首页|基于UMAP流形特征提取和KELM的非侵入式负荷监测方法研究

基于UMAP流形特征提取和KELM的非侵入式负荷监测方法研究

扫码查看
非侵入式负荷监测是"坚强智能电网"用户侧智能数据挖掘的关键技术.针对现有辨识算法对叠加态负荷辨识准确率低的问题,提出了一种基于均匀流形逼近与投影(UMAP)和KELM结合的非侵入式负荷辨识模型.首先利用UMAP对原始负荷特征作嵌入,提取负荷的类内流形结构,并结合随机梯度下降法优化负荷的全局结构,在保留负荷原始相邻位置信息的前提下有效增大负荷特征的区分度;然后,采用径向基函数搭建核映射网络,利用ACO算法对映射网络的径向范围和模型的惩罚系数寻优,建立最优辨识模型.与多种基于机器学习的辨识方法相比,所提模型对叠加态负荷的辨识准确率提升显著,在TIPDM和BLUED数据集上的辨识准确率分别达到了 98.48%和 99.44%.
Research on Non-Intrusive Load Monitoring Method Based on Manifold Feature Extraction of UMAP and KELM
Non-intrusive load monitoring is a key technology for smart data mining on the user side of the"strong smart grid".To address the problem of low accuracy of existing identification algorithms for superimposed state load,a non-intrusive load identification model based on the combination of uniform manifold approximation and projection(UMAP)and KELM is proposed.Firstly,UMAP is used to embed the original load features,extract the intra-class manifold structure of the load,and combine with stochastic gradient descent to optimize the global structure of the load,which effectively increases the distinguishability of the load features while retaining the original adjacent position information of the load.Then the kernel mapping network is constructed using radial basis functions,and the ACO algorithm is used to optimize the radial range of the mapping network and the penalty coefficients of the model to establish the optimal identification model.Compared with other machine learning-based identification methods,the proposed model achieves significant improvement in the identification accuracy of superimposed state load,reaching 98.48%and 99.44%on the TIPDM and BLUED datasets,respectively.

non-intrusive load monitoringsuperimposed state loadUMAPACOKELM

张瀚文、李鹏、郎恂、沈鑫、梁俊宇、苗爱敏

展开 >

云南大学信息学院,云南 昆明 650504

云南电网有限责任公司,云南 昆明 650217

仲恺农业工程学院自动化学院,广东 广州 510225

非侵入式负荷监测 叠加态负荷 均匀流形逼近与投影 蚁群算法 核极限学习机

国家自然科学基金云南省中青年学术和技术带头人后备人才培养项目工业控制技术国家重点实验室开放基金云南大学专业学位研究生实践创新项目

62163036202105AC160094NoICT2022B24ZC-22222774

2024

电子器件
东南大学

电子器件

CSTPCD
影响因子:0.569
ISSN:1005-9490
年,卷(期):2024.47(2)
  • 24