@derek-m Dont know, I presume i would have to had lowered ft to 28 which would also increase cycling, was cycling at 2 an hour from 7am to 9pm yesterday so dont want to add more cycling. Im happy with performance as is so im going to finish trying to eke out more savings.

Posted by: @newhouse87

↑

@derek-m Dont know, I presume i would have to had lowered ft to 28 which would also increase cycling, was cycling at 2 an hour from 7am to 9pm yesterday so dont want to add more cycling. Im happy with performance as is so im going to finish trying to eke out more savings.

You are in the very enviable position of having a well insulated home with a large thermal mass, so you can heat up slowly during the warmer daytime period, and allow your home to cool slowly overnight. Unfortunately not everyone is in the same boat, so running their system in a similar manner to you, could possibly lead to a cold home and/or even higher bills. This is why I may question the claims that are being made, so that the full facts are known, and others can then make a balanced judgement as to what may work best for them.

I am very pleased that you have discovered a mode of operation that works best for you.

@cathoderay

I hope that you will forgive me for annoying you, by not paying sufficient attention to your plots and bar charts, but at the time I was still processing the remaining results. I did note that there was a difference between the two methods, which is precisely what I expected, since my objective was to identify the differences so that further investigation may then highlight the reason for such differences. I have been carrying out some of that analysis today and hope to be able to post some results probably tomorrow.

It would be helpful if you could provide details of the underlying formula's used within the raw data file, also how the summing range is set, i.e. when summing 'minute' data into an hourly total, does the summing start with say the 09:00 value and end with the 09:59 value, or the 09:01 value and end with the 10:00 value.

I do hope that we can all work together in a friendly and co-operative manner, and try to tone down the rhetoric to more acceptable levels.

Posted by: @derek-m

↑

I do hope that we can all work together in a friendly and co-operative manner, and try to tone down the rhetoric to more acceptable levels.

I very much agree. I hope now that we have said our bits, we can leave them behind us, and carry on forwards in a friendly and productive manner.

I will do what I can to answer your questions tomorrow. Earlier I bar plotted the hourly figures for the setback periods and there is something not quite right, not sure yet if it is my plotting or in the data. It is also not the easiest of plots to make easy on the eye, too many bars. Will see what I can do tomorrow.   

I'm posting these charts for discussion, not as a conclusion. They have not been easy to produce. I wanted to get all the data on one chart, but it ended up far too busy, so I have divided it into two charts. What they show is the observed energy in (blue line) with a setback, and then the predicted energy in without a setback. I have also added the OAT (ambient temp) as it is the major determinant of energy in. Recall my observed minus expected methodology uses these two variables, and determines the energy saved (or not saved) as the cumulative difference between the two over a 24 hour period. In the upper chart, the expected (predicted) values come from @derek-m's model, in the lower chart, the expected (predicted) values come from my regression line (regression of observed energy in on OAT). In the upper chart, 'the information' (that which determines the values) comes from theory, in the lower, 'the information' comes from observations. All the usual caveats (small sample, one building etc etc) apply, as ever.

First, the things that make sense. The observed use (blue line) makes sense, goes inversely up and down with the changing OAT as expected. You can also see the recovery boosts immediately after the two setbacks (where the blue line drops to zero). Both the expected (predicted, red) lines follow the observed blue line fairly closely in the pre-setback period (R squared values are 0.84 for the model based prediction, 0.91 for the observations based predictions). The model based line is a bit spikier, the observations based line a bit smoother, but that is probably neither here not there.

The differences, and they are marked, occur during the setback days, when the model based predictions (upper charts) fall clearly below the predictions based on the observed data (lower chart). This is how and why the model predicts trivial or no savings, or even worse, extra cost, when using a setback, the (possible/likely) implication being that the heat pump has to play significant catch up during the all the time when it is running. But what if for some reason the model predictions for running without setback are artificially low? The model predictions will then wipe out any actual savings achieved by running with setback.

Equally, of course, it is possible the observations based predictions are artificially high. If that is the case, then any savings will be artificially inflated. The only thing we do know with reasonable certainty is is the actual energy in when running with setback: what we do not know is which (if either) of the predicted series is closest to what would have been the case if running without setback.  

