3484 Ind. Eng. Chem. Res.

1996, 35, 3484-3497

Method for Assessing the Effect of Design Parameters on

Controllability
Philip R. Lyman† and William L. Luyben*
Chemical Process Modeling and Control Research Center, Department of Chemical Engineering,
Lehigh University, Bethlehem, Pennsylvania 18015

Bjorn D. Tyreus
DuPont Central Science and Engineering, Experimental Station, P.O. Box 80101,
Wilmington, Delaware 19880

The interaction between design and control can offer significant opportunities for dynamic
performance improvement if one is willing to consider building and operating a chemical process
at other than the steady-state economic optimum. This paper describes a method that a process
designer could use to identify design and control parameters that could be used to improve the
controllability of a process. The method relies upon designed experimentation in two separate
steps. The first step uses steady-state information as a screening tool to filter the important
factors from the many possible design factors. The second step uses dynamic simulation and
provides more details about how the design and control decisions affect the ultimate dynamic
performance in the presence of disturbances and production rate changes. The capital and
operating costs of the various designs are calculated to indicate the economic penalty for designing
and operating at other than the steady-state optimum. Three examples are provided to illustrate
the method and to develop several design heuristics.

1. Introduction The objective of this study is to focus on the sensitivity

of the dynamic performance of processes to variations
Many chemical processes can be classified into those in process design parameters, control structures, and
that are easily controlled and those for which even the controller tuning. A method was developed that allows
most elaborate control schemes fail to perform ad- a designer to quickly screen design parameters to find
equately. This property of robustness to disturbances, those that are important dynamically. The method
or “controllability”, is a function of the process design makes use of a mathematical model of the process to
parameters as well as the control structure and algo- calculate steady states and to simulate the dynamic
rithm used. The controllability of a chemical processing response of the process to disturbances. Application of
plant is fixed partly by the plant design and partly by this method to three process examples of increasing
the control system design. The plant design can impact complexity leads to a set of generic rules that can be
the controllability through equipment size, flowsheet used to design more easily controlled processes. The
layout, and other decisions that can affect the dynamic material in this paper is taken from and described in
behavior of the plant. Dynamic behavior is usually greater detail in Lyman, 1995.
considered only during the control system design step. This paper is arranged as follows. The method for
At this latter stage of the design sequence, it may be identifying dynamically important design and control
too costly to make desired equipment design changes, factors is described. This is followed by three example
and this may result in a limitation on the ultimate processes. The first process is a reactor/column with
performance of the control structure. The dynamic reaction A f B and recycle of A. The second process
performance of the plant to anticipated disturbances consists of a reactor and two columns with two recycle
should be considered during the initial stages of the streams with reaction A + B f C (ternary process). The
process and equipment design. Since in practice this third process (quaternary process) features reaction A
may not always be possible due to time or engineering + B h C + D with a flowsheet consisting of a reactor/
resource constraints, it would be useful to have available stripper combination followed by two columns with two
a set of generic guidelines that could be incorporated recycle streams. The last section describes general
into the design to help the designer arrive at a more results common to all three processes and finalizes the
inherently stable process. It is generally considered design heuristics.
that there is a tradeoff between dynamic performance
and steady-state economics. An example of this tradeoff 2. Method
is that a large chemical reactor, while more costly, will
be less sensitive to feed composition disturbances. We intend to examine quantitatively four properties
Therefore, the generic guidelines must include consid- of a proposed process design in order to arrive at
eration of the economic penalties associated with process conclusions regarding its controllability and how it may
design modifications. be improved. Design in this context includes equipment
sizes, specifications, and flowsheet layout as well as the
control structure and controller tuning rules.
* Author to whom all correspondence should be addressed.
Telephone: (610)758-4256. Fax: (610)758-5297. Email: wll0@ The first two properties are the installed capital cost
lehigh.edu. and the operating cost of the plant. The capital cost is
† Present address: Air Products and Chemicals Inc., 7201 determined solely by the equipment and equipment
Hamilton Blvd., Allentown, PA 18195-1501. sizes required in the design and is estimated using
S0888-5885(95)00682-8 CCC: $12.00 © 1996 American Chemical Society
