Multisource Domain Adaptation for Nonintrusive Load Monitoring Through Data Distillation

The nonintrusive load monitoring (NILM) is a technique that disaggregates household total power consumption into the usage of individual appliances, which can promote the application of home energy management through detailed monitoring of electricity usage behavior. The rapid development of deep learning has driven significant advancements in the theory and algorithms of NILM. However, most deep learning models are fed with data from multiple households during the training process, and they suffer from performance decay due to the distribution discrepancy between the source domains and target domain. To address this issue, this article proposes a novel data distillation-based multisource domain adaptation mechanism that not only considers the different distribution distances between multiple source domains and the target domain but also investigates the varying similarities between source domain samples and target domain samples. Concretely, the proposed mechanism consists of four stages: 1) pretrain the feature extractor and energy regressor using labeled data from the source domain; 2) adversarially train the feature extractor that adaptive to the target domain by minimizing the empirical Wasserstein distance between the source and target domain; 3) select the source domain samples that are closer to the target domain to fine-tune the energy regressor; and 4) aggregate power estimations from each source domain to the target domain based on each domain weight. Using data from five households in the public REFIT dataset, multiple comparative experiments validate the effectiveness of the proposed algorithm. Extensive experiments, including feature visualization and ablation studies, demonstrate that the proposed algorithm can significantly enhance the domain adaptation capabilities of the model.