Please help design a NILM competition!

Cross posted from Jack Kelly's blog:

Has disaggregation accuracy improved since the 1980s? Which algorithms are most accurate for a given use-case? Which (if any) use-cases are well served by NILM already?

It's pretty much impossible to answer any of these questions with confidence (unless you only consider the tiny number of algorithms for which you have access to executable code). We can't directly compare published results across papers because, when testing the disaggregation accuracy of NILM algorithms, each paper uses different datasets, different metrics, different pre-processing, etc.

This means that we can't measure progress over time. Nor can we decide which NILM algorithms are most promising and which might be dead-ends.

These are bad problems. Let's work towards fixing them.

Some other machine learning communities have had great success running yearly competitions. For example, the ImageNet "Large Scale Visual Recognition Challenge" has been running yearly since 2010. Some regard this competition as having played a crucial role in the recent dramatic increase in the accuracy of image classification algorithms.

The idea of running a NILM competition has been rumbling around for several years. But designing and implementing a NILM competition is hard. The community uses sample rates ranging from monthly to MHz. No single metric is informative for all use-cases. Collecting ground truth data (the power demand of individual appliances) is expensive and time-consuming.

Maybe we can pull this off. The first step is to decide on a design which will work for everyone.

To give us something concrete to debate, we'll outline one way this could work. This is not meant to be definitive! Think of this as the DNA for a clumsy, inefficient animal 500 million years ago. Together, we need to evolve this design into an elegant, efficient beast, well adapted to its environment.

Please shoot holes in this proposal! What won't work for you? What's impractical? What's unfair? What opens the competition up to cheating? How can we make the competition more attractive to researchers? How can we make the competition more informative for the community? How can we simplify the process?

The draft proposal is available on Google Docs. I've linked to a Google Doc rather than copying-and-pasting the proposal into this post so that we can update the proposal as the discussion develops. Please add your comments either to the mailing list discussion; or to the Google Doc (please sign your comment with your name; unless you deliberately want to be anonymous); or if you want to keep your comment private then email Jack directly.

Thanks, (in no particular order) Jack, Mario, Oli, Stephen, Grant, Marco, Peter

PNNL NILM vendor survey

Below is a message to NILM vendors I'm sharing on behalf of PNNL:

The Pacific Northwest National Laboratory (PNNL) continues its work to develop Non-Intrusive Load Monitoring (NILM) test protocols.  Activities to date have focused on the development of a technical working group (NILM vendors, users/potential users, and other stake holders), research and decisions on candidate performance metrics, and the development of performance protocols.  This last activity we are seeking input from the larger NILM vendor community.

Below is a short feedback form to assist in directing the protocol development.  We would appreciate your responses as soon as possible. Please send responses by this email to Joseph.Petersen@pnnl.gov.  We appreciate your time and participation.

NILM Performance Metrics Project:  Status and Feedback Request

Feedback Goal:  To better understand preferences and constraints of proposed metric implementation approaches.  Please consider the two approaches to evaluating NILM performance listed below, and then provide your feedback to the following questions.

Approach 1: Data Driven – NILM devices use a diverse set of previously collected interval data to test device performance.

Approach 2:  Laboratory Testing – NILM devices are connected to actual appliances and/or load simulation systems to test performance.

1. Is your NILM platform/product capable of accepting 1-second to 1-minute interval data as inputs for disaggregation?
2. At what sampling interval is your NILM platform/product designed to take measurements or data inputs, e.g., 1 minute, 5 minute, hourly, other?
3. What specific inputs are necessary for your NILM platform or product, e.g., interval power data, energy data, voltage, current, reactive power, other?
4. What appliances or end-uses does your NILM product target?
5. What are the target use cases for your NILM product?
6. Other comments or questions you’d like to share regarding the development of the Data Driven or Laboratory Testing protocols?

NILM 2016 presentation videos and slides now available

Stephen Makonin has just uploaded the last of the videos to the NILM 2016 Youtube playlist, meaning that you can now easily watch individual talks from the conference. In particular, I'd recommend George Hart's keynote talk: Life after NILM, covering what it was like to do energy disaggregation in the '80s, his passion for mathematics throughout his life, his more recent shift towards art and sculpture, and finally his educational workshops.

The full set of slides and papers relating to each talk are also available via the NILM 2016 program if you'd like to catch up on everything else from the conference too. Happy watching!

