变压器油中气体的多核核主元回归预测模型  

Multiple-kernel kernel principal component regression model for dissolved gas in transformer oil

作　　者：唐勇波[1,2] 桂卫华[1] 彭涛[1] 

机构地区：[1]中南大学信息科学与工程学院,湖南长沙410083 [2]宜春学院物理科学与工程技术学院,江西宜春336000

出　　处：《电机与控制学报》2012年第11期92-98,共7页Electric Machines and Control

基　　金：国家自然科学基金(61134006)

摘　　要：为了克服常用预测方法在建模时只单独考虑某种特征气体发展变化的不足,进一步提高预测的精度和可靠性,提出了一种基于多核核主元回归(multiple-kernel kernel principal componentregression,MK-KPCR)的变压器油中气体预测模型。采用不同类型核函数的线性加权组合构造新的等价核,以降低建模精度对核函数及其参数选择的依赖性;利用核主元分析(kernel principalcomponent analysis,KPCA)对变压器油中溶解气体样本数据提取核主元,进行回归计算与分析,建立同时预测变压器油中主要特征气体的核主元回归(kernel principal component regression,KPCR)模型;与灰色多变量预测模型(multivariable grey model,MGM),主元回归(principal component re-gression,PCR)及KPCR进行一步和多步预测比较。实验结果表明,MK-KPCR预测模型对核函数及参数选择的依赖性小,具有较优的预测精度和泛化能力。On the basis of analyzing available prediction models, the multiple-kernel kernel principal component regression （MK-KPCR） is proposed to forecast the related dissolved gases in power transformer oil in order to further improve accuracy and reliability of the forecasting, and make up the disadvantage that only one feature gas is considered. An equivalent kernel was built by linear-weighted combination of multiple kernels to reduce the dependence of modeling accuracy on kernel function and parameters. The kernel principal component analysis （KPCA） model of dissolved gas-in-oil samples was applied to choose the kernel principal component used to establish kernel principal component regression （KPCR） model. Compared with muhivariable grey model （MGM）, principal component regression （PCR） and KPCR, one-step and multi-step prediction experimental results show that MK-KPCR approach lies little on kernel function and parameters and has a better performance of prediction and generalization.

关 键 词：电力变压器 油中气体 预测 核主元回归 多核学习 

分 类 号：TM855[电气工程—高电压与绝缘技术] TP183[自动化与计算机技术—控制理论与控制工程]