Communication Traffic Analysis With The Study And Comparison Of Multiple Prediction Models

With the increase of the applications and size of the communication networks, network management is increasingly important. The new generation of network management system for predicting business has made new demands, in which communication traffic forecasting has become one of the most important parts. Accurate traffic forecast for network management, planning, design is of great significance. According to this issue, this thesis focuses on the communication traffic analysis, and proposes three communication traffic forecasting models. General forecasting model includes ARMA model and BP prediction model. However, it is very slow to use ARMA model to predict the future value, while it is very hard to fit the optimal parameters for BP prediction model.In this thesis, the main contributions include the following parts: 1). Designing a generic communication traffic forecasting model framework; 2). Implementation of three different forecasting models, which are two commonly used ARMA model and BP model, as well as our proposed SVM-based forecasting model; 3) Detailed comparison of the performance of various models, and description of the process how to select the optimal prediction model. The specific process is as follows:Firstly, we study the properties of the communication traffic, which is a typical time series. We also investigate the common communication traffic forecasting methods.Secondly, we introduced the theories of a variety of forecasting models: ARMA model will regard the time series as a random sequence, and use some of the mathematical model to approximate this sequence; BP model approximate the simulation of a nonlinear function for forecasting the communication traffic through the multi-layer neural network; SVM regression model is to obtain a range of support vectors, and then estimate a nonlinear optimization function to predict the traffic by a quadratic programming.Thirdly, based on the theoretical analysis about the forecasting model, this thesis illustrates the framework of the prediction model. And then we described the three foreccisting models in details: including the ARMA model, BP model and the SVM regression model. After that, we showed the process how to select the optimal prediction model. We found that the dimension of input vector would affect the performance of the prediction models significantly.Finally, we report on the results of the three forecasting models. In order to assess the performance of the forecasting models, we collected a number of communication traffic datasets. The experiments show that SVM-based forecasting model achieves the minimum average MSE error among the three different forecasting models, and the average MSE is 0.0091. The average MSE of the ARMA model is higher than that of the SVM-based model, which is 0.0114. Comparing with BP model, both of the ARMA and SVM models outperform BP model. Moreover, SVM-based forecasting model performs faster than ARMA model to forecast the communication traffic. Generally speaking, the overall performance of SVM model is optimal.