【作者】 伍长荣；

【导师】 胡学钢；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2004， 硕士

【摘要】 多因素时间序列预测是数据挖掘的一个重要研究内容，描述预测指标与影响因素之间存在的潜在关系，被广泛应用于许多领域。经典的预测方法在用于非线性系统预测时有一定的困难，而RBF神经网络具有较好的非线性特性，特别适用于高度非线性系统的处理，为多因素时间序列预测开辟了新的发展空间。本文对基于RBF神经网络的预测模型进行了深入的研究，并详细研究了对网络输入空间的降维重构。论文主要内容如下： 采用RBF神经网络进行建模训练，并将结果与BP网络比较，仿真实验表明RBF网络的训练速度比BP网络显著加快，具有较好的泛化能力，能有效地应用于多因素时间序列预测。 将灰色关联分析(GRA)引入预处理过程，以消除与预测指标关联度相对小的影响因素，提出了基于GRA的RBF神经网络预测模型的约简，简化了网络结构，提高了预测精度。 针对多因素时间序列各因素之间存在相关性，导致信息重叠的缺点，提出了基于PCA的RBF神经网络预测模型的约简。文中利用PCA方法对原有指标体系进行处理，提取主成分构成新的指标作为RBF神经网络的输入，优化了网络结构，提高了网络的泛化能力。 将上述两种约简方法相结合，提出了基于GRA-PCA的RBF神经网络预测模型的约简，减少了采集样本数目，提高了建模效率和预测精度。更多还原

【Abstract】 The multifactor time series prediction is an important part of Data Mining, which describes the potential relationships between prediction indexes and influential factors, and has a vast application in many fields. Because the general predicting methods are based on linear analysis, when they deal with non-linear cases they will meet many difficulties. However, the RBF Neural Network has excellent non-linear character, especially for non-linear proceeding. The predicting methods based on RBF Neural Network extend the space of predicting research. In this dissertation, the RBF Neural Network prediction model and the original input space reconstruction of the RBF network are studied in detail. The main work is as the following:The RBF Neural Network is applied to model training and its experiment result is compared with that of the popular BP network. The simulation shows that the training speed of RBF network is obviously faster than that of BP network and the generalization ability of this network is also better. Therefore, it is effective to apply RBF network in the multifactor time series prediction.By introducing GRA in the process of the pretreatment to remove the smaller grey relation factors, a reduction of the RBF network forecasting model based on GRA is presented. This method simplifies the ANN structure, and improves the forecasting ?precision.The reduction of forecasting model based on the PCA is put forward to solve the correlation of influence factors which cause index message redundancy. This dissertation abstracts the prime factors from the training samples as input variables of RBFNN. The method optimizes the structure of network, and improves the generalization of network.Combining the above two methods of reduction, this dissertation brings forward a reduction of the RBF network forecasting model based on the GRA-PCA. This method reduces collecting samples, and improves modeling efficiency and forecasting precision.更多还原