多任务学习在时间序列预测中的研究及应用

发布时间：2018-04-08 18:42

本文选题：时间序列预测　切入点：多任务学习　出处：《太原理工大学》2017年硕士论文

【摘要】：时间序列是一种常见的数据结构,广泛地存在于人类社会活动和客观世界中。时间序列预测旨在依据时间序列中隐含的内在变化规律建立数学模型,挖掘其隐含的时间演变关系,从而实现对序列未来发展趋势的预测。深入分析来源于不同观测事物的时间序列数据,挖掘蕴含的可用信息,对辨识事物发展趋势、刻画事物间的相关特性以及对事物变化的决策与把控等有重要的实际意义。目前,基于数据驱动建模的时间序列预测方法可分为全局模型和局部模型。由现实世界得到的时间序列数据往往具有强非线性和不确定性等特点,这就使得传统的预测模型难以有效建模,从而限制了模型的预测精度。基于全部历史数据建模的全局模型虽然方法简单,但是对于序列中的异常值有较高的敏感性。针对上述问题,本文提出了一种基于云模型相似性度量的局部建模方法,并结合BP神经网络(BPNN)和最小二乘支持向量机(LS-SVM)建立预测模型,可有效提高序列的预测精度。上述局部建模方法虽然对预测精度有一定的贡献,但是仍属于单任务学习的范畴,对时间序列中隐含的相关性信息没有进行充分挖掘,从而影响了模型的泛化性能。而多任务学习(MTL)能兼顾任务间的相关性与差异性,挖掘了任务间具有共享知识结构的共性信息,最终改善了所有任务的学习性能。因次,本文综合考虑局部模型与多任务学习的优势,提出了一种基于多任务学习的局部建模方法。该方法将时间序列的相邻时间点中蕴含的多种信息可看成是多任务学习中的不同任务同时进行学习,从而提高模型的泛化能力。本文的研究工作主要包括以下几个方面:(1)分析了几种常见时间序列相似性度量方法的优势和不足,并对两种不同类型的时间序列预测方法进行了对比;(2)针对原始时间序列数据具有非线性、复杂性以及不确定性等问题,采用云模型理论对原始序列以及一阶差分处理后的序列同时进行时间序列表示;(3)针对传统距离函数难以有效度量包含不确定性因素的数据,提出了基于云模型相似性度量的局部建模方法,并采用BPNN和LS-SVM建立预测模型进行建模和预测;(4)针对单任务学习方法在处理时间序列中存在的信息挖掘不充分、预测精度低等问题,在提出的局部建模方法的框架下,将多任务学习用于时间序列预测,提出了基于多任务学习的局部建模方法;(5)采用选自阿尔托大学理学院的机器学习应用研究组的六个真实的工程数据集对提出的方法进行了验证。
[Abstract]:Time series is a common data structure, which widely exists in human social activities and objective world.The purpose of time series prediction is to establish a mathematical model based on the inherent law of change implied in time series, and to excavate its implicit time evolution relationship, so as to predict the trend of future development of time series.It is of great practical significance to analyze the time series data from different observational objects and to mine the available information to identify the development trend of things, to depict the related characteristics of things, and to make and control the changes of things.At present, the time series prediction method based on data-driven modeling can be divided into global model and local model.The time series data obtained from the real world are often characterized by strong nonlinearity and uncertainty, which makes it difficult for the traditional prediction models to be effectively modeled, thus limiting the prediction accuracy of the models.Although the global model based on all historical data is simple, it is sensitive to the outliers in the sequence.In order to solve the above problems, a local modeling method based on cloud model similarity measurement is proposed in this paper, and a prediction model based on BP neural network (BP) and least squares support vector machine (LS-SVM) is established, which can effectively improve the prediction accuracy of the sequence.Although the local modeling methods mentioned above have a certain contribution to the prediction accuracy, they still belong to the category of single-task learning. The implicit correlation information in time series is not fully mined, thus affecting the generalization performance of the model.Considering the advantages of local model and multitask learning, a local modeling method based on multitask learning is proposed in this paper.In this method, the multiple information contained in the adjacent time points of time series can be regarded as different tasks in multitask learning, so as to improve the generalization ability of the model.The research work of this paper mainly includes the following aspects: 1) analyzing the advantages and disadvantages of several common time series similarity measurement methods.Two different kinds of time series prediction methods are compared. The original time series data are nonlinear, complex and uncertain.The cloud model theory is used to represent the time series of the original sequence and the first order difference processing sequence simultaneously. In view of the traditional distance function, it is difficult to measure the data containing uncertain factors effectively.This paper proposes a local modeling method based on similarity measurement of cloud model, and uses BPNN and LS-SVM to build prediction model for modeling and forecasting.In the framework of the proposed local modeling method, multitask learning is applied to time series prediction.A local modeling method based on multitasking learning is proposed. The proposed method is validated by six real engineering data sets selected from the Machine Learning Application Group of the School of Science of Aalto University.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O211.61

【参考文献】