@jamespa - at this stage, I am only looking at the relationship between energy in and OAT, and as the heat pump in normal WC no auto-adaption mode sets the LWT (and as a result, the energy in) based solely on the OAT, those other factors, which the heat pump knows nothing about, will not affect LWT/energy in.

The problem is I can't collect what you call robust experimental data with the other variables controlled because many of them are beyond my control. I can only collect noisy messy data, but perhaps I am more comfortable with that, because I am used to working with epidemiological data, and that is almost invariably noisy and messy, sometimes almost to the point where the only data worth its salt is mortality data, and then you realise even that is noisy and messy eg the elastic definitions of a covid death during the pandemic meant it was impossible to know how many alleged covid deaths were really covid deaths. You then end up giving up on case definitions, and look solely at excess mortality, only to find that is noisy and messy too! Was the excess due to covid, or anti-covid interventions (shutting down routine healthcare, emergency patients avoiding hospital etc). All I am just saying is that I am very used to, and familiar with, noisy messy data.

For the heating data, I am going to start by finding the longest period of steady WCC endpoints, and then pull out that data and see what it shows. The main problem is the data overall is sparse for lower OATs, which in many ways is the operating conditions we are most interested in. God forbid that I should ever want cold weather, but a bit of me says if only, just to get some data. 

If it looks as though there is something in the data, I will persevere, if not I will move on to something else.         

Posted by: @cathoderay

↑

You then end up giving up on case definitions, and look solely at excess mortality, only to find that is noisy and messy too! Was the excess due to covid, or anti-covid interventions (shutting down routine healthcare, emergency patients avoiding hospital etc). All I am just saying is that I am very used to, and familiar with, noisy messy data.

Good to hear.

By robust I don't mean low noise (I think that's impossible without a lab), as I say and as you doubtless know there are techniques for digging signal out of noise and so what I mean is data which lends itself to the application of these techniques.

The problem should be a bit easier than getting Covid deaths, because there are no ethical problems with deliberately switching on and off the variable whose effect we are trying to assess!

Posted by: @derek-m

↑

If my understanding is correct, I think that you will find that the modeling tool performs the task that is being requested, since it calculates the Energy In for each 1 hour period of the day based upon OAT and Energy Out.

It will not. I am not persuaded you have read my recent posts describing what I am trying to do, or whether you have, and I have written them so badly you failed to grasp what I was saying, but the key point and difference is I start with observed data, not a model. The only 'modelling' involved, and I wouldn't even go so far as to call it a model, is to fit, in effect, a regression line to the OAT/Energy In plot (a process which I would, unsurprisingly, call fitting a regression line), which I can then use to derive expected energy in at various OATs. I don't even need the underlying regression equation, I can instead use the regression line as a sort of look up table. Apart from the regression line, the whole point is not to use a complex model, and instead do a simple transparent and understandable comparison, between expected and observed. Epidemiologists do that sort of thing all the time, excess mortality is just the observed number of deaths compared to the expected number of deaths (and can be negative, if the observed number is less than expected). Having said that, it can also be problematic: how were the expected number of deaths determined? Exactly the same problem applies to my heat pump data, which is why I want a transparent and graspable method that does not involve complex modelling.

Posted by: @jamespa

↑

The problem should be a bit easier than getting Covid deaths, because there are no ethical problems with deliberately switching on and off the variable whose effect we are trying to assess!

It should, the main complication with the covid deaths data is human behaviour, deciding what is and isn't a covid death, and those decisions, particularly with pandemics, are often political decisions, not medical decisions, that make political sense, not medical sense. And then there are things like hot stuff bias (the tendency to be more likely to make a diagnosis when it is a hot topic). At least we by and large don't have those complications to deal with! 

Posted by: @cathoderay

↑

Posted by: @derek-m

↑

If my understanding is correct, I think that you will find that the modeling tool performs the task that is being requested, since it calculates the Energy In for each 1 hour period of the day based upon OAT and Energy Out.

