NONLINEAR MODEL PREDICTIVE CONTROL OF CHEMICAL PROCESSES

NONLINEAR MODEL PREDICTIVE CONTROL OF CHEMICAL PROCESSES

Departamento de Engenharia Química , Universidade Federal de São Carlos

Via Washington Luiz, km 235 - Caixa Postal 676, 13565-905 - São Carlos - S. P. - Brazil

Phone: [phone], Fax: [phone]

e-mail: [e-mail]

(Received: April 5, 1998; Accepted: March 15, 1999)

Abstract - A new algorithm for model predictive control is presented. The algorithm utilizes a simultaneous solution and optimization strategy to solve the model's differential equations. The equations are discretized by equidistant collocation, and along with the algebraic model equations are included as constraints in a nonlinear programming (NLP) problem. This algorithm is compared with the algorithm that uses orthogonal collocation on finite elements. The equidistant collocation algorithm results in simpler equations, providing a decrease in computation time for the control moves. Simulation results are presented and show a satisfactory performance of this algorithm.

Keywords: Nonlinear control, model predictive control, CSTR, continuous fermenter.

INTRODUCTION

Traditionally, linear controllers have been used in chemical processes. However, it is well recognized that most chemical processes present inherent nonlinearities. Linear controllers can yield a satisfactory performance if the process is operated close to a nominal steady state or is fairly linear.

During the past decade there has been an increase in the use of linear model predictive control (MPC) techniques. The most frequently cited MPC techniques are dynamic matrix control (DMC) (Cutler and Ramaker, 1979) and model algorithmic control (MAC) (Richalet et al., 1978), which are used successfully in a larger number of industrial processes because they explicitly handle constraints (Prett and Gillette, 1979; Moro and Odloak, 1995; Odloak, 1996). A survey of model predictive control, including applied and theoretical papers, was conducted by Garcia et al. (1989).

Many times the dynamic characteristics of the process will change dramatically due to a large disturbance or to significant set-point changes from an on-line optimization routine. Also, batch processes typically operate over a wide operating range, making a linear control strategy ineffective. Thus, there is an incentive to develop and implement nonlinear control strategies in chemical processes. A review was done by Bequette (1991b). Several approaches have been used to design controllers for such nonlinear processes. The most common approach is to linearize the model equations around the nominal operating point and then apply one of the many methods available for designing linear control systems. Nonlinear control methods based on differential geometric concepts have received considerable attention from chemical engineers in recent years. These strategies involve the transformation of the nonlinear process model into an equivalent linear process by means of nonlinear variable transformations. The resulting transformed process is then controlled using a linear controller, and the manipulated variables for the actual process are calculated using the inverse transformations. However, it may not always be possible to use the transformation method because suitable transformation may not always exist. Nonlinear transformations are tedious to develop and may be too sensitive to modeling error. Constraints on the states and on the manipulated variables are also difficult to implement using this approach.

Successful control of several nonlinear systems, however, is critically dependent upon knowledge of the nonlinear dynamics of the processes. An alternative is model predictive control using nonlinear programming techniques. Nonlinear model-based control algorithms can be applied to processes described by a wide variety of model equations, such as nonlinear ordinary differential/algebraic equations, partial differential equations and integro-differential and delay-differential equations. Such models are accurate over a broad range of operating conditions. Model predictive control strategies based on these nonlinear models allow tight control of the process, improved constraint handling and greater robustness in terms of handling unusual dynamics and time delays. In addition, these control strategies may permit plant operation in a region that is economically attractive, but where a linear controller will not be able to control the process satisfactorily. For these reasons, nonlinear model predictive control has become a popular research topic in chemical process control. There are two ways of performing nonlinear model predictive control (NMPC) calculations. The first method employs separate algorithms to solve the differential equations and carry out the optimization (Morshedi, 1986; Economou et al., 1986; Jang et al., 1987; Bequette, 1991a; Sistu and Bequette, 1991; Ali and Zafiriou, 1993). This is a sequential solution. However, it is difficult to incorporate constraints on state variables into this procedure. Also, the algorithm requires the solution of differential equations at each iteration. This can make them prohibitive in terms of computation time and unattractive for use in real-time applications. An attractive alternative is to use a simultaneous solution and optimization strategy (Patwardhan et al., 1990, 1992; Eaton and Rawlings, 1990; Patwardhan and Edgar, 1993; Lee and Park, 1994). The model differential equations are discretized, and along with the algebraic model equations are included as constraints in a nonlinear programming (NLP) problem. An overview of current NMPC technology and applications, as well as to propose topics for future research and development, was done by Henson (1998).

