Posted by: @editor

↑

@cathoderay, interesting. Most installers have also said it’s the outdoor sensor that’s only needed for weather compensation to work. Makes sense. But what doesn’t make sense is that if you have a super insulated house and it’s cold outside, the weather compensation instructs the heat pump to increase the flow temperature and just run indefinitely continuously heating the inside space - what happens when your target temperature is exceeded? Surely the heat pump should then stop, so the internal sensor should play a role too, otherwise the temperature will just keep rising.

But there’s nothing simple about the simplicity of weather compensation.

Hi Mars,

The indoor air temperature will not just keep increasing, because as the indoor air temperature increases, so does the heat loss, to the point where the heat energy supply and the heat energy demand balance.

If the system is working within its operating scope, and can actually supply the heat loss as it varies, and the WC curve is correctly adjusted, then the indoor air temperature should remain reasonably constant.

The problem in Cathoderay's case is that the heat pump could not supply the heat demand during the cold spell, so the indoor air temperature fell below the desired level. His heat pump is now playing 'catch-up' because it failed to perform adequately previously. If his controller has the temperature offset facility when in WC mode, then he can easily bump up the LWT until the desired indoor air temperature is achieved.

It is a pity that heat pump controllers don't use the industry standard 4 to 20 mA signals widely used within industry, since it would then be quite a simple matter to install a temperature sensor, with associated transmitter, that is ranged to provide an adjustable means of controlling the heating system based upon both the outdoor and indoor temperatures.

Another method would be to vary the response of any indoor temperature sensor by using additional resistances.

@editor and @derek-m - there is partial indirect feedback about the indoor temp, in the RWT. As the house warms, the RWT gets warmer, and that's what tells the heat pump to slow down, at least in theory. Apart from that, there is no indoor temp sensor connected to the heat pump, and the room stat (an on/off switch) is set to a high temp, 26 degrees during the heating season, and so is always on.

I'm pretty sure the weather comp curve is very basic in Midea units, with no tweaks available apart from manually tweaking the curve when catch up after a cold spell is needed. The fundamental problem in such a period is the heat pump controls are too simplistic. The heat pump reads 10 degrees ambient, and puts out heat to match heat loss at 10 degrees ambient, but the house needs more than that, because it is colder than it should be. The end result is the  catch up does happen, but only very slowly, over a period of days. During the cold spell, my kitchen was at 15.5 to 16 degrees (ie 3 to 3.5 degrees below design). As the weather started getting milder, the temp has crept up: 16 degrees on Sunday morning, 17 degrees yesterday morning, 18 degrees this morning. 

Posted by: @cathoderay

↑

The heat pump reads 10 degrees ambient, and puts out heat to match heat loss at 10 degrees ambient, but the house needs more than that, because it is colder than it should be.

Ah, but it also "reads" the return water temperature is lower (say, RWT = LWT - 8), because the house has taken more heat out of it because the house is colder than it should be, so the pump is  putting out more heat to raise the water 8 degrees than if the house was already at temperature and RWT = LWT - 5 so it would only need to heat the water by 5 degrees.

It feels like what you're asking for is the pump to go into some sort of panic-reheating mode if the house is too far below the target temperature, but you can do that already manually (for example, switch to a fixed flow temperature of 45-50 degree water) and how would the control know when to do it? It could get expensive if used needlessly, such as while you're out for the day and a slow reheat to target would be fine.

@mjr - I agree, and said earlier, it does get indirect feedback via the RWT, and yes, I agree, a bigger difference all other things being equal between the LWT and RWT does suggest the house is cooler. What I am not sure about is why the heat pump allows that bigger difference, and doesn't try to reduce it to say 5 degrees - perhaps because the only way to reduce gap is to reduce the LWT, and that is (a) counter-productive and (b) the weather curve doesn't 'allow' it to happen? 

I'm not asking for panic mode, more like fine-tuning. I can do it manually, don't need to go into fixed mode, just up the weather comp curve a bit, and as you say, that means it won't happen when it doesn't need to, but it is a bit of a faff, security codes and endless button taps, and then I also have to remember to put the curve back to its normal position. My problem is that with the standard curve and my setup, reheat speed is very slow, about one degree per day, hence the request to be able to mitigate this. I suppose I should just say I won't have to do it very often, just accept I have to do it manually when I need to, with a post it note on the wall to remind me to turn it down again once the temps have recovered. 

