TriDiffSTG: traffic flow prediction with spatiotemporal graph using triplet attention and diffusion model

Traffic flow prediction represents a core task within intelligent transportation systems, significantly influencing urban congestion mitigation and traffic efficiency enhancement. However, existing spatiotemporal graph neural network methods face substantial challenges in modeling intricate spatiotemporal dependencies. Traditional models predominantly employ decoupled sequential modeling strategies for temporal and spatial dimensions, failing to effectively capture deeper interactive patterns across temporal, spatial, and channel dimensions. To address these challenges, this paper proposes TriDiffSTG, a novel three-dimensional collaborative spatiotemporal graph prediction framework grounded in diffusion modeling. Central to this framework is TriGraphNet, a structured denoising network utilizing triple graph convolution architecture to achieve unified modeling across temporal, spatial, and channel dimensions. Experimental results across multiple real-world traffic datasets demonstrate that TriDiffSTG significantly outperforms existing mainstream approaches in prediction accuracy, exhibiting superior generalization capabilities in multi-scale pattern modeling and adaptive dynamic structural learning. Ablation studies highlight the pivotal roles of the triple attention mechanism and Convolutional Block Attention Module in enhancing model performance. This research introduces a new paradigm for spatiotemporal graph prediction tasks, featuring generative modeling capability and structural scalability, which holds considerable potential for practical engineering applications and real-world deployment.