Optimised district heating connection stations using a self-learning district heating meter - a pilot project

NN is a pure neural network whilst NNPIFF is a neural network combined with other control mechanisms, incl. a room model, a radiator model, a PI controller and two 'feed-forward' controllers. The NN controller can be constructed without detailed knowledge of the installation but requires a long training period. The NNPIFF is more complex but should be easier to train. The two controllers have been trained against a simulated room-model and results compared. Both controllers showed good ability to control energy consumption and to avoid energy peaks. There was no marked difference between the efficiency of the two controllers and therefore the NN controller was chosen for installation, training and test in a concrete installation. Before installation the controller was modified slightly, as it was decided to let the NN controller, which was to work together with the heat meter, share control with a PI controller. The result was called the MAK controller where MAK stands for a learning method, which is Model-free, Adaptive and comfort-Keeping. This now meant that the NN controller would receive its entire learning whilst installed and operative in the installation. However, to start with, when the output of the NN controller can be quite unreliable, the PI controller will carry out most of the control. Gradually, the NN controller will learn from the PI controller and in the course of time take over full control. The advantage of this form of training the NN part of the controller is that learning takes place without affecting the comfort of the users. The installed 'meter' was a computer- based solution comprising a pc, two specially developed energy meters, a motor-operated valve and a number of sensors, etc. Test runs with the MAK controller in the test installation have shown that the PI controller is capable of regulating the meter-valve combination, whilst the NN controller can receive training data 'behind the scenes'. However, the tests have shown that training of the NN controller has not been sufficient for it to take over the control of the meter-valve combination on its own. The initial results from tests with simulated data were very encouraging. However, with regard to the results from field testing the regulator in a test installation it appears that further work is required to explain whether the training time alone has been too short or whether there is the need for tuning of input and output parameters and weighting factors