Alex's data over the past two weeks

Alex has collected data from his house over the past 2 weeks, recording both instantaneous power readings and energy consumption over each period. The following feeds have been metered:

• Household aggregate
• Tumble dryer
• Washing machine
• Dishwasher
• Fridge
• Computer
• TV

The pie chart below shows each metered appliance's share of the total energy consumption.

This is similar to what we'd expected, as the metered appliances account for roughly half of the total energy consumed (if we pretend the 'always on' appliances are metered too). The most surprising thing is that the tumble dryer doesn't appear to have consumed any energy. This is most likely due to Alex simply not using the appliance, but it could also be due to the appliance being unplugged from the Plogg.

I've also run my script to subtract each metered appliance's energy consumption from the aggregate reading for the same period. The graphs below plot the energy consumption of Alex's house over one day, excluding the appliances listed above.

Day 1:
Day 2:
Day 3:

The most interesting point of these graphs is how few of the appliances appear to overlap. This is in stark contrast to plots of the household consumption without any appliances removed. This supports our assumption that subtracting known signatures from the aggregate can be useful in identifying unknown signatures.

Another interesting feature is the varying energy consumption overnight. Now that the fridge has been removed, I would have expected this to be constant. However, there are two possibilities that could account for this variation. The first of which is noise. This would be introduced if the periods over which each meter samples differ. Alternatively, the variation could also have been caused by appliances with uneven energy consumption that we simply haven't considered yet.

Return from skiing and skills course

I've just got back after my week away skiing and the 4 day generic skills course.

The generic skills course covered these topics:

1. Research
2. Planning
3. Literature searching
4. Research writing
5. Oral presentation skills
6. Presenting with visual aids
7. Poster presentations
8. Personal skills

Overall, I found the course really beneficial. I think best exercise was the opportunity to present a number of times, and getting feedback purely on how you presented it, instead of about technical aspects.

Another thing we were encouraged to do was write imaginative introductions and conclusions. Instead of introducing a presentation with "Hi, I'm someone and I'm going to be talking about something", you should try to capture your audience's interest. This is something I'd never have the confidence to try out for real without first practising it in a risk free environment.

My plan for today is to analyse the energy data from Alex's house over the period that I was away to validate my previous assertions. Next week I may get involved with some coding for BAE Systems.

Week summary

This week I've been working on three things:

1. Using preliminary data from Alex's fridge, computer and TV to estimate the fraction of his household's energy consumption. I've updated last week's pie chart to reflect this.
2. Investigating the technique of subtracting known appliance electricity readings from the aggregate readings. This does not appear to work reliably for power readings at the granularity of 1 reading/minute. However, it does work more reliably for energy readings, as is represented by the previous post. The line graph in the post shows the aggregate power readings with the dishwasher, washing machine and tumble dryer removed.
3. Categorising the features which I use to identify appliances when looking at their isolated signatures or their signature within the aggregate data. I've made a list of the features which can be used when disaggregating data captured from a smart meter, and linked them to who proposed it, and who has investigated it. This list is not complete yet, and I intend to expand it as I work through my collected literature.

I am now going away skiing for a week. Following this, I will attend a 4 day course on research and presentation skills.

Future work plan:

• Collect data over a 2-3 week period using Alex's aggregate supply, and sub-metered data from 6 appliances. I'm expecting these sub-metered appliances to make up roughly half of the total energy consumption, and will hopefully be able to clearly see the remaining high consumers easily. I'm hoping to be able to recognise some of these appliances: kettle, microwave, cooker, hob and lighting.

Household energy consumption without major consumers

I've already noted that the tumble dryer, washing machine and dishwasher account for a quarter of Alex's energy consumption. However, now that we have the data for each of these appliances through sub-metering, we can subtract these appliances from the overall demand. This produces a graph of the appliances which account for the remaining 75% of the total energy consumption.

Also as noted earlier, it's not feasible to subtract individual appliance's power demand from the aggregate demand. However, we should get reasonable results by using the energy consumption. Below is a graph of Alex's overall energy consumption over one day, with the tumble dryer, washing machine and dishwasher removed.

As before, we can see the repeating patterns overnight and peaks during the day. Interestingly, without three of the major consumers, it's easier to identify some other major consumers.

For example, a kettle draws the maximum amount of power, which is approximately:

230 V * 13 A = 3000 W

Given that our samples are recorded over minute intervals, that's the equivalent of:

3 kW * (1/60) = 0.05 kWh each minute

Therefore, jumps of 0.05 kWh or more might correspond to the kettle. We can confidently rule out all peaks below 0.05 kWh as belonging to the kettle. Since the kettle's duration of use is only likely to be a few minutes, we expect it to be represented by a thin peak of only a few points.

