基于EMD的时间序列预测混合建模技术及其应用研究

发布时间：2018-01-03 02:01

本文关键词：基于EMD的时间序列预测混合建模技术及其应用研究　出处：《华中科技大学》2014年博士论文　论文类型：学位论文

【摘要】：经验模态分解(Empirical Mode Decomposition, EMD)是对非平稳信号进行时频分析的理想工具,目前该方法在许多科学和工程领域得到了广泛的应用。然而,EMD在非线性时间序列建模和预测中的应用研究却相对匮乏,相关应用研究尚不深入,没有充分展示EMD的优越性。本文的目的是将基于EMD的混合建模框架更深入地应用到非线性时间序列分析和预测研究中,同时结合商业与金融领域中的预测问题,设计适合的预测模型与方法。本文的主要研究内容如下：本文首先提出一个基于集合经验模态分解(Ensemble Empirical Mode Decomposition, EEMD)和支持向量机(Support Vector Machines, S VM)的混合预测模型,同时结合股票价格预测这一金融市场研究中的传统热点进行实证研究。其次,现有基于EMD的混合建模框架及应用研究均未考虑端点效应,针对以上弊端,本文提出抑制端点效应的基于EMD和SVMs的预测建模框架,并从预测准确度的视角比较四种主流的抑制端点效应的方法。经大量数据实验和相关方法的比较,结果证明端点效应对基于EMD的混合建模框架的预测性能有较大的负面影响。在此基础上,本文将斜率法引入传统的基于EMD技术的混合建模框架,并运用较EMD更优的EEMD技术对时间序列进行分解,提出抑制端点效应的基于EEMD和SVM的预测模型对具有高度波动性的航空客流进行预测。再次,本文对基于EMD的时间序列多步预测及预测策略进行研究。(1)针对已有预测策略存在的诸多不足,本文提出一个基于粒子群优化算法(Particle Swarm Optimization, PSO)的变预测步长多输入多输出预测策略(PSO-MISMO),并以人工数据和NN3竞赛数据为预测对象,多种策略为对比方法,从预测准确度、收敛性和训练耗时等方面评估PSO-MISMO策略的实用性。(2)针对现有基于EMD技术的建模框架局限于单步预测应用的情形,本文提出适用于时间序列多步预测的基于EMD技术的建模框架,同时结合国际原油价格预测这一能源经济研究中的传统热点进行实证研究。(3)传统支持向量回归算法因其单输出结构的特征仅能采用迭代或直接策略,而不能采用较以上两种策略更优的MIMO策略进行时间序列多步预测。针对以上弊端,本文提出适用于MIMO策略的多输出支持向量回归算法对时间序列进行多步预测。最后,针对现有研究局限于单值预测的情形,本文从两个不同的研究视角分别提出适用于区间型时间序列预测的方法。(1)在保留区间型数据特征的情形下,凭借多输出支持向量回归算法(Multiple-output Support Vector Regression, MSVR)的多输出结构特征和萤火虫算法(Firefly Algorithm, FA)的高效优化能力,本文提出适用于区间型时间序列预测的基于MSVR和FA的混合模型,同时结合区间型股票价格指数预测这一金融市场研究中的新兴热点进行实证研究。(2)在不保留区间型数据特征情形下,凭借双变量经验模态分解技术(Bivariate Empirical Mode Decomposition, BEMD)对复值序列高效的分解性能和基于EMD技术的建模框架在单值时间序列预测中的优异表现,本文提出基于BEMD技术的建模框架对区间型时间序列进行预测,同时结合区间型电力需求预测这一电力市场研究中的新兴热点进行实证研究。
[Abstract]:The empirical mode decomposition (Empirical Mode Decomposition, EMD) is an ideal tool for time-frequency analysis of non-stationary signal, this method has been widely used in many fields of science and engineering. However, the research and application of EMD in nonlinear time series modeling and forecasting is relatively scarce, relevant applied research is not deep, not fully demonstrate the superiority of EMD. The purpose of this paper is the hybrid modeling framework EMD further applied to nonlinear time series analysis and prediction research based on combined prediction of business and the financial sector in question, prediction model and the method of design for.
The main contents of this paper are as follows:
This paper proposes a based on ensemble empirical mode decomposition (Ensemble Empirical Mode Decomposition, EEMD) and support vector machine (Support Vector Machines, S VM) of the hybrid forecasting model, combined with the prediction of stock price of traditional hot point of this financial market research in empirical research.
Secondly, the existing hybrid modeling framework and Application Research Based on EMD are not considering end effect, in view of the above problems, this paper puts forward based on predictive modeling framework of EMD and SVMs to suppress the endpoint effect, and from the suppression of the endpoint effect prediction accuracy from the perspective of comparison of four mainstream methods. By comparison of large amounts of experimental data and related methods the results prove that the endpoint effect, there is a greater negative impact on the prediction performance of hybrid modeling framework based on EMD. On this basis, this paper will be based on the hybrid modeling framework of EMD technology into the traditional slope method, and the use of better than EMD EEMD technology for time series decomposition is proposed to suppress the endpoint effect prediction model and EEMD based on the SVM of air passenger flow with a high degree of volatility forecasting.
Again, this paper focuses on the research of EMD time series prediction and forecast based on strategy. (1) aiming at existing problems of prediction strategy, this paper proposes an algorithm based on particle swarm optimization (Particle Swarm Optimization, PSO) of the variable prediction step, multiple input multiple output (PSO-MISMO), and the prediction strategy based on artificial the data and NN3 data for the prediction of competition strategy for a variety of objects, the method of comparison, from the forecast accuracy, evaluate the usefulness of PSO-MISMO strategy and convergence time. Training (2) according to the existing modeling framework of EMD technology is limited to single step prediction application case based on the proposed modeling framework based on EMD technology for time the sequence of multi step prediction, combined with the international crude oil price prediction of traditional hot this energy in economic research and empirical research. (3) the traditional support vector regression algorithm for single output structure The feature can only adopt iterative or direct strategy, and can't use more than two strategies and better MIMO strategy for multi-step prediction of time series. In view of the above drawbacks, this paper proposes a multi output support vector regression algorithm suitable for MIMO strategy for multi-step prediction of time series.
Finally, according to the existing research is limited to single value prediction, this paper from two different perspectives respectively put forward methods applied to interval time series prediction. (1) in the retention interval data characteristics of the case, with multi output support vector regression algorithm (Multiple-output Support Vector Regression, MSVR) multi output the structure and characteristics of firefly algorithm (Firefly Algorithm FA), the ability of optimization, this paper applies a hybrid model of MSVR and FA based on interval time series prediction, combined with the interval of stock price index forecasting emerging hot spots in this financial market research in empirical research. (2) in the retention interval the data type feature case with double variable EMD Technology (Bivariate Empirical Mode Decomposition, BEMD) of complex valued sequence efficient decomposition properties and based on EMD Technology The performance of modeling framework in single value time series prediction is excellent. In this paper, a modeling framework based on BEMD technology is proposed to predict interval type time series, and combined with the emerging hot spots in the research of electricity market, the empirical research is carried out based on interval power demand forecasting.

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：F830.91;F224

【参考文献】