Unsupervised energy disaggregation using conditional random fields

Abstract

The task of energy disaggregation is to break up total energy usage into usage by its component electrical appliances. This can provide home owners with useful feedback about how they use electrical energy, and motivate them to save significant amounts of energy. Our ultimate goal is to provide ways to modify human energy consumption behavior in order to conserve and optimize the use of energy. This paper focuses on stochastic modeling and energy disaggregation based on conditional random fields (CRFs) using real-world energy consumption data. The proposed disaggregation method uses a clustering method and histogram analysis to detect the ON/OFF states of selected types of energy-using devices in the home. Data labeling is not required, because the label sequences are obtained by applying a clustering method, and decoding using all of the data. Long spans of data from 21 households were used in a binary classification experiment, in which an 86.1% average classification accuracy was achieved. The proposed method was also evaluated using hidden Markov models (HMMs), but significantly higher accuracy was obtained when CRFs were applied.