基于核主成分降维的RBF网络降水预测

On RBF Network Precipitation Forecast Based on Kernel Principal Component Dimension Reduction

针对径向基函数神经网络(RBF网络)的隐层节点数、中心和宽度难以确定的问题,为提高网络性能,首先采用模糊聚类分析法对样本数据进行初始聚类,以初始分类间的最小距离作为初始宽度;然后引入正交最小二乘法训练出新的数据中心、个数及权值,修改宽度为当前数据中心间的最小距离;最后采用梯度下降法训练并调整中心、宽度及权值;几种算法进行的组合优化改进,使网络泛化性能更优.由于降水影响因子众多,采用了核主成分分析法(KPCA)对样本数据进行特征提取降维预处理.对广西5月3区的日降水量使用上述模型进行预报实验,结果表明,该模型具有较好的泛化性能,预报准确率高于T213降水预报模式,具有一定的推广价值.

Aiming at the problems that are difficult to identify the number of hidden layer nodes in the radial basis function neural network （RBF network）, centers and widths, this paper, in order to improve the network performance ,firstly used the fuzzy clustering analysis method to the sample data for the initial clustering with an initial classification of minimum distance between as the initial width;then, in- troduced the orthogonal least squares to train a new data center, the number and weights to modify the width between the current data cen- ter minimum distance;finally adopted the gradient descent to train and adjust the center,width and weight. Several algorithms were com- bined and optimized to better the generalization performance of network. Due to the rainfall affecting factors varied, The nuclear principal component analysis（KPCA） applied in the sample data pretreatment feature extraction dimension reduction. This method has been applied in 3 districts of Guangxi＇ s daily precipitation in May, the results show that the model had a good generalization performance, the forecasting accuracy was higher than that of T213 precipitation forecast model,it has certain popular value.