Modelling and power estimation of continuously varying residential loads using a quantized continuous-state hidden markov model

Hidden Markov Models (HMMs) and their extensions have broad useful applications in several fields. Energy disaggregation, or non-intrusive load monitoring (NILM), is the process of analyzing and decomposing the total aggregate energy consumption of a household into the individual consumptions by respective devices. These details were found informative and can influence occupants in a way that achieves noticeable energy savings. Hidden Markov Models (HMMs) were found efficient in modelling and detection of household devices. In this work, we propose a quantized continuous-state HMM so as to model continuously varying loads which is a challenging problem in the domain of energy disaggregation. Two core enhancements to the standard quantized continuous-state HMM are proposed. First, we propose a method that estimate the transition matrix considering potential probabilities to states neighboring that the model switches to. This method reduces the effect of domination of a state transition and achieve better simulation of switching cases in real variable loads. Second, the consumption of the variable load is estimated from the collective mean resulting from the Viterbi algorithm instead of the assigning the center value of the state with the maximum likelihood. In this way, the effect of quantization can be reduced. The proposed approach was tested on synthetic and real variable loads from the REDD public data set. It was found that the proposed models outperform the reference HMM that applies standard estimation algorithms.