基于受限波尔兹曼机的移动业务预测模型研究与应用

发布时间：2018-11-22 10:36

【摘要】：随着互联信息时代的快速发展,移动用户业务需求量迅猛增加,使得移动网络的无线资源显得日趋紧张。为了更加有效合理地利用有限的移动网络资源,准确的业务量预测具有重要意义。本文通过对基于受限玻尔兹曼机(Restricted Boltzmann Machine,RBM)的深度学习方法进行研究,进而实现对移动网络业务量的预测。RBM可以视为一个无向图双层网络模型,是一种有效的无监督特征提取方法,常用于图像数据特征的提取,而条件受限玻尔兹曼机(Conditional Restricted Boltzmann Machine,CRBM)是在RBM的基础上,进一步考虑了时间范围内的图像间关联性,因而获得了在图像识别等方面更加优越的性能。本文首先对基于RBM、CRBM的深度学习模型进行了研究。提出一种新颖的基于RBM的深度信念网络(Deep Belief Network,DBN)时间序列预测模型,该模型主要由三部分组成:底层包含两层高斯伯努利RBM模块,中间层为前馈神经网络(Feedforward Neural Network,FNN),顶层为错误校正算法模块。通过实验将该模型应用于[Energy]、[Dollar]和[TAIEX]三个时间序列数据集,并与传统的FNN模型以及基于高斯伯努利深度玻尔兹曼机(Gaussian-Bernoulli Deep Boltzmann Machine, GDBM)的FNN模型相比,结果表明,本文所提出的DBN模型的预测精度有显著提升。考虑到CRBM具有时间关联性的优点,本文进一步提出了基于CRBM的深度DBN模型,并针对原始CRBM只适合如图像等空间像素的输入,并不适合时间序列预测中的时序数据输入的特点,提出一种新颖的输入表达,使得时序数据能够适于在CRBM中的应用。本文基于上述研究,分别将基于RBM的DBN模型和适于时序数据输入的基于CRBM的DBN模型应用于移动业务量预测。选取现网热点地区的10个小区用户上报测量报告(MR),分别以20分钟时间粒度为单位构建时间序列,预测结果表明,80%的小区达到了 80%以上的预测精度。
[Abstract]:With the rapid development of interconnected information age, the demand for mobile user services is increasing rapidly, which makes the wireless resources of mobile network appear increasingly tight. In order to utilize the limited mobile network resources more effectively and reasonably, accurate traffic prediction is of great significance. In this paper, the method of depth learning based on constrained Boltzmann machine (Restricted Boltzmann Machine,RBM) is studied, and the prediction of mobile network traffic is realized. RBM can be regarded as an undirected graph bilevel network model. It is an effective unsupervised feature extraction method, which is often used to extract the features of image data. The conditional constrained Boltzmann machine (Conditional Restricted Boltzmann Machine,CRBM) is based on RBM and further considers the correlation between images in the time range. As a result, more superior performance in image recognition and so on has been obtained. In this paper, the depth learning model based on RBM,CRBM is studied. A novel (Deep Belief Network,DBN time series prediction model based on RBM is proposed. The model consists of three parts: the bottom layer is composed of two layers Gao Si Bernoulli RBM module, and the middle layer is feedforward neural network (Feedforward Neural Network,. The top level of FNN), is the error correction algorithm module. The model is applied to [Energy], [Dollar] and [TAIEX] data sets through experiments, and compared with the traditional FNN model and the FNN model based on Gao Si Bernoulli depth Boltzmann Machine (Gaussian-Bernoulli Deep Boltzmann Machine, GDBM). The results show that the prediction accuracy of the proposed DBN model is significantly improved. Considering that CRBM has the advantage of temporal correlation, a depth DBN model based on CRBM is proposed in this paper. The original CRBM is only suitable for the input of spatial pixels such as images, but not suitable for the time series prediction. A novel input representation is proposed, which makes temporal data suitable for application in CRBM. Based on the above research, the DBN model based on RBM and the DBN model based on CRBM, which are suitable for time series data input, are applied to mobile traffic prediction, respectively. In this paper, the authors select 10 community users in hot areas of the present network to report the measurement reports (MR), and construct the time series in the unit of 20 minutes of time granularity respectively. The prediction results show that 80% of the residential areas have achieved more than 80% prediction accuracy.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F626;TP181

【相似文献】