基于多核自适应网络的高速公路交通流预测方法  

Highway Traffic Flow Prediction Method Based on Multi-Kernel Adaptive Network

作　　者：习金浩 孟峰 朱凤华[2] 贺正冰 李润梅[5] 吕宜生[2] XI Jin-hao;MENG Feng;ZHU Feng-hua;HE Zheng-bing;LI Run-mei;LÜYi-sheng(School of Artificial Intelligence,University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory for Management and Control of Complex Systems,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;National Energy Group Dayan Group,Hulunbuir 021000,China;Beijing Key Laboratory of Traffic Engineering,Beijing University of Technology,Beijing 100024,China;School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]中国科学院大学人工智能学院,北京100049 [2]中国科学院自动化研究所复杂系统管理与控制国家重点实验室,北京100190 [3]国家能源集团大雁集团(神宝能源),内蒙古呼伦贝尔021000 [4]北京工业大学交通工程北京重点实验室,北京100024 [5]北京交通大学电子信息工程学院,北京100044

出　　处：《交通运输研究》2021年第4期1-9,共9页Transport Research

基　　金：国家自然科学基金面上项目(71871010);广东省基础与应用基础研究基金项目(2019B1515120030)。

摘　　要：针对高速公路各路段交通流信息差异较大这一现象,为提高交通流预测准确率,将注意力机制引入卷积神经网络,建立描述交通流时空关联特征的多核自适应网络(Multi-Kernel Adaptive Network,MKAN)。首先对输入的历史交通流数据进行多分支卷积,获得不同尺度的交通流特征;然后根据输入信息自适应调整各卷积分支权重并对各分支多通道特征图进行加权融合;最后根据融合特征图,利用多层感知机预测下一时段交通流。基于加州交通运输部性能测试系统中的高速公路交通流数据设计实验进行模型验证和对比分析。实验结果表明,在大多数站点,MKAN模型的预测均方根误差和平均绝对误差低于长短期记忆网络、门控循环单元、K近邻算法和支持向量回归模型,对140号站点进行全天交通流预测,在1d内的各时段,MKAN模型预测绝对误差均小于其他对比模型;相比于单核卷积神经网络,在绝大多数站点,MKAN模型预测结果的均方根误差和平均绝对误差降低7%以上,对31号站点进行全天交通流预测,在1d内的大多数时段,MKAN模型预测绝对误差小于其他单核卷积神经网络。实验证明,多核自适应网络可有效提高交通流预测准确率,其预测效果优于部分传统预测模型和单核卷积神经网络。Aiming at the phenomenon that the traffic flow information of each section of the expressway is quite different,in order to improve the accuracy of traffic flow prediction,the attention mechanism was introduced into convolutional neural networks,and a multi-kernel adaptive network(MKAN)was designed to model the temporal-spatial relations among traffic flow data.First,multibranch convolution was performed on the input historical traffic flow data,and the traffic flow characteristic of different scales was obtained.Then,the weight of each convolution branch was self-adaptively adjusted according to the input information,the multi-channel feature maps of each branch were weighted and summed.Finally,according to the fusion feature maps,the multi-layer perceptron was used to predict the traffic flow in the next period.To carry out model verification and comparative analysis,the experiment was designed based on highway traffic flow data from California Department of Transportation(Caltrans)Performance Measurement System(PeMS).The experiment results showed that at most sites,the prediction root mean square error and the mean absolute error of MKAN were lower than that of long short-term memory network,gated recurrent unit,K-nearest neighbor algorithm and support vector regression model.Conducting whole-day traffic flow prediction for site No.140,the absolute prediction error of MKAN was smaller than other comparison models at each time of the day.Compared with single-kernel convolutional neural networks,the prediction root mean square error and the mean absolute error of MKAN were reduced by more than 7%at most sites.Conducting whole-day traffic flow prediction for site No.31,the absolute prediction error of MKAN was smaller than single-kernel convolutional neural networks at most time of the day.Experiments have proved that MKAN can effectively improve the accuracy of traffic flow prediction and its prediction effect is better than some traditional prediction methods and single-kernel convolutional neural networks.

关 键 词：城市交通 交通流预测 多核自适应网络 高速公路交通流 深度学习 

分 类 号：U491.14[交通运输工程—交通运输规划与管理]