Training Nonintrusive Load Monitoring Algorithms Without Supervision From Submeters

Nonintrusive load monitoring allows to estimate the energy consumption of major household appliances by just analyzing the aggregated power consumption collected at the main meter of the house. Recent disaggregation algorithms based on deep learning techniques showed superior performance with respect to previous methods. However, they require a large amount of submeter data to be trained. In this work, we present a new solution for training nonintrusive load monitoring algorithms without any supervision from submeters. To achieve this goal, we divided the disaggregation algorithm into two stages-appliance detection and state-based disaggregation. In the first stage, we aim at identifying the start and stop times of the individual appliance operations within the whole-house power signal. In the second stage, we reconstruct the power signature of the target device by exploiting appliance-specific power states learned in the house. We tested our methodology on fridges, washing machines, and dishwashers in a public dataset, showing double-digit improvements with respect to previous methods trained with submeter data. Most importantly, the proposed solution allows to collect a large number of appliance power signatures with minor costs, thus helping to achieve the generalization capabilities required by a real-world disaggregation system.