网络流量预测技术的研究

发布时间：2018-11-19 10:29

【摘要】：在计算机网络技术不断发展的今天,网络所覆盖的范围越来越广,所承载的业务需求和使用的规模也日趋普遍。近年来,特别是P2P等新技术的大量涌起,更是严重地劣化了计算机网络的性能。为了加速网络运行的速率和增强网络的利用率,最重要的环节就是通过一些方法有效地预测出网络流量的趋势。假如我们能够实时监控网络的运行情况,在出现网络拥塞问题之前对流量数据进行分析,便能显著地提高网络的服务质量、有效性和安全性。作为网络行为研究的一个重要领域,流量预测在拥塞控制、准入控制以及无线和有线网络管理上发挥着重要作用,具有现实的研究意义。国内外学者将一些模型理论引入到了网络流量的预测中,如ARMA线性预测模型、神经网络等。本文比较了一些传统的预测模型和新技术的优势及不足,重点分析了最小二乘支持向量机(LSSVM)的方法,该方法是机器学习方法的典型代表,可以较好地应用于非线性预测的环境中,它克服了传统机器学习需要大量数据的特点,即使样本数据量较小,预测也能达到较好的效果。然而,随着网络流量的混沌性、非平稳性、复杂性等特性的出现,现存的单的方法已经不能对其进行高精度的预测。针对网络流量的混沌性,本文提出了一种基于相空间重构(PSR)和LSSVM的网络流量预测模型。首先计算最大Lyapunov指数来判断网络流量的混沌特性后,使用粒子群算法优化的LSSVM对相空间重构后的多维序列进行训练并预测出未来网络流量的走势。实验效果优于单一的LSSVM模型。针对网络流量的非平稳性和复杂性,本文提出了一种组合小波变换和PSR-LSSVM的网络流量预测模型。首先利用小波变换在非线性系统中发挥出来的多尺度分析的特性,将网络流量分解并单支重构为高频分量和低频分量,相当于对原始网络流量序列进行了平滑处理。然后判断各分量的混沌性,将具有混沌特性的分量通过PSR-LSSVM模型进行预测,其余分量通过粒子群算法优化的LSSVM进行预测,最后将各分量的预测结果综合计算输出,获得最终的预测流量。在Matlab中使用本文提出的新的模型对真实的网络流量进行实验并预测,预测精度高达90%以上,预测效果明显优于单一的LSSVM模型以及神经网络模型。
[Abstract]:With the development of computer network technology, the scope of network is more and more extensive, and the demand and scale of service are becoming more and more common. In recent years, especially with the emergence of P2P and other new technologies, the performance of computer network has been seriously degraded. In order to speed up the operation of the network and enhance the utilization of the network, the most important link is to effectively predict the trend of network traffic through some methods. If we can monitor the operation of the network in real time and analyze the traffic data before the network congestion problem we can significantly improve the quality of service effectiveness and security of the network. As an important field of network behavior research, traffic prediction plays an important role in congestion control, access control, wireless and wired network management, and has practical significance. Scholars at home and abroad have introduced some model theories into network traffic prediction, such as ARMA linear prediction model, neural network and so on. In this paper, the advantages and disadvantages of some traditional prediction models and new techniques are compared, and the method of least square support vector machine (LSSVM) is emphatically analyzed, which is a typical representative of machine learning methods. It can be well applied to the environment of nonlinear prediction. It overcomes the characteristics that traditional machine learning requires a lot of data. Even if the sample data is small, the prediction can achieve better results. However, with the appearance of chaos, nonstationarity and complexity of network traffic, the existing single method can not predict it with high accuracy. Aiming at the chaos of network traffic, this paper presents a network traffic prediction model based on phase space reconstruction (PSR) and LSSVM. First, the maximum Lyapunov exponent is calculated to judge the chaotic characteristics of network traffic, then the LSSVM optimized by particle swarm optimization is used to train the multi-dimensional sequence after phase space reconstruction and to predict the trend of network traffic in the future. The experimental results are better than the single LSSVM model. Aiming at the nonstationarity and complexity of network traffic, a network traffic prediction model combining wavelet transform and PSR-LSSVM is proposed in this paper. Firstly, the network traffic is decomposed and reconstructed into high-frequency and low-frequency components by using the multi-scale analysis of wavelet transform in nonlinear systems, which is equivalent to smoothing the original network traffic sequence. Then the chaos of each component is judged, the components with chaotic characteristics are predicted by PSR-LSSVM model, the other components are forecasted by LSSVM optimized by particle swarm optimization. Finally, the prediction results of each component are synthetically calculated and outputted. Get the final predicted flow. The new model proposed in this paper is used to test and predict the real network traffic in Matlab. The prediction accuracy is more than 90%, and the prediction effect is obviously better than that of the single LSSVM model and neural network model.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.06

【参考文献】