Implementation of a commercial PI-neural controller for building services controls

In heating, ventilation and air-conditioning (HVAC) applications a large variety of components must be controlled to maintain comfort conditions. Current building controls are often implemented in stand-alone multitasking microprocessor based 'outstations' in which up to around 30 control loops may be active simultaneously. The majority are PI loops, these are used to control such components as cooling coils, heating coils and dampers. The non-linear nature of the controlled systems lead to commissioned controllers with typically sluggish action. This paper discusses an implementation of a hybrid PI-neural controller specifically designed to combat this problem. To ensure commercial application it is necessary that such a controller is no more complex to commission than a PI controller, and that it is robust enough to operate unsupervised for periods of years. Issues discussed include tolerance to faults and maintenance periods, computational cost and auto-configuration of neural network parameters. Results are presented from validation testing performed in simulation and on a full scale air-conditioning test facility.