This work describes a new algorithm for model predictive control using the simultaneous solution and optimization strategy. The idea was to use equidistant collocation because it is a simpler alternative than orthogonal collocation on finite elements for discretization of the model differential equations. A nonlinear programming approach was developed and in order to demonstrate the effectiveness of the strategy, three examples were simulated.

NONLINEAR PROGRAMMING APPROACH

In model predictive control, the control problem is posed as a nonlinear programming problem: to optimize some objective function of the inputs and outputs such that (1) the model equations are satisfied and (2) other constraints (if any) on the states, outputs and manipulated variables are met. This approach can be established as:

(1)

The process is described by the following differential/algebraic equations:

(2)

(3)

where

: Controlled variable vector

: Manipulated variable vector

: State variable vector

: Set of model parameters which may include disturbances

: Integral square error (ISE)

The objective of NMPC is to select a set of future control moves (control horizon) to minimize a function based on a desired output trajectory over a prediction horizon, as shown in Figure 1.

Figure 1:
Predictive control approach.

Orthogonal Collocation on Finite Elements

In order to use orthogonal collocation on finite elements, the prediction horizon of R sampling periods must correspond to R finite elements - one element for each sampling period (Figure 2). The simultaneous solution and optimization approach using orthogonal collocation on finite elements (Eaton and Rawlings, 1990; Patwardhan et al., 1992; Lee and Park, 1994) is applied to Equations 1-3:

(4)

where

t_k: Current time

T_R: Prediction horizon time

M: Number of outputs

R: Prediction horizon

N: Number of internal collocation points

y_desired: Desired output

y_predicted: Predicted output

with

(5)

subject to:

(i) Model differential equations ( contains the first-derivative weights at the collocation points):

i = 1,2,...,R (6)

Figure 2:
Collocation on finite elements.

Since is known from the estimator or the previous element, the first equation in Equation 6 is redundant and is thus not used as a constraint.

(ii) Model algebraic equations:

i = 1,2,...,R; j=1,2,...,N+2 (7)

(iii) Initial condition and continuity of the state variables:

(8)

(iv) Definition of control horizon:

i = L,...,R-1; R> L (9)

(v) Bounds on state variables:

i = 1,2,...,R; j=1,2,...,N+2 (10)

(vi) Bounds on outputs:

i = 1,2,...,R; j=1,2,...,N+2 (11)

(vii) Bounds on manipulated variables:

i=1,...,R (12)

(viii) Bounds on changes in the manipulated variables:

i=1,...,L-1 (13)

The objective function is the sum of the squares of the residuals between the model predicted outputs and the desired values over the prediction horizon of time steps (Equation 4). It should be noted that the objective function can be much more general, including economic criteria or final conditions relationships. The optimization decision variables are the control actions time steps into the future; after the Lth time step, it is assumed that the control action is constant (Equation 9).

The constrained NLP problem is solved using a nonlinear optimization code such as Successive Quadratic Programming (SQP) or Generalized Reduced Gradient (GRG) to calculate the control law. Although optimization is based on a control horizon, only the first action control is implemented and optimization is performed again. This type of control algorithm is also known as moving horizon control. After the first control action is implemented, plant output measurements are obtained. This process is repeated at every sampling instant. The effect of modeling error and unmeasured disturbances is treated as an additive, unmeasured disturbance and is estimated at the kth sampling instant in a manner similar to the DMC approach:

(14)

This constitutes the feedback portion of the algorithm. If a perfect process model is available, d is equal to the additive disturbance in the process output.

Proposed Algorithm

The objective of this paper is to present a simpler formulation of the nonlinear programming approach using a simultaneous strategy. The characteristics of the NMPC algorithm were preserved while reducing the computational effort.

In the proposed algorithm, the prediction horizon of R sampling periods corresponds to R-1 internal collocation points for discretization of the model differential equations (Figure 3).

Figure 3:
Equidistant collocation.

