Using Deep Learning and Knowledge Transfer to Disaggregate Energy Consumption

Abstract

NILM or Non-Intrusive Load Monitoring is the task of disag-gregating the energy consumed by a building in the energy con-sumed by its constituent appliances. With the increase in energy demand, governments started searching for solutions to reduce energy wastage on the demand side, and the deployment of smart meters was one of them. Their purpose was to give users information about the aggregated energy consumed in a given household at any given time. Since the smart meters collect the aggregated readings, the interest in NILM grew, as researchers could focus their attention on the disaggregation algorithm. Regarding the disaggre-gation algorithms, deep learning models have shown remarkable results surpassing the previous state-of-the-art models. With this in mind, this paper proposes three different deep learning models: a convolutional neural network with residual blocks, a recurrent neural network, and a multilayer perceptron that uses discrete wavelet transform as features. These models are trained on the UK-DALE and REFIT datasets and compared with the state-of-the-art models present in NILMTK-Contrib. The models are evaluated on their generalization and transfer learning ability, as these are two critical factors for a broad NILM deployment. Our models have shown competitive results compared to the state-of-the-art, achieving lower errors than 2 out of the three models used, getting closer to the performance of the third. The main advantage of our models is the ability to do real-time disaggregation, while the best model has a 30 min delay.