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3485
installed capital cost correlations (Douglas, 1988). The alternative selections of flowsheet layouts or different
operating cost of the design under steady-state condi- control structures, or they may be different values of a
tions can be quantified by assigning a cost for energy design parameter such as vessel size. The factor levels
and utilities. These first two properties of a proposed should be chosen to cover a large design space in the
plant design can be combined into a discounted cost of initial studies of a new process. This allows the effects
capital such as the net present value (NPV) (Peters and of factors on the controllability to be more easily
Timmerhaus, 1980). identified. As more is learned about the interactions
The third property is the steady-state disturbance of design and control for a specific process, the studies
sensitivity (DSA) of the plant to known disturbances and should cover a smaller number of factors and a smaller
production rate changes. This concept was described range of levels. If the feasible design region for the
initially by Luyben (1975) and more recently was particular process allows it, the study should be cen-
extended by Yi (1994). This sensitivity measure uses tered about the steady-state economic optimum design.
the deviations of the manipulated variables that control This allows the controllability to be studied as a function
product quality. In order to perform this analysis, it is of deviations from the optimum and can provide the
necessary to have a steady-state process model that can basis for the economic justification of design changes
be converged to new steady states that represent the to improve the controllability.
process after step disturbances or production rate Depending upon the number of factors and the level
changes have occurred. It is not necessary to define a of information desired regarding factor interactions, the
control structure at this stage with the exception of designer can select a full or a partial factorial designed
identifying the manipulated variables that control the experiment (Box and Hunter, 1961a,b). The partial
quality of the final product. In order to solve the steady- factorial design allows you to alias factors with factor
state model for given operating conditions, one must interactions and greatly reduces the number of cases
have an itemized list of controlled variables and the required in exchange for loss of some factor interaction
group of manipulated variables that will be used to information.
maintain them at their setpoints. At the end of the second step all of the information is
The fourth property is the controllability, which is known to allow the user to design each of the processes
typically deviation in product quality. It will depend represented by one case in the designed experiment. A
upon the disturbance(s) used in its evaluation as well complete process design includes known values for all
as the control system used in the process. The method vessel inventories and compositions and all stream
we propose uses dynamic simulation of each design and flowrates and compositions. Step 3 includes calculation
control combination in order to account for this depen- of the NPV using the capital and steady-state operating
dence. We will track a single or a small number of cost associated with each design. This measure is a
process variables whose deviations we desire to mini- comparison between the series of future cash flows
mize. provided by the initial capital outlay (sales of product
The proposed method consists of eight steps as minus the operating costs) and the rate of return
summarized below. possible from an alternative investment. The calcula-
1. Select design and control factors. tion of net present value requires some assumptions of
2. Assemble designed experiment. the lifespan of the project and the prevailing rate of
3. Conduct the steady-state design for each case. return. The values chosen are ROR ) 15% and span )
4. Perform disturbance sensitivity analysis (DSA) for 5 years. These two parameters do not affect compari-
each disturbance and setpoint change and if possible sons between the cases but act like scaling factors on
eliminate some of the design and control factors from the NPV’s.
further study. Step 4 is the calculation and comparison of steady
5. Recast the designed experiment based upon the states after disturbances have occurred (DSA). If the
results of the DSA or previous designed experiment dynamic transition of the process is viewed in a space
results. defined by the value of the manipulated variables used
6. Tune each control loop. to control the product quality variables, those processes
7. Simulate dynamics with disturbances. that have the endpoints of the transition close to the
8. Evaluate effect of factors. initial values will have better dynamic controllability.
The first step is to examine the process and select a This relationship will be shown by example. The
small number of design and control factors. These distance between the quality manipulated variables is
factors are typically such things as conversions, recycle calculated as
flowrates, overdesign factors, number of vessels or vessel
sizes, and flowsheet layout. Typical control factors can
vary the location of the throughput manipulator, select
dual or single-ended composition control, and choose
x∑ m
(Vm,new - Vm,old)2 (1)

