Friday 9 September 2011

Top 10 papers for Non-Intrusive Appliance Load Monitoring (NIALM) - updated for 2011

Update: Top papers of 2012

About a year ago I posted a short list of some of the most useful papers for NIALM. Since then I have learnt a lot, found much more material and read many newly published papers. This post is intended as an update to the original list, containing what I now believe to be the most relevant and useful academic papers. Below is my top 10 papers for NIALM (in alphabetical order). Any comments and discussions are welcome!

Wednesday 7 September 2011

The Reference Energy Disaggregation Data Set

Accurately comparing the performance of energy disaggregation (Non-intrusive Appliance Load Monitoring) methods is almost impossible unless the same test data is used. For this reason, a team at MIT have built The Reference Energy Disaggregation Data Set (REDD). A few weeks ago, an initial (v1.0) snapshot of this data set was released for use in academic work. The paper describing the equipment used for data collection is:

Kolter JZ, Johnson MJ. REDD : A Public Data Set for Energy Disaggregation Research. In: Workshop on Data Mining Applications in Sustainability (SIGKDD). San Diego, CA; 2011

The dataset contains data for 6 houses. For each house:
  • The current and voltage of the two mains circuits are monitored at 15kHz. This is also available for download as power readings down sampled to 1Hz
  • The power of each circuit in the home is also monitored to a frequency of 1 reading every 3-4 seconds. Each house contains 11-26 circuits, with large appliances often operating on their own dedicated circuit
  • The power drawn by individual appliances is also being monitored, although not enough data had been collected to warrant its release in v1.0
Collecting energy consumption data is a time consuming and expensive process. The lack of available data about appliance behaviour and energy consumption often limits the understanding necessary to design energy disaggregation methods. Through the study of this dataset, I hope to provide some insight into how the energy is consumed in the home and how energy disaggregation methods can take advantage of this.

Energy Breakdown

The aim of energy disaggregation methods is often to maximise the disaggregation accuracy, or minimise the difference between the predicted and actual energy consumption of each appliance. Therefore, the penalty for not recognising an appliance has consumed any energy increases with the energy consumption of the appliance. Consequently, it is more important for energy disaggregation methods to be able to disaggregate an appliance which consumes a high amount of energy than one that consumes a low amount of energy. So which appliances are most important? The figure below shows the relative contributions of each household circuit for house 1 of REDD.
Energy breakdown by circuit of house 1
The above figure shows that just 7 circuits collectively contribute over 75% of the house's total energy consumption. Admittedly, not all circuits contain a single appliance, although I doubt any circuit contains more than a handful of appliances. I'm a little surprised that the fridge comes out as the highest energy consumer, but I don't think it's unusual that it would contribute a fair portion of a house's energy consumption.

Appliance Behaviour

In order to disaggregate a household's energy consumption into the contributing appliances, it is often necessary to build models of such appliances. Analysing the power demand of each appliance is often a good place to start to decide what information such a model should contain. Since the fridge came out as the highest consumer, I've chosen to take a closer look the fridge's power demand in 5 of the REDD houses. Below is a plot of the power demand over a 45 minute period.
Almost all fridges have a different power demand, duration of use and periodicity. However, interestingly, they all share an identical model of usage, in that they all exhibit similar cyclic behaviour. For each fridge, the cycle begins with a single increase in power, which gradually decreases during the cycle and finally ends sharply. Although not ideal, this is encouraging for the purpose of general appliance models. Such a general model of fridges could be built, and would only need to be tweaked in order to match a specific fridge's signature.