Non-Intrusive Load Monitoring Based on Residual U-Net and Conditional Generation Adversarial Networks

Abstract

A non-intrusive load disaggregation method based on residual U-Net and conditional generation adversarial networks (RUCGAN) model is proposed to address the low decomposition accuracy and poor generalization of traditional load disaggregation algorithms. The method is based on a conditional generative adversarial networks (CGAN), which is a variant of the encoder-decoder model that is suitable for processing time-series data and overcomes the limitation of requiring a manually designed feature extractor in traditional encoder-decoder structures. By introducing the U-Net structure as the encoder of the CGAN network, the size of the feature map can be gradually reduced through convolution and pooling operations, and gradually restored through deconvolution and upsampling operations. The U-Net structure also has skip connections that effectively preserve feature information and accelerate gradient propagation, thus improving model stability and generalization. Furthermore, combining the residual structure with the U-Net structure further enhances the model’s performance, as the residual connections can effectively reduce the number of network parameters and computation. Experimental results show that the MAE value of the model on the UK-DALE dataset decreased by at least 20.5%, and the MAE value of the model on the REFIT dataset decreased by at least 9.9%. Moreover, while improving the decomposition accuracy, the model size decreased by at least 5.6%.

Keywords

• Conditional generative adversarial networks
• NILM
• ResU-net