基于轻量时空图卷积模型的路网交通流预测

Traffic flow forecasting based on lightweight spatial-temporal graph convolution networks model

交通流预测是智能交通系统的重要组成部分。针对路网交通流天然具有的时空依赖性,结合交通流时序因果卷积与路网空间拓扑结构图卷积,提出一种基于递增式丢边的轻量时空图卷积神经网络模型,实现时空特征的有效融合,建立路网交通流高精度预测模型,提高交通流预测精度的同时降低其计算资源消耗、缩短预测响应时间。模型以单“三明治”式时空卷积模块为核心组件,减少时间卷积与空间卷积间的高计算消耗交互,有效提取交通流时空特征的同时保持整体结构轻量,其中的“厚夹心”空间图卷积采用多层图卷积网络以捕获远程高阶邻居节点信息、扩大空间感受野,并引入递增式丢边策略分阶处理邻居节点边,消解其潜在的过平滑。在模型训练中引入动态初始学习率,随模型训练进程演进动态调适学习率,进一步提升优化器性能,保证模型整体上的优越性。以真实基准交通流数据开展实验,对比分析本文所构建模型与多种相关基线模型的训练时间、预测精度等指标,并分析讨论所建模型在路网各节点上预测结果的离散性及其精度,解析多层图卷积可能具有的过平滑现象以及递增式丢边策略的消解能力。研究结果表明,本文所构建模型能有效捕获路网交通流的时空特性,以更少的训练时间获得更高的预测精度。

Traffic flow forecasting is one of the core components of intelligent transportation systems.Considering the fact that the dynamical evolution of network traffic flow naturally depends on the spatial structure of the road network and the temporal distribution of traffic flow,a lightweight spatio-temporal graph convolutional network model with incremental DropEdgewas proposed for accurate and timely traffic flow forecasting.It could integrate effectively the spatio-temporal characteristics of traffic flow,by combining the methods of causal convolution and graph convolution in a way with low computation resource consumption.In the proposed model,the crucial component was a single“sandwich”type of spatio-temporal convolution module.It could extract spatio-temporal characteristics from the traffic flow with a compact structure.To enlarge the conceptive filed and capture the information contained in long-range high-hop neighbors,a“thick decker sandwich”was adopted by using a multi-layer GCN and meanwhile.An incremental DropEdge strategy was proposed to eliminate the potential risk of over-smoothing.A scheme that dynamically adjusts the initial learnings rates with the evolution of training process of the proposed model was introduced to further improve the performance of the optimizer.It could ensure the overall superiority of the proposed model.A series of experiments were carried out on real traffic flow benchmark datasets,where the proposed model was compared with several baselines in terms of prediction accuracy,training time cost and so on.The dispersion degree and accuracy of prediction results of the proposed model at each node of the road network were analyzed and discussed as well,to validate the effectiveness of the proposed incremental DropEdge strategy to eliminate the potential risk of over-smoothing caused by the multi-layer GCN.The results show that,the proposed model can effectively capture the spatio-temporal characteristics of network traffic flow,and,compared with baselines,achieve higher prediction accuracy with lower training time cost.