Study On Network Traffic Analysis And Prediction Model

Study on Network Traffic Analysis and Prediction ModelWith the rapid development of Internet, the number of Internet users increased sharply. Network QoS becomes more and more important as well as network traffic management. The process of network traffic management includes collecting data related to network, finding irregularities in network through analysis and prediction of data. Then do some corresponding operations. For example, sending alarm information to network administrator.Recent years, there has been a lot of researches on network traffic model. According to characteristics of models, it can be divided into two types. One is stationary network traffic model, and the other is non-stationary model. Stationary network traffic model is usually assumed to satisfy the linear relationship and is described by linear recursive and combination method. Conventional stationary network traffic model (such as Auto Regression Integrated Moving Average Model, ARIMA ) which establishes on the basis of the strict mathematical theory has simple architectures, high speed of predicting and is capable of describing the features of short or long range dependence but the non-stationary characteristics. In the transition from non-stationary data of network traffic to stationary data, it necessitily lost some information of network traffic. In addition, the confidence interval of results increases with the extended cycle of prediction. This is also a defect of conventional stationary network traffic model.Neural network is a non-stationary model which could make up for the defect of stationary model. Therefore, the method that applying neural network to network traffic prediction is proposed. The NN(neural network) by way of a non-stationary model could make up for the defect of stationary model; however, there has not been a specific theory to select parameters of network yet. Such as the number of neurons, training method and neural network structure. These parameters are generally decided by experience. This is the defect as well as difficulties of neural network. In order to improve prediction accuracy, network traffic model should draw advantages of stationary model and non-stationary model.With the research of the traditional time series forecasting model and neural network model, we know that all these models, to some extent, can achieve the purpose of modeling the network traffic. But they all have their own shortcomings. In this paper, we introduce a multiple combination model based on ARIMA model and Elman model in order to fully and accurately characterize network behaviors and features of traffic and enhance the accuracy of predictioaCompared ARIMA model with Elman model, the proposed model always chooses the best output value of forecasting to be the input value of the next training network. This method can efficiently avoid the defect of ARIMA that it is incapable of describing non-stationary characteristics and the step of forecasting increases while accuracy falls and the defect of Elman that there has not been a theory to select parameters of the network yet. At last, simulation experiment is given to show the comparison of performance of all kinds of network traffic models.In this paper, the contents are as follows:1) Optimize the data of network traffic. According to Grabs Rule and the similarity oftime series data, abnormal data can be identified and optimized smoothly based onmathematical statistics and smoothing principle.2) Select a reasonable time series model. Studying the feasibility of network traffic dataprediction, and analyzing the characteristics of traffic information.Then select the seasonalARIMA time series model for traffic modeling. The model can reflect the cyclical and self- similar characteristics of network traffic more accurately. The advantage of this model is that the prediction accuracy in short term is relatively high.3)Select a reasonable Elman neural network model. According to the non-linearcharacteristics of network traffic, neural network is applied to network traffic modeling.Elman neural network is a powerful learning system which ensures non-stationarycharacteristics of network traffic and provides us with highly non-linear mapping from inputto output. In this paper, neural network topology and its parameters can be decided by themethod of combination theoretical with experimental approach. A better prediction model isacquired according to the experiment results.4) Propose the multiple combination model. In multiple combination model, we take theabsolute value of the margin caculation between the output sequence of ARIMA model orElman model and real value of network traffic respectively and we always choose the valuewhich is closest to real traffic to be the input value of BP network in order to exert maximumadvantage of the two models, and achieve the purpose of reducing the prediction error. Thismethod combines several prediction models to form a new prediction model. In this paper,the description of the corresponding algorithm and simulation experiment are given. Theexperimental results show that this model improves effectiveness and has not only goodprediction accuracy, but also a better stability.5) Update network traffic model. Take the prediction value of every detection cycle andthe real traffic value into update formula to get the reference value of the next cycle. Thepurpose of improving the update formula is to make the prediction value to fit the real trafficmore exactly.6) Detect abnormal network traffic. Detect network using adaptive threshold value which iscalculated by the standard deviation of the network traffic. Through this method, abnormalnetwork signs can be found on time in order to take preventive measures to reduce or avoidthe impact on the network.Realistic network environment is changing constantly, the relationship between the flow variables are complex. Modeling network traffic comes down to very extensive field of research work, this paper is only a small part. Simulation results show that the multiple combination model proposed in this paper is effective and obtains satisfactory accuracy. This method can be used as a integral part of overall solutions of network security and also contact with other common solutions to solve security problem of network management.In this paper, the research work is selected from National doctor fund of the Ministry of Education: Neural network fusion learning mechanism and research of its application (20060183043). The paper named Study on Applying Multiple Combination to Network Traffic Analysis has been accepted by Mini-Micro Systems.