Electrical Load Disaggregation using a two-stage deep learning approach

Abstract

Electrical Load Disaggregation is an important area of research for demand-side energy management, especially in residential buildings or units. This problem has therefore received significant attention and especially in the context of high-sampled smart meter data, a range of deep-learning based algorithms exist in the literature. However more often than not, learning for these architectures incurs considerable computational costs as models for each appliance need to be learnt separately. Such models have also to be re-trained each time the data changes as the models get fixated to the given aggregate data, irrespective of the size of the latter. We address these problems in this paper by proposing a two-stage learning approach comprised of a) representational learning which learns patterns implicit in the aggregate data in the first stage and, b) a regression technique which uses these representations to regress with the individual appliance class labels. We observe that the proposed architecture is computationally simple which in turn makes it more flexible in handling changes in the smart meter data.