Neural network-based control designs for complex industrial process applications.

摘要

Neural networks have successfully transitioned from an academic interest into a viable technology which is now being used in everyday products. To date, neural networks have been predominantly applied to forecasting or modeling applications. Based on their success in such applications, there has been significant interest in using neural net works in control applications, creating a new field called neurocontrol. Although there have been significant advances in the theory of neurocontrol, there are very few successful commercial applications using neurocontrollers. Commercial applications often provide the most challenging problems because the controllers are required to function robustly in complex and unknown environments. Real-world processes are complex and difficult to control because they contain a large number of highly interdependent variables, have highly nonlinear responses to these variables, and change their response over time.; This work identified two significant reasons why neurocontrol designs fail in real-world applications: First, the controllable parameters over most industrial processes are highly correlated, often not for physical reasons but because of our process control strategies. Second, intermediate process states that affect the process output, which are also affected by the controllable parameters, have a significant impact on controller performance. When the controller changes the controllable parameters, the impact that this has on the process states, which will in turn affect the process output, is not accounted for in most neurocontrol designs in the literature.; This dissertation advances the field of neurocontrol by providing the following solutions: first, the use of statistical significance testing on the local linearized relationships extracted from nonlinear neural network models to avoid problems with correlated controllable parameters; second, augmenting neurocontrol designs to incorporate dependent state models. These enhancements have been applied to four distinct neurocontrol architectures. The new control architectures have been applied to the novel application of controlling NOx emission from an oil and gas-fired electric power plant.