to precisely predict the future outputs as well as being simple to implement and to understand.

As MPC is not a unique technique but a set o$ different methodologies, there are many types of models used in various formulations. One of the most popular in industry is the Truncated Impulse Response Model, which is very simple to obtain as it only needs the measurement of the output when the process is excited with an impulse input. It is widely accepted in industrial practice because it is very intuitive and can also be used for multivariable processes, although its main drawbacks are the large number of parameters needed and that only open-loop stable processes can be described this way. Closely related to this lund of model is the Step Response I Model, obtained when the input is a step. The Transfer Function Model is, perhaps, most widespread in the academic community and is used in most control design methods, as it is a representation that requires only a few parameters and is valid for all kind of processes. The State-Space Model is also used in some formulations, as it can easily describe multivariable processes. The optimizer is another fundamental part of the strategy as it provides the control actions. If the cost function is quadratic, its minimum can be obtained as an explicit function (linear) of past inputs and outputs and the future reference trajectory. In the presence of inequality constraints the solution has to be obtained by more computationally taxing numerical algorithms. The size of the optimization problems depends on the number of variables and on the prediction horizons used and usually turn out to be relatively modest optimization problems which do not require sophisticated computer codes to be solved. However the amount of time needed for the constrained and robust cases can be various orders of magnitude higher than that needed for the unconstrained case and the bandwidth of the process to which constrained wc can be applied is considerably reduced. Notice that the MPC strategy is very similar to the control strategy used in driving a car. The driver knows the desired reference trajectory for a finite control horizon, and by taking into account the car characteristics (mental model of the car) decides which control actions (accelerator,brakes and steering) to take in order to follow the desired trajectory. Only the first control actions are taken at each instant, and the procedure is again repeated for the next control decisions in a receding horizon fashion. Notice that when using the classical control schemes, such as PIDS, control actions are taken based on past errors. If the car driving analogy is extended, as has been done by one of the commercial MPC vendors (SCAP) [71] in their publicity, the PID way of driving a car would be equivalent to driving the car just using the mirror as shown in figure 1.3. This analogy is not totally fair with PIDS, because more information (the reference trajectory) is used by MPC. Notice that if a future point in the desired reference trajectory is used as the setpoint for the PID,the differences between both control strategies would not seem so abysmal:

10

Chapter 1. Introduction to Model Based Predictive Control Model uncertainty: although model identification packages provide uses estimates of model uncertainty, only one product (RMPCT) this information in the control design. All other controllers can be detuned in order to improve robustness, although the relation between performance and robustness is not very clear. Constant disturbance assumption: although perhaps the most reasonable assumption is to consider that the output disturbance will remain constant in the future, better feedback would be possible if the distribution of the disturbance could be characterized more carefully. Analysis: a systematic malysis of stability and robustness properties of MPC is not possible in its original finite horizon formulation. The control law is in general time-varying and cannot be represented in the standard closed loop form, especially in the constrained case.

The technology is continually evolving and the next generation will have to face new challenges in open topics such as model identification, unmeasured disturbance estimation and prediction, systematic treatment of modelling error and uncertainty or such an open field as nonlinear model predictive control.

a: 1.4 Outline of the Chapters

The book aims to study the most important issues of MPC with regards to its application to process control. In order to achieve this objective, it is organized as follows. Chapter 2 describes the main elements that appear in any MPC formulation and reviews the best known methods. A brief reviewal of the the most outstanding methods is made. Chapter 3 focuses on commercial Model Predictive controllers. Because of its popularity, Generalized Predictive Control (GPC) treated in greater detail in chapter 4. Two related methods which is have shown good stability properties (CRHPC and SGPC) also described. are Chapter 5 shows how GPC can easily be applied to a wide variety of plants in the process industry by using some Ziegler-Nichols types of tuning rules. By using these, the implementation of GPC is considerably simplified, and the computational burden and time that the implementation of GPC may bear, especially for the adaptive case, is avoided. The rules have been obtained for plants that can be modelled by the reaction curve method and plants having an integrating term, that is, most of the plants in the process industry. The robustness of the method is studied. In order to do this, both structured and unstructured uncertainties are considered. The closed loop is studied, defining the uncertainty limits that preserve stability of the real process when it is being controlled by a GPC designed for the nominal model.