| 引用本文: | 崔昊杨,吴轶凡,江友华,江超,韩韬,许永鹏.非侵入式负荷识别的电流序列可视化方法[J].电力自动化设备,2022,42(7): |
| CUI Haoyang,WU Yifan,JIANG Youhua,JIANG Chao,HAN Tao,XU Yongpeng.Current sequence visualization method of non-intrusive load recognition[J].Electric Power Automation Equipment,2022,42(7): |
|
| 摘要: |
| 以单一特征为标签的用电设备识别，因特征携带的信息量不足，在区分性质相似的负荷时易产生误判，为此，提出一种将电流序列编码为图像的二维可视化方法，通过计算机视觉技术对负荷进行分类识别。利用Fryze功率理论提取电流的非有功分量，通过格拉姆角场(GAF)将一维电流序列转换成二维图像，借助数据扩充的方式进行升维，并赋予矩阵颜色特征来提升负荷标签的辨识度；基于迁移学习的思想，利用预训练模型Inception_v3提取并学习GAF图像特征，并以该特征为标签对负荷类型进行分类识别。在2个公开数据集上的实验验证了所提方法在高频采集场景下的准确性和有效性。 |
| 关键词: 非侵入式 负荷监测 二维可视化 计算机视觉 格拉姆角场 迁移学习 |
| DOI：10.16081/j.epae.202203002 |
| 分类号:TM73 |
| 基金项目:国家自然科学基金资助项目(52177185)；上海市科委地方院校能力建设项目(20020500700) |
|
| Current sequence visualization method of non-intrusive load recognition |
|
CUI Haoyang1, WU Yifan1, JIANG Youhua1, JIANG Chao1, HAN Tao2, XU Yongpeng3
|
|
1.College of Electronics and Information Engineering, Shanghai University of Electric Power, Shanghai 200090, China;2.State Grid Electric Power Research Institute, Nanjing 211106, China;3.School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
|
| Abstract: |
| The recognition of electrical equipment with single feature as the tag is prone to misjudgment when distinguishing loads with similar characteristics due to insufficient information carried by the feature, for which, a two-dimensional visualization method is proposed to encode current sequence into image, and the loads are classified and recognized by computer vision technology. Fryze power theory is used to extract the non-active component of current, one-dimensional current sequence is transformed into two-dimensional image by GAF(Gramian Angular Field),the dimension is increased by means of data expansion, and color feature is given to the matrix to improve the identification of load labels. Based on the idea of transfer learning, the pretrained model Inception_v3 is used to extract and learn features of GAF images, and the features are taken as the labels to classify and recognize load types. The experiments on two public data-sets verify the accuracy and effectiveness of the proposed method under high-frequency acquisition scenarios. |
| Key words: non-intrusive load monitoring two-dimensional visualization computer vision Gramian angular field transfer learning |