These charts cannot answer these questions on their own, a longer, ideally much longer, run of data is needed. But they do contain some intriguing hints from the setback days. First, the observations based predictions (lower chart) do follow the actual energy in closely when the pump is running, although they miss the recovery boost (which is OK, without a setback, there is no recovery boost). On the other hand, the model based predictions (upper chart) do seem low, all the more so whenever comparisons can be (very roughly) made between similar OATs. For example, in the period before the first setback, the OAT stays around 11 degrees, and the energy in is pretty stable at around 1 kWh per hour (and both the model and observations based predictions agree). However, in the period between the setbacks, the OAT drops to around 10 degrees, but the model predictions remain at or even slightly below the 11 degree OAT predictions, while the observations based predictions increase slightly (as expected). Even more curious, in the period after the second setback, the model predictions fall even lower (to their lowest point), despite the fact the average OAT during that period being around 11 degrees - as it was in the period just before the first setback.

As I said, and now repeat, this very short run on one system cannot provide definitive answers, but, at least for me, it does raise interesting questions.    

Posted by: @derek-m

↑

It would be helpful if you could provide details of the underlying formula's used within the raw data file, also how the summing range is set, i.e. when summing 'minute' data into an hourly total, does the summing start with say the 09:00 value and end with the 09:59 value, or the 09:01 value and end with the 10:00 value.

All the calculations (formulas) are in the python scripts, the raw data file is a csv file and so can't contain formulas. There are two scripts, the minute data script and the hour data script. The minute data script just gets the data once a minute, and apart from some minor formatting/extraction/scaling (eg flow rate sent as an integer say 143 needs division by 100 to get the actual value of 1.43), does no processing per se at all on the data, instead, it just dumps it to the minute data file. The only exception to this is the allocation of the energy in and out to space heating and DHW heating, which is done using the position of the three port two way valve, using conditional logic eg if valve is in space heating position, allocate to space heating etc. 

The calculations are done by the hour data script. This script starts by getting the last 60 rows from the minute data file. The script runs at one minute past the hour, every hour, and I guess, but can't know, because I never actually see those 60 rows, that it gets from one minute past the previous hour to on the hour for the current hour eg the 10:01 run will get 09:01 to 10:00. There rows end up in a pandas data frame (in effect a virtual spreadsheet), with rows being the minutes and the columns being the variables. The calculations are then done on the data frame.

The actual calculations use means, sums and subtractions. For example, the hourly ambient is the mean of the ambient column in the data frame. The calculated energy in and out use both sums and means. The calc_kWh_in (all energy in), for example, uses the means of the amps and volts columns. I did various tests eg means vs calculating for each minute and summing, and even did some fancy AUC (area under curve) calculations, and they all gave pretty much the same answer, and used the mean based method as it is simplest, ditto for the calculated energy out. The space/DHW specific energy in/out values need a more complex approach to achieve the correct allocation, as the change from space to DHW happens within the hour, meaning an hours average will lose when the changeovers happen. To get these allocated energy in/out values, I calculate the energy in/out for each minute, and then allocate the value to new columns, depending on whether the system is in space or DHW heating mode. The actual allocation is done by multiplying by the htg_on_off and dhw_on_off values, which either 1 (that mode is on) or 0 (that mode is off). Multiplying by 1 means all that value goes into that new column, multiplying by 0 means none of that value goes into that new column. Probably easier to show the actual code:

df['htg_kWh_in'] = ((df['amps_in'] * df['volts_in'] * (1/60) * df['htg_on_off'])/1000) * dcf
df['dhw_kWh_in'] = ((df['amps_in'] * df['volts_in'] * (1/60) * df['dhw_on_off'])/1000) * dcf

The various df['name'] entries are the columns (df = the name of the data frame, followed by the column name). Thus, the new df['htg_kWh_in'] column gets the amps x volts divided by 60 ( * (1/60), hours to minutes) multiplied by the binary htg_on_off (either keeps it as it is or makes it zero), then divide by 1000 to get kWh, and finally apply the data correction factor (dcf, the 1.18 correction factor). 

