Distribution-free uncertainty quantification and calibration for energy disaggregation

Buildings account for approximately one-third of global energy consumption. Research shows that providing an appliance-wise energy consumption breakdown can lead to up to 15% energy savings. Energy disaggregation or Non-Intrusive Load Monitoring (NILM) offers a scalable method by using machine learning to estimate individual appliance usage from total household energy consumption data. In this article, we investigate methods for uncertainty quantification and calibration in the context of energy disaggregation. We utilize homoscedastic and heteroscedastic likelihood for the state-of-the-art Sequence-to-Point (S2P) models and apply various distribution-dependent uncertainty quantification methods, including Monte Carlo (MC) Dropout, Deep Ensemble (DE) and Bootstrap (BS). To address distribution assumptions limitations, we introduce Quantile Regression as a distribution-free uncertainty quantification method for NILM. Additionally, we employ Conformal Prediction—a novel approach for energy disaggregation—and Isotonic Regression for uncertainty calibration. Our experiments are conducted on four publicly available datasets: REDD, Pecan Street, REFIT, and UK-DALE. Detailed analysis is provided for each uncertainty quantification and calibration method, including a comparison of time complexity for training, calibration, and inference. Conformal Prediction and Quantile Regression offer performance better than or comparable to other methods, while significantly reducing computational costs, with up to 10x faster training and 50x faster inference. The code to reproduce our results is available at https://github.com/haikookhandor/conformalNILM.