Hidden Markov models for nonintrusive appliance load monitoring

A method of device modeling for nonintrusive appliance load monitoring (NIALM) is presented. The proposed method uses hidden Markov models to describe device behavior. Unlike previous methods of device modeling, observations are associated with instantaneous power measurements as opposed to step changes in power use or on-off transients. The training procedure for individual devices is discussed. Accuracies of seven different device models are assessed using k-fold cross validation. In this assessment, the correlations between sequences of known state transitions and calculated Viterbi sequences representing predicted transitions are determined. This process is repeated for power use profiles collected at different sampling rates. Individual devices' Viterbi sequences are shown to be able to accurately approximate the actual device power use.