ESTGFormer: A spatio-temporal graph transformer with embedding and structure-aware loss for traffic forecasting.

Area of Science:

• Artificial Intelligence
• Graph Neural Networks
• Transportation Engineering

Background:

• Traffic forecasting is crucial for intelligent transportation systems.
• Complex spatiotemporal dependencies and non-stationary dynamics in road networks pose significant challenges.
• Existing models struggle with accurate multi-horizon forecasting.

Purpose of the Study:

• Introduce ESTGFormer, a spatiotemporal graph Transformer for accurate multi-horizon traffic forecasting.
• Address challenges of complex dependencies and non-stationary dynamics in road networks.
• Enhance robustness and temporal consistency of forecasts.

Main Methods:

• Developed ESTGFormer, integrating multi-head temporal self-attention with a serial spatial pathway (global spatial self-attention + learnable graph convolution).
• Employed lightweight embeddings combining periodic priors (time-of-day, day-of-week), node identity, and adaptive components.
• Proposed StructureAwareLoss, augmenting Huber loss with a structure-aware regularizer to align prediction errors across horizons.

Main Results:

• ESTGFormer achieved state-of-the-art accuracy on five benchmark datasets (METR-LA, PEMS-BAY, PEMS04, PEMS07, PEMS08).
• Demonstrated competitive computational efficiency.
• Ablation studies confirmed the effectiveness of temporal attention, serial spatial modeling, learnable embeddings, and StructureAwareLoss.

Conclusions:

• ESTGFormer provides stable and generalizable multi-step traffic forecasts.
• The model effectively captures latent, time-varying dependencies and road network topology.
• Offers a promising solution for intelligent transportation applications requiring accurate traffic prediction.