基于Attention和残差网络的非侵入式负荷监测

何健明; 李梦诗; 张禄亮; 季天瑶

何健明,李梦诗,张禄亮,季天瑶.基于Attention和残差网络的非侵入式负荷监测[J].电测与仪表,2024,61(6):173-180.
HE Jianming,LI Mengshi,ZHANG Luliang,JI Tianyao.Non-intrusive load monitoring algorithm based on attention and residual networks[J].Electrical Measurement & Instrumentation,2024,61(6):173-180.

基于Attention和残差网络的非侵入式负荷监测

Non-intrusive load monitoring algorithm based on attention and residual networks

DOI：10.19753/j.issn1001-1390.2024.06.023

中文关键词: 非侵入式负荷监测深度学习 BiGRU 残差网络注意力机制

英文关键词:non-intrusive load monitoring, deep learning, BiGRU, residual network, attention mechanism

基金项目:国家自然科学基金资助项目（ 52077081）

作者中文名	作者英文名	单位
何健明	HE Jianming	华南理工大学
李梦诗	LI Mengshi	华南理工大学
张禄亮	ZHANG Luliang	华南理工大学
季天瑶	JI Tianyao	华南理工大学

摘要点击次数: 509

中文摘要:

非侵入式负荷监测（Non-Intrusive Load Monitoring, NILM）可以从家庭电表的总功率读数，估算出各用电器的功率。由于对于同一类用电器，其状态种类、各状态持续时长、各状态的功率波形都不同，这使得基于特征工程和聚类的模型的泛化能力不强；回归模型的分解功率难以迅速跟踪真实功率。针对这些问题，文中将回归问题转化为在序列每个时刻的多分类问题，并提出基于Attention和残差网络的非侵入式负荷监测模型。该模型基于具有编码器和解码器的seq2seq框架，首先通过嵌入矩阵将高维稀疏one-hot向量映射为低维稠密向量；在编码部分，通过双向GRU从前后两个方向提取序列信息，引入Attention机制计算序列中当前时刻最重要的信息，引入残差连接学习残差部分输入输出之间的差异；在解码部分，用回归层组合BiGRU解码结果，取经过softmax函数处理的最大概率功率类别作为结果。该模型在选取REFIT数据集中表现良好，其中测试集与训练集完全独立，表明训练好的模型可以直接应用在新的住宅用户中。

英文摘要:

Non-Intrusive load monitoring (NILM) is a technique to disaggregate the power consumption of the appliances from the total power reading of the household electricity meter. Even for the same type of appliances, their state types, the duration of each state and the power waveform of each state are different, which requires high generalization ability of the model based on feature engineering and clustering. Meanwhile, the disaggregate power of the regression model is difficult to quickly track the ground true power. To solve these problems, the regression problem is transformed into a multi-classification problem for each moment in the sequence, and a non-intrusive load monitoring model based on attention mechanism and residual networks is proposed in this paper. The proposed model is based on the seq2seq framework with encoder and decoder. Firstly, the high-dimensional sparse one-hot vector is mapped to the low-dimensional dense vector through the embedding matrix. In the encoder, BiGRU is used to extract the sequence information from the front and back directions, an attention mechanism is introduced to calculate the most important information at the current time in the sequence, and a residual connection is introduced to learn the difference between the input and output of the residual part. In the decoder, the regression layer is used to combine the BiGRU decoding results, and the maximum probability power category processed by a softmax function is taken as the result. The model performs well in the REFIT dataset, and the testing dataset and training dataset are completely independent, which indicates that the trained model can be directly applied to new households.

查看全文查看/发表评论下载PDF阅读器

关闭