Abstract: To address the limitations of existing traffic flow prediction models in effectively utilizing the fusion of traffic region and node features, as well as the lack of dynamic representation of spatiotemporal features. A Double-layer Multi-scale Dynamic Graph Convolution Network (DM-DGCN) was proposed for urban traffic prediction. First, a double-layer network architecture was adopted to fuse the spatiotemporal features of regions and nodes, with both node and regional traffic flow data being processed. Second, in the spatial dimension, a Spatial-dynamic Graph Convolution Nerwork(S-GCN) was constructed to capture dynamic spatial correlations. Next, in the temporal dimension, a Multi-Scale Temporal Convolution Network (MSTCN) was designed to capture potential temporal dependencies under different semantic environments. At the same time, the idea of spatial relationship learning was introduced into the temporal domain by designing a Temporal-dynamic Graph Convolution Network(T-GCN) to construct a time-varying temporal relationship matrix. Finally, a dynamic fusion module based on the attention mechanism was designed to integrate regional and node features, and the final prediction results being generated through a fusion layer. Experiments conducted on the Jinan and Xi’an traffic datasets show that, for the 60-minute prediction task, the DM-DGCN model reduces the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) by 5.7% and 5.56%, respectively, compared to the Spatio-Temporal Pivotal Graph Neural Network (STPGNN) model on the Jinan dataset; and by 3.7% and 4.18%, respectively, compared to the Hierarchical Spatio-Temporal Graph Ordinary Differential Equation network (HSTGODE) model on the Xi’an dataset. The DM-DGCN model outperforms existing baseline models and effectively captures dynamic multi-scale spatiotemporal dependencies in traffic data, accurately predicting future traffic flow.

Key words: traffic flow prediction, spatiotemporal feature fusion, Graph Convolutional Network (GCN), dynamic spatiotemporal graph, multi-scale time dependence

摘要： 针对现有交通流量预测模型不能有效利用交通区域特征与节点特征融合信息，以及缺乏对时空特征的动态表达的问题，提出一种用于城市交通预测的双层多尺度动态图卷积网络(Double-layer Multi-scale dynamic graph convolution network，DM-DGCN)模型。首先，采用双层网络架构，融合区域与节点的时空特征，同时处理节点和区域的交通流量数据。其次，在空间维度上，构建空间动态图卷积（S-GCN），提取动态空间相关性。再次，在时间维度上，考虑到不同语义环境下的潜在时间依赖关系，设计多尺度时间卷积模块（MSTCN），同时将空间关系学习思想引入到时间域，设计时间动态图卷积模块（T-GCN），构建随时间动态变化的时间关系矩阵。最后，设计基于注意力机制的动态融合模块，将区域特征与节点特征整合，通过融合层生成最终预测结果。在济南交通数据集和西安交通数据集上的实验结果表明，在60min预测任务中，DM-DGCN模型在济南数据集上的平均绝对误差（MAE）、均方根误差（RMSE）相较于时空关键图神经网络（STPGNN）模型分别降低了5.7%、5.56%；在西安数据集上，相较于层次时空图常微分方程（HSTODE）模型的MAE、RMSE分别降低了3.7%、4.18%。以上验证了DM-DGCN模型优于现有基线模型，能有效挖掘交通数据中动态多尺度时空依赖关系，准确预测未来交通流量。

关键词: 交通流量预测, 时空特征融合, 图卷积网络, 动态时空图, 多尺度时间依赖