Adaptive neural control with stable learning

The paper considers the problem of training on-line a neural network model of non-linear heater battery for implementation in a model-based control scheme. A stable learning scheme is proposed which reduces parameter drift due to process-model mismatch in radial basis function (RBF) networks. A network of pre-defined structure is trained and shown to exhibit finite model mismatch errors, which can produce parameter drift or ''de-training'' of the network, resulting in inferior control performance. A deadzone approach, similar to ones used in linear dynamic system identification, is applied to RBF network adaption, successfully reducing the degree of ''de-training''. The learning scheme is used in a neural controller which is capable of compensating for plant non-linearities, and adapting on-line to degradation in the plant. Experimental results are presented which have been obtained from a flow-controlled heating coil, on a full-size air conditioning plant at the UK Building Research Establishment.