Research On Network Traffic Prediction

As the rapid development and application of internet, the scale of internet is becoming larger and larger, the application of the internet is becoming more and more complicated, the task of management becomes more and more heavily. The problem of network appears again and again. In order to realize reliable data transfer and reasonable internet resource distribution, it is very important to comprehend the control mechanism and complicated behavioral character of network. Traffic prediction has significant meanings for management, layout and design of large scale network. Traffic prediction with high quality is getting more and more important and exigent.In this paper, the research is focused on exploring new network traffic prediction models with high accuracy and speed.Firstly, the paper comprehensively narrates the current situation about network traffic prediction at home and abroad,it offers the foundation for the following researches.Secondly, the paper analyzes some main character about network traffic in the actual network environment, which presents quite obvious multi-scale character, then, analyzes and compares the advantage and disadvantage of some network analytic models based on the character of self-similarity.Thirdly, a prediction model based on improved BP wavelet neural network is established. Considering that traditional BP wavelet neural network (BPWNN) is easy to take local convergence and has slowly learning convergent velocity,a method based on adaptive learning rate is used to optimize it in accelerating the learning convergence velocity.Fourthly, a prediction mode based on BP neural network trained by improved QPSO is established. Proposes a new adaptive parameter control method for QPSO to avoid the particle prematurely and improve the ability of global convergence.Fifthly, a prediction mode combining wavelet transform and Bayesian LS-SVM is established. The original network traffic time series is decomposed into approximate series and several detail series. The result of single branch reconstruction of each decomposed series is more unitary than the original series in frequency, and it can be built traffic model with LS-SVM. The Bayesian evidence framework is applied to LS-SVM in order to determine the regularization parameters and kernel parameters effectively.Finally, compares the speed and precision between the three models in the case of one-step and multi-step network traffic prediction.