alternative inventory control structures. The number for m columns when the vapor boilup rates are the
of factors and the number of levels of each factor manipulated variables used to control product quality.
determine the size of the problem. Since the method Step 4 generates a vector of the distance of the quality
includes the closed-loop dynamic simulation of each case manipulated variable moves, one for each case in the
for each disturbance, the number of factors should be designed experiment. The vector is then analyzed using
kept small; six or fewer is recommended. If the factors analysis of variance techniques (ANOVA) in order to
that are expected to interact can be grouped together, reach conclusions regarding the relative importance of
such as recycle flowrate and composition, then a series design factors on the controllability. This step serves
of executions of the method can be performed to handle as a filter on the factors allowing the elimination of some
a large number of factors. of them from further study. Because of the large
Step 2 is to organize the factors in a designed number of cases resulting from the inclusion of a single
experiment framework consisting of a set of cases in factor in the designed experiment, the elimination of one
which each factor is varied between two or three distinct or two factors can greatly reduce the number of cases
levels. Depending on the factor, the levels may be required in the dynamic simulation step.
3486 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996

One drawback of the DSA measurement is that it is Table 1. Process Parameters

dependent upon the scaling of the manipulated vari- constant description value
ables. This could be a problem when the manipulated F0 fresh feed flowrate 239.5 lb mol/h
variables are different or when the production rates Z0 fresh feed composition 0.9 mole fraction of A
differ. Attempts to modify the DSA measurement by A0 preexponential factor 70000 h-1
using the open-loop deviation of product compositions EA activation energy 15000 Btu/lb mol
resulted in a measure that was not indicative of control- U overall heat transfer coefficient 200 Btu/h/ft2/F
Xb bottoms composition 0.0105 mole fraction of A
lability. For this reason, the closed-loop form of the DSA Md0 holdup of reflux tank 69.4 lb mol
is used in this paper despite the scaling limitation. Mb0 holdup of column bottoms 89.4 lb mol
The factors that remain important after the DSA, R relative volatility 2.0
hw weir height 1 in.
including all of the control factors, are recast into a
smaller designed experiment in step 5.
purity of 0.0105 mole fraction A. The process param-
Step 6 involves controller tuning. The tuning proce- eters are given in Table 1.
dure must be done consistently for all of the designs in 3.2. Design Variables. The number of design
order to compare the results. The procedure used degrees of freedom available for this process is found
throughout this paper is pole placement for the inven- by subtracting the total number of equations from the
tory loops and dynamic simulation with a relay feedback total number of variables. For a detailed degree of
signal followed by a set of tuning rules for the temper- freedom analysis for the binary process, the interested
ature and composition loops. The relay feedback test reader is referred to Lyman, 1995. The binary process
provides an estimate of the ultimate gain and frequency has 7 design degrees of freedom for a specified number
for the loop (Astrom and Hagglund, 1984). The tuning of reactors. The number of trays in the column and the
rules used are the Tyreus-Luyben settings (Tyreus and feed tray location consume 2 degrees of freedom. The
Luyben, 1992). production rate is set with 1 degree of freedom. One
Step 7 is dynamic simulation of each of the cases. In degree of freedom is used to fix the reactor temperatures
order to obtain a controllability measure, it is necessary and one is used for product composition control. At this
to include disturbances in the simulation. The selection point there are 2 degrees of freedom remaining. Two
