Recent Non-Intrusive Load Monitoring (NILM) approaches consider probabilistic graphical models and statistical inference algorithms such as Hidden Markov Model (HMM). One interesting HMM based approach towards unsupervised energy disaggregation proposes prior models of general appliance types which are tuned to specific instances using only aggregated electrical consumption measurements. An essential step of this approach is the subtraction of the estimated usage from the aggregated load before the disaggregation of a new load. Then wrongly-detected states of a given device lead to errors that are disseminated by the subsequent disaggregation. In this paper we aim at investigating an unsupervised HMM based approach that overcomes this limitation. First, the general models are tuned for a selection of suitable periods of the signal in analysis. Second, the load disaggregation of each appliance is carried out. Third, given the inferred consumption of the whole set of devices and that the measured aggregated electrical usage is a linear combination of these loads, an optimization step regarding the adjustment of the calculated sources is performed. Although experiments yielded in the REDD data set show the adequacy of the approach, still further improvements are required.

--

Electrical Signal Disaggregation, Non-Intrusive Load Monitoring, Non-negative Tensor Factorization, Sparse Coding, Single-Channel Source Separation