稀疏建模方法在时间序列预测中的应用

发布时间：2018-06-04 07:16

本文选题：时间序列 + 稀疏表示　；参考：《兰州交通大学》2017年硕士论文

【摘要】：近年来,机器学习算法在时间序列预测方面的应用逐渐受到了国内外学者的广泛关注。SR(Sparse Representation,稀疏表示)是一种典型的稀疏建模机器学习方法,作为一种不同于以往方法的基于内存的建模方法已成为具有重要理论与应用价值的研究热点。尤其是随着风力发电技术的日益成熟,并网风电场规模不断增加,风力发电对电网的影响越来越显著,风电功率时间序列的精确预测对于电力系统的发展规划意义重大。目前基于神经网络、SVM(Support Vector Machine,支持向量机)等方法的预测模型具有网络结构参数难以确定,泛化能力有限等不足。因此,探究稀疏建模方法在风电功率时间序列预测中应用具有重要意义。本文主要研究了稀疏建模方法在混沌时间序列预测和风电功率时间序列预测方面的应用,以满足现实应用对预测精度的要求,为基于内存的机器学习方法在时间序列预测方面的应用提供了新的思路。本文的主要研究内容包括如下几个方面:(1)研究了SR的基本理论,探究其中基于贪婪算法和松弛算法的两类稀疏向量求解思路,以及超完备字典的构造算法的基本原理和算法实现。(2)将SR方法引入混沌时间序列预测模型,通过将时间序列输入数据的分解重构为超完备字典和稀疏向量的乘积形式,以提取历史序列中的隐含信息。并将求解得到的稀疏向量和输出数据代入SVM(Support Vector Machine,支持向量机)方法中,建立SR-SVM组合预测模型,并在基准混沌时间序列中与单一SVM等方法进行对比,验证方法的可行性。(3)提出一类基于自适应数据字典的稀疏编码预测模型,将历史时间序列数据的输入输出数据构建以原子形式分别构成输入和输出字典,组成字典对;再针对待预测的时延输入数据向量,直接使用稀疏编码方法借助字典求得稀疏向量,即可由输出字典与稀疏向量的内积求得待预测值。与此同时,还考虑了字典的自适应更新策略,以实现在线预测,进一步提高精度。将所提出方法分别应用于混沌时间序列预测以及不同地区的短期风电功率直接和间接预测中,通过与现有方法在同等条件下的对比,验证方法的有效性。
[Abstract]:In recent years, the application of machine learning algorithm in time series prediction has been paid more and more attention by scholars at home and abroad. SRS parse representation (sparse representation) is a typical sparse modeling machine learning method. As a kind of memory based modeling method, which is different from previous methods, it has become an important research hotspot in theory and application. Especially with the development of wind power generation technology the scale of grid-connected wind farm is increasing and the influence of wind power generation on power grid is becoming more and more significant. The accurate prediction of wind power time series is of great significance for the development planning of power system. At present, the prediction models based on neural network support Vector machine (SVM) have some disadvantages, such as difficult to determine the network structure parameters, limited generalization ability and so on. Therefore, it is important to explore the application of sparse modeling method in wind power time series prediction. This paper mainly studies the application of sparse modeling method in the prediction of chaotic time series and wind power time series, in order to meet the requirement of prediction accuracy in practical applications. It provides a new idea for the application of memory-based machine learning in time series prediction. The main contents of this paper are as follows: 1) the basic theory of SR is studied, and two kinds of sparse vector solutions based on greedy algorithm and relaxation algorithm are explored. And the basic principle and algorithm realization of the construction algorithm of the supercomplete dictionary. The SR method is introduced into the chaotic time series prediction model, and the decomposition of the input data of the time series is reconstructed into the product form of the supercomplete dictionary and sparse vector. To extract hidden information from a historical sequence. The sparse vector and output data obtained from the solution are substituted into the SVM(Support Vector Machine (support Vector Machine) method, and the SR-SVM combination prediction model is established, and compared with the single SVM method in the benchmark chaotic time series. The feasibility of the method is verified. (3) A sparse coding prediction model based on adaptive data dictionary is proposed. The input and output data of historical time series data are constructed into an input and output dictionaries in atomic form to form dictionary pairs. For the delay input data vector to be predicted, the sparse vector can be obtained directly by using the sparse coding method, and the predicted value can be obtained from the inner product of the output dictionary and the sparse vector. At the same time, the adaptive updating strategy of dictionary is considered to realize online prediction and improve accuracy. The proposed method is applied to the prediction of chaotic time series and the direct and indirect prediction of short-term wind power in different regions. The validity of the proposed method is verified by comparing with the existing methods under the same conditions.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM614

【参考文献】