Model predictive control based on nonlinear autoregressive and neural network models

This research was dedicated to investigating the feasibility of using nonlinear autoregressive with exogenous inputs (NARX) models and artificial neural networks (ANN) for the identification and the closed-loop control of chemical processes. Within this research, the existing theory of NARX models was extended by incorporating measured disturbance terms to the model structure thereby increasing the flexibility and applicability of NARX models for a wide range of processes. For selecting parsimonious submodels, the modified Gram-Schmidt orthogonalization procedure was adapted. The concept of a pointer vector was introduced. With this concept it is possible to fully automate the identification and closed-loop control algorithm.; For the closed-loop control, an adaptive model predictive control (MPC) approach was selected, resulting in a novel solution to the nonlinear programming problem. Since the extended NARX model is linear in the parameters, a recursive constant trace identification algorithm could be applied to adapt the model to possible time-variations of the process which guarantees offset-free set point tracking.; As a comparative study, the application in the aforementioned context of a special case of artificial neural networks, feedforward neural networks (FNN), was investigated. The feasibility of applying different FNN topologies within a MPC algorithm was discussed (multiple predictions with one network or repeated networks with one-step-ahead predictions). Since FNN models can not be adapted on-line the estimated plant/model mismatch was used to correct the model predictions used for the MPC portion. The modified Marquardt algorithm for solving the nonlinear programming problem can be applied without changes.; For the verification of the proposed nonlinear identification and control structures, the model of a MIMO bioreactor and a model as well as an experimental set-up of a waste water neutralization process were available.; By numerous identification experiments it was shown that the extended NARX model is very suitable for approximating nonlinear process dynamics and suggested a better performance than the FNN model. The closed-loop control results are more comparable and show only slight advantages of the NARX model based MPC approach compared to the slightly different MPC structure used for the FNN based approach.