Robustness issues in long-range predictive control

A long range predictive control (LRPC) law is designed on the assumption, that the given model truly reflects the plant dynamics. However in reality there is always a model-plant mismatch (MPM) which gives rise to stability and performance problems when the designed LRPC is implemented on the actual plant. This thesis addresses the crucial question of designing a LRPC that is robust to MPM and looks at was to enhance the actual performance of the LRPC control loop in the presence of MPM.;The small gain theorem is used as a tool to design a LRPC that is robust to MPM. Such a robust design method requires the knowledge of the MPM, which can be estimated from the plant data using signal processing methods. This robust design method has been applied to generalized predictive control (GPC) and Markov-Laguerre based model predictive control (MPC) laws. It is shown that, irrespective of the model types, the robustness bounds of both these LRPCs behave similarly under the influence of tuning parameters. In the case of GPC: (a) the robustness properties associated with important tuning parameters are established analytically and verified experimentally; (b) it is shown that the model and MPM can be estimated from closed loop data; and (c) an optimization problem is formulated within the small gain framework to select some of the controller tuning parameters. For the Markov-Laguerre MPC: (a) designed stability is improved by incorporating steady state weighting; and (b) faster disturbance rejection is obtained by including a structured noise model in the controller design.;Ideally, a stable LRPC with satisfactory performance can be obtained by estimating a model with minimal MPM. Therefore significant emphasis has been given to system identification methods and their applications have been illustrated through two industrial case studies. In the context of model estimation, an extension based on the augmented UD identification method to simultaneously estimate parameters of different orders of the orthonormal function models is developed.;Three control relevant identification methods are reviewed and suitably modified to enhance the achieved performance of GPC. The key issue in control relevant identification is to bring the designed and the achieved closed loop performances as close as possible. This is accomplished by: (a) designing a suitable model estimation filter; and (b) appropriately modifying the designed controller objective function, and using them to obtain control relevant models and thereby upgrading the performance.