【作者】 王欣冉；

【导师】 邢永丽；

【作者基本信息】 中国地质大学（北京） ， 应用数学， 2011， 硕士

【摘要】 时间序列预测作为时间序列预测领域及非线性时间动态复杂系统的重要组成部分,需要在随机性和不确定性的变化中总结规律,建立预测模型,从而为决策者提供有效的决策依据。本文以时间序列预测研究中时间序列变化特征的提取、预测模型的构建及模型参数的优化这三个关键方面为研究重点,提出了一种基于小波包与最小二乘支持向量机的时间序列预测新方法,并主要从以下几个方面进行了探讨与研究:首先,系统地阐述了时间序列预测方法、小波分析及支持向量机的理论基础。分析比较了小波变换与小波包变换的优缺点,探讨了小波包在对信号特征提取时所表现出的优越性与适应性。讲解了最小二乘支持向量机推导机理,确立了基于最小二乘支持向量回归机来构建预测模型;其次,建立了基于贝叶斯推断的小波包与最小二乘支持向量机时间序列预测模型。对时间序列进行小波包分解与重构,并将分解得到的近似序列和各细节序列分别单支重构到原级别上,对各个重构后的序列分别用最小二乘支持向量机进行训练。采用贝叶斯推断进行模型参数优化,训练并选取最优模型进行预测,合成而得到原序列的预测结果;第三,基于所建的小波包与最小二乘支持向量机预测模型,针对标准QPSO算法的缺陷,提出了自适应调整收缩扩张因子β方法来优化模型参数,使得改进的QPSO算法的搜索能力更强、收敛速度更快,从而实现加快模型运算速率、提高预测精度的效果;第四,对所建立的预测模型以及所提出的优化方法,分别以上证综指和美国纽约商品交易所原油价格为例,进行了实证分析,并对实验结果进行了多层次、多角度的分析与评测。分析结果表明,本文所建预测模型的运算速度与预测精度都有显著提高,并表现出很好的稳定性和适用性,具有良好的应用前景。更多还原

【Abstract】 As an important part of nonlinear time series prediction and dynamic complex system,the methods of time series prediction sum up the law and establish prediction model in the change of randomness and uncertainty, so as to provide an effective decision basis for decision-makers. Focusing on the extraction of time series variation feature, the establishment of the prediction model and the optimization of the model parameters, a new method which based on wavelet packet and least squares support vector machines for time series prediction is proposed. This paper is mainly discussed and investigated in such aspects as follows:Firstly, the basic theory of time series prediction method, wavelet analysis and support vector machine is elaborated. Then, the advantages and disadvantages between wavelet transform and wavelet packet transform are analyzed, and the adaptive advantage of wavelet packet in signal feature extraction is discussed. The mechanism of least squares support vector machines is interpreted, and the regression model based on least squares support vector machine is established;Secondly, on the basis of Bayesian Inference, a time series prediction model based on wavelet packet and least squares support vector machines is established. By using wavelet packet decomposition and reconstruction, the time series are decomposed into the approximate sequence and the details sequences, and then reconstructed to the original single level. Separately, the reconstructed sequences are trained and predicted via LSSVM prediction model. By using Bayesian Inference, the model parameters are optimized, and the optimal model is selected to predict. The final prediction result of the original time series is the composition of the respective predictions;Thirdly, According to WP-LSSVM prediction model, and the defect of standard QPSO algorithm, a new method which adjusting the factorβadaptively to optimize the model parameters is proposed. This proposed method makes the search ability of QPSO algorithm more effective, and the convergence faster, so as to improve the prediction accuracy of the results;Fourthly, the proposed method is applied in the Shanghai stock index and the New York Mercantile Exchange crude oil prices respectively, and the experimental results are analyzed and evaluated completely, which shows that the computational speed and prediction accuracy are improved significantly. Above all, the proposed method shows good stability and applicability, which will perform a good prospect.更多还原