Unsupervised energy disaggregation with factorial hidden Markov models based on generalized backfitting algorithm

This paper proposes an approximate inference and parameter estimation in factorial hidden Markov models (FHMMs), a generalization of hidden Markov models (HMMs) in which the state is factored into multiple Markov chains. Earlier research has proved exact inference and parameter estimation can be computationally intractable. The proposed method in this paper considers FHMMs to be an alternative generalization of generalized additive models (GAMs). We use backfitting algorithm to estimate the parameters in FHMMs instead of exact but complex derivation and make the approximate inference more efficient. This method is motivated by the problem of energy disaggregation which is the process of decomposing a whole household´s electricity consumption into individual appliances in order to improve energy utility efficiency both in load terminal and electricity supplier. Numerical simulations indicate the effectiveness of the proposed method for energy disaggregation.