基于偏最小二乘神经网络的大电机定子绝缘击穿电压混合预测方法

应用混合计算智能方法，进行大型发电机定子绝缘击穿电压预测以对其剩余寿命进行评估。为了解决在样本数较少及自变量间存在多重相关性时的击穿电压预测问题，文中通过将RBF神经网络和偏最小二乘(PLS)集成在一起，来计算PLS输入的外部模型权值，利用PLS方法提取变量成份来降低输入变量维数，这样消除了变量建模时的共线性，从而大大提高PLS的建模能力，同时利用RBF神经网络的非线性拟合能力来捕获变量投影空间的非线性关系。另外，在建模过程中对原始数据进行了中心化和压缩处理，以消除参数的量纲效应。大电机定子击穿电压预测试验结果表明：混合模型的预测结果明显优于传统的预测模型。

A Hybrid Prediction Approach for Stator Insulation Breakdown Voltage of Large Generator Based on PLS Combined With Artificial Neural Network

A hybrid computing intelligence approach was used in the residual life prediction of large generator stator insulation.Aimed at the problems of few samples and multi-coUinearity of variables in complicated data modeling, RBF neural network was embedded into the regression framework of Partial Least Square (PLS) method.The PLS method was used to extract variable components from sample data and the dimension of input variables was then reduced. Moreover,RBF neural network was used to fit the non-linearity between input and output variables in projection space,and the disadvantages of traditional modeling methods were overcome. In addition,in order to avoid parameter dimension effects in modeling,a centered-compress data preprocessing method was adopted.Finally,this approach was applied to the prediction of residual breakdown voltage of the large generator stator winding.The test results show that that of the hybrid model has better prediction ability than traditional prediction methods.