Thus, the control problem can be posed as the following nonlinear programming:

(15)

with

(16)

subject to:

(i) Model differential equations ( contains the first-derivative weights at collocation points):

i = 1,2,...,R (17)

(ii) Model algebraic equations:

i = 1,2,...,R (18)

(iii) Initial condition of the state variables:

(19)

(iv) Control horizon :

(20)

(v) Bounds on state variables:

i = 1,2,...,R (21)

(vi) Bounds on outputs:

i = 1,2,...,R (22)

(vii) Bounds on manipulated variables:

(23)

(viii) Bounds on changes in the manipulated variables:

(24)

In this case, the control horizon is restricted to a time step into the future (L=1), but this is not a limitation in moving horizon control. Maurath et al. (1988) found that closed-loop performance for linear model predictive control can easily be adjusted using a control horizon of one and varying the prediction horizon. Using many simulations of a number of chemical processes, Bequette (1991a) has found this to be true for nonlinear systems as well, so the control horizon can generally be set to one time step with the prediction horizon varied for performance and robustness.

SIMULATION EXAMPLES

To demonstrate the effectiveness of the new algorithm, three control problems of continuously stirred tank reactors were simulated. In the first example, an adiabatic CSTR with an exothermic, first-order irreversible reaction was used as a SISO nonlinear process. In the second example, a stirred tank reactor was used as an example of a MIMO nonlinear process. In the other example, regulatory control of a continuous fermentor subject to disturbances in the model parameters was simulated.

In the first and second examples two versions of the nonlinear MPC algorithm were considered: (1) the NMPC1 algorithm (using orthogonal collocation on finite elements) and (2) the proposed NMPC2 algorithm (using equidistant collocation). In the third example, a linear MPC (DMC) and the two versions of nonlinear MPC algorithm were compared. The control horizon was a sampling interval for all controllers. In the present study, the resulting nonlinear constrained optimization problem was solved using the MATLABâ Optimization Toolbox routine CONSTR while MATLAB routine ODE23 was used to integrate nonlinear model equations. All the examples were performed on a PC Pentium 200 MHz MMX microcomputer.

Example 1. Exothermic CSTR

The control problem of the continuously stirred tank reactor (CSTR) was simulated. It consisted of an irreversible, exothermic first order reaction, in a constant volume reactor cooled by a coolant stream. In the example, the objective was to control the measured temperature by manipulating the coolant flow rate. Sistu and Bequette (1991) studied the performance of nonlinear predictive control in handling chemical processes with parametric and structural uncertainty. Also, Embiruçu et al. (1995) used this problem to compare the performance of three MPC algorithms.

The dimensionless modeling equations for this CSTR are given by

(25)

(26)

(27)

where

x₁: Dimensionless concentration

x₂: Dimensionless temperature

x₃: Dimensionless cooling jacket temperature

The dimensionless coolant flow rate, q_c, was used to control x₂. The relationships between the dimensionless parameters and variables and the physical variables are shown in Table 1; the nominal parameter values are shown in Table 2. This system has three steady states, as shown in Table 3. The lower and upper temperature steady states are stable, while the middle temperature steady state is unstable. All the results presented are based on reactor temperature x₂ control.

Table 1:
Dimensionless variables and parameters for the CSTR model.

Table 2:
Parameter values

b = 8.0	d = 0.3	x1f= 1.0	x3f = 1.0	q = 1.0	qcs = 1.65
g = 20.0	x2f = 0.0	f = 0.072	d 1 = 0.1	d 2 = 0.5

Table 3:
Steady states.

	State 1 (stable)	State 2 (unstable)	State 3 (stable)
x1	0.8933	0.5528	0.1890
x2	0.5193	2.7517	5.1373
x3	-0.5950	0.0000	0.6359

The performances of NMPC1 and NMPC2 controllers were compared in this simulation study. The sampling interval was 0.25t and the prediction horizon of the two control strategies was 5; the dominant time constant was approximately 2.5. In all simulations, two internal collocation points on each finite element were used for the process model for NMPC1. The length of each finite element was equal to the sample time of 0.25t .

The reactor temperature behavior was analyzed over 80 sampling times for the following perturbations on the steady states 1 (stable):

(1) +/- 10% on x₂ set-point;

(2) step of -0.1 on feed temperature, x_2f

Figure 4 shows the control performances of NMPC1 and NMPC2 for positive and negative set-point x₂ changes of 10% of the stable 1 operating point.

