多时间尺度小波变换和LSTM自编码器电弧故障检测方法
Multi time scale wavelet transform and LSTM autoencoder arc fault detection method
邱婷婷 1曹文平 1刘孝宇 1漆星1
作者信息
- 1. 安徽大学 电气工程与自动化学院,安徽 合肥 230001
- 折叠
摘要
在光伏发电系统中,电弧故障检测是维持系统安全运行的关键问题.以往的电弧故障检测方法大多基于单时间尺度的故障特征,然而单一时间尺度特征往往会受到环境变化的干扰,导致检测精度降低,针对这一问题,提出一种多时间尺度小波和长短时记忆(LSTM)自编码器电弧故障检测方法,该方法首先在机理分析的基础上找到电弧3 个特性,即电弧初始阶段电流发生突变、燃弧阶段电流均值降低、燃弧阶段高频分量变大.再基于上述电弧特性进行小波变换提取对应多尺度特征,然后使用LSTM自编码器进行端到端的自动检测.与以往方法不同,该方法提取了电弧特性的多种时间尺度特征,增加了故障信号的检测依据,降低了受外界干扰时检测结果出现误报漏报的可能性.理论分析和实验结果表明,所提出的方法降低了故障电弧检测的误报率,提高了其准确率.
Abstract
In the photovoltaic power generation system,arc fault detection is the key to maintain the safe operation of the system.Most of the previous arc fault detection methods are based on the fault features of single time scale.However,the single time scale features are often disturbed by environmental changes,resulting in the reduction of detection accuracy.To solve this problem,a multi time scale wavelet and long-short memory(LSTM)autoencoder arc fault detection method was proposed.Firstly,on the basis of mechanism analysis,the method finds three characteristics of the arc,namely,the sudden change of cur-rent in the initial stage of the arc,the average value of current in the arcing stage decreases and the high frequency component in the arcing stage increases;Based on the above arc characteristics,wavelet trans-form was performed to extract the corresponding multi-scale features,and then LSTM self-encoder was used for end-to-end automatic detection.Unlike the previous methods,this method extracts the multi-time scale features corresponding to the arc characteristics,which increases the detection basis of the fault signal and reduces the possibility of false alarm and missing alarm in the detection result when the fault signal is interfered by the outside.Theoretical analysis and experimental results show that the proposed method reduces the false alarm rate of fault arc detection and improves its accuracy.
关键词
光伏发电/电弧故障/单类/小波变换/长短时记忆自编码器/多时间尺度特征Key words
photovoltaic power generation/arc fault/one-class/wavelet transform/long-short memory utoencoder/multi-time scale features引用本文复制引用
基金项目
科技合作专项国际科技合作项目(2022h11020023)
出版年
2024
NETL
NSTL
维普
万方数据