That is a very good point, and that is also why it can't work with the slow flow meter that is used. To work, the flow must be correctly measured before the heat buffer in the boiler runs out, and the smallar the boiler the harder it will indeed be.
You say that the temperature will fall as soon as the flow increases, that is correct, but it does not tell the whole picture. It will *start falling* as soon as the flow increases. How quickly it falls depens on the size of the heat buffer in the boiler compared to the output power. Normally the controller uses this falling temperature as it's only reference for balancing input power with output power, ie. it knows absolutely nothing about the process.
If however the increased flow can be measured before the temperature has fallen substantially, then an almost perfect estimate of the required power input can be made without waiting for the feedback loop to react at all. If the temperature falls so fast (small boiler) that the power output becomes limited by the input power, the estimate for required increase of input power will be too low and the boiler temperature will still fall, leaving more work for the PI feedback loop. The worst case will be that the forward component is zero and all work is left to be done by the PI feedback loop, ie. it will work exactly as it does now.
My point is, whenever an estimate for forwad feed can be made that is better than nothing, it will improve control. The more that is known about the process, the better the estimate can be. For instance the feedforwad algorithm can be made to compensate for the fact that an increase in measured output power is soon followed by a decline if input power can't be increased fast enough but not the other way, since that is a known quality of the process.
It is however possible that the forward feed estimate is worse than nothing, for example if it comes too late. And that is why it can't work with a slow flow meter, I won't try to guess how fast it would need to be with the small blackstar boiler, but I fail to see any theoretical reason why it couldn't work given fast enough reaction time. It's completely possible that a fast enough flow meter simply doesn't exist.
(this is obviously no longer a "Ønsker til funktioner i chip" subject, more of a theoretical discussion on process control in general, please regard as such)