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

Accurate traffic forecasting is a challenging task due to the complex spatiotemporal dependencies and nonstationary dynamics present in irregular road networks. In this work, we introduce ESTGFormer, a spatiotemporal graph Transformer for multi-horizon forecasting. The model integrates multi-head temporal self-attention with a serial spatial pathway. This pathway first applies global spatial self-attention and is followed by learnable graph convolution, which effectively separates non-local context encoding from topology-aware diffusion. To capture latent, time-varying dependencies, ESTGFormer employs lightweight embeddings that combine periodic priors, such as time-of-day and day-of-week, with node identity and adaptive components. To improve robustness and temporal consistency across time horizons, we propose StructureAwareLoss, which augments Huber loss with a structure-aware regularizer. This regularizer aligns prediction errors over different horizons, reducing training fluctuations. Our experiments on five benchmark datasets (METR-LA, PEMS-BAY, PEMS04, PEMS07, PEMS08) show that ESTGFormer achieves state-of-the-art accuracy while maintaining competitive computational efficiency. Ablation and sensitivity studies confirm the complementary roles of temporal attention, serial spatial modeling, learnable embeddings, and the proposed loss function, making the model capable of providing stable and generalizable multi-step forecasts for intelligent transportation applications.