Abstract: The accuracy of traffic flow prediction directly affects the dispatching efficiency of intelligent transportation systems and indirectly determines the quality of travel experience. In response to the deficiencies of existing models in complex spatio-temporal dependency modeling, dynamic road network relationship processing, and handling of uncertain disturbances, a Multiscale Uncertianty-aware Dual-Graph Collaborative Learning(MUDGCL) is proposed. Firstly, the traffic sequence is decoupled into trend and disturbance terms through wavelet decomposition, and multi-scale features are captured by time attention mechanism and temporal convolutional network respectively, with adaptive fusion achieved through gated units. Meanwhile, a dual-graph collaborative framework is constructed, combining the probabilistic modeling of variational graph generator and the dynamic graph structure of periodic time embedding to jointly model stable topology and time-varying spatial correlations. Additionally, node importance masks and Kullback-Leibler Divergence(KL) constraints are introduced to enhance model robustness. Experimental results on PEMS04, PEMS08, METR-LA, and PEMS-BAY datasets demonstrate that MUDGCL achieved reductions of 1.57% in Mean Absolute Error(MAE) and 3.07% in Root Mean Square Error (RMSE) compared to Multi-scale Time-Enhanced Graph Convolutional Recurrent Network(MTEGCRN) on PEMS04. On PEMS08, decreases of 5.98% in MAE, 5.64% in RMSE, and 2.61% in Mean Absolute Percentage Error(MAPE) were observed relative to Temporal-Frequency Self-supervised Guided Graph Learning(TFSGGL). For multi-step prediction tasks on METR-LA (3/6/12-step), MAE was reduced by 1.47%, 2.59%, and 1.45% respectively compared to Random Graph Diffusion Attention Network(RGDAN). On PEMS-BAY, MAE decreased by 2.99%, 2.41%, and 3.55% respectively relative to Wavelet Decomposition and Structure-Enhanced Graph Attention Method(WD-SEGAM). Furthermore, MUDGCL outperforms the existing mainstream models in all evaluation metrics, effectively improving the ability to model complex spatio-temporal dependencies, the accuracy of uncertainty quantification in prediction, and generalization performance.

Key words: traffic flow prediction, spatiotemporal dependency, graph neural network, uncertainty modeling, multiscale analysis

摘要： 交通流量预测精度直接影响智能交通系统的调度效率，并间接决定出行体验质量。针对现有模型在复杂时空依赖建模、动态路网关系处理及不确定扰动应对方面的不足，提出了一种多尺度不确定性双图协同学习模型(MUDGCL)。首先，通过小波分解将交通序列解耦为趋势项与扰动项，分别采用时间注意力机制和时序卷积网络捕获多尺度特征，并通过门控单元实现自适应融合；同时构建双图协同框架，结合变分图生成器的概率建模与周期性时间嵌入的动态图结构，共同建模稳定拓扑与时变空间关联；此外引入节点重要性掩码与KL散度约束以增强模型鲁棒性。在PEMS04、PEMS08、METR-LA和PEMS-BAY数据集上的实验结果表明，MUDGCL在PEMS04数据集上较多尺度时间增强图卷积循环网络(MTEGCRN)的平均绝对误差(MAE)、均方根误差(RMSE)分别降低了1.57%、3.07%；在 PEMS08 数据集上，相较时频自监督引导图学习方法(TFSGGL)的MAE、RMSE和平均绝对百分比误差(MAPE)分别下降了5.98%、5.64%、2.61%；在METR-LA数据集的3/6/12步多步预测任务中，相较随机图扩散注意力网络(RGDAN)，MAE分别降低了1.47%、2.59%、1.45%；在PEMS-BAY数据集的多步预测任务中，相较小波分解-结构增强图注意力方法(WD-SEGAM)，MAE分别下降了2.99%、2.41%、3.55%。MUDGCL在各项评价指标上优于现有主流模型，有效提升了复杂时空依赖性建模能力、预测不确定性量化精度及泛化性能。

关键词: 交通流预测, 时空依赖, 图神经网络, 不确定性建模, 多尺度分析