Skip to content Skip to navigation

Disaggregation Algorithms

Performance of different disaggregation algorithms on high-frequency data
Algorithmic approaches to energy disaggregation have traditionally been very simple, and focused solely on detecting changes in a limited number of device states, (e.g., off, on high, on low, sleep mode), in a power signal.
Performance of different disaggregation algorithms on high-frequency data

Such detection alone does not give a breakdown of power consumption in homes and is fundamentally limited to monitoring frequencies where such events are obvious. The methods would be unusable, for example, with just hourly data from smart meters. This project developed new algorithmic techniques that can breakdown electricity use more accurately than previous approaches, using data at a variety of different time scales. The researchers accomplished this goal through two studies: the first developed models of typical fingerprints for numerous types of appliances and devices; the second study used these models to develop algorithms to break down consumption from the aggregate signal.

The first study recorded power each hour on average for about 10,000 individually monitored devices to develop models for appliances and separate the signals. Researchers examined a fixed period of time, (in this study, one week), and expressed each appliance's entire energy trace over that time in terms of some linear combination of basis functions, which capture typical usage patterns. Activations specify which of these basis functions makes up any given signal. Given a collection of example usage patterns for an individual device, the study learned both the bases and activations using a method known as sparse coding. Researchers also developed additional algorithmic extensions that tailor sparse coding to this source separation setting.

A portion of a circuit-level data from a report showing data collected over one day
A portion of a circuit-level data from a report showing data
collected over one day

The study built models using the appliance-level data from 413 homes, and then evaluated the models to see if they could successfully identify appliances from 117 homes. On average the method was able to correctly assign 55 percent of the energy correctly into one of 10 device categories, (divided evenly so that random guessing would give about 15 percent accuracy).

Although the previous approach is promising in its ability to use the existing monitoring infrastructure, new sensing modalities offer the promise of much higher frequency data. Thus, the next project used the REDD data set to disaggregate energy in a home using ~1Hz data. The researchers used a factorial hidden Markov model, which captures a time varying process with several devices that can take on some discrete number of power states (e.g. on, off, standby). While not all devices have discrete power levels, several common appliances do, and the method is able to accurately model most of the devices in a common home. The particular algorithmic contribution for this work was to develop an "inference" procedure—an algorithm that determines the state of appliances given an aggregate signal—that was many times more accurate and faster than existing approaches.

This second study evaluated the potential of higher-frequency data, (1Hz whole-home power), and compared it to more traditional event-based detection methods. Researchers built models for appliances and then separated out the different end-uses. The algorithm correctly assigned about 87 percent of the energy in a home, (assigning to one of 10 categories of end-use), whereas the event-based approach assigned about 49 percent correctly. This highlighted the potential benefit of higher-frequency sampling and demonstrated the advantage of using more advanced algorithms over the previous simple approaches. The publication from this study has been cited by 40 other published papers.

Publications and Presentations

Approximate inference in additive factorial HMMs with application to energy disaggregation (745KB PDF)
In International Conference on Artificial Intelligence and Statistics (pp. 1472-1482)
Kolter, J. Z., & Jaakkola, T. (2012)

REDD: A public data set for energy disaggregation research
In proceedings of the SustKDD workshop on Data Mining Applications in Sustainability (pp. 1-6)
Kolter, J. Z., & Johnson, M. J. (2011)

Energy disaggregation via discriminative sparse coding (274KB PDF)
In Advances in Neural Information Processing Systems (pp. 1153-1161)
Kolter, J. Z., Batra, S., & Ng, A. (2010)

Future Work

Next steps and future/ongoing work for these approaches include: developing methods that can build models using aggregate data alone, rather than both aggregate and individual device level; combining small numbers homes monitored at high frequency with large data sets of smart meter data; integrating the approaches into deployed systems in building energy management solutions.