Abstract
Non-intrusive load monitoring (NILM) as the process of extracting the usagepattern of appliances from the aggregated power signal is among successfulapproaches aiding residential energy management. In recent years, high volumedatasets on power profiles have become available, which has helped makeclassification methods employed for the NILM purpose more effective and moreaccurate. However, the presence of multi-mode appliances and appliances withclose power values have remained influential in worsening the computationalcomplexity and diminishing the accuracy of these algorithms. To tackle thesechallenges, we propose an event-based classification process, in the firstphase of which the $K$-nearest neighbors method, as a fast classificationtechnique, is employed to extract power signals of appliances with exclusivenon-overlapping power values. Then, two deep learning models, which considerthe water consumption of some appliances as a novel signature in the network,are utilized to distinguish between appliances with overlapping power values.In addition to power disaggregation, the proposed process as well extracts thewater consumption profiles of specific appliances. To illustrate the proposedprocess and validate its efficiency, seven appliances of the AMPds areconsidered, with the numerical classification results showing markedimprovement with respect to the existing classification-based NILM techniques.