基于递进式模型结构和时间信息嵌入的非侵入式负荷分解  

Non-intrusive Load Decomposition Based on Progressive Model Structure and Temporal Information Embedding

作　　者：孙睿晨 董坤 赵剑锋[1,2] 毛妍纯 SUN Ruichen;DONG Kun;ZHAO Jianfeng;MAO Yanchun(School of Electrical Engineering,Southeast University,Nanjing 210096,China;College of Information Science and Technology,Nanjing Forestry University,Nanjing 210037,China;State Grid Nanjing Power Supply Company,Nanjing 210019,China)

机构地区：[1]东南大学电气工程学院,江苏南京210096 [2]南京林业大学信息科学技术学院,江苏南京210037 [3]国网江苏南京供电公司,江苏南京210019

出　　处：《智慧电力》2024年第2期55-62,70,共9页Smart Power

基　　金：国家自然科学基金资助项目(51936003 )。

摘　　要：现有的非侵入式负荷监测算法对多时间尺度的用电规律缺乏关注,且存在电器状态误判率高和功率预测误差大的问题。对已有模型在学习框架、信息嵌入和损失函数3个方面进行优化,提出一种基于递进式模型结构和时间信息嵌入的负荷分解方法。模型由预分解模块和功率预测模块构成,递进地完成判断电器开关状态与估计功率值2个任务。2个模块的网络结构均基于Transformer设计,使用不同的复合损失函数优化。另外,提出了多尺度时间信息编码及嵌入方法,增强模型对用电行为特征的提取能力。基于居民用电数据集REDD和UKDALE的测试结果验证了该方法的有效性。Existing non-intrusive load monitoring algorithms have no regard for multiple-time-scale electricity consumption patterns,and have a high rate of misjudgment of electric appliance conditions and a big power prediction error.This paper optimizes existing models in terms of learning framework,information embedding,and loss functions,and proposes a load decomposition method based on progressive model structure and temporal information embedding.The model consists of a pre-decomposition module and a power prediction module,thereby progressively completing the determination of the on/off states of the electric appliances and the estimated power value.The two modules use Transformer-based network structures,and are optimized by different composite loss functions.Furthermore,the multi-scale time information encoding and embedding methods are proposed to enhance the feature extraction ability of power consumption behavior.The effectiveness of the proposed method is verified on REDD and UKDALE datasets.

关 键 词：非侵入式负荷分解 深度学习 自注意力机制 Transformer模型 

分 类 号：TM714[电气工程—电力系统及自动化]