psychology e-book

Although predictive control techniques such as Dynamic Matrix Control and Model Algorithmic Control have received much attention in recent years, systematic methods for on-line updating of the process models and predictive control laws raely have been rported. Such adaptation is desirable if the process conditions change significantly over a period of time. The same approach also could be used to generate a discrete convolution model for the initil controller design. In ftis paper, several altrative stategies for on-line aptation during closedloop opertion are evaluated and compared via simulations.


INTRODUCTION
Industrial process-control groups introduced the predictive control techniques of Dynamic Matrix Control (DMC) [1] and Model Algorithmic Control (MAC) [2,3], respectvely, nearly a decade ago. They were developed specifically for multipleinput, multiple-output (MIMO) control problems in which the process could be adequately described by linear dynamic models. These model-based techniques minimize predicted process deviations from the set point using quadratic performance indices. Thus, these methods were the first optimal controllers to be routinely used in the process industries. The evaluation and extension of these techniques spread quickly among academic circles as industry revived dying enthusiasm among researchers for studying and developing optimal controllers.

Since DMC and MAC were introduced, industrial and academic researchers have reported extensions and modifications of both control techniques. In addition, researchers have developed several new predictive controllers and have shown that they are robust under many circumstances. For example, Ogunnaike [4] gives a statistical interpretation of the robustness of DMC. Maurath et al [5] illustrate stability regions for a single-input, single-output (SISO) predictive controller. Garcia and Moari [6] show the stabilizing effect of filtering feedback errr for such controllers.

If the process characteristics change significantly, the predictive controller may need to be retuned or redesigned. This, naurly, leads one to consider adaptng the predictive controler on-line to accommodate the changing process conditions. In fact, one merely needs to estimate the process models on line to make many predictive controllers adaptive. The dearth of publicized industrial experiences extolling adaptive predictive controllers, though, suggests that implementation of these controllers may not be straightforward. Some of the pertinent questions that point to possible implementation problems are:
1. How does one decide if the current controller's performance is unsatisfactory?
2. How should a new process model and controller be obtined, if neded?
3. How should a good process model for the initial controller design be obtined?
This paper helps to answer these questions by presenting systematic methods for the on-line updating of the process models and control laws. The proposed strategies also can be used to develop the convolution model for the initial controler design. In the next section, existing predictive controller techniques are discussed briefly.