Sensitivity Analysis Applied To Reactive

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

SENSITIVITY ANALYSIS APPLIED TO REACTIVE SYSTEMS FOR

IMPROVING CONTROL PERFORMANCE

João T. Manzi1 and Heleno Bispo


Catholic University of Pernambuco – Department of Chemistry

Abstract. This paper is devoted to sensitivity analysis for nonlinear multivariable processes, particularly
applied to chemical reactive systems. The presence of uncertainties or shift in the system’s behavior has been
investigated, based on deterministic models and also by direct application of the disturbance or uncertainty in
the model. Such an approach describes the behavior of sensitivity, and since a reduction of sensitivity is
desired, it can reveal how the minimum of sensitivity can be obtained. If the principal aim is concerned with
process control, the results can also provide a strategy for improving performance, using the idea of fictitious
smoothing reactors, meaning a partition and scheduling of variables of interest. Such a strategy can be
inserted into software for process control, without additional cost or appreciable computational load.

Keywords: modeling, reactive system, sensitivity analysis, nonlinear PI

1- Introduction
Despite there being a new class of advanced controller, the PI or PID controller is still widely used in more
than 95% of all control requirements for most industrial control systems. Their well-known qualities such as
robustness in a wide range of industrial process with satisfactory performance, the simplicity of their structure
and the low cost of implementation and maintenance are the reasons for such a long life. Due to the linear
structure of PID, this class of feedback controllers cannot yield satisfactory results for some application in the
industrial environment since chemical real-world processes have some degree of nonlinearity. However, to
enable a controller such as this to deal with nonlinearity, efforts have been made to overcome the difficulties,
using strategies which can include models of the process (Manzi and Odloak, 1998) and consequently to reduce
the degrees of freedom of the system (Rivera et al., 1986), or even to implement the process control in such a
way as to reduce dynamically sensitivity to disturbance.
Reactive systems have embodied structures of advanced control techniques for solving problems especially
those associated with their nonlinearities, but, in practice, in view of the need to have qualified staff and the
consequently increase in the cost of production, such control systems are not yet ordinarily used.
It can also be observed in industrial practice that the results obtained with the PID controller applied mainly
to neutralization processes can be improved or its performance further justified, as in the case of temperature
control, when sensitivity analysis is considered as a tool which can provide a better understanding of how the
process responds in the presence of perturbations in the variables. For a neutralization system it is common to
use a typical control strategy known as smoothing tanks. However, such a practice may not be reasonable either
from the technical or economic point of view, since it can affect the plant configuration or compromise some
operational conditions or the sanitary stability of the process. In this case, sensitivity analysis can contribute to
improving the control performance, providing valuable insights and alternative control configurations.
For reaction temperature control, sensitivity analysis can reveal that the use of the so-called smoothing
reactor is unnecessary. Such a procedure is already ordinarily applied in practice.
A review of the literature shows a widespread interest in stability analysis but relatively little has appeared on
sensitivity, especially with regard to nonlinear systems. Skinskey (1973) presented a sensitivity analysis for
cyclic variations in the influent pH, which does not reflect the fundamental details of the sensitivity. Other
1
To whom all correspondence should be addressed.
Address: Departamento de Química - Universidade Católica de Pernambuco (UNICAP)
Rua do Príncipe 526, Boa Vista, Recife – PE – Brasil - CEP: 50050-900, e--mail: jmanzi@unicap.br
interest in an analysis of sensitivity has been focused on controllability aspects (Skogestad, 1996; Faanes and
Skogestad, 2004), a different concept introduced by Kalman, which examines the implementation of buffers
tanks for neutralization purposes. The results have confirmed the necessity for carrying out neutralization
process in several steps in order to reduce dynamically sensitivity to disturbance, a usual practice in the industry.
Another helpful way to use sensitivity analysis is concerned with optimization procedures by means of the
classical Lagrange multipliers (Heyen et al., 1996). However, by being linear the use of such an operator is
restricted to the linear case.
The aim of this paper is to provide some insights into aspects of sensitivity analysis used in the design of
reactive control systems. In particular, the analysis deals with the impact of uncertainties or disturbance on the
control device, taking into account a minimum of sensitivity. By being concentrated on the fictitious smoothing
idea, such analysis can provide a strategy to improve control performance in some cases, and in others it can also
justify some procedures which are already well-established in practice.
For all purposes, consider the following schematic diagram which represents a continuous, well-stirred
reactor.
Reagent, CA,e , Te , Fe Reagent (F2) for pH process
or
Acid feed stream (F1)
Fe , Tc,s

Coolant, Fc , Tc,e
Product, CA , T, F or
neutralized effluent

Fig.1. Schematic of general reactive system.

