Neural network control of a non-linear heater battery

Abstract

The unsatisfactory performance of conventional building services controls is frequently due to poor commissioning, and to the inability of conventional controllers to deal with non-linearities and to adapt to long-term changes in the behaviour of the plant. The paper describes a hybrid neural control scheme which is capable of compensating for plant static non-linearities and of adapting on-line to degradation in the plant, but avoids the instability problems that can arise when neural networks are introduced into the feedback control loop. The hybrid controller uses a neural network, which learns the non-linear static characteristics of the plant, to generate feed-forward control action, and a conventional proportional controller, acting as a feedback trimmer, to deal with unmeasured disturbances. Results of a detailed computer simulation of a heater battery are presented; these show that the hybrid scheme can produce more consistent control, and is less sensitive than a conventional PI algorithm to initial tuning and to variations in the temperature of the water supplied by the boilers.