Figure 4:
Responses to positive and negative set-point change on stable 1 point.

Similar behavior was observed for both the NMPC1 and the NMPC2 algorithms in the output responses to set-point changes.

When an unknown disturbance affected the process, the control responses of the two control strategies were simulated. To show the simulations of the unknown disturbance rejection, feed temperature was used as an unknown disturbance. The feed temperature x_2f =0 was changed to x_2f = -0.1. Figure 5 shows the responses of reactor temperature and coolant flow rate. The disturbance rejection of NMPC1 and that of NMPC2 were approximately the same.

Figure 5:
Comparison of the disturbance rejection performances in NMPC1 and NMPC2.

Table 4 summarizes the results for this example in terms of the integral of the absolute value of the error (IAE) criterion, where it shows a better performance for NMPC1 algorithm. However, according to the computation time, the NMPC2 algorithm was faster, as shown in Table 5.

Table 4: Integral of the absolute value of the error (IAE).

	+10% x2	-10% x2	-0.1 x2f
NMPC1	0.103	0.088	0.829
NMPC2	0.127	0.144	0.911

+10% x2-10% x2-0.1 x2fNMPC1273s339s450sNMPC219s25s28s

	+10% x2	-10% x2	-0.1 x2f
NMPC1	273s	339s	450s
NMPC2	19s	25s	28s

Table 5: Computation time.

Example 2. CSTR

A stirred tank reactor was used as an example of a MIMO nonlinear process (Li and Biegler, 1988; Lee and Park, 1994; Rodrigues and Odloak, 1996). For the reaction:

A + B ® P

where A is in excess, the reaction rate of consumption of B can be written as follow:

(28)

The concentrations of B in the two inlet flows were assumed to be fixed: C_B1=24.9, C_B2=0.1. Both inlet flows contained an excess amount of A. The tank was well stirred with a liquid outflow rate determined by the liquid level in the tank h; i.e., F(h)=0.2h^0.5. The cross-sectional area of the tank was 1.

The modeling equations for this CSTR are given by:

(29)

(30)

(31)

(32)

where

u₁ = F₁: Inlet flow rate with concentrated B

u₂ = F₂: Inlet flow rate with dilute B

x₁ = h: Liquid height in the tank

x₂ = C_B: Concentration of B in the reactor

When the liquid level (y₁) is at its set-point (y₁^* = 100.0) and two inlet flow rates are at steady state (u₁ = 1.0; u₂ = 1.0), there are three different equilibrium points (a , b , g ) for the system, a and g which are stable equilibrium points and b which is unstable (Table 6). In addition, from this unstable equilibrium point a small disturbance can drive the system to one of two stable equilibrium points, a or g , depending on the increase or decrease of B concentration in the reactor caused by the disturbance.

The control objective is to minimize the difference between reactor output and set-point. Two sets of initial conditions were used. Set 1 is in the region where the system automatically goes to equilibrium point a of the system without control. Set 2, on the other hand, is in the region where the system goes to equilibrium point g of the system without any control. Table 6 shows the numerical values of these initial conditions.

This example, with various control variable constraints but without any constraints on the state variable, was simulated and controlled. The u₁ and u₂ variables have the same lower and upper bounds with lower bounds of u₁ and u₂ at zero. The upper bounds varied from 5 or 10 to infinity.

Figure 6 shows y₁, y₂, u₁ and u₂ versus time, respectively, using the initial conditions of set 1 for NMPC1. Figure 7 presents similar plots with the initial conditions of set 2.

Figure 6:
Initial conditions for NMPC1 - Set 1

Figure 7:
Initial conditions for NMPC1 - Set 2.

Figure 8 shows y₁, y₂, u₁ and u₂ versus time, respectively, using the initial conditions of set 1 for NMPC2. Figure 9 depicts similar plots with the initial conditions of set 2.

Figure 8:
Initial conditions for NMPC2 - Set 1

Figure 9:
Initial conditions for NMPC2 - Set 2

The simulation results show that both NMPC1 and NMPC2 algorithms exhibit similar responses forcing the system to the set-point without violating constraints. This example was controlled and simulated with R = 5. It can be verified that NMPC1 algorithm has a better performance than NMPC2 algorithm in terms of IAE criterion, as shown in Table 7. However, it should stress that NMPC2 was faster, requiring about 17 seconds against 3 minutes of CPU time for NMPC1, as shown in Table 8.

