| 引用本文: | 王守相,赵宁,王同勋,郭陆阳,魏孟迪,赵倩宇,冯丹丹.基于轨迹颜色编码与知识蒸馏的电能质量扰动轻量化识别方法[J].电力自动化设备,2024,44(12):85-91. |
| WANG Shouxiang,ZHAO Ning,WANG Tongxun,GUO Luyang,WEI Mengdi,ZHAO Qianyu,FENG Dandan.Lightweight identification method for power quality disturbance based on trajectory color coding and knowledge distillation[J].Electric Power Automation Equipment,2024,44(12):85-91. |
|
| |
|
|
| 本文已被：浏览 5958次 下载 837次 |
码上扫一扫！ |
|
|
| 基于轨迹颜色编码与知识蒸馏的电能质量扰动轻量化识别方法 |
|
王守相1,2, 赵宁1,2, 王同勋3, 郭陆阳1,2, 魏孟迪1,2, 赵倩宇1,2, 冯丹丹3
|
|
1.天津大学 智能电网教育部重点实验室，天津 300072;2.电力系统仿真控制天津市重点实验室，天津 300072;3.国网智能电网研究院有限公司，北京 102209
|
|
| 摘要: |
| 当前电能质量扰动识别中存在特征混叠导致电压信号特征难以提取的问题，且电能质量识别装置部署在变电站等场所，算力较低，难以运行复杂的识别模型。为此，提出了基于轨迹颜色编码与知识蒸馏的电能质量扰动轻量化识别方法。通过圆轨迹转换和颜色编码，将1维电压信号转化为特征明显的彩色圆轨迹特征图，解决了特征混叠问题。使用二进制映射将特征图像降维，实现特征轻量化。构建以ResNet18为教师模型和SqueezeNext为学生模型的知识蒸馏模型，使用教师模型训练生成的软标签指导学生模型的训练，进行2个模型之间的知识迁移，实现模型压缩。为了验证所提方法的有效性和实用性，在Raspberry Pi-4B边缘设备上进行验证实验，结果表明所提方法能够有效解决特征混叠问题，且识别速度相对压缩前提高了82.84 %，能够满足轻量化部署的需求。 |
| 关键词: 电能质量扰动 轻量化 颜色编码 模型压缩 知识蒸馏 边缘设备 |
| DOI：10.16081/j.epae.202410002 |
| 分类号:TM715 |
| 基金项目:国家重点研发计划资助项目(2023YFB2407500) |
|
| Lightweight identification method for power quality disturbance based on trajectory color coding and knowledge distillation |
|
WANG Shouxiang1,2, ZHAO Ning1,2, WANG Tongxun3, GUO Luyang1,2, WEI Mengdi1,2, ZHAO Qianyu1,2, FENG Dandan3
|
|
1.Key Laboratory of Smart Grid of Ministry of Education, Tianjin University, Tianjin 300072, China;2.Tianjin Key Laboratory of Power system Simulation and Control, Tianjin 300072, China;3.State Grid Smart Grid Research Institute, Beijing 102209, China
|
| Abstract: |
| At present, there is a problem that the feature aliasing in the power quality disturbance identification makes it difficult to extract the characteristics of the voltage signal, and the power quality identification device is deployed in substations and other places, and the computing power is low, so it is difficult to run a complex recognition model. Therefore, a lightweight identification method for power quality disturbance based on trajectory color coding and knowledge distillation is proposed. Through circular trajectory conversion and color coding, the one-dimensional voltage signal is transformed into a color circular trajectory feature map with obvious features, which solves the problem of feature aliasing. Binary mapping is used to reduce the dimensionality of the feature image to achieve feature lightweight. Then, a knowledge distillation model with ResNet18 as the teacher model and SqueezeNext as the student model is constructed, and the soft labels generated by the teacher model training are used to guide the training of the student model, and the knowledge transfer between the two networks is carried out to achieve model compression. In order to verify the effectiveness and practicability of the proposed method, the verification experiments are carried out on Raspberry Pi-4B edge devices, and the results show that the proposed method can effectively solve the problem of feature aliasing, and the identification speed is increased by 82.84% compared with that before compression, which can meet the needs of lightweight deployment. |
| Key words: power quality disturbance lightweight color coding model compression knowledge distillation edge device |
|
|
|
|
