Çimen, HalilWu, YingWu, YanpengTerriche, YacineVasquez, Juan C.Guerrero, Josep M.2022-11-282022-11-2820221551-32031941-0050https://doi.org/10.1109/TII.2022.3150334https://doi.org/10.1109/TII.2022.3150334https://hdl.handle.net/20.500.13091/3155Energy disaggregation is the process of disaggregating a household's total energy consumption into its appliance-level components. One of the limitations of energy disaggregation is its generalization capacity, which can be defined as the ability of the model to analyze new households. In this article, a new energy disaggregation approach based on adversarial autoencoder (AAE) is proposed to create a generative model and enhance the generalization capacity. The proposed method has a probabilistic structure to handle uncertainties in the unseen data. By transforming the latent space from a deterministic structure to a Gaussian prior distribution, AAEs decoder transforms into a generative model. The proposed approach is validated through experimental tests using two different datasets. The experimental results exhibit a 55% MAE performance increase compared to deterministic models and 7% compared to probabilistic models. In addition, considering the predictions made when the appliances are on, the AAE improves the performance by 16% for UKDALE and 36% for REDD dataset compared to the state-of-art models. Moreover, the online analysis performance of AAE is examined in detail, and the disadvantages of instant predictions and the possible solutions are extensively discussed.eninfo:eu-repo/semantics/closedAccessHidden Markov modelsTrainingProbabilistic logicAnalytical modelsGenerative adversarial networksDeep learningData modelsAdversarial autoencoder (AAE)deep learningenergy disaggregationgenerative adversarial networksnonintrusive load monitoring (NILM)online energy disaggregationprobabilistic energy disaggregationresidential energy disaggregationLoad DisaggregationArticle10.1109/TII.2022.31503342-s2.0-85124765333