Table 7: Integral of the absolute value of the error (IAE).

	NMPC1			NMPC2
	Infinity	0£ u1,u2£ 5	0£ u1,u2£ 10	Infinity	0£ u1,u2£ 5	0£ u1,u2£ 10
C. I. 1	139.67	251.80	145.25	158.74	257.10	160.97
C. I. 2	142.38	317.38	176.88	162.01	320.65	188.29

NMPC1NMPC2 Infinity0£ u1,u2£ 50£ u1,u2£ 10Infinity0£ u1,u2£ 50£ u1,u2£ 10C. I. 1155s170s162s15s17s17sC. I. 2155s195s162s15s17s17s

	NMPC1			NMPC2
	Infinity	0£ u1,u2£ 5	0£ u1,u2£ 10	Infinity	0£ u1,u2£ 5	0£ u1,u2£ 10
C. I. 1	155s	170s	162s	15s	17s	17s
C. I. 2	155s	195s	162s	15s	17s	17s

Table 8: Computation time.

Example 3. Continuous Fermenter

Regulatory control of the productivity of a continuous fermenter, subject to disturbances in the cell-mass yield and the maximum growth rate, was investigated. The cell-mass yield and the maximum specific growth rate tend to be especially sensitive to changes in the operations conditions. From a process control perspective, these two model parameters can be viewed as unmeasured disturbances because they may exhibit significant time-varying behavior (Henson and Seborg, 1992). Consider a well-mixed constant yield continuous fermenter with product-free and sterile feed represented by the following unstructured model:

(33)

(34)

(35)

where

X: Effluent cell mass or biomass concentration

S: Substrate concentration

P: Product concentration

D: Dilution rate

S_f: Feed substrate concentration

a m +b : Specific product formation rate

Y_X/S: Cell mass yield

The three process state variables are X, S and P. The dilution rate D and feed substrate concentration S_f were employed as manipulated inputs in single input/single output control strategies.

Using the substrate and product inhibition model, the specific growth rate m is given as:

(36)

where

m _m: Maximum specific growth rate

P_m: Product saturation constant

K_m: Substrate saturation constant

K_i: Substrate inhibition constant

The values of the different model parameters are given in Table 9.

Table 9:
Nominal fermenter parameters and optimum conditions.

Variable	Nominal value	Variable	Nominal value	Optimum conditions
YX/S	0.4 g/g	Ki	22 g/l
a	2.2 g/g	D	0.15 h-1	0.1636 h-1
b	0.2 h-1	Sf	20 g/l	23.3993 g/l
m m	0.48 h-1	X	7.038 g/l	7.3044 g/l
Pm	50 g/l	S	2.404 g/l	5.138 g/l
Km	1.2 g/l	P	24.87 g/l	24.9999 g/l

For most continuous fermenter processes the control objective is to maximize the production of a desired product. The productivity Q can be defined as the amount of product cells produced per unit of time (Q = DP). The optimum steady-state operating point can be obtained by maximizing the productivity rate using the dilution rate and feed substrate concentration as independent variables with steady-state model equations as constraints. The optimization results are presented in Table 9 (Henson and Seborg, 1992). The control objective is to regulate the fermenter near optimum productivity because in practice the optimum is not known.

Figure 10 shows the effects of dilution rate and feed substrate concentration on productivity for three values of maximum growth rate and cell mass yield. Note that small changes in the m _m and Y_X/S can have dramatic effects on maximum productivity. If the fermenter is operated at high dilution rates or feed substrate concentrations, no cells are produced and the productivity drops to zero. This phenomenon is known as washout. Unfortunately, the optimum productivity is near the washout region. A well-designed controller will maintain the fermenter near the optimum while avoiding washout.

Figure 10:
The effect of dilution rate and feed substrate concentration on productivity.

Closed-Loop Behavior

The control of the productivity Q = 3.7303 using dilution rate and feed substrate concentration as manipulated variables in SISO strategies for productivity control was examined. Linear MPC (DMC) and the two versions of nonlinear MPC (NMPC1 and NMPC2) were tested for 0.46 m _max and 0.44 m _max and for 0.36 Y_X/S and 0.32 Y_X/S disturbances. As shown in Figure 10, for the 0.44 m _max and 0.32 Y_X/S disturbances, there was no dilution rate which satisfied the productivity Q = 3.7303.

