Paper accepted at AAAI on NIALM training

I recently received notification that my paper titled Non-intrusive Load Monitoring using Prior Models of General Appliance Types has been accepted at AAAI-2012. The paper will appear in the Computational Sustainability for AI track, for which I will give an oral and poster presentation at the conference. The abstract for the paper is below:

Non-intrusive appliance load monitoring is the process of disaggregating a household's total electricity consumption into its contributing appliances. In this paper we propose an approach by which individual appliances can be iteratively separated from an aggregate load. Unlike existing approaches, our approach does not require training data to be collected by sub-metering individual appliances, nor does it assume complete knowledge of the appliances present in the household. Instead, we propose an approach in which prior models of general appliance types are tuned to specific appliance instances using only signatures extracted from the aggregate load. The tuned appliance models are then used to estimate each appliance's load, which is subsequently subtracted from the aggregate load. This process is applied iteratively until all appliances for which prior behaviour models are known have been disaggregated. We evaluate the accuracy of our approach using the REDD data set, and show the disaggregation performance when using our training approach is comparable to when sub-metered training data is used. We also present a deployment of our system as a live application and demonstrate the potential for personalised energy saving feedback.

Full details can be found on my publications page.

1st International Workshop on Non-Intrusive Load Monitoring

Mario Bergés recently pointed me towards this really exciting workshop on Non-Intrusive Load Monitoring:
http://www.ices.cmu.edu/psii/nilm/agenda.html
Some important information:

• When: May 7th 2012
• Where: Pittsburgh, PA, USA
• Objective: Unite researchers from a variety of backgrounds working on Non-Intrusive Load Monitoring

The agenda contains overview talks about the state of the art and also some spotlights from industry partners. There's also an afternoon poster session to present and gain feedback on any ongoing work. 

Overall, I think the event will be a great opportunity of learn about current research in both academia and industry, and I would thoroughly encourage anyone working in this area to consider attending. I'm planning to present some of my current work at the non-event based poster session, and hope you see some of your work at the workshop too!

Unsupervised learning for NIALM

I've recently been thinking about various training methods for NIALM systems, specifically those which can be applied to unlabelled aggregate power demand data sampled once per minute (or less frequently). Assuming no prior information of the appliances or their usage patterns, this clearly falls into the category of unsupervised learning.

In unsupervised learning, the goal is often to determine the unknown structure of unlabelled data. However, in our case we don't simply want to construct a model which represents the aggregate power data. In fact, we want to build a model of the data in which appliances are explicitly represented. This way, once the learning process is complete, we can form the disaggregation task as an inference problem.

Previous unsupervised approaches to this problem have used clustering to identify unique behaviour of appliances. These approaches have been shown to work well when applied to multiple features extracted from high granularity data (sampled at kHz). However, in the case of low granularity data, there is no way to extract features such as reactive power, power factor, etc. and we are instead left with a single feature; (real) power.

To give a visual representation of how clustering might perform on real aggregate data sampled at 1 minute intervals, I ran some experiments on the REDD dataset. To do so, I did the following:

1. Down sampled all data to 1 minute resolution
2. Subtracted the power of each circuit from the household mains circuit to calculate the unallocated, or 'unknown', power
3. Calculated the difference between consecutive power readings for each circuit
4. Excluded any change in power less than 100 W
5. Counted the power differences into bins for each circuit
6. Plotted these bins as a stacked bar graph for each household

As an example, here's the chart for house 1:

You might want to click on the image to enlarge it since the inline resolution isn't great.

There are two key points to take from this plot:

1. There are two unique clusters at the higher end of the power axis (labelled washer dryer and oven I think). These clusters would be easily identified by a clustering algorithm due to their clear separation from the other appliances.
2. There are two clusters around the 1500 W mark (corresponding to the microwave and kitchen outlets I think). One cluster completely subsumes the other, making it very difficult or even impossible for a clustering algorithm to separate the two.

This is just one example, and although the appliances and their usage will be different across houses, I believe this trend will continue. There's always likely to be appliances with high power demands that are easily clustered, however, for appliances with lower power demands the corresponding clusters are increasingly more likely to overlap.

Although at first glance this might seem okay, because we're more interested in the appliances that consume the most energy. However, power demand and energy consumption are not always correlated. This is because power demand represents the rate of energy consumption, and therefore energy consumption depends of both the appliance's power demand and its duration of use. Two examples of appliance types with low power demands but high energy consumptions are the refrigerator and lighting. Because these appliances are on for such a long time, their energy consumption might turn out to be similar or even greater than kitchen white goods with the highest power demands.

