基于BP神经网络模型的复合绝缘子伞裙优化技术  

Optimization Technology of Composite Insulator Shed Based on BP Neural Network Model

作　　者：李晓刚 谢敏 刘祝鸿 周振安 张楚岩 LI Xiaogang;XIE Min;LIU Zhuhong;ZHOU Zhen’an;ZHANG Chuyan(Peral River Delta Research Institute of Tsinghua,Guangzhou 510700,China;Guangzhou Guanghua Electric Technology Limited Cooperation,Guangzhou 510700,China;China University of Geosciences(Beijing),Beijing 100083,China)

机构地区：[1]清华珠三角研究院,广州510700 [2]广州广华智电科技有限公司,广州510700 [3]中国地质大学(北京),北京100083

出　　处：《高压电器》2022年第11期98-105,共8页High Voltage Apparatus

基　　金：广东省重点领域研发计划资助(2021B0101420002)。

摘　　要：伞裙参数对复合绝缘子的性能有着决定性的影响,为在确定的绝缘距离下得到污闪性能优异的复合绝缘子,文中分别采用人工污秽试验和机器学习的方法对复合绝缘子伞裙结构参数的优化问题开展了研究。机器学习的基础是数据,文中首先对不同伞裙结构参数的12种复合绝缘子试品完成了人工污秽试验,获得了关键伞裙参数对交流污闪电压的影响以及具有优异污闪性能的伞裙参数需满足的条件;其次,为深度挖掘人工污秽试验数据的价值,文中使用RMSprop梯度下降算法建立了以复合绝缘子伞裙结构参数和表面污秽度为输入的4层BP神经网络模型,可实现对交流污闪电压的预测以及性能比较,结果表明模型预测的平均误差小于6%,最大误差小于10%,采用机器学习算法与采用人工污秽试验所得的复合绝缘子伞裙优化结果是一致的,结论可为复合绝缘子的产品设计和性能校验提供参考。Parameters of shed have a decisive impact on the performance of composite insulators.In order to obtain composite insulators with excellent flash⁃over performance under the determined insulation distance,the optimiza⁃tion of shed structure parameters of composite insulator is studied in this paper by artificial contamination tests and machine learning method,respectively.The base of machine learning is data.First,the artificial contamination test on 12 kinds of composite insulator samples with different shed structure is performed in this paper,the influence of key shed parameters on AC flash⁃over voltage and condition that the shed structure with superior flash⁃over perfor⁃mance shall meet are obtained.Then,in order to deeply mine the value of artificial contamination test data,the RM⁃Sprop gradient descent algorithm is used in this paper to set up a 4⁃layer BP neural network model with the shed pa⁃rameters and surface pollution severity of composite insulator as the input,which can achieve AC flash⁃over voltage prediction and performance comparison.The results show that the average error predicted by the model is less than 6%and the maximum error is less than 10%.The optimization results of composite insulator shed structure,obtained by machine learning method and by artificial contamination,is identical.And the conclusion can provide a reference for the product design and performance verification of composite insulator.

关 键 词：复合绝缘子 污闪电压 伞裙优化 机器学习 人工神经网络 

分 类 号：TM216[一般工业技术—材料科学与工程] TP183[电气工程—电工理论与新技术]