How smart are smart meters? The hidden manual effort behind machine learning

Reports that autonomous vehicles have been seen driving around with drivers dressed as car seats reminded me of a recent article in the Financial Times by Tim Bradshaw about the intensive manual activity required for autonomous vehicles.

In the piece, he reports that “Most companies working on this technology employ hundreds or even thousands of people, … manually marking up or “labelling” thousands of hours of video footage“.

Most of the companies prefer that this stays hidden behind the curtain. “They all like to say it’s machine-learning ‘magic’” according to Dan Weld, professor of computer science and engineering at the University of Washington in Seattle.

And despite machine learning, deep learning, and artificial intelligence researchers "chasing this goal of “unsupervised learning”, when machines can teach themselves unaided" substantial human involvement is still required.

Machine learning for smart meters

This hidden activity was highlighted to me after I received an invitation to take part in a trial for an advanced smart meter. I had read some years ago that the voice recognition algorithms behind the Sony Aibo robotic dog had been acquired by a company looking to apply them to pattern recognition for electrical appliance consumption.

Having spent two years trialing their solution with households in Japan, the company - Informetis - is now repeating the trial in the UK to gather appliance signatures. The reason for this is that UK appliances are "somewhat different in their ‘profile’ to the ones in Japan."

Their vision is to create cost effective technology solutions that help home consumers reduce their electricity bills, starting with a single sensor that enables an ‘itemised view’ of the electricity bill. This shows how much electricity each of the main appliances uses, and calculates the cost.

The "first step in the deployment of our service, we need to collect data from multiple appliances (e.g. Fridge, Washing Machine, Microwave etc) in order to “train” our algorithms". This is their equivalent to marking up video footage.

My part in the trial lasted a little more than 2 weeks, and involved installing a smart sensor in the consumer unit, "black boxes" between the socket and appliances and a router to manage communications.

It took two people around 2 hours to install the equipment, register the appliance make and models, and ensure that the router was communicating with the server. An electrician is required to fit the smart sensor in the consumer unit, although the preferred method for the trial is to install it in a separate unit on the wall next to the consumer unit. While the uninstall was quicker, an electrician would be required to take the unit out of the consumer unit. As he wasn't available on the day my black boxes were removed, I was happy to leave it in for now.

Participants in the trial receive a John Lewis gift voucher, so with the man-hours also required, the overall cost to build a library of profiles and train the machine learning algorithm will add up.

At the end of the trial, I received a range of consumption analysis (however this will be available real time via an app, once it has been translated from Japanese).

Pattern recognition for domestic appliances

The first chart breaks down total usage by appliance over the period. I hadn't anticipated that the freezer would consume the most, or that the dishwasher would exceed the washing machine's demand. The reasons for this become clear from the detailed analysis.

The daily usage chart breaks this down further. While average consumption from the devices was a little over 3 kWh per day, minimum consumption was as little as 1 kWh.

Time-of Use tariffs have been available for businesses for many years, and the adoption of smart meters will enable domestic retailers to offer rates that incentivise or disincentivise consumption at different times of the day or week.

A study produced for Citizen's Advice by Brattle Group found that the system value was limited, and that modest annual savings of ￡5 per household would be achieved. Increasing the value to consumers would depend on higher electrical consumption from heat pumps and electric vehicles.

Having commented extensively in the media about the opportunity for industrial demand side flexibility in a smart energy system, I wanted to understand the potential in the domestic sector.

Fridge and freezer

A weekend away camping highlighted the appliances which are always on, and is a great chart to see the distinctive signature of individual appliances - fridge and freezer - that cycle on and off in a repeated, regular pattern. This chart also revealed that the coffee machine was consuming electricity even when switched off.

Refrigeration is frequently cited as a key opportunity for demand response, as the thermal mass and insulated cases allow the appliance to be switched off with limited impact on the contents unless someone opens the door.

This strength may also be a weakness, as the data shows that my freezer only consumes 100W for 10 minutes in every 40 minute cycle, while the fridge consumes 70W for 15 minutes then nothing for nearly an hour.

While the rapid, brief interruption of frequency response may work well for refrigeration, nearly 50,000 similar appliances would have to be aggregated to reach the minimum volume to provide Frequency Response services to National Grid.

While there is also the potential to provide demand response, without a Time of Use tariff, the avoided cost would be just 0.2p per cycle. Since average consumption of cold appliance fell by over 60% between 1990 and 2016, it would appear that consumers would benefit more from the reduced consumption from efficient appliances than dynamic demand response.

Dishwasher, washing machine and coffee machine

Some appliances can be scheduled to run at different times, and the next chart shows the dishwasher running both overnight, and in the evening after the evening peak. The profile highlights two spikes in consumption, even in the 50° Eco mode. The first as water is heated at the start of the cycle, and the other at the end for drying. In between, consumption is significantly lower at around 30W.

In contrast, while the washing machine also has a spike in consumption at the start (although for a 30° wash this is shorter than for the dishwasher), the spin cycle at the end consumes much less electricity.

While the challenge for scheduling the dishwasher is to ensure that the full cycle avoids peak rates, both appliances can be scheduled to start during cheaper periods.

The coffee machine, on the other hand, is indispensable first thing in the morning.

Oven and TV

The way we watch TV has changed in recent years. No longer are thousands of kettles and lights switched on at the same time, as on-demand streaming and binge watching series have reduced the "TV pick-up". While screens have become larger, technology has reduced the electricity consumption, with a 42 inch LED TV consuming less than a 21 inch CRT. Compared to the other appliances, the red line of TV consumption is barely visible on the chart.

Like the other energy intensive appliances, the oven consumes most, up to 3kW, when heating up at the start. Consumption drops to around 1kW to maintain the temperature, and drops to 25W just for the fan. Rather than baking or roasting in the evening, a Sunday roast would be cheaper. Maybe the smart, flexible future will have similarities to the past after all.

While a smart meter would be required for tariffs that reward demand shifting, the current models won't display consumption by appliance, or minute by minute.

The Informetis machine learning engine has the potential to transform the insights given to consumers, by using the patterns learned in the trial to disaggregate meter reads. This level of detail would help householders to target and optimise consumption from domestic appliances.

However, even with these trials to train the machine learning algorithm, these appliances constitute only a proportion of a household's consumption. The algorithm will have to identify the learned signals while filtering out the noise caused by all the other devices and appliances.

Limits to domestic demand response

While the insights provided by this level of metering increase awareness and understanding, and will help to adapt to Time of Use rates, the ability for domestic appliances to respond dynamically to short term price signals appears to be minimal.

I want (or maybe need) coffee in morning, and to eat dinner in the evening. I can schedule the dishwasher and washing machine to respond to Time of Use tariffs, but question how practical it will be to provide dynamic response from interrupting a cycle.

The fridge and freezer are constantly cycling on and off, but modern appliances are so much more efficient that a trade-in scheme might be the best policy to reduce demand. Without this, there is a danger that only those able to afford efficient appliances, solar panels and home batteries will benefit from the transformation of the energy sector. The savings from reducing consumption far outweighs the potential savings from flexibility, but domestic storage will amplify this opportunity. This may result in network costs being borne disproportionately by those without the ability to manage their use.

I entered the trial optimistic of what I would learn about my potential to participate in domestic demand management and response. And it has helped me understand the consumption patterns of my appliances. But my experience supports the Citizen's Advice conclusion that the electrification of heat and transport may be necessary for ToU tariffs to make a meaningful contribution to system flexibility.

If you live near Cambridge, click here to apply to take part in the trial.

Update: The data displayed in this blog comes from the individual black boxes, not from the machine learning algorithm. I wonder how well it compares.