It will not. I am not persuaded you have read my recent posts describing what I am trying to do, or whether you have, and I have written them so badly you failed to grasp what I was saying, but the key point and difference is I start with observed data, not a model. The only 'modelling' involved, and I wouldn't even go so far as to call it a model, is to fit, in effect, a regression line to the OAT/Energy In plot (a process which I would, unsurprisingly, call fitting a regression line), which I can then use to derive expected energy in at various OATs. I don't even need the underlying regression equation, I can instead use the regression line as a sort of look up table. Apart from the regression line, the whole point is not to use a complex model, and instead do a simple transparent and understandable comparison, between expected and observed. Epidemiologists do that sort of thing all the time, excess mortality is just the observed number of deaths compared to the expected number of deaths (and can be negative, if the observed number is less than expected). Having said that, it can also be problematic: how were the expected number of deaths determined? Exactly the same problem applies to my heat pump data, which is why I want a transparent and graspable method that does not involve complex modelling.

Please feel free to ignore all my post.

I've got through the data, and while there were a fair number of WCC end point changes over last winter, they all happened mid winter, and the period from the time I started full data collection and the end of the heating season ran throughout on one WCC setting, with no setbacks (apart from one power cut on April Fool's day). Here is the scatter plot:

The plot has tightened up considerably, especially at higher OATs, and I feel reasonably confident in saying that when the hourly average OAT is for example 10 degrees, then the energy in will be, on average (average is OK, it will even out) 1.2 kWh. I can't explain the higher OAT outliers, always below the main line - could be hangover effects eg a transition from DHW to heating or vice versa, or more likely an hour where some of the energy in went to DHW heating, so that going to space heating was less than it would otherwise have been for that hour eg in a DHW priority heating hour, 30m is spent heating the hot water, 30m on space heating, meaning the total for space heating for that hour will be half what it should have been for the given OAT. I suspect the increasing spread at lower OATs, becoming decidedly more visible at 5 degrees and less, may well be due to defrost cycles, an hour that has more defrosts will presumably use less energy in, because it is not heating the water so much. Overall, the period was longish (a very precise scientific term) and there was a reasonable spread of OATs, but with less points at lower OATs,

The period 16 Oct 23 to 4 Nov 23 also ran with one WCC throughout, and was characterised by mostly mild OATs. Here's the scatter plot:

Perhaps slightly more scatter, but that may be a visual artefact cause by less data points. Nonetheless, again I can say that, given an average hourly OAT of 10 degrees, the average hourly energy in will be 1.1 kWh.

Are these estimates of expected energy in for given OATs good enough? I think they probably are, partly because they are based on real data, and partly because the scatter plots look credible. Personally, I am not too bothered about the spread at a given OAT, provided the spread is reasonably symmetrical (technically, they are close enough to a parametric distribution), which most of the time it is, because the average will then reflect a meaningful number - the average energy in for that OAT.  

Two important caveats: (a) these are data from the csv file, I haven't yet applied the calculated to actual correction factor and (b) the weather curve is now slightly different (both ends a bit lower) than it was during the above two periods, meaning neither of the above apply directly to the current running state, regardless of whether I have setbacks and recovery boosts or not. I can however adjust the current WCC to match that used previously, and then the estimates from the above data will apply, at least well enough to satisfy me. After all, I am not trying to find the Higgs Boson particle, instead, I am just trying to get a pragmatic answer to a simple question with a surprisingly elusive answer. 

@jamespa

Having now answered the first part of the original question, that setbacks do appear to save energy, how do you propose that the second part be answered?