2- Deriving the model equations for sensitivity analysis

2.1 For reaction temperature control


Consider the classical first order chemical reaction M? N, irreversible, exothermic, which occurs in a
continuous well-stirred reactor, depicted in figure 1.
This reaction has been the object of interest for several researchers and it also has been given particular
emphasis, especially in text books dealing with process control. Nevertheless, analysis, focused on the sensitivity
point of view, can provide a new interpretation for the results, aiming at process control.
The mathematical model for the reaction considered can be described by the following equations:
dV
= Fe − F (1)
dt

dC M Fe
= (C M ,e − C M ) − k o e − E / RT C M (2)
dt V
E
dT Fe − Q
= (Te − T ) + J k o e RT CM − (3)
dt V ρcp V

where J = ( − ∆H R ) / ρ c p . F ,T and V represent the flow rate, temperature and the reactor volume. The subscripts

M and e are related to the chemical M and inlet variables respectively.


Since the reactor volume must be held constant, the first equation may be neglected and consequently the
resulting set of nonlinear equations should be solved simultaneously. Since the linear terms make a weak
contribution to explaining the behavior of sensitivity for the process under consideration, the terms related to the
−E
Arrhenius law, k = k o e RT , present in the system equations should be essentially focused to establish such

analysis.

2.1.1 Stochastic point of vi e w


From the stochastic point of view, the concept of the expectation may be used to evaluate the mean and
variability in the observed variable. Due to the expectation being a linear operator, its application to the
nonlinear system must be preceded by a linearization procedure. Among the methods used to perform the
linearization of the Arrhenius equation, the well-known strategy based on a truncated Taylor series was chosen

by which such linearization can be found about the point x i0 , where x i0 assume the values k o0 and T 0 or even
around the mean of k and T.

The system consists of two independent random variables k and T with µ k o and µT , its expected values and

σ k2 , σ T2 its variances, respectively. Uncertainties( ε ) and perturbations are also considered here due to the
parameters and output variables as well as uncertainties that relate to the modeling. Since ε is normally

distributed with mean zero and variance σ ε2 , then the mean of the Arrhenius equation can be performed as

follows
−E E −E −E E
µ k = − k 00 e RT 0 +e RT 0 µ + k 00 e RT 0 µT
( )
0 k0 2
(4)
RT RT0
or
µk = A + B µk 0 + C µT (5)

where A, B and C are given by


E E E
− − − E
E RT 0 and C = k 0 e RT 0
A= − k0 e =
0 0

( )
RT ; B e
0 0 2
RT RT0

According to definition of variance, the following result can be obtained

σ k2 = B2σ k20 + C 2σ T2 + σ ε2 (6)

Assuming the presence of shift in the process mean from µi to µi + δ i σ i where δ i is related to uncertainty

in the process mean for an independent variable i , consequently the mean of observed variable k can be given
by
µk + δ kσ k = A + B ( µ k0 + δ k 0 σ k 0 ) + C ( µT + δ Tσ T ) (7)

The equation above can be rewritten as

δ kσk 1 δ k0 σ k0 1 δT σ T
= + (8)
µk  C  µ   A  1  µ k0  B  µ k   A  1  µT
1 +   T  +    1 +    0  +    
 B   µk0   B   µ k0 
  C   µT   C   µT 
If the linearization procedure is taken around the point µk 0 and µT , then the mean is given by

E

R µT
µk = µk 0 e (9)

and the variance


2 2
 − E   −
E 
σ k =  e R µT
2  σ 2 +  µ e Rµ T E  σ 2 + σ 2 (10)
ε
  k0
 k0
RµT2 
T
   
Taking into account the same degree of uncertainty previously assumed, the following considerations can be
made
E

R (µ T + δ T σ T )
µk + δ kσ k = ( µk 0 + δ k 0 σ k 0 ) e (11)

 δ Tσ T 
 
E  µT 
RµT  δ T σ T 
σ k  δ k 0 σ k0   1+ µ



δk = 1 + e T
− 1 (12)
µk  µk 0 

2.1.2 Deterministic considerations


There are in the literature several definitions concerning sensitivity. However, one of them seems to be more
general and it can be defined by the fractional change of the observed variable produced by a fractional change
of the independent variables. It is straightforward to observe that such a mathematical statement has its
foundation in the definition of the total differential. Then, considering the Arrhenius equation and solving the
partial derivatives inserted into the total differential, the following equation can be obtained
dk dko E dT
= + (13)
k ko RT T

