Process optimization and control of continuous and batch processes with consideration of process/model mismatch

The importance of process optimization and control has been recognized for many years. Recently, nonlinear dynamic process models in conjuction with constrained large-scale optimization methods have been used for on-line optimization and advanced process control.; The primary objective in this thesis was to address the problem of the integration process optimization and advanced process control using recent advances in optimization technology. Current practice in on-line applications is to separate the dynamic control problem and the steady-state optimization problem. This is an artificial suboptimal partition, and requires that the new process reach steady-state before next optimization step is attempted. To address this problem, a generic model optimum control (GMOC) technique is developed by simultaneously considering both economic and control objectives. This technique can coordinate the process optimization and control performances, while compromising the data requirements for model updating and optimization search speed. Simulation studies demonstrated that the GMOC technique can effectively improve the process economics for different kinds of operating situations. Also, universal dynamic matrix control (UDMC) algorithm is re-examined to consider the on-line optimization and feedback control by using a dynamic neural net model. There are two important advantages which this approach offers over conventional UDMC. One advantage is that a dynamic neural net model can be developed from process data and used for optimization calculations, instead of using a first principles model. This neural net model is easily updated on-line. With this technique, the detrimental effects of process-model mismatch can be reduced. The other significant advantage is that our neural net model based UDMC algorithm greatly reduces the computation time required for the nonlinear dynamic matrix used for the Successive Quadratic Programming (SQP) algorithm.; A secondary objective of this thesis is to explore the use of the neural networks learning and adaptive abilities in order to overcome the difficulties in optimizing control of chemical processes. These difficulties arise from modeling uncertainty and changes in the operating environment. A recursive backpropagation neural network (RBPN) algorithm with a forgetting factor is proposed for the on-line modeling of the chemical processes. With the development of this adaptive modeling technique, on-line optimization approach based on gradient information from the neural net model is proposed to update the optimal operating conditions when the system has process/model mismatch. A baker's yeast continuous fermentation process was studied, and the potential of the adaptive modeling and optimization technique was demonstrated. Also, a neural net trajectory learning technique is developed for solving a feedback optimization problem for a fed-batch process. Using the ability of neural networks, feedback optimization can be realized as a nonlinear function of the measurable state variables through an iterative learning process. The advantage of feedback optimization over open-loop optimization is illustrated when system has either model/process mismatch or initial condition error.