Exploring the Data Requirements of Energy Disaggregation

By Abhishek Balakrishnan

Our project is focused on non-intrusive load monitoring, or energy disaggregation, in order to gain deeper insights from the appliance level information as to how we can reduce our energy consumption and save money. In a previous post, we discussed at a high level the benefits of non-intrusive load monitoring over plug-level monitoring and what types of sensors we can use to collect data for disaggregation. Here, we explore the actual process of disaggregation. A researcher from Carnegie Mellon University points out that energy disaggregation is analogous to face recognition and elaborates on the complexities of automatically identifying the focal points of a photo (as opposed to other faces and objects in the background) and extracting further information on what the face or object is after its detection. The challenge is similar in energy disaggregation to break down an aggregate power signal into each of its constituent devices. This challenge is even harder because there are limited reference datasets to use, so much of our analysis relies on new data we collect. In order to understand our approach to disaggregation algorithm development, we have to start with the data – how do we collect data for machine learning algorithms to use?

Our team is using the Watts Up? PRO power meter to collect energy data from individual appliances in order to collect a database of device signatures. The Watts Up is able to transfer real power, current, and voltage, among other data via USB at once-per-second intervals. Below are examples of power data collected by the Watts Up meter for two different lights and a household fan in each of its modes of operation (i.e. speeds).

This plot clearly shows two sets of curves:– one for a compact fluorescent lamp (CFL) and one for an incandescent lamp (INC). These two devices have considerably different average values, as the incandescent consumes much more power than the compact fluorescent. This difference is a feature that can be used to distinguish between these appliances from an algorithmic perspective.

Using these power signatures, we can begin to train algorithms to differentiate loads using statistical learning techniques. To do this, we will collect many sample readings for various appliances. For example, having more data samples will improve the training of our models and better discriminate between a fan on low speed (state 1) as opposed to medium speed (state 2).

Moving beyond appliance detection and recognition, we are analyzing aggregate load data, or power time series that reflect using multiple devices simultaneously. Ideally, we would like to develop a comprehensive dataset for training our algorithms that includes measurements all of the electrical equipment in a residential or commercial space. For instance, take a dormitory, which includes many of the typical appliances you find in a household such as a fridge, microwave, fan, lighting, HVAC, etc. This discussion points to our overall research goals, which are to use extensive training data to develop better-performing disaggregation algorithms. The system we are developing will eventually be required to analyze new devices for which we have not collected data. This can only be made possible through a generalized, robust approach to disaggregation from extensive training data.