The equation developed above describes the behavior of sensitivity for a multiple input system.
Assuming the same degree of perturbations previously present in the process and inserting them into the
Arrhenius equation, the following approach can be developed
E

R (T + dT )
k + dk = (k0 + dk0 ) e (14)

or
 dT 
 
E  T 
RT  dT 
dk  dk0   1+ 
= 1 + e  T  −1 (15)
k  k0 

Using γ , θ and ρ to denote dk k , dk o k o and dT T , respectively, and given that the variables are

statistically independent, it can be easily shown that for small values of θ and ρ near the origin, the following

relationship is valid
E
γ =θ + ρ (16)
RT
It can be clearly observed that the sensitivity given by Eq.(16) is the same result expressed by Eq.(13).
2.2 For pH process control
Several researchers have contributed to the development of dynamic models applied to the pH process.
McAvoy was the first to use the concept of reaction invariant to overcome the difficulties in working with mass
balances for reactive systems. A methodology with a systematic form was developed and well-structured by
Waller and Mäkilä (1981) to obtain the set of reaction invariants. Manzi et al. (2004) has made concerted efforts
to apply such methodology in order to develop models for a neutralization system, and recently, they have also
shown how the system can be modeled for use in sensitivity analysis.
Consider the reactor for a neutralization process, as illustrated in Figure 1. Then, the following equations can
be easily derived which represent the dynamic behavior of the weak acid-strong base system, as reported by
Manzi et al.(1998)

VR
d (ξ − ε )
dt
[ ]
= − F1 (ξ − ε ) − F2 (ξ − ε ) − ξ 2 − F1 ε 1 (17)

where ξ = C B + , the concentration of the titration means, and ε = C Af − + C HAf represent the balance of carbon.

VR is reactor volume. The superscripts 1 or 2 denote the inlet stream.


Assuming the steady state condition, the following relationship can be obtained:

(C H )3 + (K a + ξ )(C H )2 +  F K+aF (F2ξ 2 − F1ε 1 )− K W  − KW K a = 0


+ + (18)
 1 2 
Equation (18) describes the titration curve and reveals the gain of the process for the open system considered.

Taking into account the titration means present in sufficiently high concentration such that F1 >> F2 ,
neglecting the contribution of the small terms and when perturbations present in the process variables or
disturbance are introduced into Eq.(18), the result is

dC H + 1 Ka ε
1  dε 1  1 K a 2
1+ = 1 + − +
( )2
(19)
C H+ 3 C +  1  3 C +
ε  3
H  H

for a weak acid-strong base system.


Using the considerations previously mentioned to solve Eq.(18), and assuming that k a << ε then

CH = kaε 1 for the weak acid-strong base. Then, substituting this relationship into Eq.(19), the following

expression can be obtained for the systems under study.


dC H + dε 1
1+ =1+ (20)
CH + ε1
Using the definition of p H and inserting the same degree of perturbations into the equation, the result is:
pH + dpH = − log (C H + + dC H + ) (21)

or under another form

  dC + 
log  1 + H 
  C 
dpH   H
+

= (22)
pH log C H +

Finally, substituting Eq.(20) into Eq.(22) gives


 dε 1 
log  1 + 1 
 ε 
dpH
=  (23)
pH log  Ka ε1 
 

The equation above governs the sensitivity behaviour for the system under consideration as a function of
disturbances.

3- Results and Discussion

3.1- For reaction temperature control


To investigate the application of methodology for sensitivity analysis related to reaction temperature control,
some conditions must be specified to carry out the simulation procedure. In this case, the nominal operating
conditions established for the reaction are the same as those used by Barron et al.(1999). For describing the
nonlinear behaviour of sensitivity in accordance with Eq.(12) or Eq.(15) and shown in Figure 2, it assumes the
values 32, 33 and 34 for E RT which correspond to T = 333 K , T = 323 K and T = 313 K , respectively. For

small disturbances, it is clear that the sensitivity presents a linear behaviour. Furthermore, the presence can also
be observed of an infinite number of surfaces of sensitivity for specified temperature limits. If a temperature
disturbance has occurred, in this case the return to the steady state must be led through surfaces to reach the
minimal trajectory of sensitivity. Since the permanence on level surfaces of sensitivity can correspond to a
residence time in the reactor and the surface transition related to temperature and consequently to the number of
reactor involved in the process, such an approach may be used to determine the number of reactors and the time
of residence in each one of them for minimizing the sensitivity.

Fig.2. The behaviour of sensitivity expressed by eqs. 12 and 15.

