基于图小波注意力门控循环神经网络的交通流预测  

TRAFFIC FLOW PREDICTION BASED ON GRAPH WAVELET ATTENTION-GATED RECURRENT NEURAL NETWORK

作　　者：李松江[1] 黄小莉 王鹏[1] Li Songjiang;Huang Xiaoli;Wang Peng(College of Computer Science Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机应用与软件》2022年第12期89-95,共7页Computer Applications and Software

基　　金：吉林省科技发展计划技术攻关项目(20190302118GX)。

摘　　要：针对现有交通流预测方法无法对局部空间及动态时间建模的问题,提出一种图小波注意力门控循环神经网络模型(GW-AGRU)。将道路网络的空间信息以图的形式表示,运用基于小波变换的图卷积神经网络从图节点中提取邻近特征;在门控循环单元中融入注意力机制,充分挖掘交通数据的时间相关性;融合时空特征进行回归预测。在真实数据集上的实验结果表明,所提方法的预测性能均优于其他模型,能够有效地预测长期的交通流量。To solve the problem that the existing traffic flow prediction methods cannot model the local space and dynamic time, this paper proposes a graph wavelet attention-gated recurrent unit(GW-AGRU). The spatial information of the road network was represented as a graph. We used the graph wavelet neural network to extract the neighboring features from graph nodes, and incorporated the attention mechanism in the gated recurrent unit to fully mine the time correlation of traffic data. Spatio-temporal features were integrated for regression prediction. Experiments on real traffic flow datasets show that the predictive performance of the proposed method is better than other models, which can effectively predict the long-term traffic flow.

关 键 词：小波变换 图卷积网络 注意力机制 门控循环单元 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]