of these disturbances is quite important in determining design parameters (D and zn or D and Vr) can be selected
the final result; hence, their selection should be carefully where D, zn, and Vr are the recycle flowrate, composition
considered. It is important to select disturbances that of component A in the last reactor, and the reactor
are realistic in their location, frequency content, and inventory, respectively. With the degrees of freedom
magnitude. If random fluctuations can be anticipated consumed in this manner, the design of the process is
in a specific process variable, then this can also be fixed.
simulated. Once the design is complete, different control struc-
The next step is evaluation of the results of the tures result from fixing different variables than the ones
designed experiment. The goal in this step is to identify fixed in the design. For example, D and zn can be used
which factors influence which of the three properties to fix the design (including the calculation of xd, F, and
described above. This step uses an analysis of variance Vr, where xd and F are the distillate composition and
technique to assign a single value to each of the direct reactor effluent flowrate). However, in the operation
factor effects and each interaction. These measures can of the plant we can control F and Vr and let D and zn
be compared to determine the relative importance of the vary. This example produces the control structure
factors and interactions in determining the controllabil- shown in Figure 1.
ity of the process. Time-dependent results from indi- 3.3. Design Procedure. The design of the process
vidual simulations can be compared to provide expla- can proceed once values are specified for each of the
nations for the differences observed. assumed known variables and the 7 degrees of freedom
If desired, the entire procedure can be repeated either have been satisfied. In the procedure described below
to further refine initial findings (using larger or smaller the number of reactors in series (all having the same
ranges for the same factors) or to investigate additional holdup and temperature), the fresh feed composition
factors. (z0), the Arrhenius preexponential and exponential
constants (A0f, Ef), the heat exchanger area and heat-
transfer coefficient (Ah, U), the temperatures of the fresh
3. Example 1sBinary Process feed, cooling water at the heat exchanger inlet, and the
3.1. Process Description. The process is derived recycled stream (T0, Tcw,in, and Td, respectively), and the
from the Papadourakis system (Papadourakis, 1985) reflux ratio design factor (RR factor) are known. The
and will be referred to as the binary process in this reactor temperatures (T) and the product purity (xb) are
paper. The process is shown with a specific control also fixed. The production rate (B) is fixed. The
structure in Figure 1. Two components exist in the example described here fixes D and zn, and the reactor
process, A and B. Component A is fed to a series of volumes required to achieve the production rate are
CSTR’s where a first-order exothermic liquid phase calculated.
reaction occurs, producing component B. Arrhenius The method used to generate the steady-state design
kinetics are used to model the reaction rate constant is as follows:
as a function of the temperature within the reactors. 1. The fresh feed flowrate is equal to the product
The exothermic heat of reaction is removed by recircu- flowrate since the reaction is equimolar.
lating a portion of the reactor contents through an
external heat exchanger. The effluent stream from the F0 ) B (2)
last reactor (a liquid mixture of A and B) is fed to a
distillation column. Component A, the light component, 2. The reactor effluent flowrate is calculated from an
is withdrawn as the distillate stream and is recycled overall material balance on the reactor section.
back to the first reactor. The product stream is the
column bottoms. There is a specification on product F ) F0 + D (3)
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3487

Figure 1. Process diagram for CSTR/column system.