As you say much tighter.  This allows you, as you say, to get an average expected value.   Your actuals will have the same scatter so you will need to average a good number of actuals to get the average actual to a similar degree of accuracy, in order that the difference between actual with setback and expected without setback is meaningful.  I'm not sure exactly what you are planning to compare, but whatever it is principles should be the same, the measurements will be roughly as noisy as the data above so you will need quite a few to average out the noise so that the difference is meaningful)

Posted by: @jamespa

↑

I'm not sure exactly what you are planning to compare

I described above what I plan to compare: expected vs observed energy in during periods with setback and recovery boost. The key thing I am trying to measure is the extra energy in, over and above what would have been used anyway, during the recovery boost. Thus, simple example during a recovery period: OAT = 10C, expected energy in (from historical data) = 1.2 kWh, observed energy in = 1.7 kWh => extra (boost) energy in = 0.5kWh.

Expected = expected energy in, based on aggregate historical data, collected as described above. For example, if hourly average OAT = 10, expected energy in for that hour = 1.2 kWh (as in above example).

Observed = the actual observed energy in value for the hour.

I'm trying to find the extra energy in you say is getting used by the system to achieve the recovery, the energy in missed by the Mk 1 Eyeball Integrator. The current expected energy in needs to be based on the current OAT, to deal with the common situation that the OAT changes between the setback and recovery period, eg setback 2100 to 0300, temp steady at 8 degrees, recovery boost starts 0300 and at the time the OAT starts to fall, to 5 degrees, and the energy in rises - as expected - but how much of the rise is expected, would have happened anyway, because of the fall in OAT, and how much is the boost? In other words, what does the boost cost? That then gets set against savings during the setback, to determine overall savings. 

@cathoderay Yes understood, in general terms.

But given that you cant measure the 'actual observed energy in value for the hour' multiple times in the same conditions (to do the necessary averaging to reduce the noise), what I meant, but did not say sufficiently clearly, is - I'm not sure exactly what set of observations (of actual observed energy in value for the hour) you are planning to aggregate to do the sum whilst taking out the noise on those observations. 

Its almost certainly much easier to do than describe, so I suggest you may want to do it rather than attempt further to describe it!

@jamespa - the 'actual observed energy in value for the hour' is the value in the csv file for that hour. Its the equivalent of the observed deaths over the week year or whatever period you are using. An epidemiological worked example might help (data from ONS weekly deaths):

In the week ending 3 Nov 2023 (Week 44, this is the latest available data), the number of observed registered deaths (all causes, all ages, all sexes) in E&W was 10,983. It's a one off value that applies to that week, you can't take multiple readings (well there are exceptions, eg late registrations, but I want to keep things simple), you can't average anything, there is no noise as such, it is just the count of deaths for that week.

The five year average for the same deaths (E&W, all causes ages and sexes) for the same week 44 is 10,604. ONS currently use a rather dodgy five year average, that includes some covid years (2017 to 2019 and 2021, 2022), and has used at various times used others: (I quote) 2016 to 2019 and 2021 for comparisons with 2022; and 2015 to 2019 for comparisons with 2020 and 2021, which generally has the effect of conveniently reducing excess mortality, I prefer to use 2015 - 2019 throughout, but that's another story for another time, lets run with the ONS number: 10,604. You will appreciate this is an extremely crude way of determining the expected number: just a crude average of only five years.

We thus have, for the week ending 3 Nov 2023, 10,983 observed deaths, compared to expected 10,504 deaths, giving an excess mortality of 10,983 - 10,604 = 379 excess deaths. This is actually quite a small number, over the past couple of years it has been much higher, but there is a legitimate question about how much of the excess is due to population changes, ie more people, and more older people (in the trade, AKA air thieves, and I should know, I am one of them), meaning that even if death rates remain constant, you will still see more deaths. 

In my proposed heat pump energy in analysis, the total observed and recorded in the data file energy in kWh over the hour corresponds to the observed deaths, the expected number of energy in kWh at the current OAT derived from the historical data corresponds to the expected number of deaths based on the five year average, and the difference (the extra (excess) energy) corresponds to the excess (extra) deaths.