The control performances of three controllers were compared in this simulation study. The sampling interval was 1.0 h and the prediction horizon of the two control strategies was 5.0 h; the dominant time constant was approximately 10.0 h. Results are shown in Figures 11 to 13. Note that the NMPC1 and NMPC2 controllers presents better performance disturbance rejection than the DMC controller.

Figure 11:
NMPC1, NMPC2 and DMC control with dilution rate (Y_X/S and m _m disturbances).

Figure 12: NMPC1, NMPC2 and DMC control with feed substrate concentration (m _mdisturbance).

Figure 13: NMPC1, NMPC2 and DMC control with feed substrate concentration (Y_X/S disturbance)

CONCLUSIONS

In this paper, NMPC algorithms for some aspects of solution of the model's differential equations were investigated. The algorithm proposed (NMPC2) uses a simultaneous solution and optimization strategy. For discretization of the model's differential equations, the sampling times for the prediction horizon correspond to the internal collocation points. Thus, a significative reduction in the vector dimension of the optimization variables was obtained. The advantage of dimension size reduction was related to the computational effort required to calculate the control actions. Although the control horizon is restricted to one sampling period, this is not a limitation on moving horizon control. The simulation results using this control algorithm were compared with those from an NMPC algorithm using simultaneous strategy and model differential equations discretized by orthogonal collocation on finite elements (NMPC2) for two nonlinear reactor control problems. It was observed that both algorithms showed satisfactory control behavior and they also presented similar output responses. However, NMPC2 was better than NMPC1 in terms of IAE criterion. The fact is that orthogonal collocation on finite elements method for solving differential equations is more precise than equidistant collocation. Thus for control problems that do not require a high performance, NMPC2 is more advantageous, since it does not have to use a sophisticate control algorithm as NMPC1 demands. Results for a continuous fermenter show that the two versions of nonlinear MPC (NMPC1 and NMPC2) algorithms presents a better performance than does DMC. These results show the effectiveness of the proposed algorithm.

ACKNOWLEDGEMENT

The authors acknowledge the financial support received from CNPq in the form of a fellowship for one of the authors (R.G.S).

NOMENCLATURE

d Additive disturbance D Dilution rate D Collocation matrix e Error L Control horizon M Number of outputs N Number of internal collocation points p Parameter vector R Prediction horizon or number of elements x State variable vector t Time t_k Current time T_R Prediction horizon time u Manipulated variable vector y Output variable vector D u Changes in manipulated variables F Performance index

Example 1

C Reactor composition C_f Feed composition q Dimensionless reactor feed flow rate q_c Dimensionless coolant flow rate q_cs Steady-state value of Q Feed flow rate Q_c Coolant flowrate T Reactor temperature T_c Coolant temperature UA Heat transfer coefficient x heat transfer area V Reactor volume V_c Cooling jacket volume x₁ Dimensionless reactor concentration x₂ Dimensionless reactor temperature x₃ Dimensionless cooling jacket temperature x_1f Dimensionless feed concentration x_2f Dimensionless feed temperature x_3f Dimensionless coolant feed temperature b Dimensionless heat of reaction d Dimensionless heat transfer coefficient d ₁ Dimensionless volume ratio d ₂ Dimensionless density x heat capacity ratio f Damkohler number (at nominal flow rate) g Dimensionless activation energy r C_p Density x heat capacity r _cC_pc Density x heat capacity of coolant t Dimensionless time

Example 2

u₁ Inlet flow rate with concentrated B u₂ Inlet flow rate with dilute B y₁ Liquid level y₂ Concentration of B in the reactor a Equilibrium point b Equilibrium point g Equilibrium point

Example 3

K_i Substrate inhibition constant K_m Substrate saturation constant P Product concentration P_m Product saturation constant Q Feed flow rate or productivity S Substrate concentration S_f Feed substrate concentration X Biomass concentration Y_X/S Cell-mass yield a Yield parameter for the product b Yield parameter for the product m Specific growth rate m _m Maximum specific growth

REFERENCES

Ali, E. and Zafiriou, E., Optimization-Based Tuning of Nonlinear Model Predictive Control with State Estimation, J. Proc. Cont., 3(2), 97-107 (1993).

