Modeling and control of a continuous crystallization process using neural networks and model predictive control

Continuous crystallizers are distributed dynamical systems. Physical modeling of these systems using basic principles results in partial and integro-differential equations. To exploit the physical models, in the analysis of the system behavior and the design of an appropriate controller, requires complicated measurement techniques especially in the spatial domain (crystal size distribution or crystal population density). Therefore, obtaining a lumped model structure is desirable. The lumped model of a continuous crystallizer can be obtained either from the physical model, using conventional techniques such as the discretization or function separation methods, or from input and output measurements using system identification approaches. Studies of the crystallization process have indicated that in order to improve the control performance, expressing the process dynamics using single-input, single-output models is insufficient. The aim of this thesis was to investigate the process behavior in a multivariable framework. In this regard, the dynamics of a continuous cooling KCl crystallizer were identified using three-input, three-output linear and nonlinear model structures. The autoregressive exogenous model structures were employed in linear modeling of the process. The nonlinear modeling was performed using several architectures of feedforward and recurrent neural networks. Simulation results demonstrated that the linear modeling, using a single model for the entire dynamics, is not adequate. Either multi-model or nonlinear modeling is recommended. The performance of different neural network structures in the nonlinear modeling of the process was illustrated and, based on the results, some comparisons were made between these networks. The next step in the study of the crystallization process as a multivariable system was to design and apply a multivariable control scheme. Simulation results from the modeling of the process indicated that strong interactions are present among different loops of the system. The process is nonlinear and some of the outputs exhibit inverse or non-minimum-phase responses. The model predictive control strategy is known to perform well in the control of the systems with the behaviors found in the crystallization process. To ensure a feasible solution, the feasible sequential quadratic optimization algorithm was successfully exploited in a model predictive controller. Computer simulations of the controller were performed in order to demonstrate control of the crystal size distribution, crystal purity, and production rate. The effects of different control parameters were illustrated using the simulation results. A brief discussion on how to select these parameters was also provided. Robustness of the model predictive controller was studied in the presence of mismatch between the model and the process