网络流量预测中的时间序列模型比较研究

Research on the Comparison of Time Series Models for Network Traffic Prediction

网络流量预测在新一代网络协议设计、网络管理与诊断、设计高性能路由器等方面都具有重要意义.目前通常采用ARMA和FARIMA时序模型对网络流量序列进行拟合与预测,但没有对时间尺度的大小与模型选择的关系进行研究.本文对实际网络流量在不同时间尺度(毫秒、秒、分)下进行了流量预测建模并对预测性能进行比较,分析表明使用时序模型进行流量预测时,大时间尺度(分)流量预测较小时间尺度(毫秒、秒)具有更小的预测误差.并且,对于小时间尺度上的自相似流量序列,自相似模型FARIMA并没有较其他时序模型有更好的预测性能.

Network traffic prediction is very important in network protocol designing,network management and high performance routers designing etc. Currently, ARMA and FARIMA time series are the main models used to fit and predict the network traffic. But the relation between time scale and time series models hasn＇t been studied. The network traffic was modeled according to different time scales using the traffic trace data taken from the Intemet traffic archive, and the prediction performance of those models was compared. The results show that time series models perform better under the large scale（ minute）than under small time scales（millisecond and second）, and the performance of self similarity model FARIMA shows no advantage over other models.