基于改进GMM算法的非侵入式负荷监测方法研究  

Research on Non-Invasive Load Monitoring Method Based on Improved GMM Algorithm

作　　者：张媛 王飞[2] 崔秀华[1] 翟琳 张照锋[1] ZHANG Yuan;WANG Fei;CUI Xiuhua;ZHAI Lin;ZHANG Zhaofeng(School of electronic Information,Nanjing Vocational College of Information Technology,Nanjing Jiangsu 210023,China;College of Artificial Intelligence and Automation,Hohai University,Nanjing Jiangsu 210024,China)

机构地区：[1]南京信息职业技术学院电子信息学院,江苏南京210023 [2]河海大学人工智能与自动化学院,江苏南京210024

出　　处：《电子器件》2024年第5期1382-1388,共7页Chinese Journal of Electron Devices

基　　金：江苏省高等学校基础科学(自然科学)面上项目(21KJB520014);南京信息职业技术学院校级基金项目(YK20230102)。

摘　　要：非侵入式负荷监测是实现电力负荷精细化管理的重要技术手段,负荷分解及特征提取是非侵入式负荷监测的关键环节。针对常规算法负荷特征提取弱或者过度拟合造成泛化能力低的问题,提出了基于拟合和聚类思想的负荷分解及特征提取算法,即改进的高斯混合模型(GMM)。首先通过引入AIC和BIC信息准则,分析不同混合成分个数下的GMM模型的拟合程度,选取最优的混合成分个数;其次研究和分析电力负荷功率信号的特征曲线,结合混合成分个数,改进初始值的选取;最后进行优化后的功率信号聚类分解,输出能充分体现时域功率曲线特征的特征矩阵。基于智能洗衣机洗涤模式下的实测功率信号,验证了改进GMM算法的可行性,并对模型改进前后的特征矩阵进行差异性分析,验证了采用改进高斯混合模型进行负荷分解得到的特征矩阵具有更好的信息表征能力及更好的训练识别特性。结果表明提出改进算法具有较强的自适应性以及较高的准确性。Non-intrusive load monitoring is an important technical means to achieve the fine management of electric power load,and load decomposition and feature extraction are the key links of non-intrusive load monitoring.Targeting at the problem of weak load feature extraction and low generalisation ability of conventional algorithms,a load decomposition and feature extraction algorithm based on the idea of fitting and clustering is proposed,i.e.,the improved Gaussian mixture model(GMM).Firstly,by introducing the AIC and BIC information criterion,the fitting degree of the GMM model under different numbers of mixing components is analysed,and the optimal number of mixing components is selected.Secondly,the characteristic curve of load power signal is studied and analyzed,and the selection of initial value is improved by combining the number of mixed components.Finally,the optimised clustering decomposition of power signals is carried out,and feature vectors that can fully embody the characteristics of the time-domain power curve are output.The feasibility of the improved GMM algorithm is verified based on the measured power signals in the washing mode of a smart washing machine,and the feature matrices before and after the improved model are analyzed for differences,which verifies that the feature matrices obtained by load decomposition using the improved GMM have a better ability to characterize the information and a better training recognition property.The results show that the proposed improved algorithm has strong adaptivity and high accuracy.

关 键 词：非侵入式 GMM 电力负荷分解 特征提取 

分 类 号：TM744[电气工程—电力系统及自动化]