Abstract: In intelligent transportation systems, traffic flow prediction plays a crucial role. However, traffic data often exhibit significant dynamic spatial correlations and sudden pattern changes, making it difficult for traditional models to comprehensively capture spatiotemporal dependencies in real-world transportation networks. To address this challenge, a bidirectional diffusion dynamic graph convolutional traffic flow prediction model, BDDGNet, was proposed by integrating multi-source information and graph-level attention. The model integrates spatiotemporal embedding and various dynamic features, and utilizes a multi-head attention mechanism to construct dynamic adjacency relationships between nodes, capturing temporal evolution of spatial connectivity in traffic network. For spatial modeling, a bi-directional diffusion graph convolution module was designed to learn propagation dynamics of node states from both forward and reverse directions, enabling the collaborative modeling of spatial dependencies in traffic flow. In addition, a graph-level contextual attention mechanism was introduced to extract global semantic representations of key nodes, thereby enhancing model’s understanding of overall topological structure of traffic network. Experimental results on two real-world traffic flow datasets, PEMSD4 and PEMSD8, demonstrate that proposed BDDGNet achieves superior numerical performance for 15-minute, 30-minute, and 60-minute prediction tasks. The overall average results show that on PEMSD4, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) reach 18.26, 30.26, and 12.16%, respectively, representing improvements of 25.2%, 19.3%, and 26.9% compared with Diffusion Convolutional Recurrent Neural Network (DCRNN) based on static graph structures. On PEMSD8, MAE, RMSE, and MAPE reach 14.07, 23.44, and 9.24%, corresponding to improvements of 16.3%, 11.1%, and 15.4% over DCRNN. These results indicate that proposed model can more accurately capture spatiotemporal dependencies in dynamic traffic environments and significantly enhance prediction accuracy.

摘要： 在智能交通系统中，交通流预测至关重要。然而，交通数据存在显著的动态空间相关性与突发性模式，这使得传统模型难以全面建模真实交通网络的时空依赖特性。为解决该问题，提出了一种融合多源信息与图级注意力的双向扩散动态图卷积交通流预测模型BDDGNet。该模型融合时空信息与多源动态特征通过多头注意力机制构建节点间的动态图邻接关系，捕捉交通网络中随时间演化的空间连接强度。在空间建模方面，使用双向扩散图卷积模块，从不同路径方向学习节点状态的传播动态，实现对交通流中空间依赖的协同建模，同时，模型引入图级注意力机制，以提取关键节点的全局语义信息，增强对整体交通网络拓扑结构的感知能力。在PEMSD4和PEMSD8两个真实交通流数据集上的实验结果表明，本文所提 BDDGNet 在 15 分钟、30 分钟和 60 分钟预测任务中均取得了更优的数值表现。综合平均结果显示：在 PEMSD4 上，平均绝对误差（MAE）、均方根误差（RMSE）和平均绝对百分比误差（MAPE）分别达到18.26、30.26和12.16，相对于基于静态图结构的扩散卷积循环神经网络（DCRNN）提升了25.2%、19.3% 和 26.9%；在 PEMSD8上，MAE、RMSE和MAPE分别达到14.07、23.44和9.24，相对于 DCRNN 提升了16.3%、11.1%和15.4%。结果表明，所提模型能够在多变的交通环境下更准确地刻画时空依赖关系，并显著提升预测精度。