Posted by: @majordennisbloodnok

↑

My preference would be to keep it simple. I'd say that if two days share similar maximums, minimums and averages then it's a close enough profile.

Somehow Google matches images, but I suspect the techniques are quite sophisticated.

Building on @majordennisbloodnok suggestion could it be as simple as average temps equal to within (say) 2C AND Max-Min equal to within (say) 3C.  Numbers chosen at random but this would be relatively easy to try whilst making the limits adjustable to see what it turns up.

It strikes me that the fact we are having this discussion at all illustrates the extreme difficulty of answering the question experimentally (unless of course you have access to a house in a lab).

Im still not convinced that comparing similar days is sufficient, I think you may well need to include the previous day at least.  Of course that still doesn't account for solar gain or wind.   

Posted by: @majordennisbloodnok

↑

My preference would be to keep it simple.

Always my preference too, but I fear it is going to be too simple! I first considered much the same as you did, min, max, mean, standard deviation (though strictly speaking meaningless, but nonetheless two similar profiles should generate similar SDs), easy to calculate, and maybe useful as a screening tool. But consider two imaginary profiles, one where the OAT goes from 5 to 10 degrees in a straight line, and the the other goes from 10 to 5 degrees in a straight line: the min, max, mean and SD will all be the same, but the profiles are rather different!

I'm currently thinking about things like imageJ which I have used before for image processing, possibly ImageMagick and duplicate image finders. The latter may work, especially with the right similarity setting, but any image processing approach needs images, meaning I will need to write some code to extract hundreds of images...

There is another thing that bothered me a while ago, which I then forgot, but it has now come back to me. Because I use a pivot table to get the daily sums and means, they are conventional days, midnight to midnight, but that means they contain two half setback periods, the last three hours of one, and the first three hours of the second. I don't think this matters if all we are doing is comparing energy use - but maybe it does? Any thoughts?

Posted by: @jamespa

↑

Somehow Google matches images, but I suspect the techniques are quite sophisticated.

See above - there are desktop programs that can do this (duplicate image finders). But it means generating the images first. Whenever I have tried google similar image searches, I have not been impressed. It's that blasted AI rubbish again.

Posted by: @jamespa

↑

Im still not convinced that comparing similar days is sufficient, I think you may well need to include the previous day at least.  Of course that still doesn't account for solar gain or wind. 

I can but try. I think the initial values of a 24 hour period will probably capture at least the ending state of the previous day. Weather effects, again, see above. There will be ways and means... 

Posted by: @cathoderay

↑

I can but try. I think the initial values of a 24 hour period will probably capture at least the ending state of the previous day. Weather effects, again, see above. There will be ways and means... 

I completely agree, just sounding a note of caution!

Posted by: @cathoderay

↑

Because I use a pivot table to get the daily sums and means, they are conventional days, midnight to midnight, but that means they contain two half setback periods, the last three hours of one, and the first three hours of the second.

I cant think of a reason why it would matter but, that said, when I did something similar (I cant remember exactly what and it may even have been on your data) I did use an offset day.  If it makes a difference at all I would guess that the best approach would be a day that starts as setback occurs (or maybe one hour before to capture the starting IAT/OAT) and ends 24hrs later.  But Im not saying I have an argument for why it would make a difference!  

I have just tried a program called Visipics (olden but golden...) on nine manually grabbed 24 hour period images and using loose match criteria, it matched these two images:

They are not the same, but they are similar...

Posted by: @cathoderay

↑

Always my preference too, but I fear it is going to be too simple! I first considered much the same as you did, min, max, mean, standard deviation (though strictly speaking meaningless, but nonetheless two similar profiles should generate similar SDs), easy to calculate, and maybe useful as a screening tool. But consider two imaginary profiles, one where the OAT goes from 5 to 10 degrees in a straight line, and the the other goes from 10 to 5 degrees in a straight line: the min, max, mean and SD will all be the same, but the profiles are rather different!

I agree, but the “similarity categorisation” could just as easily include first/last values as well to cover that. That would also at least partly address @jamespa’s concern about effect of previous day.

Posted by: @jamespa

↑

Im still not convinced that comparing similar days is sufficient, I think you may well need to include the previous day at least.  Of course that still doesn't account for solar gain or wind.

Actually, if the OpenWeatherMap API data is included, solar gain and windchill wouldn’t be difficult to include. That historical data includes cloud cover percentage, sun elevation and wind direction/speed. It couldn’t calculate the effect of solar gain/windchill on a particular home but would allow for comparison of similar solar gain/windchill conditions.

Posted by: @majordennisbloodnok

↑

It couldn’t calculate the effect of solar gain/windchill on a particular home but would allow for comparison of similar solar gain/windchill conditions.

Exactly. We only need to compare conditions, to decide whether they are similar or not. 

I now have the python code to get a simple 24 hour plot of OAT, I just need to run a loop to get a series of plots. It is a simple matter of changing a couple of numbers to change the start and end hour:

But it is also Christmas Eve, and I have chestnuts to roast, final presents to deliver, and holly to cut for some decorations! Python will have to wait!

Merry Christmas all!