相对变换KFCM的变压器油击穿电压预测

Prediction Model of Breakdown Voltage for Transformer Oil Based on Relative Transformation and KFCM

针对变压器油击穿电压在线测量困难,基于多模型结构可以提高预测精度和鲁棒性的思想,提出基于相对变换核模糊C均值聚类（Kernel Fuzzy C-means, KFCM）算法的变压器油击穿电压预测建模方法。首先,采用相对变换将原始数据空间变换到相对空间,抑制数据噪音,提高数据之间的可区分性;在相对空间中利用KFCM算法将样本划分成不同的子类,同时,对KFCM核参数和聚类数采用差分进化算法进行优化;然后,利用核主元分析（Kernel Principal Component Analysis, KPCA）对相对空间进行特征提取,降低数据维数、提取数据非线性主元作为各子类构建的最小二乘支持向量机（Least Squares SVM, LSSVM）模型的输入;最后,对子类LSSVM综合加权得到最终输出。将所提出的方法与KFCMLSSVM方法进行比较,实验结果表明所提方法具有良好的预测精度和泛化性能。

Aiming at the difficulty of measuring the breakdown voltage of transformer oil in real time, based on the idea that the combination of multi models can improve the model accuracy and robustness, a modeling method based on relative transformation（RT） and kernel fuzzy C-means（KFCM） clustering algorithm is proposed. Firstly, the original data space is converted to the relative data space by RT in which data noise can be suppressed and the transformed data is more distinguishable; Secondly, samples in relative space are divided into subsets with different cluster centers by using KFCM, meanwhile, the differential evolution algorithm is used to search the optimal kernel parameter and the clustering categories; then, KPCA is employed in the relative space for the purpose of data dimension reduction, and extracting nonlinear features as the input of least squares support vector machines（LSSVM） sub-model; Finally, the system output is obtained by summing each sub-model＇s weighted output comprehensive. Utilizing average relative error indexes to analyze the performance of the proposed method and KFCM-LSSVM, simulation results illustrate that the proposed prediction method has better prediction and generalization performance.