In fact, close to the 840th minute, we can see the energy consumption quickly jumps from approximately 0.035 to 0.085; exactly the signature we're looking for. The fact that this signature is identifiable even during the use of another unknown appliance is encouraging.

Summary of appliance signatures and plan for this week

We can account a third of Alex's energy consumption to the tumble dryer, dishwasher, washing machine and 'always on' appliances.

The signatures generated by individual appliances contain the following distinguishing features:

• Alternating/constant power draw
• Shape of energy consumption

Although signatures containing an alternating power draw are not identical, this is still a useful feature to identify a signature by. This might still be visible even amongst the noise generated by overlapping energy draws from other appliances.

The reason why the shape of the energy consumption should be used instead of the power consumption is due to the repeatability of appliance signatures. This is due to some appliances drawing an alternating power with a frequency different to the sampling frequency.

However, the shape of the energy consumption only works well when appliances have a fixed length. Such appliances account for approximately a third of total energy consumption. Therefore this feature is only applicable to categorise the minority of energy consumption.

In addition, when appliances overlap this shape matching approach corresponds to picking the combination of signatures which minimises the error. However, this approach is not robust to noise generated from other appliances which do not have a fixed length.

As a first step towards shape matching, it is beneficial to subtract the consumption of all appliances known to be on from the aggregate consumption. This will be my main aim of work this week.

Breakdown of Alex's energy use

This pie chart shows the percentages of Alex's total energy consumption of broken down by the metered appliances.

The tumble dryer, washing machine and dishwasher were monitored over a two week period and the energy consumption measured. The fridge-freezer, computer and TV were monitored over a 12 hour period and the energy consumptions scaled up. The 'always on' was calculated by measuring the minimum power demand, subtracting that of the computer, and calculating its consumption scaled up over the period. The consumption of all remaining appliances was calculated by subtracting the consumption of each of these categories from that of the household.

These values are similar to what you'd expect from an average household.

Appliance survey - tumble dryer, washing machine and dishwasher

To gain a better understanding of how to recognise appliances from their power/time signature, I studied the plots of three appliances: tumble dryer, washing machine and dishwasher.

Tumble Drier: (power/time)


The signature is made up of two distinctive sections; the main cycle and the end of the cycle.

During the main cycle, the power draw alternates between 2000 and 200 W over a duration of between 100 and 140 minutes. The reason this is not represented consistently between the plots is due to the periodicity of sampling. Power values are only sampled at minute intervals, so any alternations between samples are not recorded.

The second section constitutes the final sixth of the appliances operation. Although this section is consistent on three of the four graphs, one plot still captures alternation of the power drawn.

Washing machine: (power/time)

This appliance also has two distinct sections; a main cycle and an end section. The main cycle is between 10 and 30 minutes in duration, with a constant power draw of 2000 W. The end period is generally much longer, with an alternating power draw of either 200 or 0 W.

Dishwasher: (power/time)

In this appliance signature, there are four distinct sections; two main washes and two idle periods.

The first main wash takes 25 minutes while the second takes 30 minutes, during both of which a constant power of 1600 W is drawn.

The first idle period lasts 20 minutes during which an almost constant power of 50 W is drawn. The deviations from this power are represented consistently on each of the plots.

Generalising the plots:

The overall shape of each graph containing an alternating power draw is better generalised from a smoothed energy/time graph. This effectively represents an average over the minute interval, instead of samples at minute intervals. An example of a smoothed energy/time graph for the tumble dryer is shown below.

Tumble Drier: (energy/time)

Smart Grid Enabled Value-Added Services for Residential and Small Commercial Buildings

Gopal just sent me this on home appliance monitoring: http://www.pointview.com/data/2010/05/38/pdf/Chellury-(Ram)-Sastry-6003.pdf

It's a good overview of the NIALM field, and raises some interesting web 2.0 issues rarely covered by academic literature.

In particular it describes the issues with data ownership and privacy now that cloud computing used to remotely store and process data.

It also mentions smart appliances, which can self report and/or be controlled by a central system. There's a clear application of agents here, and in particular human-agent collectives.

December 2010 Activity Log

Last month I had planned to:
Collect data from a variety of households, metered using different techniques.
Use Matlab to apply increasingly complex machine learning techniques described in literature.
Read up on relevant parts of HCI and Ubiquitous computing fields.

Last month I achieved:
Everything I'd planned.
Improved my general knowledge of energy characteristics of appliances.

Next month I plan to:
Look into fields other than NIALM for methods to recognise repeating patterns (fridge, washing machine etc).
Complete research and presentation skills training.
Keep reading the Bishop book.