Figure 2 shows a family of curves which present overlapping of surface for meaningful values of dk o k o

and dT T . However, it should be noted that such values represent a large disturbance from a practical point of

view. Therefore, only one level surface can be considered for any disturbance presented, allowing the answer to
the critical challenge of how many reactors would be necessary to reach the requirements of performance
objective with minimum cost or computational load. In this case, only one reaction may be used to satisfy the
requirements for a minimum of sensitivity. This is already evidenced in practice.
For assessing the control performance as a function of the residence time and the number of reactors and then
to verify the above conclusions, the results have been examined by simulation and Figures 3a and 3b depict the
response of the process of the reaction temperature control in close loop with a classical PI in the presence of
disturbance in the inlet temperature, assuming the integral of absolute value of errors (IAE) as the performance
criterion.

a b

Fig.3. Performance of the reaction temperature control in the presence of disturbance in the inlet temperature from 350 to
300K e 350 to 400K.
Observing the results obtained above for the system, it is possible to conclude that no significant
contributions exist for improving the response for any pair of points in the horizontal plane. Hence, the
statements already mentioned are quite satisfactory.

3.2- For pH process control


Related to applying such an approach to a pH process control, specifically for weak acid-strong base systems,
consider a 3L reactor as shown in Figure 1. The inlet flow of 1.667 L/min is a solution of weak acid in water
neutralized by a strong base (1N). The tuning parameters for the classical PI controller are k c = 117 , τ I = 0 .01

and the set point remains pH set = 7 for all experiments.

a b

Fig.4. Response of a weak acid-strong base system in closed loop with a PI controller when submitted to disturbance in the
feed stream. The return is carried out in one or several stages.
It is instructive to observe that when the neutralization process is submitted to a disturbance in the inlet pH ,
the conventional PI controller displays a poor performance, similar to an offset, as shown in figure 4a. It can be
further observed that if the return to steady state is conducted through several smoothing stages, a better
performance will be obtained, as depicted in figure 4b. In the industrial environment, such an approach is
recognized as a smoothing tanks strategy.
From Figure 4b, the relationship can also be verified between the step disturbance and the smoothing tanks as
well as the width of each step and the residence time in each vessel. Furthermore, the response reveals a strong
dependence on the number of smoothing tanks and on the residence time in each tank.
Once the neutralization process is sufficiently sensitive to disturbance, then the success of such a strategy can be
analyzed and explained in the light of the typical behaviour of sensitivity expressed by Eq.(23) and depicted in
Figure 5.
0,00 0,20 0,40 0,60 0,80 1,00
0,000

-0,010

-0,020

-0,030
dpH/pH

-0,040

-0,050

-0,060

-0,070

-0,080

d ε /ε
1 1

dpH/pH(4) dpH/pH(5) dpH/pH(6) dpH/pH(7)

Fig.5. The typical behaviour of sensitivity expressed by Eq. (23) when the pH process is submitted to a change in influent pH
from 4, 5 or 6 to 7.

Figure 5 presents a family of curves given by the solution of a differential x y ' = y which can be considered
approximately straight lines. Figure 5 shows also the behaviour of sensitivity of the process when it is submitted
to a disturbance in the inlet pH , indicating a clear reduction of sensitivity, strongly desired, expressed by the
slope of each curve, when returned to the original state. If the same degree for all the disturbance is preserved,
this return must be led through the radial lines, and, based on a directional derivative concept, sensitivity will

decrease most rapidly to reach the minimal value in the direction ∇ φ = − ∂φ ∂ [ ( ) ] j , where φ represents the
dpH
pH

angular coefficient of the curves shown in Figure 5.


Since a correspondence exists between the residence time in each radial line and the residence time in each
smoothing tank, as well as the transition of a radial line to another being related to the number of tanks involved
in the smoothing process, it can be verified that the optimal strategy would consist of infinite transitions on
infinite radial lines. Thus, the residence time would be infinitesimal if infinite smoothing tanks were used.
From the process control point of view, this approach presents a critical challenge, namely: how many tanks
would be necessary to reach the required performance objective with minimum cost or computational load? The
answer to this question was obtained by simulation, assuming the integral of the absolute value of the error
(IAE) as the performance criterion for the response of the process in closed loop with the classical PI controller,
and the results of which have been shown in Figures 6a and 6b.
a b

Fig.6. Performance of the neutralisation process (weak acid-strong base) in closed loop with a PI controller when influent pH
changes from 7 to 5 (a) and 4 (b) and then returns to 7 by a step function for several residence times and smoothing tanks.
IAE was used as the performance criterion.

