PhD Thesis

The task of breaking down a household's electricity consumption into individual appliances is often referred to as non-intrusive appliance load monitoring (NILM), or energy disaggregation. The field was introduced in the late 80s and early 90s, although it has recently gained momentum as a result of national smart meter deployments. Smart meters are deployed primarily for billing purposes, and therefore generally only measure the energy consumption of the home at low resolution (e.g. up to 30 minutes). However, such smart meters often make higher resolution (e.g. 10 second) power data available over a local home area network. My PhD focused on the disaggregation of 10 second power data as measured by such smart meters.

A household's power demand disaggregated into appliances

In order to perform such a disaggregation task, it is necessary to first build a mathematical model of a household. My work adopted an approach based on hidden Markov models (HMM), in which appliances within a household are represented by HMMs. Each HMM contains a number of parameters which describe the behaviour of a particular appliance, such as its power demand or frequency of use. However, every household is different, and two appliances of the same type (e.g. two fridges from different manufacturers) will behave slightly differently. This motivated a Bayesian treatment of HMMs, which would describe both the similarities and differences between different appliances of the same type.

A Bayesian hidden Markov model

Many existing NILM approaches require the collection of appliance-level sub-metered or manually labelled data from the household in which disaggregation is applied. In contrast, my work focused on learning specific appliance models from only aggregate electricity data. To do so, appliance behaviour which generalises between households was learned from data sets containing many examples of appliances of the same type (e.g. many different fridges). These general appliance models could then be tuned to match an appliance in a specific household using only aggregate data. This approach is similar to transfer learning, in that sub-metered appliance-level data from other households was combined with aggregate-level data from the target household to produce highly accurate appliance models.

Comparison of a general model of an appliance type and specific model of an appliance instance for a microwave's power demand

During my PhD, I presented my work at both the first and second International NILM Workshop, and at the AAAI-2012 conference. I also published a paper in the Journal for Artificial Intelligence towards the end of my PhD. See my publications page for more details.

1st International Workshop on Non-Intrusive Load Monitoring