基于注意力机制和LSTM-LightGBM的特高压直流输电线路可听噪声无效数据清洗方法  

Invalid Data Cleaning Method of Audible Noise for UHV HVDC Transmission Lines Based on Attention Mechanism and LSTM-LightGBM

作　　者：吴海荣 李振华[1] 程紫熠 张传计 WU Hairong;LI Zhenhua;CHENG Ziyi;ZHANG Chuanji(College of Electrical Engineering and New Energy,China Three Gorges University,Yichang,Hubei 443002,China;State Grid Chongqing Electric Power Company Construction Branch,Chongqing 401120,China;Wuhan National High Magnetic Field Center,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443002 [2]国家电网重庆市电力公司建设分公司,重庆401120 [3]华中科技大学国家脉冲强磁场科学中心,武汉430074

出　　处：《南方电网技术》2024年第8期115-123,140,共10页Southern Power System Technology

基　　金：国家自然科学基金资助项目(52311530337);武汉强磁场学科交叉基金资助(WHMFC202202)。

摘　　要：特高压直流输电线路可听噪声试验过程中,外界环境的突发性干扰会使实验数据中掺杂较多的无效数据,严重影响后续的数据分析。提出了一种基于注意力机制(attention mechanism,AM)和长短时记忆网络-轻量级梯度提升机(long short-term memory network-light gradient boosting machine,LSTM-LightGBM)的输电线路可听噪声无效数据清洗方法。首先,针对可听噪声数据的非线性、高维时序冗余特征等特点,以LSTM神经网络为基础进行特征提取;同时,引入特征维度注意力机制,自适应地分配权重来刻画关键特征信息的表达能力;进而,利用LightGBM对提取到的特征进行分类,检测出无效数据;然后,以某特高压直流输电线路实测可听噪声数据试验分析,结果表明该方法的检测精准率为95.55%,召回率为97.73%,F1分数为0.9663,均优于对比实验模型;最后,将无效数据删除并使用均值插补法填补,无效数据清洗后数据的50%值和95%值基本不变,仅降低无效数据的最大值和5%值。该算法对提高输电线路可听噪声数据的可靠性具有一定参考意义。During the audible noise test of UHV HVDC transmission lines,the sudden interference of the external environment will make the experimental data doped with more invalid data,which seriously affects the subsequent data analysis.In this paper,a method based on attention mechanism(AM)and long short-term memory network-light gradient boosting machine(LSTMLightGBM)is proposed to clean the invalid data of the transmission lines with audible noise.Firstly,feature extraction is carried out based on LSTM neural network,aiming at the characteristics of nonlinear and high-dimensional temporal redundancy of audible noise data.At the same time,the feature dimension attention mechanism is introduced,and the weights are allocated adaptively to describe the expressive ability of key feature information.Then,LightGBM is used to classify the extracted features and detect invalid data.Then,the measured audible noise data of an UHV HVDC transmission line is analyzed experimentally.The results show that the detection accuracy rate of this method is 95.55%,the recall rate is 97.73%,and the score of F1 is 0.9663,which are superior to the comparison experimental model.Finally,the invalid data is deleted and filled with the mean interpolation method.After the invalid data is cleaned,the 50%value and 95%value of the data remain basically unchanged.Only the maximum value and 5%value of the invalid data are reduced.This method has certain reference significance for improving the reliability of audible noise data of transmis⁃sion lines.

关 键 词：输电线路 可听噪声 长短时记忆网络 注意力机制 轻量级梯度提升机 无效数据 

分 类 号：TM75[电气工程—电力系统及自动化]