It is straightforward to observe from the Figures above, a plane surface indicating the best control
performance, meaning also a relative stability in the sensitivity. For more severe disturbance, a gradual reduction
of the surface can be verified too. It must be emphasized that such an analysis provides the minimum number of
tanks and residence time to reach the minimal sensitivity at the minimum cost.
Insights from sensitivity analysis suggest the idea of fictitious smoothing tanks which can be inserted in the
computer codes for process control, resulting in the use, in practice, of only one tank. This strategy also allows
the disadvantages and restrictions to be overcome when several smoothing tanks must be used, since using
several of them cannot be reasonable either from the technical or economic point of view.
Since Figure 6b shows how the performance of the process depends on the number of smoothing tanks and
the residence time for a disturbance in the inlet pH value from 7 to 4, the optimal values for both parameters can
be chosen as 9 and 0.06 min respectively. Then, using the same data previously mentioned for the process and
considering the imp lementation of fictitious smoothing tanks strategy, the response of a such system is described
in figure 7.

Fig.7. Response of a neutralisation process (weak acid-strong base) in closed loop with a conventional PI when the fictitious
smoothing tanks strategy is implemented. The pH was changed from 7 to 4.
Figure 7 shows the dynamic behaviour of the neutralization process under the fictitious smoothing tanks
strategy, revealing the ability of the classical PI controller to treat nonlinear features and dis turbance, without
additional cost of equipment, instrumentation or indeed without appreciable computational charge.

4- Conclusion
The remarkable improvement in the performance of the classical feedback control, particularly for PI
controller, was obtained by using sensitivity analysis applied to the pH process. Relative to reaction temperature
control, the analysis has shown that even for oversized values of disturbance, an overlapping of level surfaces of
sensitivity can be observed. The results presented in Figure 3 allow the conclusion that, in practice, the
smoothing reactors strategy does not have considerable effect on the control performance, that is, only one
reactor may be used to satisfy the requirements for a minimum sensitivity.
Based on the minimization of IAE as performance criteria, the responses for a weak acid-strong base system
for several feed conditions were obtained and the optimal number of smoothing tanks as well as the residence
time were easily found for each inlet pH condition. A new approach has been proposed for reducing the
sensitivity, making use of the so-called fictitious smoothing system, whose the results described in Figure 7,
show the good performance of the controller. Moreover, this approach can be introduced into the software for
process control without any additional cost or computational load.

References
Barron, M.A., González, J., Aguilar R. (1999). Disturbance Rejections of a Feedback Linearization Controller with
Control Input Constraints. Chem. Eng. Technol, 22, a3, 241-245.
Faanes, A, Skogestad S. (2004). pH-Neutralization: Integrated Process and Control Design. Computer and Chemical
Engineering, 28:1475-1487.
Heyen, G., Maréchal, E., Kalitventzeff, B. (1996). Sensitivity Calculations and Variance Analysis in Plant Measurement
Reconciliation. Computer and Chemical Engineering, 20, S539-S544.
Manzi, J T, Odloak D. (1998). Control and Stability Analysis of the GMC Algorithm Applied to pH Systems. Brazilian
Journal of Chemical Engineering, 15, 3, 247-264.
Manzi, J. T, Odloak D, Langer O. U (1998). Control of the Neutralization Process with the Generic Model Algorithm.
Chem. Eng. Technol, 21, 4, 369-376.
Manzi, J. T., Rocha, A. M., Bispo, H.(2004). pH Process Control with Fictitious Smoothing Tanks. Chem. Eng. Technol,
27, 12, 1309-1317.
McAvoy T J, Hsu E, Lowenthal, S.(1972). Dynamics of pH in Controlled Stirred Tank Reactor. Ind Eng. Chem. Process
Des. Dev., 11, 68-70.
Rivera, D. E., Morari, M., Skogestad, S.(1986). Internal Model Control. 4. PID Controller Design. Ind Eng. Chem.
Process Des. Dev., 25, 252-265.
Shinskey F.G.(1973). pH and pION Control in Process and Waste Streams. John Wiley & Sons. New York, NY.
Skogestad, S.(1996). A Procedure for Siso Controllability Analysis – with Application to Design of pH Neutralization
Processes. Computers Chem. Engng, 20, 4, 373-386.
Waller K V, Mäkilä, P.M. (1981). Chemical Reaction Invariants and Variants and Their Use in Reactor Modelling. Ind.
Eng. Chem. Process Des. Dev. 1981, 20, 1-11.

Acknowledgments
The authors would like to thank the Catholic University of Pernambuco for financial support .

You might also like