Automatic multi-state load profile identification with application to energy disaggregation

Abstract

Non-Intrusive Appliance Load Monitoring can greatly benefit the Smart Buildings for energy awareness, while reducing cost and avoiding intrusive technology. This paper presents a generic algorithm for extracting the main power states of electrical appliances. The method is based on iterative K-mean clustering that is applied on historical plug-level active power data. The resulting multi-state load profile identification module is then integrated within an existing Building Management System for outlet-level energy disaggregation. Factorial Hidden Markov Modelling models the plugged appliances for low-frequency power disaggregation purposes, and incorporates the extracted set of appliances states. The solution was validated using the ECO dataset and NILM-Eval toolbox, allowing a comparison with standard binary ON/OFF modelling. It showed that the multi-state modelling significantly reduces the RMS error of the inferred power signals, yet at the expense of a higher computing time. Moreover, given a small set of appliances, the total inferred energy may be evaluated more precisely, leading to an enhancement of the quality of user energy feedback.