Neural network based adaptive control of a non-linear system: application to a thermal process

Multilayered neural networks are used to model a variable capacity heat pump. Two approaches are investigated. The first is modular and allows the representation of the heat pump using four sub-models: compressor, evaporator, condenser and the expansion valve. The second gives an overall representation of the heat pump as a function of its inputs and outputs only. Both models account for the dynamics of the heat pump. For each of the two approaches the neural network models are used as either emulators for testing and developing controllers or as online one-step predictors. Measurements on an air/water heat pump prototype were run and used to train the various neural networks used. The networks obtained are then compared with the process. Extensive comparison results are presented