I will get on with the analysis, but don't hold your breath, because I need to collect more data, and a day's worth of data takes a day to collect.

In the meantime, I did manage to run some of @kev-m's data (extracted using the Mk 1 Integrator) through The Model ™, which I have been meaning to do for a while, as he has, as I understand it, an Ecodan running with setbacks and auto-adaption ie recovery boost, with some interesting results. 

Posted by: @derek-m

↑

@jamespa

Having now answered the first part of the original question, that setbacks do appear to save energy, how do you propose that the second part be answered?

Im not sure why this was directed at me especially, but since you ask here are my thoughts:

For the purpose of the second half of the question I think we can discard any 'results' (whether they are simulated or experimental) where the house is not returned to its original state within 24 hrs.  Longer setbacks eg for vacation are a different case altogether

Beyond that I think that, for many people, there are two interesting cases for a night time set back namely:

• recovers (sufficiently) by breakfast time (let us say 7.30am)
• recovers (sufficiently) by teatime (let us say 5pm)

Individuals have their own thresholds and timings but I would say recovery to (say) within 0.5C by teatime is probably good enough to qualify as sufficient (subject to the proviso that full recovery happens within 24 hrs), perhaps one could stretch this to 1-1.5C at breakfast time?

So I would be looking for results where the control strategy achieves one of the above (as well as the basic requirement to recover completely within 24hrs) as the ones which meet the second part of the question.  Obviously this constrains the scenarios for the first part.

Those are my initial thoughts but there are clearly an infinite number of possibilities.  Understanding these would be an excellent start though.

I think the other thing of interest is the answer to the question - are there scenarios (which are likely to occur) where setback costs money.  There clearly are, a very high thermal mass house where the house barely responds will be more expensive to run with setback than without, because it acts essentially as an integrator.  Are they likely to affect many people - dont know.

This all suggests some more playing with the model (because this really is too difficult to do by experiment) to get a more quantitative feel.  The table you posted is an excellent start, but there is more to explore.  If I get a few hours I may do this but at present Im concentrating on other matters related to, and also not related to, heat pumps!  

@jamespa

Thanks James.

I posted it to you since you are familiar with the subject matter, and would provide an unbiased reply.

Posted by: @cathoderay

↑

In the meantime, I did manage to run some of @kev-m's data (extracted using the Mk 1 Integrator) through The Model ™, which I have been meaning to do for a while, as he has, as I understand it, an Ecodan running with setbacks and auto-adaption ie recovery boost, with some interesting results.

       

Life has got in the way a bit but I'm going to pull together some proper data to correspond with the charts I produce. The current issue is that the Melcloud data dump is now too big (about 600,000 rows) to load into Excel on my laptop and I need to find a way to extract parts of it rather than Melcloud's approach of giving you all data since your system was installed.  

In the meantime this is one example of setback and recovery on a cold day in January, along with the data. Below zero and regular defrosts; just about the worst conditions possible.  I'll pull together some examples of milder weather and using WC only rather than AA when I get my data issue sorted. BTW energy is recorded as Wh, each row is one minute. 

@Derek-m is this the sort of data you need for your model? 

Posted by: @derek-m

↑

I posted it to you since you are familiar with the subject matter, and would provide an unbiased reply.

I suspect a more accurate answer may be that D wants to paint a picture that has D and J as sage old chaps having a learned chinwag in the library while C is the child in the garden having a tantrum. 

