in Energy, Methodology

Changing energy behavior with real-time feedback

In this blog post I try to put household energy usage in a control theory context to see if we can find better ways to influence behavior.

A lot of people are thinking about how we can reduce household energy use these days. Changing household energy use can be a challenge because to most people there is no working feedback loop available. Energy use is invisible. My own feedback loop consists of getting a monthly invoice displaying the cost I had for the previous month (e.g. “278 SEK”). That is a lousy feedback loop if you want to change behavior. I can not even see how much energy I used. Not that it matters because that information is useless without knowing what would be a reasonable level to compare with. And getting the information once a month does not provide a feedback loop for me to see the effect of my immediate behavior.

A basic model of a feedback loop for a system looks like this:

A controller operates on aystem and adjust control behavior based on the output form a sensor.

Based on a desired state (a reference value) a controller operates on a system and adjust control behavior based on the output from a sensor to make the system align the output to the desired state. For this to work there are a couple of things required:

  1. We must know what the desired state is. Without it we don’t know if the output from the system is good or bad.
  2. We must know the system output. Without the output we can’t calculate the measured error between output and the desired state.
  3. The controller must be able to provide system input to control the system and its parts in a predictable way.
  4. The delay between the actions of the controller and the sensor feedback needs to be short so that the result of different actions are visible.

Applying this basic model to the typical feedback loop for household energy use we get:

Monthly feedback through the electricity bill without a reference value makes it impossible to take action.

The “system” in this model would be your household. It consists of several things that use energy (refrigerator, TV, other appliances…). You control it by operating various appliances, setting the refrigerator temperature and making purchasing decisions for new hardware that affects energy use. You get the energy bill once a month with the amount of energy used.

As you can see there are several problems. I don’t know the desired energy level. Is my household below or above the average for similar households? If I switch off the TV instead of keeping it on standby, would it matter? Will setting a higher temperature in the refrigerator impact my energy use? Since I can’t see the immediate effects of my behavior there are no drivers to change it.

Real-time feedback and knowledge about the desired state

A first step to  influence behavior would be to provide real-time feedback and information about a desired state. That would make it possible for you to try various control actions and see what implications they have on the system. There are a lot of third party solutions to get real-time feedback e.g. Wattson, TED-5000 and so on, but from what I have seen there are very few options to get information about a desired state. Getting more data about desired state can in part be accomplished by sharing information about your household and its energy use with others.

With an energy meter a person can see the immediate effect of behavior on the household energy consumption

But, to be able to really affect behavior we need to increase the resolution of the parts of the system (at least sections or types of energy use in the household) by using multiple sensors and comparing the output with multiple desired states. This may quickly become complicated for an individual and a reasonable next step is to automate the controller function (you) by allowing automated control actions on parts of the system (e.g. adjusting your refrigerator temperature based on data about its contents, model, and similar household data).

I guess we have some work to do before we can get a better feedback loop in place. What’s in it for the energy companies? Will we have to install our own real-time meters?

(Thank you to Daniel Brolund who provided interesting insights into control theory last week).

Write a Comment here on the real web


  1. Thanks Peter, for an interesting topic that really gets me going. I just can’t use twitter to give you “feedback” on this, it gives me too small amount of data to send to you in one go, and hence to small amount of data for you to make a good control-decision on ;-)

    Feedback loops that spans both the technical (automation) and social behaviour are always interesting beucause they are “wicked” problems. I would love to sit down with you and a whiteboard to discuss this but I guess this will do for a first session. A couple of things that struck me while reading your text was:

    1) I read into this that what we actaully want to do is to behave in such a way (control) that our household either minimize energy consumption or minimize the households cost of energy while still maintaining a set of habitat-parameters, such as indoor temperature, humidity, TV-set on following my TV-show time schedules, etc. 

    What we do have then is a multi-in-single-out-system exactly as you depicted it. But what you also could have is a faster feedback loop based on a model/models of your system (or part of your system). This model then gives you measurements (predictions) of your systems states at a faster pace. You can in that way get a measured error at a higher frequency than before without introducing any new sensors etc. and get a better control of your system. When you then at a slower rate get feedback on some of the states (if at all) you then make corrections to your model. Creating a model for some electrical applicanes or social behaviour might sometimes be easy, sometimes hard. The idea is that it only needs to be correct to a certain degree. You can then later adjust it with real life measurements to bump it into the correct state.

