Iterative identification and predictive control

Identification and control have so far been treated mainly as two independent exercises. However, several researchers have shown that identified models that are commensurate with the control algorithm can provide better closed loop performance. Clearly, if the true process model is known then there can not be a better model for any purpose whether it be control, prediction or fault detection. All the identification methods developed for compatibility with the control algorithm, assume that the true model is not known and/or is not identifiable accurately. The central theme of this thesis is about identifying models that are commensurate with model predictive controllers. A commensurate model is expected to provide an overall better closed loop performance.; A new metric for model quality based on the prediction horizon of the controller is defined. A number of properties of this metric under both open and closed loop conditions are derived in this thesis. The effect of this new metric on the bias distribution of the identified models is studied in detail. Expressions for optimal input spectra for obtaining the model predictive control relevant models are also presented in this section.; The dual control problem provides the optimal solution to the problem of online model adaptation and control. However, it is a difficult problem to solve. An approximation of the dual control problem is provided and its extension to model predictive controllers is outlined. With this as motivation, a new iterative algorithm for improving the closed loop performance is developed based on the metric defined earlier. An improvement of the iterative feedback tuning method is also provided in this part of the thesis. Most of these methods are developed for linear systems. For nonlinear systems, a robust backstepping based iterative adaptive algorithm is also developed.; Every identified model must be qualified with a characterization of its quality. Most of the available methods for quantifying model uncertainty are based on asymptotic statistical properties. Using learning theory, finite sample upper bounds on rates of convergence of identified models are derived for prediction error models.; An offshoot of the research done in this thesis are methods for identification of delay dominated recycle systems and identification of multi-rate systems. An industrial application of the multi-rate identification method is also presented.