基于线性复杂度自注意力机制的非侵入式负荷识别方法

廖耀华; 常艳平; 王恩; 魏龄; 潘国兵; 王海鹏

廖耀华,常艳平,王恩,魏龄,潘国兵,王海鹏.基于线性复杂度自注意力机制的非侵入式负荷识别方法[J].电测与仪表,2025,62(8):197-205.
LIAO Yaohua,CHANG Yanping,WANG En,WEI Ling,PAN Guobing,WANG Haipeng.A non-intrusive load monitoring method based on linear complexity self-attention mechanism[J].Electrical Measurement & Instrumentation,2025,62(8):197-205.

基于线性复杂度自注意力机制的非侵入式负荷识别方法

A non-intrusive load monitoring method based on linear complexity self-attention mechanism

DOI：j.issn1001-1390.2025.08.023

中文关键词: 智能电网边缘侧非侵入式负荷监测深度学习 UFO模块

英文关键词:smart grid, edge-side, non-intrusive load monitoring, deep learning, UFO module

基金项目:浙江省重点研发计划项目（2022C01244）

作者中文名	作者英文名	单位
廖耀华	LIAO Yaohua	云南电网有限责任公司电力科学研究院; 云南省绿色能源与数字电力量测及控保重点实验室
常艳平	CHANG Yanping	云南电网有限责任公司玉溪供电局
王恩	WANG En	云南电网有限责任公司电力科学研究院; 云南省绿色能源与数字电力量测及控保重点实验室
魏龄	WEI Ling	云南电网有限责任公司电力科学研究院;云南省绿色能源与数字电力量测及控保重点实验室
潘国兵	PAN Guobing	浙江工业大学机械工程学院
王海鹏	WANG Haipeng	浙江工业大学机械工程学院

摘要点击次数: 54

中文摘要:

非侵入式负荷监测(non-intrusive load monitoring，NILM)技术对于实现智慧用电与管理具有重要意义。针对现有的非侵入式负荷监测方法在高噪声环境下对特征相似电器以及微小负荷变化监测精度不足的难题，提出了一种基于单位力操作视觉变换器的非侵入式负荷监测(non-intrusive load monitoring based on unit force operated vision transformer, UFONILM)模型的非侵入式负荷监测的深度学习框架。UFONILM模型的单位力操作（unit force operated, UFO）模块通过层归一化和一系列卷积层有效地提取和利用了多尺度的时间序列数据，特征。在标准的UK-DALE数据集上进行的实验显示，UFONILM模型在准确性和F1得分上均优于现有方法，特别是在细粒度的负荷监测场景中。研制了基于UFONILM模型的嵌入式系统，实现了边缘计算的非侵入式负荷监测，可实时监测和响应电网中的异常用电行为，如违规充电事件。实验检测证明，UFONILM模型嵌入式非侵入式负荷监测方法在监测效率方面具有显著的提升，具有高效、便捷安装、可扩展等特点。

英文摘要:

Non-intrusive load monitoring (NILM) technology is of great significance for achieving smart power consumption and management. Aiming at the problems of insufficient recognition accuracy of existing non-intrusive load monitoring methods for feature-similar appliances and small load changes in high-noise environments, a deep learning framework for non-intrusive load monitoring based on the non-intrusive load monitoring based on unit force operated vision transformer（UFONILM）model is proposed. The unit force operated（UFO）module of the UFONILM model effectively extracts and utilizes multi-scale time series data features through layer normalization and a series of convolutional layers. Experiments on the standard UK-DALE dataset show that the UFONILM model outperforms existing methods in terms of accuracy and F1 score, especially in fine-grained load monitoring scenarios. An embedded system based on the UFONILM model is developed, which realizes edge computing non-intrusive load monitoring, which can identify and respond to abnormal electricity consumption behaviors in power grid in real time, such as illegal charging events. Experimental tests prove that the UFONILM model embedded non-intrusive load identification method has a significant improvement in identification efficiency, and it is characterized by high efficiency, convenient installation, and scalability.

查看全文查看/发表评论下载PDF阅读器

关闭