摘要
非侵入式负荷监测(NILM)通过分析电力总线数据估计单个负荷的功率波形,是电力系统能耗管理的关键技术之一。随着用户对设备能耗管理需求的增加,NILM的准确性成为研究的重点之一,但它容易受到功率类型、功率水平和负荷变化的影响。单一NILM模型面对不同类型的负荷时准确性差异较大,使用单一方法难以在各类负荷上均取得理想效果。因此,提出一种基于堆叠集成学习的非侵入式负荷高精度辨识方法 AMEL(Aggregation Method based on Ensemble Learning)。首先,选择在各种类型的负荷中表现最优的几种方法构建NILM模型库;其次,建立一个基于多层感知机(MLP)的NILM模型偏好框架,以实现对不同负荷的高精度监测。在UK-DALE数据集上的实验结果表明,与典型的NILM方法相比,所提方法的平均绝对误差(MAE)平均降低了35.6%,F1、召回率和马修斯相关系数(MCC)分别平均提升了33.5%、30.6%和32.1%。此外,通过比较现有的堆叠集成方法和各类设备的辨识波形,验证了所提方法的有效性。
As a key technology in energy consumption management in power systems,Non-Intrusive Load Monitoring(NILM)estimates individual load power waveforms by analyzing overall power bus data.With increasing user demand for device energy consumption management,the accuracy of NILM has become a research focus,which is susceptible to influence of power types,power levels,and load variations.A single NILM model often shows significant accuracy differences across various load types,making it challenging to achieve ideal results with a single method for all loads.Therefore,a high-precision non-intrusive load identification method based on stacking ensemble learning was proposed,called AMEL(Aggregation Method based on Ensemble Learning).Firstly,several methods that perform the best in different types of loads were selected to construct the NILM model base.Then,an NILM model preference framework based on MultiLayer Perceptron(MLP)was established to achieve high-precision monitoring for different loads.On UK-DALE dataset,the proposed method was compared with typical NILM methods.The results show a decrease of 35.6%in Mean Absolute Error(MAE)averagely,and improvements of 33.5%,30.6%and 32.1%in F1,recall,and Matthews Correlation Coefficient(MCC),averagely and respectively.The effectiveness of the proposed method was validated by comparing with the existing stacked ensemble learning methods and identification waveforms for various devices.
作者
黄宇
何耿生
刘西卓
刘玺
牟景艳
陈学艳
曾金灿
HUANG Yu;HE Gengsheng;LIU Xizhuo;LIU Xi;MOU Jingyan;CHEN Xueyan;ZENG Jincan(Electric Power Research Institute of Guizhou Power Grid Company Limited,Guiyang Guizhou 550002,China;Energy Development Research Institute of China Southern Power Grid,Guangzhou Guangdong 510630,China;Anshun Power Supply Bureau,Guizhou Power Grid Company Limited,Anshun Guizhou 561000,China)
出处
《计算机应用》
北大核心
2025年第S1期323-328,共6页
journal of Computer Applications
基金
南方电网公司科技项目(066600KK52222044/GZKJXM20222165)。
关键词
非侵入式负荷监测
集成学习
堆叠方法
序列到序列
双向长短期记忆网络
去噪自编码器
Non-Intrusive Load Monitoring(NILM)
ensemble learning
stacking method
Sequence-to-Sequence(Seq2Seq)
Bi-directional Long Short-Term Memory(BiLSTM)network
Denoising AutoEncoder(DAE)