Some more to the point comments on the above:

(1) I don't know why @derek-m thinks the first part of the original question (do setbacks save money?) has been answered. His own posted hypothetical modelling suggests savings can be achieved, but at the cost of comfort; if a recovery boost is added, the savings become trivial. I don't recall @jamespa making a definitive statement either way, and the general thrust of his most recent comment is that nothing is settled, experiments are doomed because the subject is just too complicated ("this really is too difficult to do by experiment") and our best chances lie with "some more playing with the model...to get a more quantitative feel". None of that suggests the matter is settled. Nor do I think the matter is settled. My hunch, based on the simple idea that it must cost more to maintain a house at a higher average temperature, is that setbacks do save money, but I have yet to demonstrate that, because the experiments or rather observations and analysis are difficult (but not impossible) to do. I say simple because I appreciate that the thermodynamics of recovery and boost are much more complex that those of maintaining a steady state. I also have, as is known only too well, a profound distrust of models, and the depth of that distrust increases the further the model is from observed data. I stress this is a personal view, if others want to fly model aeroplanes, and then find themselves wondering why they are crashing to the ground, that's fine by me, though I might be inclined to add please don't crash in my back garden, because then I might have a real tantrum. 

(2) @derek's "second part [of the question]" isn't really a second part to be answered, it is instead a condition on the first part, that the setback should not compromise comfort. Clearly I could turn the heating off for 23 out of every 24 hours, rather severe to say the least, but it is still a setback, and save a whole lot of money, but it fails on the condition "without compromising comfort". The single question (there aren't two parts) is, to put it another way, 'can a setback that doesn't compromise comfort save money?'.

(3) comfort (ie personal comfort) is subjective, and for that reason I don't think we can put a number on it, like 18 degrees Celsius. Instead, we can leave it as an arbitrary and formally unstated number, and instead say that the actual applied condition, for answering the question, is that the house must over time stay at the same average temperature, whatever that average temperature, chosen by the occupier, is. This is close to, but not identical, to @jamespa's returning to the original state within 24 hours, they are just different ways of testing for the same thing, ie a steady state over time. We now have a simple global test for whether the condition has been met: the average IAT over time must stay constant.

(4) This global test is necessary but not sufficient. It could be satisfied by a running state in which the house is chilly at breakfast time, but does recover by early evening. There needs to be an additional requirement, and this again is a personal choice. For me, it is that the actual IAT should have recovered to within one degree of the desired IAT by 0700. Others can choose whatever works for them. Again, this is a simple test, either the house is or isn't within one degree of the desired IAT at 0700. It is by this simple test that setback without a recovery boost fails in my house for me: simple observation tells me the house is more than one degree below the desired IAT on most (but not all, eg if it is very mild) mornings. To meet the comfort condition, if I have a setback, I must also have a recovery boost.

As noted above, I remain sceptical of the supremacy of modelling over observation (others are perfectly entitled to hold opposing views, we don't need or benefit from having a spurious row about it), and for that reason I am going to continue to collect observational data, and try to find out what it can tell us. In the meantime, I mentioned I had nonetheless attempted to run @kev-m's real world data through @derek-m's model. I did this by reading the (approximate) numbers of @kev-m's charts and putting them into the model. I've probably got this wrong, as the instructions on how to use the model are not exactly clear. What I did is enter the initial starting IAT, the hourly OAT, and then adjusted the heat pump running state (top row, 1 or 0 for on or off) to add the setback. The results, top half with setback, bottom half with no setback, appear to suggest a substantial saving by using a setback, around 12 kWh over a 24 hour period (note I had to convert PI (W/h) to PI (W/m) to match @kev-m's data format). However, the setback and recovery fails my comfort test, the actual IAT is more than one degree below the desired IAT at the end of the modelled period, and I suspect over longer periods the settings would fail the global average IAT must stay the same test, but I am not sure how I can increase the recovery boost to meet the comfort test. All that said, these are the results I got:

I also plotted the model's setback results (lower plot), and compared them to @kev-m's original plot (upper plot):

The right hand side of the model results plot is really rather good, as is the IAT throughout. The left hand side of the energy in (PI) on the other hand appears to be a long way out. A Mk 1 Eyeball averaging out of the cycling in the real world data suggest an average PI somewhere around 10 or maybe 12, while the model results have it running between 15 and a little over 20 W/m. Perhaps @derek-m would be kind enough to explain where I have got it wrong, to get such contrary results.