Keywords
non‑intrusive load disaggregation, adaptive sliding data window, convolutional neural network, nested long and short‑term memory network, improved attention mechanism
Abstract
Non?intrusive load disaggregation technology can effectively mine the appliance information of customers, which is the basis to carry out interactive customer load response by the grid company. The conventional non?intrusive load disaggregation technology has several drawbacks, such as limited scope of application and low accuracy. In this paper, a non?intrusive load disaggregation model with multiple optimization selection of appliance characteristics is proposed. First, an adaptive sliding data window is designed for appliance operation characteristics to obtain a more complete power segment and to adjust the network input and output dimensions. Second, the appliance features can be extracted and deepened by fusing shallow convolutional neural networks (CNN) with two?layer nested long and short?term memory networks (NLSTM), which is further fed into an improved attention mechanism to obtain the optimum appliance feature sequence by adjusting the feature weights. Finally, experimental analysis on the REDD dataset shows that the multiple selection, deepening and reusing of appliance features can significantly improve the accuracy of load decomposition while reducing training time.
DOI
10.19781/j.issn.1673-9140.2023.01.017
First Page
146
Last Page
153
Recommended Citation
WANG, Jiaju; WANG, Junping; BAI, Tai; ZHANG, Ran; DING, Yihui; YANG, Lin; and ZHANG, Shu
(2023)
"Non‑intrusive load disaggregation based on multiple optimization of appliance features and CNN‑NLSTM model,"
Journal of Electric Power Science and Technology: Vol. 38:
Iss.
1, Article 17.
DOI: 10.19781/j.issn.1673-9140.2023.01.017
Available at:
https://jepst.researchcommons.org/journal/vol38/iss1/17