Modeling And Optimization Approaches For Nonlinear Model Predictive Control Of Chemical Processes

Motivated by the current increased standards in higher quality of product, stricterenvironment regulations and tighter safety requirements, analysis and synthesis of nonlinearmodels based model predictive control schemes has beening gained more and more attentionNonlinear model predictive control (NMPC) possesses a strong potential for improving thecontrol performance through the improvement of the model accuracy. However, it hasrecognized that the practical application of NMPC is limited by the difficulty in nonlineardynamic model development as well as the computational burden associated with the solutionof the on-line optimization problem. Under this situation, this dissertation focuses on the themodeling and optimization strategies of nonlinear model predictive control for chemicalprocesses, and the following contributions are obtained.One of the critical open issues in the NMPC scheme is the computational burdenassociated with the solution of the optimization problem, since at each sampling time anonlinear dynamic optimization problem must be solved in real-time. To alleviate theaforementioned problem, a new NMPC algorithm using the global orthogonal collocationmethod is proposed. Higher order interpolation polynomial is used to simultaneously discretestate variables and control variables over the optimization horizon, and therefore the originalcontinuous dynamic optimization is transcribed to a nonlinear programming problem (NLP).The NLP problem has a fixed structure with certain computational advantages and can besolved by an appropriate numerical optimization algorithm. Taking full advantage of thefeatures of the global orthogonal collocation, the proposed algorithm provides the potential toreduce the scale of NLP and thereby reduce computational burden efficiently, even it workswith a long optimization horizon. The effectiveness of the proposed algorithm isdemonstrated by its application to a continuous polymerization process. It is found thealgorithm achieves a smooth transition for large-magnitude setpoint changes and behaves wellin the presence of disturbances.The optimal solution of a dynamic optimization problem generally consists of different kinds of arcs, control variables are continuous and differentiable within each interval, but canjump from one interval to the next at the so-called switching times. This inherentdiscontinuous nature of optimal control profiles may pose problems to numerical solutionmethods. To solve this kind of problems, a partitioning and simultaneous strategy for dynamicoptimization problem based on the global orthogonal collocation method is proposed. Thoughpartitioning, the ill-condition problem caused by the discontinuity of optimal control profilecould be avoid. In addition, the optimization problems obtained after the partition would besolve in a simultaneous way, the final solution would satisfy the linkage constraints.Simulation study on the dynamic optimization of two batch process shows the feasibility andvalidity of the proposed strategy.A very important issue for the implementation of advanced control scheme is thedevelopment of a proper process model. Sometimes, data-driven models are considered as agood alternative to the detailed first principles model. The performance of data-driven modelsrelies heavily on the amount and the quality of the training samples and hence it mightdeteriorate significantly in the regions where samples are scarce or nonexistent. The objectiveof this paper is to develop an on-line support vector regression (SVR) model updating strategyto track the changes in the process characteristics efficiently with affordable computationalburden. This is achieved by adding new sample that violates the KKT conditions of theexisting SVR model and deleting old sample that has the maximum distance with respect tothe newly added sample in feature space. The benefits offered by such an updating strategyare further exploited to develop an adaptive model-based control scheme, where the modelupdating task and the control task perform alternately. The effectiveness of the adaptivecontroller is demonstrated by conducting simulation studies on an exothermic continuousstirred tank reactor. The results reveal that the adaptive MPC scheme outperforms itsnon-adaptive counterpart, for both large-magnitude setpoint changes and variations in processparameters.A methodology that incorporates the polynomial ARX model with an output errorstructure into NMPC framework is presented, which retains many favorable characteristics ofconventional linear MPC while extending its capabilities to nonlinear systems. For MPCdesign purpose, the polynomial ARX model is developed based on the output error method,with emphasizing its long range prediction capability. Regarding the computationalcomplexities associated with the control law calculation, the linear parameter varying (LPV)state space model interpretation of the polynomial ARX model is exploited and the onlineoptimization problem is formulated as a standard quadratic programming problem at each time instant, thus avoiding the time consuming numerical search procedures and theuncertainty in convergence to the global optimum which is typically seen in conventionalNMPC strategies. The proposed NMPC scheme is applied to a highly nonlinear pHneutralization process over a wide range of operating conditions, with the comparison to otherMPC implementations.The control of polymerization processes during grade transition imposes greatdifficulties due to the lack of on-line measurements of polymer properties and the rapidlychanging product requirements. Under this situation, our motivation is to present a specificcontrol strategy that could achieve optimal transition control for an industrial scale propylenepolymerization process with dual-loop reactor. To this end, inferential models for polymermelt index (MI), which can trace the changeovers of the product, is derived by incorporatingthe specific characteristics of the process with the just-in-time modeling approach. Theoptimal grade transition control problem is reformulated as a dynamic optimization problemwith state path constraints and is solved by the global orthogonal collocation method. Thesimulation results confirm the effectiveness of the method presented here.