Closed-loop re-identification of an industrial debutanizer column

CATEGORY: DEBUTANIZER
Preprints of the 8th IFAC Symposium on Advanced Control of Chemical Processes The International Federation of Automatic Control, Furama Riverfront, Singapore, July 10-13, 2012
Closed-loop re-identification of an industrial debutanizer column
Renato N. Pitta*, Darci Odloak**
renato.pitta@petrobras.com.br
odloak@usp.br
*Petrobras – Revap – São José dos Campos – São Paulo, Brazil
**Department of Chemical Engineering, Polytechnic School of the University of São Paulo (EPUSP), Brazil
Abstract:
In this work it is described the industrial implementation of an approach to the re-identification of multivariable processes being controlled by a class of MPC controllers in which the control work is performed in two distinct layers. In the upper layer a steady-state optimization problem is solved and produces a set of optimum targets for the process inputs and/or outputs. These targets are passed to the dynamic layer of the controller that determines the best input trajectory that will drive the process system to these targets while preserving the process constraints. The two layer structure is found in several commercial MPC packages and so, when the practical application is concerned, the approach described here may be useful. In the case considered here, it is assumed that the continuous operation of the process systems cannot be interrupted and the normal process specifications should be maintained inside well defined bounds. These conditions point to the need of a closed-loop procedure. The industrial process selected to illustrate the method considered here is the debutanizer distillation column of a large oil refinery in Brazil. The practical results indicate that the approach can be effective and show a good potential to become the standard procedure to the closed-loop upgrade of process models that have been affected by equipment deterioration or large process excursions into new operating conditions.
INTRODUCTION
Nowadays Model Predictive Control is mostly based on a linear time invariant model of the process system, which is usually obtained at the implementation stage of the controller and based on open-loop input-output data. It is usually recognized that the model identification is the most time demanding step of the MPC implementation (Qin & Badgwell, 2003). One of the drawbacks of this procedure is that the resulting model may represent adequately the process only at operating conditions near to the conditions considered in the identification test. Model deterioration may result from changes in the dynamics of the plant produced by persistent disturbances not considered in the design stage of the controller and that drive the plant to different operating conditions (Conner & Seborg, 2005). Other causes of model deterioration are loss of performance of heat transfer equipment, separation systems, etc, or changes in product specification. Even small revamps in the process system may invalidate in various extensions the available model. Also, aging of equipment tends to intensify these problems and accelerate the loss of performance of MPC unless the re-commissioning of the controllers is carried out periodically. This procedure involves the model re-identification and retuning of the controller parameters (Gugaliya et al., 2005).
In a large scale industrial process, re-identification based on open-loop tests as usually performed in the implementation phase of the system is prohibitive in most of the cases. Besides of the need to involve a large number of operators, instrument technicians and engineers to perform the long lasting open-loop tests that may take weeks in a system of large dimensions, these tests may result in unacceptable losses in the production rate or product quality. Thus, in general, no one in the plant site is keen to authorize this sort of test in a competitive scenario. Then, in this environment, closed-loop identification is a subject that needs to be taken seriously.
An ideal situation in closed-loop identification would be the case where routine operating data could be used in the identification algorithm without the need of introducing additional external disturbances. But, usually, these data shows a large noise to signal ratio and the process system is not sufficiently excited. Thus, in order to guarantee the necessary conditions for process identification, an external signal needs to be introduced in the closed-loop system. This dither signal should be designed with the objective of guaranteeing the persistent excitation of the process system. It may be introduced into the controlled variable set-point or added to the manipulated variable. However, adding such a signal to the set-point may result in loss of product specification and adding the dither signal to the input may result in infeasibility of the MPC control problem. On the other way, insufficient excitation may compromise the identification results.
The purpose of this work is to implement in an industrial distillation column a new approach to the persistent excitation of the process system in closed-loop with a MPC controller and to compare the performance of the resulting model to the existing model in the MPC in order to evaluate the need to replace the available model.
Free Full Text Source: http://www.nt.ntnu.no/users/skoge/prost/proceedings/adchem-2012/files/0064.pdf