摘要
研究表明,历史流量数据可以用于移动网络流量的预测,同时周边区域的流量信息可以提高流量预测的准确性。为此,文中提出一种基于时空特征的移动网络流量预测模型STFM。STFM模型利用目标区域及周围区域的历史移动网络流量对目标区域的流量进行预测。其核心思想是,首先利用三维卷积网络(3D CNN)从流量中提取移动网络流量空间上的特征,再利用时间卷积网络(TCN)提取移动网络流量时间上的特征,最后全连接层对提取的特征与实际的流量值建立映射关系,产生预测的流量值。根据实验的验证与分析,STFM在移动网络流量预测上的标准均方根误差(NRMSE)相比TCN,CNN和CNN-LSTM分别减少了28%,21.7%和10%。因此,STFM模型能够有效提高移动网络流量预测的准确率。
Research shows that historical traffic data can be used for the prediction of mobile network traffic,and traffic information in surrounding areas can improve the accuracy of traffic prediction.To this end,this paper proposed the traffic prediction model STFM for mobile network based on spatio-temporal features.STFM uses the historical mobile traffic of the target area and surrounding areas to predict the traffic of the target area.Firstly,3D convolutional neural network(3D CNN)is used to extract the spatial features of the mobile network traffic,then time convolutional network(TCN)is used to extract the temporal features of the mobile network traffic.Finally,fully connected layers establish a mapping relationship between the real traffic and extracted features and generate a predicted traffic value.Validation and analysis of experiments show that the STFM reduce the normalized root mean square error(NRMSE)by 28%,21.7%and 10%,compared to TCN,CNN and CNN-LSTM Consequently,STFM can effectively improve the accuracy of mobile network traffic prediction.
作者
张杰
白光伟
沙鑫磊
赵文天
沈航
ZHANG Jie;BAI Guang-wei;SHA Xin-lei;ZHAO Wen-tian;SHEN Hang(College of Computer Science and Technology,Nanjing Tech University,Nanjing 211816,China;National Engineering Research Center for Communication and Network Technology,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
出处
《计算机科学》
CSCD
北大核心
2019年第12期108-113,共6页
Computer Science
基金
国家自然科学基金项目(61502230,61073197,61501224)
江苏省自然科学基金项目(BK20150960)
江苏省普通高校自然科学研究项目(15KJB520015)
南京市科技计划项目(201608009)
南京大学计算机软件新技术国家重点实验室资助项目(KFKT2017B21)
江苏省研究生科研与实践创新计划项目(KYCX18_1074)资助
关键词
移动网络
流量预测
时空特征
卷积网络
全连接层
Mobile network
Traffic prediction
Spatio-temporal feature
Convolution neural network
Fully connected layers
