Non-Intrusive Load Monitoring using Multi-Output CNNs

Non-Intrusive Load Monitoring (NILM) is generally framed as a supervised learning problem whose input is the time series for aggregated power load of a household and whose output is the time series for the consumption of an individual appliance. Often the interest lies in predicting whether an appliance is ON or OFF, rather than its power usage. In this paper we propose a modification of a state-of-the-art convolutional neural network architecture to allow for multi-output channels, solving the regression and classification problems with relative weights simultaneously. We analyze the performance of this multi-output model and study the interplay between the two approaches on NILM.