基于EMD的相空间重构极限学习机的短期气象要素预测方法及其应用研究

发布时间：2018-02-14 06:16

本文关键词： 经验模态分解相空间重构极限学习机气象要素预测　出处：《南昌大学》2016年硕士论文　论文类型：学位论文

【摘要】：自古以来,气象科学的时效性和准确性影响到人类生产生活的方方面面,决定着社会的发展和稳定。随着时代的进步,人类对于气象科学的需求不断提升,但是某些方法或多或少存在着不同程度的问题,例如模型准确率不高,建模时间过久,受噪音影响太大等等,如何建立更加准确,有效的预测模型,是现今气象领域的一个重点科研话题。针对这一话题,本文提出了基于经验模态分解技术的相空间重构极限学习机的预测模型,对时序序列数据进行建模,采用经验模态分解技术来对数据进行去噪,利用相空间技术进行模式抽取,使得重构后的数据更加适合模型的构建,最后运用极限学习机技术进行建模。并且在真实的气象数据集上对所采用的三种技术以及所建立的模型进行性能测试,验证模型的有效性,对比于其他模型的,分析其优越性。本文所做的主要内容有以下三点:1)针对目前气象要素中的单要素时序序列数据的预测模型所存在的噪音问题,以及预测模型准确率不够高的问题,本文提出了基于经验模态分解方法的相空间重构极限学习机的预测模型的构建。利用经验模态分解技术,对原始数据进行去噪处理,降低其他不确定因素对于观测数据的影响,为后续处理提供更加有效准确的数据;对于时序序列数据,直接进行模型的构建存在一定的难度并且不容易获得成功的模型,为了有效恢复其动力学系统,从而使得重构后的数据更加适合规则的提取和模型的构建,本文采用相空间重构技术对气象单要素的一维时序序列的原动力学系统进行研究与拓扑重现;对于重构后的数据,采用极限学习机技术进行模型的构建,能够快速构建预测模型并且能有效保证模型的泛化能力。2)通过对经验模态分解方法原理的研究,本文分析了本征模函数分量筛选算法中阈值设定所存在的不足并进行了相应的改进。相对于传统的方式,本文给出一个新的定义公式,实现对于不同的分量能够凭借当前本征模函数分量的相关信息,设定动态筛选阈值,有效的选择相关性符合的本征模函数分量进行数据重组。动态的阈值设定能够更加有效的判定临界本征模函数分量的具体归属问题,改善去噪之后的数据质量,克服了主观性。3)在真实的气温要素这种一维时序序列数据上对预测模型中相关参数的选定进行了测试分析,并对所构建的预测模型与改进的基于动态筛选阈值的经验模态分解方法所生成的模型(称为改进后的预测模型)进行了性能测试和结果分析,并与几个其他模型性能的相关数据进行了对比分析,测试结果与对比分析表明了本文所提出改进后的预测模型具有较好的预测精度,基本能够达到人们生产生活对于气象预测的需求。最后,本文给出了目前所阐述的模型仍然存在的可改进之处,确定了后续的研究方向。
[Abstract]:Since ancient times, the timeliness and accuracy of meteorological science have affected all aspects of human production and life, and decided the development and stability of society. But some methods have more or less different degree of problems, for example, the accuracy of the model is not high, the modeling time is too long, the noise is too big, and so on, how to establish more accurate and effective prediction model, It is a key research topic in the field of meteorology nowadays. In this paper, a prediction model of phase space reconstruction extreme learning machine based on empirical mode decomposition technology is proposed to model the time series data. The empirical mode decomposition (EMD) technique is used to de-noising the data, and the phase-space technique is used for pattern extraction, which makes the reconstructed data more suitable for the construction of the model. Finally, the model is modeled by extreme learning machine technology, and the performance of the three techniques and the established models are tested on the real meteorological data set to verify the validity of the model, compared with other models. The main contents of this paper are as follows: 1) the noise problem existing in the prediction model of the single factor time series data in the meteorological elements and the problem that the accuracy of the prediction model is not high enough. In this paper, the prediction model of phase space reconstruction limit learning machine based on empirical mode decomposition method is proposed. Using empirical mode decomposition technique, the original data is de-noised to reduce the influence of other uncertain factors on the observation data. To provide more effective and accurate data for subsequent processing, for sequential data, it is difficult to construct the model directly and it is not easy to obtain a successful model, in order to recover its dynamic system effectively. In order to make the reconstructed data more suitable for rule extraction and model construction, this paper uses phase space reconstruction technology to study and topology reconstruction of one-dimensional time series of meteorological single elements. By using extreme learning machine technology to construct the model, the prediction model can be constructed quickly and the generalization ability of the model can be ensured effectively. (2) through the research on the principle of empirical mode decomposition method, In this paper, the deficiency of threshold setting in the intrinsic mode function component selection algorithm is analyzed and the corresponding improvement is made. Compared with the traditional method, a new definition formula is given in this paper. For different components, the dynamic filtering threshold can be set by virtue of the relevant information of the current eigenmode function components. The dynamic threshold setting can effectively determine the specific attribution of critical eigenmode function components and improve the data quality after denoising. It overcomes subjectivity. 3) on the basis of the real temperature factor, which is one dimensional time series data, the selection of relevant parameters in the prediction model is tested and analyzed. The performance test and result analysis of the proposed prediction model and the improved empirical mode decomposition method based on the dynamic screening threshold (called the improved prediction model) are also carried out. The test results and comparative analysis show that the improved prediction model has a better prediction accuracy, and compared with several other model performance related data, the test results and comparative analysis show that the improved prediction model has a better prediction accuracy. It can basically meet the demand of people's production and daily life for weather forecast. Finally, this paper gives the improvement of the model which still exists at present, and determines the future research direction.
【学位授予单位】：南昌大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：P456;TP18

【相似文献】