I now have two new columns with energy values for each minute allocated according to operating mode (space or DHW) which are then summed to get the totals for each column for that hour. Usefully, this, by being a different calculation to the combined energy in/out calculations, provides a simple way of  checking that each is correct (ie they should be very close, as they are). It also demonstrates that using simple means and summed minute values get very similar results.

The midea_kWh_in and out use the Midea total lifetime kWh values, and get the hour's value by subtracting the first value in the data fame from the last (done using min and max, as these will always locate the first and last value).

Finally (for the hour data), the various COP values are calculated in the usual way, energy out divided by energy in, giving the trailing 60 minute COP for the last hour, and then the whole lot are written to the hour data file.

Having done that, it then goes on to do something very similar on the hour data file, to get 24 hour values. However, I don't think I have ever published those 24 hour values, so no need to go into the detail here, which in any event is very similar to the way the minute data is processed to get the hour data: start by getting the last 24 rows from the hour data file, start calculating etc etc.  

Hope that makes some sort of sense, and answers your questions!     

@cathoderay

Thank you for your time and effort in producing the graphs, I always find graphs highly useful in compressing lots of data points into a format that is much more understandable to our human brain.

Graphs allow us to view the 'big picture', but may not be fully beneficial when it is necessary to look at the detail.

I am a little confused though, as the upper graph shows the predicted (red) line staying above zero during the two setback periods, whereas the results tables I provided clearly show the Power In (PI) value is zero during both 6 hour setback periods. Could you please clarify?

Posted by: @cathoderay

↑

I'm posting these charts for discussion, not as a conclusion. They have not been easy to produce. I wanted to get all the data on one chart, but it ended up far too busy, so I have divided it into two charts. What they show is the observed energy in (blue line) with a setback, and then the predicted energy in without a setback. I have also added the OAT (ambient temp) as it is the major determinant of energy in. Recall my observed minus expected methodology uses these two variables, and determines the energy saved (or not saved) as the cumulative difference between the two over a 24 hour period. In the upper chart, the expected (predicted) values come from @derek-m's model, in the lower chart, the expected (predicted) values come from my regression line (regression of observed energy in on OAT). In the upper chart, 'the information' (that which determines the values) comes from theory, in the lower, 'the information' comes from observations. All the usual caveats (small sample, one building etc etc) apply, as ever.

-- Attachment is not available --

 

First, the things that make sense. The observed use (blue line) makes sense, goes inversely up and down with the changing OAT as expected. You can also see the recovery boosts immediately after the two setbacks (where the blue line drops to zero). Both the expected (predicted, red) lines follow the observed blue line fairly closely in the pre-setback period (R squared values are 0.84 for the model based prediction, 0.91 for the observations based predictions). The model based line is a bit spikier, the observations based line a bit smoother, but that is probably neither here not there.

The differences, and they are marked, occur during the setback days, when the model based predictions (upper charts) fall clearly below the predictions based on the observed data (lower chart). This is how and why the model predicts trivial or no savings, or even worse, extra cost, when using a setback, the (possible/likely) implication being that the heat pump has to play significant catch up during the all the time when it is running. But what if for some reason the model predictions for running without setback are artificially low? The model predictions will then wipe out any actual savings achieved by running with setback.

Equally, of course, it is possible the observations based predictions are artificially high. If that is the case, then any savings will be artificially inflated. The only thing we do know with reasonable certainty is is the actual energy in when running with setback: what we do not know is which (if either) of the predicted series is closest to what would have been the case if running without setback.  

These charts cannot answer these questions on their own, a longer, ideally much longer, run of data is needed. But they do contain some intriguing hints from the setback days. First, the observations based predictions (lower chart) do follow the actual energy in closely when the pump is running, although they miss the recovery boost (which is OK, without a setback, there is no recovery boost). On the other hand, the model based predictions (upper chart) do seem low, all the more so whenever comparisons can be (very roughly) made between similar OATs. For example, in the period before the first setback, the OAT stays around 11 degrees, and the energy in is pretty stable at around 1 kWh per hour (and both the model and observations based predictions agree). However, in the period between the setbacks, the OAT drops to around 10 degrees, but the model predictions remain at or even slightly below the 11 degree OAT predictions, while the observations based predictions increase slightly (as expected). Even more curious, in the period after the second setback, the model predictions fall even lower (to their lowest point), despite the fact the average OAT during that period being around 11 degrees - as it was in the period just before the first setback.

