基于迁移学习策略的非侵入式负荷监测方法研究  

Research on non-invasive load monitoring method based on transfer learning strategy

作　　者：方晨宇 王贤睿 杨子龙 高梓祥 陈鑫昊 贾添淇 王娟[1] FANG Chenyu;WANG Xianrui;YANG Zilong;GAO Zixiang;CHEN Xinhao;JIA Tianqi;WANG Juan(Nanjing Institute of Technology,Nanjing 211167,China)

机构地区：[1]南京工程学院,南京211167

出　　处：《黑龙江电力》2024年第4期318-324,共7页Heilongjiang Electric Power

基　　金：南京工程学院大学生实践创新训练计划项目(项目编号:202311276063Y)。

摘　　要：现有的非侵入式负荷监测算法在同一数据集中功率分解准确率上有了很大的提高,但模型泛化性差且跨数据集分解负荷误差较大。为此,探究负荷监测模型是否具有可迁移性,并研究一种基于迁移学习策略的负荷监测模型。该模型固定部分卷积网络层,实现负荷监测模型的微调迁移学习,通过试验验证固定层数对预测结果的影响,以及模型对不同电器的预测效果。相比传统的直接迁移模型,该模型有效降低各项设备的负荷分解误差,并且降低了模型对负荷训练样本数量的要求。The existing load monitoring algorithms have greatly improved the power decomposition accuracy in the same data set,but the model generalization is poor and the decomposition accuracy across data sets is low.Therefore,it experimentally explores whether the load monitoring model is portable,and studies a load monitoring model based on transfer learning strategy,which fixes part of the convolutional network layer and realizes the fine-tuning transfer learning of the load monitoring model.It verifies the influence of fixed layers on prediction results through experiments,as well as the effectiveness of the model on different electrical appliances.Compared with the traditional direct transfer model,the proposed model effectively reduces the load decomposition error of each device and reduces the requirement of the model on the number of load training samples.

关 键 词：非侵入式负荷监测 深度学习 迁移学习 

分 类 号：TM714[电气工程—电力系统及自动化]