摘要
为实现船舶电力系统电能质量扰动准确识别,结合深度学习提出基于二维残差网络(2D-ResNet)的电能质量扰动识别方法。首先将电能质量一维时间序列通过距离矩阵转化为二维平面图,随后将图像送入所提二维残差网络中提取特征。最终输出特征图通过线性层分类器得到识别结果,实现船舶电力系统电能质量扰动的在线识别。与现有特征提取方法相比,不同信噪比下该方法扰动识别准确率均最高。信噪比为20 dB时,单标签分类平均准确率为93.86%,多标签分类平均F_(1)-score为96.52%,证明了2D-ResNet能有效提取扰动特征且对噪声具备鲁棒性。对于未知复合扰动,单标签分类器识别失败,而多标签分类器准确识别出扰动中的未知成分,且F_(1)-score达到93%,证明了多标签分类适用于未知复合扰动识别。
For accurate classification,a power quality disturbance recognition method of a shipboard power system based on a two dimensional residual network(2D-ResNet)is proposed.First,the one-dimensional power quality time series is transformed into a two-dimensional image by a distance matrix,and then the image is sent to the proposed 2D-ResNet to extract features.Then an output feature map is used to obtain the recognition results through the linear layer classifier to realize on-line recognition of power quality disturbances in a shipboard power system.Compared with existing feature extraction methods,this method has the highest accuracy of disturbance recognition under different signal-to-noise ratio(SNR).When the SNR is 20 dB,the average accuracy of single-label classification is 93.86%,and the average F_(1)-score of multi-label classification is 96.52%.This proves that the 2D-ResNet can effectively extract features and is robust to noise.A single-label classifier fails to recognize unknown compound disturbance,while the multi-label classifier accurately recognizes the unknown components in the disturbance signal,and the F_(1)-score reaches 93%,which proves that the multi-label classification is suitable for the recognition of unknown compound disturbance.
作者
宋铁维
施伟锋
毕宗
谢嘉令
SONG Tiewei;SHI Weifeng;BI Zong;XIE Jialing(Department of Electrical Automation,Shanghai Maritime University,Shanghai 201306,China)
出处
《电力系统保护与控制》
EI
CSCD
北大核心
2022年第10期94-103,共10页
Power System Protection and Control
基金
上海市科技计划项目资助(20040501200)。
关键词
船舶电力系统
电能质量
二维残差网络
扰动识别
单标签分类
多标签分类
shipboard power system
power quality
two dimensional residual network
disturbance identification
single-label classification
multi-label classification