5 things I learned from NILM 2016

The 3rd International Workshop on Non-Intrusive Load Monitoring was held on the 14-15 May 2016 at the beautiful mountaintop campus of Simon Fraser University, Vancouver, Canada. I thoroughly enjoyed the event, and just wanted to summarise a few of my thoughts having had the flight home to digest the weekend's activities.

1. NILM is not dead

Despite the jokes that were thrown around at the European workshop last year, the NILM community is very much as alive as it has ever been. The Vancouver workshop was the first 2-day energy disaggregation conference, the best attended, and in my opinion the most stimulating in terms of presentations and discussions. That being said, NILM is still far from a solved problem. Michael Baker from SBW Consulting presented a paper which evaluated the performance of three NILM vendors, and concluded that "further development of disaggregation algorithms is needed before they are sufficiently accurate to provide customers with accurate estimate of how much they spend on most end uses."

2. Different customers want energy breakdowns for different reasons

Most academic researchers (I'm including myself here) see NILM as a tool to encourage energy efficiency, but utilities also see it as a tool for customer engagement. For example, a perfect monthly energy breakdown might be useful for both, but the disaggregation methodology might be quite different. In the case of energy efficiency the aim might be to identify the rare cases where large energy savings are possible, while in the customer engagement case it might be preferable for the disaggregation result to be more conservative when identifying such edge cases, since false positives are likely to be much more costly in terms of reputation than correct identification.

3. The assumption that disaggregation leads to significant energy efficiency has never been concretely demonstrated

Jack Kelly gave an excellent presentation of his paper which summarised the literature regarding whether disaggregated feedback leads to energy savings. One of the most shocking messages was that "the four studies which directly compared aggregate feedback against disaggregated feedback found that aggregate feedback is at least as effective as disaggregated feedback" (see full paper for details). Jack was very careful to clarify that this does not mean that disaggregated data is useless, but rather that the community desperately needs a large, well-controlled, long-duration, randomised, international study to confidently quantify energy reductions as a result of disaggregated data.

4. An academic energy disaggregation competition is badly needed

The panel discussion following the two algorithm sessions brought a lively debate around what a NILM competition might look like. Phrases like "bring it on!" were thrown around, though it also became clear that defining a scenario (e.g. sample rate, scale) which encouraged broad participation is a real challenge. Furthermore, such a competition would need a strong investment in time from an impartial organiser as well as an expensive process of data collection.

5. NILM researchers love puzzles

George Hart gave an excellent keynote talk describing what it was like to perform energy disaggregation research in the 1980s. He then went on to talk to a transfixed audience about his more recent interests in mathematical sculpture and puzzle solving. However, nothing prepared me for the silence which dropped over dinner when he handed out a series of puzzles (mostly physical blocks which had to be separated or assembled) as every researcher forgot the topic of the conference and indulged in some more traditional problem solving.

Watch NILM 2016 Day 2 Livestream

The stream for day 2 of NILM 2016 is now live:

https://www.youtube.com/watch?v=KlqkP3EVXUY

I'll try to add links to the videos of each talk once they're available.

NILM 2016 Livestream

Update 08:46 14.05.2016: new youtube link

Update 09:06 14.05.2016: new youtube link

The 3rd International Workshop on Non-Intrusive Load Monitoring will be held in Vancouver, Canada, from May 14 to 15 2016, at the beautiful mountaintop campus of Simon Fraser University. We are hoping to livestream the event on YouTube via the following link:

https://www.youtube.com/watch?v=3YHBC-xvm4c

This livestream will be a little experimental, so please keep an eye on NILM2016 on Twitter if we encounter any technical difficulties. For those of you who are not able to attend or watch the livestream, we are hoping to upload videos of the presentations to YouTube to be viewed at a later date.

Data Management guest lecture

Yesterday I gave a guest lecture on Gopal Ramchurn's Data Management course at the University of Southampton. It was great fun exposing the first year Computer Science undergrads to technologies like Cassandra, Elastic Search and Kafka, while also contradicting a lot of what I was taught about database design on the same course 9 years ago. My favourite questions from the (surprisingly attentive) audience were:

• Doesn't all that data replication negate the point of using a database in the first place?
• Why wouldn't you pay for supported packages of open source technologies?

If you're interested in a similar guest lecture on your course please don't hesitate to get in touch!