Multi-Granularity Temporal Embedding Transformer Network for Traffic Flow Forecasting.

Area of Science:

• Transportation Engineering
• Artificial Intelligence
• Data Science

Background:

• Traffic flow forecasting is crucial for intelligent transportation systems to mitigate congestion and accidents.
• Current models struggle with the heterogeneous and nonlinear nature of traffic data, often underutilizing temporal information.
• Effective prediction requires analyzing time period influences and integrating traffic flow characteristics across various time granularities.

Purpose of the Study:

• To address the limitations of existing traffic flow prediction models.
• To propose a novel network that effectively integrates multi-granularity temporal information and spatial characteristics.
• To enhance the accuracy and reliability of traffic flow forecasting.

Main Methods:

• Proposed a multi-granularity temporal embedding Transformer network (MGTEFormer).
• Developed an embedding input to merge complex temporal embeddings.
• Incorporated a temporal encoder for rich temporal information and a spatial encoder for sensor characteristics.
• Utilized an attention mechanism encoder and a linear regression layer for prediction.

Main Results:

• MGTEFormer demonstrated superior performance on real-world traffic datasets.
• The model achieved a reduction in mean absolute error by over 1.7% compared to existing benchmarks.
• Experimental and ablation studies validated the effectiveness of the proposed approach.

Conclusions:

• MGTEFormer significantly improves traffic flow forecasting by effectively leveraging multi-granularity temporal and spatial data.
• The proposed temporal embedding strategy reduces information loss, leading to more accurate predictions.
• The network offers a promising advancement for intelligent transportation systems.