Bidirectional Spatial-Temporal Adaptive Transformer for Urban Traffic Flow Forecasting.

Area of Science:

• Intelligent Transportation Systems (ITS)
• Traffic Forecasting
• Deep Learning

Background:

• Existing traffic forecasting methods overlook task complexity variations and the value of historical traffic data.
• Accurate urban traffic forecasting is crucial for efficient intelligent transportation systems.

Purpose of the Study:

• To propose a novel bidirectional spatial-temporal adaptive transformer (Bi-STAT) for enhanced urban traffic forecasting.
• To address the nonuniform complexity of forecasting tasks across different times and locations.
• To leverage past traffic conditions to improve future predictions.

Main Methods:

• Developed a Bi-STAT model with an encoder-decoder architecture incorporating spatial-adaptive and temporal-adaptive transformers.
• Implemented a dynamic halting module (DHM) for adaptive processing of traffic streams based on task complexity.
• Utilized a dual-decoder approach for present-to-past recollection and present-to-future prediction tasks.

Main Results:

• Demonstrated the effectiveness of individual modules within the Bi-STAT architecture.
• Achieved superior performance compared to state-of-the-art baselines on four benchmark traffic datasets.
• Showcased improved generalization through the integration of past traffic condition recollection.

Conclusions:

• Bi-STAT effectively models the intrinsic properties of traffic forecasting, namely task complexity and temporal dependencies.
• The proposed adaptive mechanisms and dual-decoder strategy significantly enhance prediction accuracy.
• Bi-STAT represents a significant advancement in accurate and reliable urban traffic forecasting.