摘要
油色谱在线监测对于电力变压器的运行维护、健康状态分析具有重要意义,然而受监测设备异常、外界环境干扰、运行状态变化等因素影响,监测数据中难免存在不同类型的异常数据。为此,提出了一种基于多层架构的油中溶解气体数据清洗与异常识别方法。首先利用变分模态分解去除时间序列中的趋势项,结合3σ准则对时序数据中的噪声值、缺失值、暂时性迁移数据等进行异常识别;然后根据关联分析结果对可清洗的异常数据利用长短期记忆神经网络进行重构清洗;并结合时间序列分段以及改进SAX算法实现对时序数据中的趋势性异常状态检测。最后,结合实例分析表明本文所提方法能够实现异常数据的有效清洗以及对趋势异常状态的准确识别。
Oil chromatography online monitoring is of great significance in the operational maintenance and health status analysis of power transformers.However,due to factors such as abnormal monitoring equipment,external environmental interference,and changes in operating status,different types of abnormal data are inevitable in the monitoring data.Therefore,we propose a data cleaning and abnormal recognition method for dissolved gas in oil based on multi-layer architecture.First,we used the variational mode decomposition to remove the trend items from the time series,and identified the noise values,missing values,and temporary migration data in the time series data with the 3σcriterion.Then,according to the results of the correlation analysis,the cleanable abnormal data was reconstructed and cleaned with the long short-term memory neural network.With time series segmentation and improved SAX algorithm,we realized the detection of trending abnormal state in the time series data.Finally,we analyzed the examples and the result shows that the method proposed in this paper can effectively clean abnormal data and accurately identify the abnormal trend state.
作者
刘云鹏
王权
许自强
刘一瑾
和家慧
韩帅
LIU Yunpeng;WANG Quan;XU Ziqiang;LIU Yijin;HE Jiahui;HAN Shuai(Hebei Provincial Key Laboratory of Power Transmission Equipment Security Defense,North China Electric Power University,Baoding 071003,China;State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources(North China Electric Power University),Beijing 102206,China;State Grid Nanjing Power Supply Company,Nanjing 210019,China;China Electric Power Research Institute,Beijing 100192,China)
出处
《华北电力大学学报（自然科学版）》
CAS
北大核心
2022年第1期81-89,共9页
Journal of North China Electric Power University：Natural Science Edition
基金
国家电网有限公司总部科技项目(5500-202055419A-0-0-00)。
关键词
变压器
数据清洗
变分模态分解
3σ准则
长短期记忆神经网络
SAX
transformer
data cleaning
variational modal decomposition
3σcriterion
long short-term memory neural network
SAX
