最小最大概率机在时间序列预测中的应用研究

发布时间：2018-04-27 12:38

本文选题：时间序列 + 最小最大概率回归机　；参考：《兰州交通大学》2014年硕士论文

【摘要】：时间序列预测伴随时代的进步日益重要，应用研究领域无处不在，常见的例如经济预测、天气预测、交通流预测和网络流量预测等。智能交通系统预测研究是路网交通流量实时在线控制规划的重要信息源泉。互联网网络流量实时帧数据合理分配对网络拥塞的缓解和网络安全的管理提供了便捷帮助。最小最大概率回归机(Minimax Probability Machine Regression,MPMR)是一种将概率分类机器学习用于解决回归问题的新型预测方法，在掌纹识别、图像分割、数据挖掘、电力预测等领域中得到了广泛的应用。文中结合混沌理论、递归图可预测性分析，将MPMR方法用于交通流和网络视频流的单步预测和直接多步预测实验中，通过核函数映射，在最优核参数条件下获取能够最大概率容纳预测点落入最小回归管道内的epsilon管道值，并与支持向量机(Support Vector Machine,SVM)预测方法、人工神经网络预测方法进行预测实验比较，验证了该方法的优越性。本文主要研究内容包括以下几个方面： (1)在贝叶斯学习的基础上，研究了线性最小最大概率机分类(Minimax ProbabilityMachine,MPM)方法、非线性最小最大概率机分类(Minimax Probability MachineClassification,MPMC)方法，并将其延伸至MPMR回归方法。 (2)针对非线性时间序列，研究了相应的混沌理论，进一步利用最大李雅普诺夫指数判别三组时间序列的混沌特性，并研究了确定最优嵌入维数m的Cao方法，确定最优延迟时间τ的互信息法，和判断时间序列可预测性的递归图方法。 (3)将概率学习机MPMR方法应用在Mackey-Glass混沌时间序列、短时交通流及网络视频流预测应用中，，并与现有同等条件下的预测方法比较实验效果，验证该方法的先进性和有效性。 (4)基于RBF核函数，MPMR研究了相应预测回归管道选取不同值时对预测精度的影响，验证了该方法的鲁棒性。
[Abstract]:Time series prediction is becoming more and more important with the development of the times. The applied research fields are ubiquitous, such as economic forecasting, weather forecasting, traffic flow forecasting and network traffic forecasting. Intelligent Transportation system (its) prediction is an important source of information for real-time and on-line control planning of road network traffic flow. The reasonable allocation of real-time frame data of Internet traffic provides convenient help to alleviate network congestion and manage network security. Minimax Probability Machine regression machine (MPMRs) is a new prediction method which uses probabilistic classification machine learning to solve regression problems. It has been widely used in palmprint recognition, image segmentation, data mining, power prediction and so on. Combined with chaos theory and recursive graph predictive analysis, MPMR method is applied to single step prediction and direct multistep prediction of traffic flow and network video flow. Under the condition of optimal kernel parameters, the value of epsilon pipeline with maximum probability of accommodating the predicted point into the minimum regression pipeline is obtained, and compared with the support vector machine support Vector machine prediction method and the artificial neural network prediction method. The superiority of this method is verified. The main contents of this paper include the following aspects: 1) on the basis of Bayesian learning, the minimax probability machine classification method and the nonlinear minimum maximum probability machine classification method are studied, and the method is extended to the MPMR regression method. (2) for the nonlinear time series, the corresponding chaos theory is studied, and the chaos characteristics of the three groups of time series are judged by using the maximum Lyapunov exponent, and the Cao method for determining the optimal embedding dimension m is studied. The mutual information method for determining the optimal delay time 蟿 and the recursive graph method for judging the predictability of time series. 3) the probabilistic learning machine (MPMR) method is applied to the prediction of Mackey-Glass chaotic time series, short-term traffic flow and network video flow, and the experimental results are compared with the existing prediction methods under the same conditions. The results show that the proposed method is advanced and effective. 4) based on the RBF kernel function, the influence of different values on the prediction accuracy is studied, and the robustness of the method is 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】：2014
【分类号】：U491.14;TP18

【参考文献】