| The neural network paradigm offers specific advantages as a methodology for providing intelligent control of complex, nonlinear plants. The ability of neural nets to approximate nonlinear functions, adapt to changing conditions, and rapidly compute input/output mappings, makes them ideally suited for handling the algorithmic complexity of real time, multidegree of freedom control problems. This dissertation proposes a neurocontrol methodology for optimal temperature control of a thermal plant with multiple chambers and will show that this relatively untouched area is replete with possibilities for such an approach. Thus the emphasis here is multidisciplinary and spans three areas: refrigeration, control, and neural networks.; Commercial household refrigerators, typical of those making up the bulk of the domestic appliance market today, use a simple cost-effective temperature control approach that provides chamber-dependent temperature control with minimized energy consumption at only a single control point. Technological advances in control hardware, including airflow control devices such as automatic thermal dampers, variable speed fan motors and compressors, and electronic expansion devices, can add control degrees of freedom to achieve optimal control with respect to temperature, humidity, and noise control and energy consumption at all control points. Control of such a plant, with increased control component functionality, presents a complex problem requiring new control algorithms to mesh hardware and control logic synergistically.; A control methodology is developed which uses the generalized learning approach for mapping, the plant's inverse dynamics to desired control signals using several control models. Two unconventional control strategies are examined: variable temperature bandwidths, and uncoupled compressor and evaporator fan operation. A plant model, representing the behavior of a conventional, dual chamber, top mount style refrigerator, was used to generate results for both strategies in combination with manual and automatic thermal damper configurations. The neural net was trained using plant outputs from various combinations of these plant control configurations and strategies. An optimal control model was defined and its neural net implementation could serve as a springboard for providing more unified control of all plant functions. The results of this research testify to this and show that neural networks, and other nontraditional paradigms, will have interesting implications for the future performance and marketability of the plant. |