Bequette, B. W., Nonlinear Predictive Control Using Multirate Sampling, Can. J. Chem. Eng., 69, 136-151 (1991a).

Bequette, B. W., Nonlinear Control of Chemical Processes: A Review, Ind. Eng. Chem. Res., 30, 1391-1413 (1991b).

Cutler, C. R. and Ramaker, B. L., Dynamic Matrix Control - A Computer Control Algorithm, AIChe National Meeting, Houston, Texas (1979).

Eaton, J. W. and Rawlings, J. B., Feedback Control of Chemical Processes Using On-Line Optimization Techniques, Computers Chem. Eng., 14(4/5), 469-479 (1990).

Economou, C. G., Morari, M. and Palsson, B. O., Internal Model Control. 5. Extension to Nonlinear Systems, Ind. Eng. Chem. Process Des. Dev., 25, 403-411 (1986).

Embiruçu, M., Tauhata, T., Wanderley, A. and Lima, E., Model Predictive Control of a Nonlinear Process, Latin American Applied Research, 25, 177-180 (1995).

Garcia, C. E., Prett, D. M. and Morari, M., Model Predictive Control: Theory and Practice - A Survey, Automatica, 25(3), 335-348 (1989).

Henson, M. A. and Seborg, D. E., Nonlinear Control Strategies for Continuous Fermenters, Chemical Engineering Science, 47(4), 821-835 (1992).

Henson, M. A., Nonlinear Model Predictive Control: Current Status and Future Directions, Computers Chem. Eng., 23(2), 187-202 (1998).

Jang, S., Joseph, B. and Mukai, H., Control of Constrained Multivariable Nonlinear Process Using a Two-Phase Approach, Ind. Eng. Chem. Res., 26, 2106-2114 (1987).

Lee, J. and Park, S., Robust Nonlinear Predictive Control Using a Disturbance Estimation, Chem. Eng. Comm., 128, 43-64 (1994).

Li, C. W. and Biegler, L. T., Process Control Estrategies for Constrained Nonlinear Systems, Ind. Eng. Chem. Res., 27, 1421-1433 (1988).

Maurath, P. R., Mellichamp, D. A. and Seborg, D. E., Predictive Controller Design for SISO Systems, IEC Research, 27, 956-963 (1988).

Moro, L. F. L. and Odloak, D., Constrained Multivariable Control of FCC Converters, Journal of Process Control, 5(1), 29-39 (1995).

Morshedi, A. M., Universal Dynamic Matrix Control, Third International Conference on Chemical Process Control, Morari, M. and McAvoy, T. J., Eds., CACHE, Elsevier: New York (1986).

Odloak, D. A., New State Space Approach to Model Predictive Control, Brazilian Journal of Chemical Engineering, 3(3), 152-167 (1996).

Patwardhan, A. A. and Edgar, T. F., Nonlinear Model-Predictive Control of a Packed Distillation Column, Ind. Eng. Chem. Res., 32, 2345-2356 (1993).

Patwardhan, A. A., Rawlings, J. B. and Edgar, T. F., Nonlinear Model Predictive Control. Chem. Eng. Commun., 87, 123-141 (1990).

Patwardhan, A. A., Wright, G. T. and Edgar, T. F., Nonlinear Model-Predictive Control of Distributed-Parameters Systems, Chemical Engineering Science, 47(4), 721-735 (1992).

Prett, D. M. and Gillette, R. D., Optimization and Constraints Multivariable Control of a Catalytic Cracking Unit, Paper No. 51b, AIChE National Meeting, Houston, Texas (1979).

Richalet, J., Raoult, A., Testud, J. L. and Papan, J., Model Predictive Control: Application to Industrial Processes, Automatica, 14, 413-428 (1978).

Rodrigues, M. A. and Odloak, D., Controle Não-Linear Via Operadores de Estado, 11^o COBEQ, 1, 517-522, Rio de Janeiro (1996).

Sistu, P. B. and Bequette, B. W., Nonlinear Predictive Control of Uncertain Processes: Application to a CSTR, AIChE Journal, 37(11), 1711-1723 (1991).

Publication Dates

• Publication in this collection
  23 Apr 1999
• Date of issue
  Mar 1999

History

• Accepted
  15 Mar 1999
• Received
  05 Apr 1998