Neuro-fuzzy predictive control of an information-poor system
While modern engineering systems have become increasingly integrated and complex over the years, interest in the application of control techniques which specifically attempt to formulate and solve the control problem in its inherently uncertain environment has been moderate, at best. More specifically, although many control schemes targeted at Heating, Ventilating and Air-Conditioning (HVAC) systems have been reported in the literature, most seem to rely on conventional techniques which assume that a detailed, precise model of the HVAC plant exists, and that the control objectives of the controller are clearly defined. Experience with HVAC systems shows that these assumptions are not always justifiable, and that, in practice, these systems are usually characterized by a lack of detailed design data and a lack of a robust understanding of the processes involved. Motivated by the need to more efficiently control complex, uncertain systems, this thesis focuses on the development and evaluation of a new neuro-fuzzy model-based predictive control scheme, where certain variables used in the optimization remain in the fuzzy domain. The method requires no training data from the actual plant under consideration, since detailed knowledge of the plant is unavailable. Results of the application of the control scheme to the control of thermal comfort in a simulated zone and to the control of the supply air temperature of an air-handling unit in the laboratory are presented. It is concluded that precious resources (as measured by actuator activity, for example) need not be wasted when controlling these systems. In addition, it is also shown that a very precise (and sometimes not necessarily accurate) control value computed at each sample is unnecessary. Rather, by defining the system and its environment in the fuzzy domain, the fuzzy decision algorithms developed here may be employed to get an "acceptable" control performance.