基于ARMA过滤器的时空图卷积网络短时交通流预测  

Short-Term Traffic Flow Prediction Based on Spatial-Temporal Graph Convolutional Network with ARMA Filter

作　　者：肖培成 曹阳 沈琴琴 施佺 XIAO Peicheng;CAO Yang;SHEN Qinqin;SHI Quan(School of Transportation and Civil Engineering,Nantong University,Nantong,Jiangsu 226019,China)

机构地区：[1]南通大学交通与土木工程学院,江苏南通226019

出　　处：《计算机工程与应用》2024年第21期308-314,共7页Computer Engineering and Applications

基　　金：国家自然科学基金(61771265);江苏高校“青蓝工程”;南通市科技计划(JC2021198)。

摘　　要：针对大多数现有的时空融合图卷积网络模型在分析交通流数据所采用的过滤器提取空间特征时,可能会导致网络节点特征过于平滑从而丢失原始信息、计算量大等问题,将一种能有效逼近任何所需响应的基于自回归移动平均(ARMA)过滤器的图卷积网络与门控循环单元(GRU)相融合,提出一种基于ARMA过滤器的时空图卷积网络短时交通流预测模型。在该模型中,采用GRU提取交通流数据的时间特征;利用基于ARMA过滤器的图卷积网络提取空间特征;通过平均池化层得到预测结果。在美国加州高速公路公开数据集PeMS04、PeMS07和PeMS08上进行了实验。结果表明,相比于切比雪夫多项式等传统的过滤器,ARMA过滤器能够更好地捕获交通流数据中的空间特征且能有效提升计算效率,在三组数据集上相比于GRU、DCRNN、T-GCN、ASTGCN和MSTGCN等基准模型,新模型的平均绝对误差指标平均降低了26.28%、28.25%、39.59%、26.24%和19.07%,训练时间平均降低了26.54%、95.00%、78.87%、93.58%和89.03%。In view of the problems that the filter used by most existing spatial-temporal fusion graph convolutional network models in the analysis of traffic flow data may lead to the loss of original information when extracting spatial features due to too smooth network node features and large amount of computation,an auto-regressive moving average(ARMA)filter based graph convolutional network which can effectively approximate any required response is integrated with the gated recurrent unit(GRU).An ARMA filter based spatio-temporal graph convolutional network short-term traffic flow prediction model is proposed.In this model,the GRU is used to extract the temporal features of traffic flow data.Then,a graph convolutional network based on the ARMA filter is used to extract the spatial features.Finally,the prediction results are obtained by means of the average pooling layer.Experiments are conducted on California Highway open datasets PeMS04,PeMS07 and PeMS08 in the United States.The results show that compared with Chebyshev polynomial and other traditional filters,ARMA filter can better capture the spatial correlation in traffic flow data and effectively improve the computational efficiency.Compared with GRU,DCRNN,T-GCN,ASTGCN and MSTGCN,the proposed model performs much better.The average absolute errors are decreased by 26.28%,28.25%,39.59%,26.24%and 19.07%on average,and the training times are decreased by 26.54%,95.00%,78.87%,93.58%and 89.03%on average.

关 键 词：交通流预测 图卷积网络 ARMA过滤器 门控循环单元 时空特征提取 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] U491[自动化与计算机技术—控制科学与工程]