148q

Robust and Reliable Modelling for a Distillation Column

Yahya Chetouani, Département Génie Chimique, Université de Rouen, Rue lavoisier, Mont Saint Aignan, France

In the last few years ever-growing interest has been shown in production quality standards and pollution phenomena in industrial environments. However process development and continuous request for productivity led to an increasing complexity of industrial units. The dynamic nature and the nonlinear behavior of such units pose challenging control system design when products of constant purity are to be recovered. In chemical process control, the processes typically exhibit nonlinear behavior. The intrinsic highly nonlinear behavior in the industrial process, especially when a chemical reaction is used, poses a major problem for the formulation of good predictions and the design of reliable control systems (Chetouani, 2007; Fortuna et al., 2005; Cammarata et al., 2002). Similar problems arise also from the uncertainty for the parameters of the process, such as the reaction rate, activation energy, reaction enthalpy, heat transfer coefficient, and their unpredictable variations. Most large-scale process models derived from first principles are represented by nonlinear differential–algebraic equation (DAE) systems. Since such models are often computationally too expensive for real-time control. Even if global nonlinear models are available, they may not be appropriate for control and monitoring or fault detection purpose (Huang et al., 2000).

Distillation is one of the most important separation processes used in many chemical industries. It can be classified into two ways namely binary distillation for separation of mixture of two substances and multi-component distillation for separation of mixture of more than two substances. Distillation is an energy-intensive separation process in which a liquid or vapor mixture of two or more substances is separated into its component fractions of desired purity by the application and removal of heat. It has extensively been studied (Eden et al., 2000; McAvoy, 1990). There are essentially two approaches by which nonlinear models can be developed for a distillation column; from first principles by using the process knowledge or empirically from input/output data. The advantages and disadvantages of each approach are well known. In industrial practice, it is not always possible in general to obtain accurate first principles models for high-purity distillation columns. Most industrial columns are used to separate multi-component mixtures whose constituent elements are often not known completely; the fundamental thermodynamics of multi-component vapor-liquid equilibrium, the physical property data, and other essential constitutive relations required for the successful development of a first principles model are not always available. And even when such knowledge is available, the resulting models usually occur in the form of a very large system of coupled nonlinear ordinary differential equations, and may therefore not always be the most convenient for controller or fault detection (FD) design. On the other hand, when fundamental process knowledge is unavailable or incomplete, or when the resulting model may not be particularly suitable for FD applications, the input/output models identified from plant data may be more useful. Nonlinear autoregressive models have been used (James et al., 2002), as well as models that use a set of polynomial basis functions (e.g. Volterra functions). Also ANNs provide an excellent mathematical tool for dealing with severe nonlinear problems. Another attractive property is self-learning ability. As a result, nonlinear systems can be modeled with a great flexibility. These features allow one to employ artificial neural networks to model complex, unknown and nonlinear dynamic processes.

The main aim of this paper is to establish a reliable model both for the steady-state and unsteady-state regimes of a nonlinear process. The use of this model should reflect the normal behavior of the process and allow distinguishing it from an abnormal one. In order to obtain this reliable model, the artificial neural network (ANN) is used to model plant input-output data by means of a NARX model. Also this paper shows another technique for neural model reduction into account the physical knowledge of the process. The study shows the choice and the performance of the ANN in the training and test phases. An analysis of the inputs choice, time delay, hidden neurons and their influence on the behavior of the neural estimator is carried out. Three statistical criteria are used for the validation of the experimental data. The model is implemented by training a Multi-Layer Perceptron Artificial Neural Network (MLP-ANN) with input-output experimental data. After describing the system architecture, a realistic and complex application as a distillation column is presented in order to illustrate the proposed ideas concerning the dynamics modelling and model reduction. Satisfactory agreement between identified and experimental data is found and results show that the reduced neural model successfully predicts the evolution of the product composition.

Keywords: Reliability; modelling; artificial neural network; NARX; distillation column; product quality.

References

Chetouani, Y. (2007). Modelling and prediction of the dynamic behavior in a reactor-exchanger using NARMAX neural structure, Chemical Engineering Communications Journal, 5, 691-705.

Fortuna, L., Graziani, S., Xibilia, M. G. (2005). Soft sensors for product quality monitoring in debutanizer distillation columns, Control Engineering Practice, 4, 499-508.

Cammarata, L., Fichera, A., Pagano, A. (2002). Neural prediction of combustion instability, Applied Energy, 72, 513-528.

Huang, Y., Reklaitis, G.V., Venkatasubramanian, V. (2000). Dynamic optimization based fault accommodation, Computers and Chemical Engineering, 24, 439–444.

Eden, M.R., Koggersbøl, A., Hallager, L., Jørgensen, S.B. (2000). Dynamics and control during start-up of energy integrated distillation column, Computer Chemical Engineering, 24, 1091–1097.

McAvoy, T.J. (1990). Modeling chemical process systems via neural computation, IEEE Control Systems Magazine, 3, 24-29.

James, S., Legge, R., Budmann, H. (2002). Comparative study of black-box and hybrid estimation methods in fed-batch fermentation, Journal of Process Control, 1, 113–121.