SOM在时间序列预测中的应用研究

发布时间：2018-04-04 17:02

本文选题：时间序列预测　切入点：自组织映射　出处：《兰州交通大学》2015年硕士论文

【摘要】：近年来,自组织映射(Self-organizing Map,SOM)神经网络在时间序列预测方面的应用逐渐受到国内外研究者的广泛关注,已成为具有重要的理论与应用价值的研究热点。作为一种非监督竞争学习型神经网络,SOM神经网络的构造简单直观,其联想记忆技术避免了传统方法易陷入局部最优的问题。本文研究了SOM神经网络的改进方法在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,为非监督神经网络在时间序列预测方面的应用扩展了新的空间。本文的主要研究内容包括如下几个方面:(1)研究时间序列预测理论,以及SOM神经网络的结构和算法。将SOM神经网络的联想记忆技术推广到时域,矢量量化临时联想记忆(Vector Quantized Temporal Association Memory,VQTAM)建模技术可实现时间序列预测。(2)提出一类递推SOM方法,包括递推的自组织映射(Recursive Self-organizing Map,RecSOM)方法和适用于结构化数据的结构化数据自组织映射(Self-organizing Map of Structured Data,SOMSD)方法。递推SOM方法利用上下文信息反映数据集的统计特性,其中,RecSOM方法用带时延的反馈表现递推的概念,SOMSD方法利用获胜神经元的网格坐标表示上下文信息,更适用于结构化数据。递推SOM方法应用于交通流预测实例,并在同等情况下与其他预测方法进行对比,结果验证提出的方法是可行的、有效的。(3)在VQTAM建模技术的基础上,提出一类基于SOM神经网络的局部自回归(Auto-regressive,AR)方法。给出具有多个局部线性AR模型的AR-SOM方法,基于前K个获胜神经元用权值代替输入向量建立单一时变局部AR模型的K-SOM方法,以及在完成数据向量聚类的同时,更新多个局部AR模型系数的LLM(Local Linear Map)-SOM方法。相对于全局模型,所提出的方法能够灵活给出有效的监督神经结构,降低了计算复杂度。将其应用于不同的混沌时间序列预测典型实例中,进一步还将其应用于网络流预测实例和视频流预测的实例中,在同等条件下与已有方法比较,实验结果表明,所提出的方法能有效改善预测精度,且性能更好,验证了其有效性与应用潜力。
[Abstract]:In recent years, the application of Self-Organizing Map SOM (SOM) neural network in time series prediction has been paid more and more attention by researchers at home and abroad, and has become an important research hotspot in theory and application.As an unsupervised competitive learning neural network, SOM neural network has a simple and intuitive structure, and its associative memory technology avoids the problem that traditional methods are prone to local optimum.In this paper, the application of the improved method of SOM neural network in time series prediction is studied in order to meet the requirement of prediction accuracy in practical applications and to extend a new space for the application of unsupervised neural networks in time series prediction.The main contents of this paper are as follows: 1) study the theory of time series prediction and the structure and algorithm of SOM neural network.In this paper, the associative memory technique of SOM neural network is extended to the time domain. A kind of recursive SOM method is proposed, which can be used to predict the time series by using vector quantization (VQ) Quantized Temporal Association memory (VQTAM) modeling technique.It includes the recursive Self-organizing mapping (RecSOM) method and the self-organizing Map of Structured data (SOMSD) method for structured data.The recursive SOM method uses the context information to reflect the statistical characteristics of the dataset. The RecSOM method uses the recursive concept of feedback with time delay to represent the context information using the grid coordinates of the winning neurons, which is more suitable for structured data.The recursive SOM method is applied to traffic flow forecasting example and compared with other forecasting methods in the same situation. The results show that the proposed method is feasible and effective on the basis of VQTAM modeling technology.A class of local autoregressive autoregressive ARs based on SOM neural network is proposed.The AR-SOM method with multiple local linear AR models is presented. The K-SOM method of establishing a single time-varying local AR model based on the weights of the first K winning neurons instead of the input vector is presented. At the same time, the clustering of the data vectors is completed.LLM(Local Linear Map)-SOM method for updating the coefficients of multiple local AR models.Compared with the global model, the proposed method can provide an effective supervised neural structure flexibly and reduce the computational complexity.It is applied to different typical examples of chaotic time series prediction, and it is also applied to network flow prediction examples and video stream prediction examples. The experimental results show that, under the same conditions, the proposed method is compared with existing methods.The proposed method can effectively improve the prediction accuracy and the performance is better. The validity and application potential of the proposed method are verified.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495

【共引文献】