Accurate spatiotemporal forecasting is of great significance in fields such as public health, environmental monitoring, and smart cities. In recent years, researchers have widely adopted spatiotemporal graph neural networks to model spatiotemporal dependencies in data and have achieved promising performance. However, existing methods still face challenges in characterizing dynamic spatial dependencies and modeling spatiotemporal interactions. These problems mainly stem from dependence on fixed graph structures, insufficient ability to model long-term series, and the lack of mechanisms for co-modeling global and local spatiotemporal features. We propose a novel spatio-temporal forecasting framework, namely the SpatioTemporal Dynamic Transformer Graph Convolutional Network (ST-DTGCN). In the temporal dimension, a Dynamic Transformer (DyTrans) module is introduced, which effectively enhances the model’s ability to model long-term dependencies through DyTanhNorm (DyT) and causal masking mechanisms. In the spatial dimension, an adaptive joint graph structure is constructed to characterize complex spatial dependencies. A temporal convolutional branch is introduced, and gated fusion is used to achieve co-modeling of local dynamics and global trends. Experimental results on four real datasets show that the proposed model outperforms existing state-of-the-art methods in terms of prediction accuracy and generalization ability. Meanwhile, in the efficiency analysis, ST-DTGCN achieved better performance while maintaining lower computational overhead.
Shi et al. (Wed,) studied this question.