterogeneity.

Main Results:

• DSTMAN outperformed benchmark models like MTGNN, DCRNN, and AGCRN on public datasets (METR-LA, PEMS-BAY).
• Achieved significant Mean Absolute Error (MAE) reductions: 4% vs. MTGNN, 6.9% vs. DCRNN, and 5.8% vs. AGCRN on METR-LA.
• Demonstrated superior performance in managing spatio-temporal correlations and spatial heterogeneity.

Conclusions:

• DSTMAN offers a robust solution for complex traffic flow prediction tasks.
• The model's novel architecture effectively captures intricate spatio-temporal relationships.
• DSTMAN represents a significant advancement for intelligent transportation systems and smart city initiatives.