3. The recycle composition is calculated from a reflux ratio, and the vapor boilup rate can then be
component balance on the column. calculated. Saturated liquid feed to the column is
assumed.
Fzn - (F - D)xb
xd ) (4)
D yf )
Rxf
(11)
1 + (R - 1)xf
4. The reactor volumes and reactor effluent concen-
trations are calculated using equations for CSTR’s in (xd - yf)/(xd - xf)
series: RR ) (RR factor) (12)
1 - (xd - yf)/(xd - xf)
F(1 - (zn/zin)1/n)
Vr ) (5) 7. A column design program is used to determine the
k(zn/zin)1/n number of trays and the optimum feed tray location.
This program starts at the bottom of the column with
zn ) zin ( F
F + Vrk ) n
(6)
the known vapor boilup rate and bottoms composition
and calculates liquid and vapor compositions for each
tray of the column. The procedure is repeated until the
5. The cooling water flowrates and cooling water vapor composition is greater than the distillate composi-
outlet temperatures are calculated from the known tion (Luyben and Wenzel, 1988). The resulting number
reaction rates for each reactor and assumed values for of trays is rounded up to an integer value.
the heat-transfer coefficients and the cooling water inlet 8. A column rating program is used to find the reflux
temperature for each. ratio required to obtain the desired split with an integer
number of trays (Luyben and Wenzel, 1988). This
Reactor 1 adjusts the tray compositions slightly.
3.4. Control Structures and Tuning. The control
Q1 ) -MCp(FT1 - F0TF0 - DTd) - λkVrz (7) structure for this system is shown in Figure 1. The
reflux drum and the column base liquid level controllers
Reactor n are assumed to be perfect controllers, as is the flowrate
controller on the reactor outlet stream. The liquid level
in each CSTR is controlled by using the inlet flowrate
Qn ) -MCpF(Tn - Tn-1) - λkVrzn (8) as the manipulated variable. The first reactor utilizes
the fresh feed flowrate for this purpose. The setpoints
-Qn for these level controllers are used to control the
Tcw,n ) + Tn (9) throughput of the process. This control structure was
UAh
found to minimize the snowball effect (Luyben, 1993a-
-Qn c). The controller tuning parameters for the level loops
Fcw,n ) (10) are calculated directly from the linear model for this
McwCp,cw(Tcw,in - Tcw,n) subsystem using a controller reset time of 0.5 h and
choosing the proportional gain such that the roots of the
6. The minimum reflux ratio is calculated from the closed-loop characteristic equation are located on the
known column feed and top compositions using the negative real axis on a root locus plot. The composition
Underwood equations. Multiplying the minimum reflux of the column bottoms is controlled using the vapor
ratio by the RR factor (typically 1.2) gives the actual boilup rate, V. Control structure 1 controls the distillate
3488 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996

Table 2. Design Factors

factor
number symbol description level value
1 CS control structure -1 structure 1, dual
1 structure 2, single
2 RR factor ratio of actual to minimum reflux ratio -1 1.2
1 1.4
3 n number of CSTR’s -1 1 reactor
1 2 reactors
4 D distillate flowrate -1 260.5 lb mol/h
1 260.5 × 1.1 lb mol/h
5 zn composition of column feed -1 0.4 mole fraction
1 0.5 mole fraction
6 Ah heat exchanger area -1 1000 ft2
1 3000 ft2

