基于时空图卷积网络的高速路交通流多步预测  

Traffic flow multi-step prediction of expressway based on Space-Time Graph Convolutional Network

作　　者：高铭 梅朵[1] GAO Ming;MEI Duo(School of Information Science and Technology,Bohai University,Jinzhou 121000,Liaoning Province,China)

机构地区：[1]渤海大学信息科学与技术学院,辽宁锦州121000

出　　处：《信息技术》2024年第10期30-36,共7页Information Technology

基　　金：国家自然科学基金项目(62172057);科技部国家科技重大专项项目(2019YFD0901605)。

摘　　要：针对图卷积网络容易限制模型在交通流预测上有效的学习时空依赖问题,文中提出了一种时空图卷积循环网络(ST-GCRN)模型,首先通过时间卷积层以消除冗余的时间信息,其次,将图卷积网络与改进的门控循环网络相结合以获取时空依赖,最后通过加入残差的编解码结构解决模型训练梯度消失等问题,从而提高预测准确率,实现多步预测。在加利福尼亚州高速公路数据集上进行了实验,结果表明,该模型的平均绝对误差与均方根误差对比基准模型分别减少了11%、7.5%。To address the problem that graph convolutional network is easy to restrict the model’s effective learning of space-time dependence in traffic flow prediction,this paper proposes a Space-Time Graph Convolution Recurrent Network(ST-GCRN)model.Firstly,the time convolution layer is used to eliminate redundant time information.Secondly,the graph convolutional network is combined with the improved gated recurrent network to obtain the space-time dependence.Finally,the residual encoding and decoding structure is added to solve the problem that the model training gradient disappears,thus improving the prediction accuracy and realize multi-step prediction.The experiment results on the California highway dataset show that the mean absolute error and the root mean square error of the model compared with the benchmark model are reduced by 11%and 7.5%,respectively.

关 键 词：交通流预测 图卷积网络 时间卷积网络 门控循环网络 编解码结构 

分 类 号：U495[交通运输工程—交通运输规划与管理]