Denoising Autoencoder for Appliance-Specific Energy Disaggregation in Residential Settings

As technology advances, household energy consumption rises, impacting the environment and grid stability. Nonintrusive load monitoring (NILM) offers a cost-effective solution for appliance-specific energy monitoring without dedicated sensors. This paper proposes a denoising autoencoder-based NILM approach for residential energy disaggregation. Using real energy data from REDD and UKDALE datasets, the model is evaluated against combinatorial optimization (CO) and factorial hidden Markov models (FHMM) across multiple performance metrics, including recall, precision, accuracy, F1 score, and root mean square error (RMSE). Additionally, cross-dataset testing is conducted to assess model generalization on unseen homes. The results show that the proposed model achieves recall up to $\mathbf{9 9. 9 9 \%}$ for certain appliances and reduces RMSE by up to 52.6% compared to FHMM and 43.1 % compared to CO, demonstrating its potential for smart home energy management.