中国电力 ›› 2025, Vol. 58 ›› Issue (5): 33-42.DOI: 10.11930/j.issn.1004-9649.202403063

• 面向新型配电系统的人工智能与新能源技术 • 上一篇    下一篇

融合SENet注意力机制和GA-CNN的非侵入式负荷识别方法

沈鑫1(), 王钢2, 赵毅涛1, 骆钊2(), 李钊2, 杨晓华1   

  1. 1. 云南电网有限责任公司,云南 昆明 650217
    2. 昆明理工大学 电力工程学院,云南 昆明 650500
  • 收稿日期:2024-03-18 发布日期:2025-05-30 出版日期:2025-05-28
  • 作者简介:
    沈鑫(1981),男,高级工程师,从事智能电网技术研究,E-mail:23755803@qq.com
    骆钊(1986),男,通信作者,副教授,从事区块链、电力市场与电力监管、能源互联网与电力规划及人工智能在电力系统中的应用等方面研究,E-mail:waiting.1986@live.com
  • 基金资助:
    国家重点研发计划(2022YFB2703500);国家自然科学基金资助项目(52277104);云南省重点研发计划资助项目(202303AC100003);云南电网有限责任公司科技项目(YNKJXM20220173)。

A Non-invasive Load Recognition Approach Incorporating SENet Attention Mechanism and GA-CNN

SHEN Xin1(), WANG Gang2, ZHAO Yitao1, LUO Zhao2(), LI Zhao2, YANG Xiaohua1   

  1. 1. Yunnan Power Grid Co., Ltd., Kunming 650217, China
    2. Faculty of Electric Power Engineering, Kunming University of Science and Technology, Kunming 650500, China
  • Received:2024-03-18 Online:2025-05-30 Published:2025-05-28
  • Supported by:
    This work is supported by the National Key R&D Program of China (No.2022YFB2703500); National Natural Science Foundation of China (No.52277104); National Key R&D Program of Yunnan Province (No.202303AC100003) and the Science and Technology Project of Yunnan Power Grid Co., Ltd. (No.YNKJXM20220173).

摘要:

随着智能电表的普及,电网信息化、数字化水平逐渐提高,需求侧的非侵入式负荷监测(non-intrusive load monitoring,NILM)逐渐成为供电企业实现能效提升的关键技术。针对目前非侵入式负荷识别算法存在特征冗余度、计算开销大、识别性能差等问题,提出一种融合压缩-激励网络(squeeze and excitation networks,SENet)注意力机制和基于遗传算法优化卷积神经网络(genetic algorithms-convolutional neural network,GA-CNN)的非侵入式负荷识别方法。首先,将SENet注意力机制嵌入CNN,提高关键特征的表征能力,降低特征冗余度;其次,提取居民负荷U-I轨迹图,对其进行加权像素化处理,通过计算得到WVI(weighted pixelated VI)特征矩阵,并以此为特征参量训练SENet-CNN模型;最后,利用遗传算法优化SENet-CNN模型的超参数,提高模型负荷识别性能和计算效率。实验结果表明,所提方法能够降低非侵入式负荷识别计算开销,准确识别出居民负荷类别,显著提升非侵入式负荷识别效率。

关键词: 居民负荷识别, 卷积神经网络, NILM, SENet注意力机制, V-I轨迹图

Abstract:

With the popularization of smart meters and the gradual improvement of grid informatization and digitization, non-intrusive load monitoring (NILM) on the demand side of residential customers' energy usage is becoming one of the key technologies for power supply companies to boost energy efficiency. Regarding the problems of the current non-intrusive load recognition algorithms, such as feature redundancy, high computational overhead, and low recognition performance, the paper proposes a non-intrusive load recognition method integrating SENet attention mechanism and GA-CNN. Firstly, the SENet attention mechanism is embedded in a convolutional neural network (CNN) to improve the characterizaion of key features and reduce feature redundancy. Secondly, the U-I trajectory map of the residential load is extracted and weighted pixelated to obtain the WVI (Weighted pixelated VI) feature matrix through computation, which is applied as the feature coefficient to train the SENet-CNN model. Finally, by virtue of the genetic algorithm, the SENet-CNN model is trained and the hyperparameters of the CNN-SENet model are optimized to improve the model load recognition performance and computational efficiency. The experimental results show that the proposed method can reduce the computational overhead of non-intrusive load identification, accurately identify the residential load categories, and significantly improve the efficiency of non-intrusive load identification.

Key words: residents load identification, convolutional neural network, NILM, SENet attention mechanism, U-I trajectory diagram