As I said, and now repeat, this very short run on one system cannot provide definitive answers, but, at least for me, it does raise interesting questions.    

Those are useful graphs which nicely illustrate the difference between the two predictive models.  To add to the above I observe two other other features in the graph/data

a) the rise in OAT at the beginning of setback.  I think we know that this is an artefact of the location of the sensor however its important continually to bear in mind what I am sure we all know, namely that, while the heat pump would be expected to respond to the 'OAT' as measured (ie the sensor temperature), the fabric of the house will respond to the actual OAT.  The most likely way this would manifest is in a divergence between predicted and actual IAT, however that would depend on the assumptions particularly in @derek-m s model.

b) (from the data) the reduction in IAT during the setback days relative to the pre-setback days.  During the pre-setback days the IAT is typically 19.7C.  During the setback days the IAT appears to peak (shortly before the next setback) at 18.9-19.1.  Depending on how various calculations have been done this could also be significant.  0.7C doesnt sound like much but with a heat capacity of ~13kWh/C it could account for a ~10kWh discrepancy somewhere.

It would be useful to plot IAT (measured and predicted) on the graph as this might give a further clue, but I appreciate its a lot of work.

@derek-m - thanks, and I agree, images can convey a lot of information that our human brains can take in very easily.

The two charts compare the actual observed energy in without (left hand side) and with (right hand side) setback, shown by the blue lines, with the predicted energy in without the setback, shown by the red lines, the upper chart using predictions from your model (the no set back tables), the lower chart using my regression line equation from the OAT vs energy in regression. I've done this because my observed vs expected method for determining savings (if any) during setbacks needs both the actual (observed) energy in during the setback period and the predicted (expected) energy in had there not been a setback. The saving (if any) is the difference between the two values summed over 24 hour periods, and, clearly, the predicted values need to be as close as possible to what the actual values would have been, had there been no setback.

As it is now lunchtime on Christmas Eve, I for one am going to put all this on one side for now, and attend to the a thousand and one things that need doing right now. I'll be back before too long... In the meantime, I wish you all a very Merry Festive and above all Warm Christmas!  

@jamespa - thanks. We cross posted - as I said, I'll be back on the case before too long!

Posted by: @cathoderay

↑

@derek-m - thanks, and I agree, images can convey a lot of information that our human brains can take in very easily.

The two charts compare the actual observed energy in without (left hand side) and with (right hand side) setback, shown by the blue lines, with the predicted energy in without the setback, shown by the red lines, the upper chart using predictions from your model (the no set back tables), the lower chart using my regression line equation from the OAT vs energy in regression. I've done this because my observed vs expected method for determining savings (if any) during setbacks needs both the actual (observed) energy in during the setback period and the predicted (expected) energy in had there not been a setback. The saving (if any) is the difference between the two values summed over 24 hour periods, and, clearly, the predicted values need to be as close as possible to what the actual values would have been, had there been no setback.

As it is now lunchtime on Christmas Eve, I for one am going to put all this on one side for now, and attend to the a thousand and one things that need doing right now. I'll be back before too long... In the meantime, I wish you all a very Merry Festive and above all Warm Christmas!  

May I wish all forum members a Merry Christmas and a Happy & Warm New Year.

So it would appear that in graphs you are not actually comparing like for like, since the 'with setback' predictions show PI at zero during the setback period, but the PI values are then higher during the operating period to make up for the reduction in thermal energy supplied. For clarity, perhaps the setback values should also be added to the graphs.

Couldn't resist one final peek...and this really is the last one, just to, I hope, clear this up:

Posted by: @derek-m

↑

So it would appear that in graphs you are not actually comparing like for like, since the 'with setback' predictions show PI at zero during the setback period, but the PI values are then higher during the operating period to make up for the reduction in thermal energy supplied. For clarity, perhaps the setback values should also be added to the graphs.

Exactly, they compare observed with setback against expected without setback, which is exactly what I want for the observed vs expected comparison. 

I did have all the data, including your and my with setback predictions, on one chart but it was just too busy - too much information, and some bits obscuring others. I will see what I can come up with once I return to the thread after the break.