- We present a refined algorithm disaggregating appliance-level electricity consumption data from smart meters

- Consumption data with a resolution of 1 Hz allows recognizing several appliances with good accuracy

Information about a household’s electricity consumption at the level of individual appliances is valuable for energy efficiency services and for marketing purposes alike. Installing sensors at each outlet to retrieve such appliance-level data, however, comes at high cost and is rather impractical. An alternative is to measure the household’s overall consumption at a central point (e.g. via a smart electricity meter) and to infer appliance-level information by means of machine learning. This, however, is a challenging task, partly due to the large number of appliances per household and the high variability of the load patterns per device.   

We implemented a prominent non-intrusive load monitoring (NILM) algorithm first published by Hart in 1992, improved its filtering technique and adapted it to electricity consumption data available at sampling frequencies of 1 Hz. These steps make it applicable to many smart electricity meters. Moreover, to support the learning phase in individual households, we developed a smartphone app that̶  based an automatic event detection ̶ enables users to label on-off events of appliances. 

The algorithm performs well for devices that have only one or two distinct consumption modes such as lights and fridges. The smartphone app enables the user to comfortably label on-off events of appliances and thereby enables  a training phase in individual households. The latter helps to bridge the gap between applications in the lab carried that are carried out by engineers, and in real homes by providing an easy-to-use tool that ordinary people without particular training and technical expertise can set up and use.