Personally I think we can get a lot more mileage out of human feedback mechanisms in addition to sensors. For example, when my wife is cold, she turns the thermostat up. Right now that does literally nothing, because I've already set it to be 2 degrees higher than the desired room temp like a good little boy (it's a dumb On/Off thermostat). But her turning the thermostat up is, in principle, a useful piece of feedback as it signals that the people in the house are a bit on the cold side. This might be in spite of the actual measured room temperatures being at the right level, due perhaps to less solar gain, it being a bit windy so more draughty, etc. Now, what I do when my wife tells me she's cold is turn the flow temperature up on the FTC (a simple positive offset on the Mitsubishi Ecodan FTC5). But there's no reason that the FTC couldn't do this itself, in response to the thermostat being turned up a notch (other than that they don't talk to each other).

The difference between this and Room Adaptation for example is that the objective function of Room Adaptation is "set the flow temperature such that the room temperature is 22 degrees C". However, the objective function of the AI/ML model would be "set the flow temperature such that people stop fiddling with the thermostat".

Now it's obvious that this can't be the only part of the objective function - what happens at night or when we're out of the house, for example? - but it would be straightforward to include "nobody fiddles with the thermostat" within your machine learning model. You could bake in what the TRV is set to and how often people fiddle with that too. And you could make a smart thermostat (or one in each room) that encouraged users to feed back data to the ML model, e.g. by allowing the user to tell it easily whether they're "too hot", "too cold" or "just right".

A more sophisticated objective function would be "set the flow temperature such that your humans feel comfortable", and the AI would try and use sensors alongside human feedback to anticipate what flow temperatures are required for the resident humans to feel comfortable. Perhaps this would work by taking historic "too hot/too cold/just right" feedback and regressing that against all the sensor data (and forecast data as mentioned by others), creating a regression model that relates the sensor data (including things like cost, time of day/night, whether there are any humans home, calendar/diary data, maybe even phone sensors, smart watch data, etc.) to the likelihood that the humans will have bashed the "just right" button the most and the "too hot/too cold" buttons the least.

There are obvious privacy concerns here, which is why, in my view, this should all be done locally and not in the cloud, with no personal data being sent to cloud servers. Of course, there is no way any company would be able to (a) train the model and (b) do this profitably, without the data being sent to their servers, so it either won't happen on a large scale or I won't want to use it because of the privacy concerns. But I do think there's room for technically-minded enthusiasts to step into this space and create something that works for them personally. The "data collection" part will be the hardest, but projects like Home Assistant are able to do a lot of the heavy lifting there, so I have hope.

Regarding the conversation about rules-based or heuristic models vs actual "learning" (i.e. via regression), in most AI systems you need both of them for the system to operate effectively. And I wouldn't consider rules-based models to be "inferior" or not "smart" - often they work better than the regression-based approaches, especially where the goal is human-oriented.

This thread is a real testament to how behind the heating industry is in this country. There is little that can be done to improve existing systems already designed and working in other countries.

Really aññ we need to do is understand what is already available but poorly promoted and appreciated in the UK

Its all been done I’m afraid, it’s available too!

Posted by: @scrchngwsl

↑

A more sophisticated objective function would be "set the flow temperature such that your humans feel comfortable", and the AI would try and use sensors alongside human feedback to anticipate what flow temperatures are required for the resident humans to feel comfortable. Perhaps this would work by taking historic "too hot/too cold/just right" feedback and regressing that against all the sensor data (and forecast data as mentioned by others), creating a regression model that relates the sensor data (including things like cost, time of day/night, whether there are any humans home, calendar/diary data, maybe even phone sensors, smart watch data, etc.) to the likelihood that the humans will have bashed the "just right" button the most and the "too hot/too cold" buttons the least.

Thanks @scrchngwsl that and the rest of your post is a very clear description of how a 'learning' system might work. I'm not sure why I put learning in quotes, maybe it's because I am a human (or at least I think I am a human)...