The Prediction For CDN Traffic Based On Multivariable Time Series

With the rapid improvement of modern information network technology,the EPG system in IPTV business is also developing rapidly.The video traffic required by the EPG system is increasing,which may bring the hidden danger of network shock.In order to reduce the impact and loss of network instability,it is necessary to predict CDN traffic in future time points and find out the traffic amplitude of the increase or decrease in advance,so as to make server resource reservation and adjust the corresponding strategy of storage,scheduling and distribution to achieve a better traffic load balancing effect.The traditional statistical methods used to predict the CDN traffic are unsatisfactory.With the development of machine learning methods,such as random forest and neural network are good at dealing with the problem with high dimensional input,and can deal with the mutagenicity and fuzziness in traffic prediction,which is helpful to solve the problem of multivariate time series prediction,so they are gradually applied in the field of traffic prediction.In this paper,basing on the data of CDN traffic from November 1,2018 to December 4,2018 in Fujian province,the linear relationship between 7 influence variables and CDN traffic are explored,and Ridge regression,Lasso regression and ordinary linear regression were established to describe the quantitative relationship between CDN traffic and variables.Then the random forest,Long Short Term Memory Network of single variable and Long Short Term Memory Network of multiple variables are established to predict CDN traffic in the coming moment,and the prediction results of three models are compared.The results show that the accuracy of traffic value predicted by Random forest model is relatively low,the prediction deviation of CDN traffic value with large local fluctuation is large,the effect of LSTM model with single variable and LSTM model with multiple variables is better than that of Random forest,and the predictive value and real value of LSTM model with multiple variables are closer.The result of CDN traffic prediction with large local fluctuation is relatively more stable.