Joint Neural Network-based Traffic Forecasting Model

From the beginning of 1990's, Internet has achieved rapidly development, and Internet development changing people's lives and traditional information industry frame work. Whether a traffic control solution is effective depending on the through understanding of the characteristics and the ability of prediction. Self-similar property is one of the key statistical characteristics in the computer networks (including LAN and WAN). Self-similar traffic has a significant impact on network management and control as well as performance evaluation. With the increase of network application and scale, the task of network management becomes more and more heavily. The problem of network appears again and again, such as bottleneck, appears because of one do not understand network operation conditions of a system, one could not find out problem while network appears failure, and also network configure and management complex for so many equipments, and even more and more network security are threatened. The appearance of these new problems is challenge and assignment of network research. This paper first introduce self-similar origin, definition, representation and the measurement method of Hurst parameter, the reason of network self-similar and the impact on network performance are discussed. The next job of network study is predicting traffic model with accurate network model based on self-similar theory analysis. In chapter 5, we establish a new prediction model. Firstly, in this model we pretreat the internet traffic using the wavelet method. Secondly, we make use of linear neural network (LNN) and Elman neural network (ENN) to predict respectively, in order to make sure that it can describe the correlation and non-stationary characters of the traffic. Finally, combine two prediction results into the final result through four combiners (averaging, LNN combiner, ENN combiner, BPNN combiner). Through prediction simulations on TCP traffic and video traffic respectively, the results indicate that the proposed combined model outperforms the individual models and the wavelet can make the performance better. The results also show that the prediction performance depends on the traffic nature and the considered timescale.