Optimal neural network-based controller for ice storage systems

This dissertation demonstrates that a self-learning neural network (NN) controller can be used to optimally control a thermal energy system for least cost. By freezing the thermal storage, namely water, during low cost hours and then melting it to cool buildings during high cost hours, electrical expenses can be significantly lowered. The optimal control of ice storage systems presents a difficult problem that relies on a series of complex interactions between components and a large combination of control setpoints.; The two primary pieces of equipment requiring accurate models are the chiller and ice storage tank, the devices where the major energy transfers occur. Self-calibrating, non-linear neural network equipment models were developed using plant-operating data. The equipment models are different from other models in that they do not require extensive effort and detailed information to develop, and can be easily updated.; A neural network-based optimal controller, consisting of a training network and a prediction network, was developed to operate equipment for least cost. The training network collects operational data and compares it to the NN predictions based on the same operational conditions. If predicted and actual performance varies by a specified amount, equipment models are retrained. The predictor network uses information from the training network and cascades the information forward in time, adjusting setpoints until a least cost solution is reached.; Results demonstrate that a neural network-based controller is capable of regulating an ice storage system for least cost given any price structure. The NN controller learns how equipment components will react to control settings and determines the settings that optimally operate equipment. Since the controller self-learns the process, it does not require experienced personnel to implement.; Results were verified through computer simulation and then by operating an actual ice storage system. The laboratory verification test was conducted by operating the Larson Laboratory using the neural network controller under two situations. The first situation tested the controller's ability to adapt to equipment as it underwent change and a second situation tested the simulated start-up of a new TES system.