Contextual Sequence-to-Point Deep Learning for Household Energy Disaggregation

Abstract

This paper examines a contextual paradigm for energy disaggregation using Non-Intrusive Load Monitoring (NILM). Due to numerous issues including low sampling rates, missing data, misaligned readings, and diverse combinations of nonlinear and multi-state appliances, this problem is challenging and complex. We proposed two different deep learning models for household energy disaggregation with shared parameter learning based on Convolutional Neural Networks (CNNs) and Gated Recurrent Units (GRUs). The proposed models utilize a sliding window of the main aggregate power readings to predict the per-appliance consumption at the end point of the sequence; using the entire input sequence gives more contextual information and reduces the prediction complexity in other problem settings. We evaluated the performance using two benchmark datasets, ENERTALK and UK-DALE, under different scenarios including sampling rates, imputation methods, cross-dataset generalization, and single and multi-target settings. The results demonstrate that the proposed models show better robustness and generalization capability than the other sequence-to-point models when no consumption information is discarded in the alignment process, especially for cross-domain disaggregation.