I also generated the graphs for the other 5 houses in the data set, which I've included below (click to enlarge):

NIPS - Day 6

This was the final day of the NIPS conference and the day of the workshop on Machine Learning for Sustainability; the workshop I submitted my paper to. It's a bit of a shame it came on the day when I think everyone was the most tired, but I think that's just a reality of post conference workshops. In general, the workshop mostly covered three main topics: climate modelling, energy and environmental management.

For me, the best part of this workshop was the chance to meet other researchers working on the problem of energy disaggregation. The workshop was organised by Zico Kolter, with an invited talk given by Mario Bergés, both of whom have published in this area and their papers have ended up on my reading list.

The morning invited talks were:

• Mario Bergés - Machine Learning Challenges in Building Energy Management: Energy Disaggregation as an Example
• Andreas Krause - Dynamic Resource Allocation in Conservation Planning

Before the midday break were the spotlights and poster session. These sessions were my first chance to present my work to the an external academic community and receive feedback, which was an absolutely invaluable opportunity. I tried my best to balance my time between presenting my poster to others and also to discuss the other posters with their authors. I was really impressed with the quality of the accepted papers, and hope this is indicative of the future of the MLSUST workshop.

The afternoon invited talks were:

• Drew Purves - Enabling Intelligent Management of the Biosphere
• Claire Monteleoni - Climate Informatics
• Kevin Swersky - Machine Learning for Hydrology, Water Monitoring and Environmental Sustainability
• Alex Rogers - Putting the “Smarts” into the Smart Grid: A Grand Challenge for Artificial Intelligence

The concluding session of the workshop was a panel discussion regarding the future of MLSUST. Much of the discussion was centered around the naming of the workhop, and whether sustainability represented the variety of topics in the workshop, but also whether the term encouraged researchers from other fields to submit and attend. Other ideas to promote interest in the workshop were raised, namely: well defined problems, competitions and data sets.

For me, the workshop was a brilliant venue to receive feedback on my PhD work, and for that I owe the organisers thanks.

NIPS - Day 5

I woke up on the first day of workshops to a beautiful sunrise in Sierra Nevada. I'm not sure what it is about mountain weather that causes the sky to turn so pink, but I'm definitely not complaining.

The NIPS organisers seem to have set it in stone that the each workshop will have a 5 hour break in the middle of the day for sleeping/skiing. As much as the 7.30am workshop starts hurt, I have to say I thoroughly enjoyed the opportunity to ski. I'm not sure that the schedule helped to sync the body clock with Spanish eating times, but I think a lot of people are jet lagged enough for that not to matter.

I attended the workshop on Decision Making with Multiple Imperfect Decision Makers. The two talks I enjoyed the most were:

• David Leslie - Random Belief Learning
• Stephen Roberts - Bayesian Combination of Multiple, Imperfect Classifiers

I also spoke to Edwin during the coffee breaks about how crowdsourcing is used to aggregate multiple data sources with of different quality, reliability etc. This prompted the thought that maybe something similar could be used in NIALM to collect large sets of appliance signatures. In my opinion, this would make a far more powerful data set for building generalisable appliance models than it would be for evaluating the performance of NIALM approaches.

NIPS - Day 4

Today was the final day of the main track, with only a morning of talks scheduled. Apart from some interesting but less relevant talks about the brain, I thought this one was pretty good:

• Continuous-Time Regression Models for Longitudinal Networks

After the morning sessions was the NIPS organised tour of the Alhambra. Moving over 1000 people around in coaches was always going to be a logistical nightmare, so it was no surprise loading/unloading took longer than planned. The tour of the Alhambra was excellent, and I've put a few photos at the bottom of this post to prove it. After the tour we set off up the mountains towards Sierra Nevada for two days of workshops. As the gradient increased our bus started to struggle, and when the road widened to two lanes the other buses stormed past us. However, the bus did its job, and we arrived in Sierra Nevada not too long after the others.

NIPS - Day 3

Today was another great day at NIPS, although I think the long days are starting to take their toll. The sessions were pretty cool, with some some great invited talks. I even noticed a few crowdsourcing papers making their way into NIPS, which I've already passed on to people in my group who might be interested.

Instead of the standard restaurant lunch, I decided to try to make the most of my last full day in Granada and go on a walk around the old town. There was so much stuff I would have missed I had not have strayed far from the conference venue, so really glad I made the effort. I took a few photos on my phone which I've stuck at the bottom of this post.

In the poster session I found one paper that was particularly relevant to what I do:

• Selective Prediction of Financial Trends with Hidden Markov Models

I think the method used for deciding whether to predict outcomes in financial markets is similar to the way in which our approach ignores observations from other appliances described in our paper. Their method of expanding single states of the Markov process to add more detail was also really interesting. I'll have to read their full paper when I get a little more time.

Tomorrow morning contains the last of the sessions at the main conference, followed by a visit to Granada's famous Alhambra. We then make the bus journey up the mountain to the workshops is Sierra Nevada.