基于特征强化神经网络的交通流预测  

Research on traffic flow prediction based on feature enhancing neural network model

作　　者：李晓霞[1] 石莹洁 祁昌平[1] 林和[3] LI Xiaoxia;SHI Yingjie;QI Changping;LIN He(School of Information Technology and Media,Hexi University,Zhangye 734000,China;School of Hydrology and Water Resources,Hohai University,Nanjing 210098,China;School of Information Science and Engineering,Lanzhou University,Lanzhou 730000,China)

机构地区：[1]河西学院信息技术与传媒学院,甘肃张掖734000 [2]河海大学水文水资源学院,江苏南京210098 [3]兰州大学信息科学与工程学院,甘肃兰州730000

出　　处：《应用科技》2023年第6期21-27,共7页Applied Science and Technology

基　　金：裕固族口头文化有声数据库建设项目(19BYY186);2022年甘肃省高等学校创新基金项目(2022B-180);2021年河西学院科研创新与应用校长基金重点项目(XXZD2021001).

摘　　要：为了解决挖掘特征能力不足的问题,充分提取不同出行高峰日特征来提高交通流预测的准确性,本文提出了基于特征强化的交通流预测模型——卷积神经网络(convolutional neural network,CNN)–特征增强(feature enhancement,FE)–门控循环单元(gate recurrent unit,GRU)(CNN-FE-GRU)神经网络模型。该模型首先采用CNN来挖掘不同时间序列下路口与车流量的潜在联系,有效地捕捉时序数据的特征;其次提出FE模块来增加模型聚焦能力,刻画不同时间节点对目标时间节点交通流的相关性,同时为每个时间节点赋予相应权重;最后采用GRU对输出的时序数据进行进一步的特征提取,并在全连接层的作用下实现交通流的预测。实验结果表明,CNN-FE-GRU模型的平均绝对误差(mean absolute error,MAE)、均方误差(mean-square error,MSE)和均方根误差(root mean square error,RMSE)平均值分别为0.229、0.1308、0.3388,相较于对比模型均有不同程度下降,CNN-FE-GRU模型在精确度和预测性能方面都有了显著提升。To address the issue of inadequate ability to mine features,a feature-enhanced traffic flow prediction model—CNN-FE-GRU,namely convolutional neural network(CNN)-feature enhancement(FE)-gated recurrent unit(GRU)neural network is proposed to fully extract daily features of different travel peak days to improve the accuracy of traffic flow prediction.The CNN-FE-GRU model first uses CNN to explore the potential connection between intersections and traffic flow under time series,and then effectively captures the features of time series data.And then,FE module is proposed to increase the focusing capability of the model to characterize the correlation of different time nodes to the traffic flow at the target time node,and to assign corresponding weight to each time node.Finally,GRU is used to selectively forget the output temporal data and achieve the traffic flow prediction under the action of the fully connected layer.The experimental results showed that the evaluation metrics of CNN-FE-GRU model in terms of mean absolute error(MAE),mean-square error(MSE),and root mean square error(RMSE)are on average 0.229,0.1308 and 0.3388,respectively,which all decrease relatively to the comparison model to a varying degree.Overall,compared with the single models without FE module and other combination models,the CNN-FE-GRU model demonstrates a significant improvement in accuracy and prediction performance.

关 键 词：门控循环单元 交通流 特征增强 卷积神经网络 时间序列 标准化 长短时记忆神经网络 误差分析 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]