基于LSTM-GMM的配电干式变压器绕组故障双参数预警模型  

Dual Parameter Early Warning Model for Winding Faults in Distribution Dry-type Transformers Based on LSTM-GMM

作　　者：高煦轲 秦超 高沨 王璁[1] 屠幼萍[1] GAO Xu-ke;QIN Chao;GAO Feng;WANG Cong;TU You-ping(Beijing Key Laboratory of High Voltage and Electromagnetic Compatibility,North China Electric Power University,Beijing 102206,China;Beijing Longdeyuan Electric Power Technology Development Co.,Ltd.,Beijing 100080,China)

机构地区：[1]华北电力大学高电压与电磁兼容实验室,北京102206 [2]北京龙德缘电力科技发展有限公司,北京100080

出　　处：《科学技术与工程》2024年第31期13407-13414,共8页Science Technology and Engineering

基　　金：北京市自然科学基金(3192083)。

摘　　要：针对配电网干式变压器监测手段有限导致的监测数据种类较少、故障预警难度较大的问题,提出了一种基于长短期记忆结合高斯混合模型(long short-term memory and gaussian mixture model,LSTM-GMM)算法的配电网干式变压器绕组故障双参数预警模型。首先,将监测数据预处理后提取与变压器绕组温度强相关的特征值,作为LSTM网络的输入;然后对LSTM网络超参数调优,以绕组温度为目标进行网络训练;通过训练后网络输出的预测值计算得到残差集,使用GMM确定残差集的概率密度分布,以置信区间作为故障判据划分预警等级;最后将预警等级转化为故障率。北京益丰园居民区干式变压器绕组故障数据分析的结果表明,本文所提出的双参数预警模型预测效果优于反向传播神经网络等算法,可提前3 h实现故障预警。A dual parameter fault warning model for dry-type transformers in distribution networks was proposed in this paper.The model is based on the LSTM-GMM algorithm and aims to address the challenges posed by limited monitoring data types and the difficulty in fault warning caused by limited monitoring methods.Firstly,the monitoring data was preprocessed,and feature values strongly related to the temperature of the transformer winding were extracted,which were used as inputs for the long short term memory(LSTM)network.Then,the hyperparameters of the LSTM network were optimized,and the network was trained with the winding temperature as the target variable.The residual set was calculated based on the predicted values output by the trained network.A Gaussian mixture model(GMM)was employed to determine the probability density distribution of the residual set,and the fault criterion was defined using the confidence interval to classify the warning level.Finally,the warning level was converted into a failure rate.The analysis of winding fault data for dry-type transformers in residential areas of Yifengyuan,Beijing that the proposed dual parameter warning model exhibits better predictive performance compared to algorithms such as backpropagation neural networks and can achieve a fault warning 3 hours in advance.

关 键 词：长短期记忆网络 高斯混合模型 故障预警 干式变压器 配电网 

分 类 号：TM307[电气工程—电机]