Table 3. Partial Designed Experiment Array effects of six factors at two levels each. As a screening
case 1 2 3 4 5 6 study, a 1/4 fraction of the full 26 design is selected. This
is referred to as a 26-2 design. It is constructed as a
1
normal 24 design with an additional two factors assigned
-1 -1 -1 -1 -1 -1
2 1 -1 -1 -1 1 -1
3 -1 1 -1 -1 1 1 as aliases to interaction effects. The resulting design
4 1 1 -1 -1 -1 1 has 24 ) 16 experiments and a resolution of 4. This
5 -1 -1 1 -1 1 1 resolution will provide information about the main effect
6 1 1 1
7
-1
1 1
-1 -1
for each factor aliased with three factor interactions
only. Also, some two factor interactions are aliased with
-1 -1 -1 -1
8 1 1 1 -1 1 -1
9 -1 -1 -1 1 -1 1 other two factor interactions (Box and Hunter, 1961a,b;
10 1 -1 -1 1 1 1 Montgomery, 1984). The levels for each factor for each
11 -1 1 -1 1 1 -1 case are shown in Table 3. The numbers within the
12 1 1 1
13
-1
1 1
-1
1
-1
table reflect the level of the particular factor that will
be used in the corresponding case. Each case corre-
-1 -1 -1
14 1 -1 1 1 -1 -1
15 -1 1 1 1 -1 1 sponds to an individual design followed by loop tuning
16 1 1 1 1 1 1 and a 15 h simulation in which the process is subjected
to a known disturbance.
composition by manipulating the reflux flowrate (dual- 3.7. Design Results. Step 3 is the calculation of the
ended composition control). Control structure 2 fixes steady-state design for each case. Table 4 shows the
the reflux flowrate (single-ended composition control). results. The steady-state design allows the calculation
The distillate and the bottoms composition loops were of the steady-state operating cost and the capital cost
tuned independently and operated together with no of the equipment. The capital cost for each design is
detuning necessary.
estimated using correlations from the literature (Dou-
3.5. Design and Control Factors. The first step
glas, 1988). The maximum NPV is provided by case 1
in the proposed method is to select design and control
(small RR factor, single CSTR, small D, small zn, and
factors. In this example we wish to determine the
small Ah). The NPV is separated into operating and
influence, if any, of a group of six design factors on the
closed-loop dynamic performance of the entire process. capital costs to study the influence of each of the factors.
These design factors are described in Table 2. The first The 10% change made in the recycle flowrate has no
factor is the control structure. The second factor is a significant influence on the process capital cost. The
multiplication factor that determines the actual reflux number of reactors, n, is the main contributor to the
ratio relative to the minimum reflux ratio for each determination of the capital cost. Using more reactors
design. The remaining factors are self-explanatory. increases the capital cost because of the low conversion
3.6. Design Factor Variation. Step 2 is assembly (40 to 50%). The annual operating cost is proportional
of a designed experiment. The design factors’ influence to the steady-state vapor boilup rate and hence has a
upon the dynamic performance measurement will be minimum corresponding to the minimum in the vapor
evaluated using a fractional factorial designed experi- boilup rate. In turn, the minimum vapor boilup rate
ment. In this study, we desire information about the corresponds to a low RR factor and a low value for D.

