Cross-domain transformer spatial-temporal fusion network for traffic flow forecasting

Area of Science:

• Artificial Intelligence
• Data Science
• Transportation Engineering

Background:

• Traffic forecasting is complex due to road network interdependencies and unpredictable congestion.
• Existing models often overlook global cross-spatial-temporal dynamics, limiting predictive performance.
• Accurate traffic prediction is crucial for efficient transportation management and urban planning.

Purpose of the Study:

• To propose a novel plug-and-play fusion unit, the Cross-Domain Transformer Spatial-Temporal Fusion Network (CDTSTFN), for enhanced traffic forecasting.
• To improve the modeling of spatial-temporal dependencies by integrating cross-domain complementary information.
• To address information loss and fusion mismatches in existing spatial-temporal prediction models.

Main Methods:

• Developed a novel Cross-Domain Transformer Spatial-Temporal Fusion Network (CDTSTFN) incorporating two-stage fusion units.
• Integrated cross-domain complementary information to capture spatial-temporal dependencies.
• Augmented base spatial-temporal predictors with learned local-global spatial and short-long temporal dependencies.

Main Results:

• CDTSTFN demonstrated superior performance in traffic forecasting across six public datasets (PeMS03, PeMS04, PeMS07, PeMS08, METR-LA, PeMS-BAY).
• The model effectively compensated for information loss and resolved fusion mismatches.
• Evaluations confirmed the model's ability to learn complex cross-domain spatial-temporal patterns.

Conclusions:

• The proposed CDTSTFN significantly enhances traffic forecasting accuracy by effectively modeling cross-domain spatial-temporal patterns.
• The novel fusion unit offers a valuable tool for improving existing spatial-temporal prediction frameworks.
• This research contributes to more reliable traffic prediction systems for intelligent transportation.