| Scope and method of study. The purpose of this study was to examine use of improved models for model predictive control (MPC) and inferential measurement. The study encompasses three areas where improved models were explored: (1) in adaptive linear MPC where parameters of first-order-plus-time-delay models were identified online and used for quadratic dynamic matrix control, (2) in neural network based inferential modeling where a systematic procedure was proposed that addresses data collection, variable selection and training of multilayer perceptron networks, and (3) in neural network model based nonlinear MPC, factorability of Gaussian functions in radial basis function (RBF) networks is exploited to formulate a novel, computationally efficient nonlinear MPC algorithm.; Findings and conclusions. Simulation studies using generic FOPTD processes and a nonlinear CSTR demonstrate performance of the adaptive QDMC scheme. However, for processes with large number of unknown parameters, the use of this method may be limited. The use of neural network based inferential modeling technique was demonstrated by inferring an ASTM property of a petroleum product using actual data from a large refinery. The proposed RBF based nonlinear algorithm was tested with simulation examples. It was demonstrated that factorization greatly improved computational efficiency. The use of the approach was also demonstrated for control of the Eastman challenge problem. |
