Feature-enhanced iTransformer: A two-stage framework for high-accuracy long-horizon traffic flow forecasting.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Intelligent Transportation Systems (ITS) rely on accurate long-horizon traffic flow prediction.
• Existing Transformer-based models often couple feature extraction and prediction, limiting domain information integration.
• Complex spatio-temporal dependencies in traffic data pose significant prediction challenges.

Purpose of the Study:

• To propose a novel two-stage framework, FE-iTransformer, for enhanced traffic flow prediction.
• To address limitations of end-to-end models by decoupling feature extraction and prediction.
• To provide a deployment-ready, graph-free alternative for traffic prediction in ITS.

Main Methods:

• Developed a Feature Enhancement Module (FEM) to distill global context from spatio-temporal dynamics, periodicity, and temporal context.
• Implemented a per-step feature enhancement mechanism to enrich input sequences with a global context vector.
• Utilized an iTransformer backbone for sequence prediction on the enriched representations.

Main Results:

• Ablation studies on Freeway and Urban datasets confirmed the efficacy of the two-stage design and FEM.
• Experiments on the PEMS08 benchmark demonstrated scalability and improved long-horizon performance.
• Achieved a 19.1% reduction in Mean Absolute Error (MAE) for 120-minute forecasting compared to the vanilla backbone.

Conclusions:

• The FE-iTransformer framework effectively enhances traffic flow prediction accuracy.
• The proposed FEM and two-stage approach offer significant improvements over standard iTransformer models.
• FE-iTransformer provides a robust, graph-free solution for traffic prediction when graph data is unavailable or unreliable.