基于贝叶斯优化随机森林的变压器故障诊断  

Transformer fault diagnosis based on Bayesian optimized random forest

作　　者：王雪[1] 韩韬 Wang Xue;Han Tao(Department of Electrical Engineering,North China Electric Power University,Baoding 071000,Hebei,China)

机构地区：[1]华北电力大学电力工程系,河北保定071000

出　　处：《电测与仪表》2021年第6期167-173,共7页Electrical Measurement & Instrumentation

基　　金：河北省教育厅指导性计划项目(Z2012063)。

摘　　要：针对集成学习参数众多,缺乏高效准确的参数寻优方法的问题,文章提出了基于贝叶斯优化随机森林(RF)的变压器故障诊断方法。该方法采用了多个决策树构成RF故障诊断模型,然后将高斯过程(GP)作为概率代理模型、提升策略(PI)作为采集函数,构建贝叶斯优化(BO)算法,进行RF模型参数寻优。此外,还对支持向量机(SVM)和K最近邻(KNN)两种模型进行贝叶斯优化并对比。在RF模型上,将贝叶斯优化与随机搜索优化进行性能对比。实验结果表明:RF模型经贝叶斯参数寻优后,诊断准确率有明显提高;与随机搜索优化方法相比,贝叶斯优化搜索的模型参数更优,寻优效率更高。Aiming at the problem that the ensemble learning has many parameters and lacks of efficient and accurate parameter optimization methods,this paper proposes a transformer fault diagnosis method based on Bayesian optimized random forest(RF).The method adopts multiple decision trees to form an RF fault diagnosis model.Then,the Gaussian process is used as a probabilistic proxy model and a lifting strategy(PI)as an acquisition function to construct a Bayesian optimization algorithm for RF model parameter optimization.In addition,Bayesian optimization is performed on support vector machine(SVM)and K nearest neighbor(KNN)respectively for comparison.The performance between Bayesian optimization and random search optimization was compared on the RF model.The experimental results show that the diagnostic accuracy of the RF model is significantly improved with Bayesian parameters optimization.Compared with random search optimization,Bayesian optimization method can find better model parameters with higher efficiency.

关 键 词：变压器 DGA 故障诊断 贝叶斯优化 RF算法 

分 类 号：TM42[电气工程—电器]