【作者】 冯华丽；

【导师】 刘渊；

【作者基本信息】 江南大学 ， 计算机软件与理论， 2011， 硕士

【摘要】 随着Internet的迅猛发展和更多的计算机联入Internet,互联网环境日趋复杂。为了实现资源共享的最终目标,网络传输数据量日益扩增,随之而带来的问题、故障层出不穷,给网络管理工作带来了沉重的负担。因此深入理解网络的行为特性和控制机制显得尤为重要。通过对网络流量的预测,了解网络之间的流量情况,以至对网络流量进行比较精确的分析和预测,十分有利于网络的设计和控制;通过设置路由功能,设计负载均衡,可以尽可能使拥塞带来的信息延迟和损耗程度降之最低;而且可以对网络进行入侵检测,发掘和排斥隐藏的攻击入侵行为。因此网络流量预测的可行性和必要性显得尤为重要,特别是高质量的预测对于管理和设计大规模网络环境意义重大。本文研究的宗旨是为了提高网络流量的预测精度和稳定性,作者的工作是探索出新的网络预测模型。第一,全面叙述了网络流量预测的背景、意义以及研究现状,为后续的研究工作奠定了基础;第二,分析了网络流量的重要特性,包括自相似性、长相关性、多重分形特性等,这对掌握流量预测本质,对提高预测的精度有着重要的意义。第三,在理解网络流量特性的基础上,分析了几种传统的网络流量模型,这对于用智能算法构造流量预测模型有着重要的指导意义;第四,建立了基于小波包消噪和Elman网络的网络流量预测模型。针对传统Elman网络预测精度低,结合小波包消噪和Elman神经网络的优势,先将原始流量序列进行小波包消噪,将消噪后的序列作为Elman神经网络的输入,待预测序列作为输出。通过消噪后的前N天的流量序列,预测出后M天流量序列。第五,提出一种新的网络流量预测模型,基于小波分析和AR-LSSVM理论,首先通过小波分析的方法对原始序列进行小波分解,再进行单支重构操作,利用最小二乘支持向量机和自回归模型分别预测,最后再一次进行重构操作,最终实现原始序列的预测结果,提高了网络流量的预测精度以及稳定性。第六,建立了基于QPSO优化BP的网络流量预测模型。针对PSO算法的缺点,采用QPSO算法对BP神经网络的权值和阈值进行优化,并利用历史记录训练BP网络。更多还原

【Abstract】 With the rapid development and more computers linked into the Internet network, the scale of internet is becoming more complicated, in order to realize the goal of resource share, the amount of the network traffic is increasing, the task and stress of network management becomes heavily. And failures and questions appear again and again, it is very important to comprehend the control mechanism and complicated behavioral character. It has important significance through analysis and predictions: It contributes to understand situation, so the network traffic can be accurately predicted and simulated. It’s useful for designing and controlling. It can be more effective for optimization, and be better for routing design and load-balanced design. It could determine the network congestion control, which can reduce network congestion due to the loss and delay of information. It can make full use of network resources to improve service quality. It may realize intrusion detection, detect and exclue potential attacks and intrusions. Network traffic prediction is necessary,high quality traffic prediction has significant meanings for large scale network management and design.Purpose of the study is to improve the forecast accuracy and stability of network traffic,the main work is to research and explore new network prediction models.Firstly, in the paper background, significance and research status about network traffic prediction are comprehensively narrated, to make foundation for the following researches.Secondly, the paper analyzes some important characters, including self-similarity, long-range dependence, multifarious characteristics. It’s important for mastering traffic prediction essence and improving traffic accuracy.Thirdly,Based on understanding the characteristics about network traffic, several traditional network traffic models are analyzed, which has significant guide meanings for making use of intelligent algorithms for traffic prediction model constructionFourthly, a model based on wavelet packet de-noising and Elman is found.De-noising the traffic series with wavelet packet, then taking the de-noised series as the input of Elman while the predictive series as the output of Elman. By using the N days’de-noised traffic series to forecast the later M days’predictive series. The N days’de-noised series is token as a sliding window and mapped into the later M days’predictive series.Fifthly, a mode based on wavelet analysis and AR-LSSVM is established. Firstly, the series are decomposed, getting a low frequency signal and several high frequency signals, the approximation part and detail parts were reconstructed to the original level, the next sequence of each were predicted using least squares support vector machines and self-regression model, the final, with the various reconstructed sequence, getting the prediction sequence of original sequence, improving the forecast accuracy and stability of network traffic.Sixthly, a mode based on BPNN optimized by QPSO algorithm is established. It considers the flaw of PSO algorithm.QPSO algorithm is applied to optimize the weights and thresholds in BP neural network,and historical records are used to train BP neural network.更多还原