    2) The other part which I find really interesting is to find a way to control the upgrade of your household to lower energy consumption appliances and heating systems, typically human/fuzzy decisions. I guess your idea of a “sharing” experiences, with other households on energy uses for appliances, heaters, fridges etc. then act as a model in your system. This model then gives you faster feedback on your households benefits when doing that control action on your system.

    Well, love the idea. Love the discussion that follows this on parallel problems such as changing behaviours etc. As I said, a wicked problem ;-)



  2. Thank you! Not sure I understand everything, but I guess this is a topic for a whiteboard face to face meeting:-). With “behavior” above I mean it in the broadest sense. Behavior can be how you decide to operate an individual appliance as well as behavior in making decisions about changing parts of your household “infrastructure” such as heating, refrigerator etc.

    The idea is that feedback information will provide you with a method to make better decisions and get feedback on the outcome of those decisions as quickly as possible. By sharing the result of individual decisions on a larger scale over the internet you can get better values for desired states and further improve the behavior. But at some point you will have too much information and it may become difficult to make use of it.

    I guess this is where I would like a part of the system to follow a modeled behavior like you describe? Sharing models that work with others may further contribute to the single-out value and contribute to an efficient feedback loop.

  3. Hi, well I wrote the piece above in a state of excitement so the “understanding” part might be on my downside here.

    Anyway, to try to put it short. Modelling gives us a way of getting sensory input quicker than waiting for real life sensors. How the model is built is another thing, it can be mathematical, it can be a social model like sharing experiences with friends etc. Which is usually how life works when you make decisions.

    But I also think I misunderstood your behaviour thing in my writing above. Now I see that the control of the output, i.e. energy consumption is not the main focus here it is changing the controllers behaviour (you, me and a computer) over time based on various influences. This is something called an adaptive controller. The tricky part here is that some of them are usually very hard to prove they are stable. But then again, thats life.



  4. Also, the book Heat [] discusses some interesting ways to lower the world’s energy consumption on a larger, systemic scale.

  5. Does not work. That is more for a snapshot figure for a single appliance. You can not aggregate the data from those over time in any simple way. The feedback loop I am after is for continuous improvement.

  6. Interesting. But can a model be efficient to change behavior if you know it is only a model? 

  7. Hej Peter

    ser fram emot att resonera med om detta i höst. Utan tvekan kommer vi få möjlighet att betala per timme för vår el framöver, från 2013 tror jag. Olika pris per timme, men ackumulerad faktura per månad fortfarande. för besparingars skull handlar det då kanske inte primärt om utnyttjad energimängd, utan om man kan omfördela över dygnet för att få ned snittpriset genom att flytta laster över dygnet, kanske specifikt ‘tunga’ laster. Visst vill vi automatisera, och annat vill vi opportunt bara ‘köra’ om energipriset är under en eller annan gräns vi satt som acceptabel.

    Igår lyssnade jag på en podcastintervju med grundaren av – en site full med öppna sensordata. jag tänker mig att koppla dit sensorer som jag kopplar till olika variabler i min miljö, och sedan använda deras ‘triggers’ för att slå på eller slå av vissa saker över dygnet beroende på pris, temperatur eller andra variabler. en härlig soppa kan det bli!

    /peter, semesterluddig i tanken 

  8. Total cost or total energy use are two different “desired states” in the feedback loop but they still depend on feedback and data about what a “reasonable” desired state is. Some behavior can not be changed over the course of 24 hours (it may be difficult to use your energy in the middle of the night). But you are right i that there are multiple ways to act depending on the type of “state”.

  9. Hi Peter, Do you know of any studies that talk about the effect of upgrading equipment on behavior. A local utility found that people who bought high efficiency equipment often used more energy because they turned the temperature up in the home and took longer showers. Our company does the upgrade of equipment and we are starting to get some complaints about energy usage and I suppect that the customers behavior is changing after we do the wook. They may not be using the wood stove as much as before. Thanks for your blog!