Table 4. Process Design Results

case xd (mol %) Vr (lb mol) RR NT trays NF feed tray V (lb mol/h) R (lb mol/h) NPV MM$
1 0.76 1883.28 1.28 19 14 595.05 334.55 4.950
2 0.95 1506.63 2.03 22 13 789.93 529.43 4.563
3 0.95 1506.63 2.32 19 12 864.14 603.64 4.286
4 0.76 1883.28 1.48 16 12 645.43 384.93 4.718
5 0.95 638.15 2.03 22 13 789.93 529.43 4.006
6 0.76 772.77 1.28 19 14 595.05 334.55 4.393
7 0.76 772.77 1.48 16 12 645.43 384.93 4.648
8 0.95 638.15 2.32 19 12 864.14 603.64 4.216
9 0.73 1883.28 1.04 19 15 585.04 298.49 4.809
10 0.91 1506.63 1.70 21 13 774.60 488.05 4.447
11 0.91 1506.63 1.91 18 12 833.97 547.42 4.526
12 0.73 1883.28 1.19 16 13 628.63 342.08 4.915
13 0.91 642.39 1.70 21 13 774.60 488.05 4.374
14 0.73 778.67 1.04 19 15 585.04 298.40 4.736
15 0.73 778.67 1.19 16 13 628.63 342.08 4.356
16 0.91 642.39 1.91 18 12 833.97 547.42 3.967
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3489
Table 5. DSAsFactor Levels, ∆V’s, and ISE’s for Load
Disturbance
factor
case 1 2 3 4 5 6 ∆V ISE (105)
1 -1 -1 -1 -1 -1 -1 8.42 0.809
2 1 -1 -1 -1 1 -1 12.86 3.176
3 -1 1 -1 -1 1 1 0.67 0.309
4 1 1 -1 -1 -1 1 10.35 1.366
5 -1 -1 1 -1 1 1 3.93 0.468
6 1 -1 1 -1 -1 1 10.17 1.178
7 -1 1 1 -1 -1 -1 10.76 1.330
8 1 1 1 -1 1 -1 12.93 3.658
9 -1 -1 -1 1 -1 1 11.42 1.516
10 1 -1 -1 1 1 1 12.73 3.029
11 -1 1 -1 1 1 -1 4.34 0.381
12 1 1 -1 1 -1 -1 10.30 1.471
13 -1 -1 1 1 1 -1 0.83 0.140
14 1 -1 1 1 -1 -1 10.18 1.138
15 1 1 1 1 14.85 3.007
Figure 2. Dynamic simulation results for case 1 (solid) and case
-1 -1
16 1 1 1 1 1 1 12.87 3.302
13 (dashed).
Table 6. Calculated Effects for Load Disturbance
effect Example simulation results are shown in Figure 2 for
two designs: case 1 (largest NPV) and case 13 (best
factor DSA dynamic (105) dynamically). The action of the distillate composition
mean 18.45 3.28 controller is shown in the top two plots. The middle
1 4.64 1.30 plots show the impurity composition in the product
2 0.81 0.421 stream and the vapor boilup. Case 13 exhibits better
3 0.68 0.270
4 0.93 0.211 controllability.
5 -0.69 -0.102 Analysis of the designed experiment results from each
6 -0.70 -0.212 set of 16 cases includes calculation of the overall effect
12 -0.99 -0.321 of each of the factors. The effects are tabulated versus
13 5.76 1.672 the factor or factor interaction in the rightmost column
14 0.98 0.163
of Table 6. The table is interpreted by picking out the
23 0.0271
most important effects and noting which factors are
-0.70
24 -3.16 0.331
34 -0.98 -0.179 responsible for that effect. Those effects that are
124 0.72 -0.116 deemed significant are displayed in boldface. Control
134 0.78 0.259 structure 1, in which dual composition control is used,
1234 -0.83 -0.401 provides better dynamic performance than control struc-
3.8. Disturbance Sensitivity Analysis (DSA). ture 2 (factor 1 is the control structure). There is also
Step 4 is performed by converging the model equations an important interaction between control structure and
to new steady states for given disturbances and compar- the number of reactors. The others are relatively
ing the distances that the manipulated variables used unimportant. This confirms the results of the DSA.
to control product quality must move. Since we have a These factors could have been eliminated prior to the
single quality manipulated variable for this process, this dynamic simulation step.
distance is the same as ∆V. Table 5 shows the DSA 3.10. Binary Process Conclusions. The case with
values for a disturbance in feed composition: 0.9-1.0 the largest NPV is case 1. This design has a small RR
mole fraction of component A. The calculated effects factor, single CSTR, small D, low zn, and small Ah. The
for each factor and interaction are shown in Table 6. design with the best controllability is case 13. The
Based upon these results, the important factors and economic penalty associated with case 13 is a difference
interactions for controllability are predicted to be factors in the NPV of $576 000. This process features dual-
1, 13, and 24 (factor 13 means the interaction between ended control and differs from the economic optimum
factors 1 and 3). DSA shows that CS 1 (dual-ended in three of the five design factors. The case 13 design
composition control, factor 1 ) -1) will give better has two CSTR’s, large D, and large zn. In order to
controllability. DSA factor interaction results must be improve the dynamic performance, dual composition
discounted based upon results to be shown later; control should be used. This recommendation entails
however, the main factors participating in the interac- an economic penalty associated with the additional
tion should be retained in the dynamic simulation. analyzer and its maintenance. A cheaper alternative
Factors 5 and 6 could be eliminated from the study after is to ratio the reflux flowrate to the column feed
this analysis. They will be retained for dynamic simu- flowrate. This does not require an additional composi-
lation to compare the simulation results with the DSA tion measurement and gives better controllability than
results. single-ended control. This alternative only makes sense
3.9. Dynamic Simulation. The dynamic model of for control structures that do not fix the reactor effluent
the process consists of a set of unsteady-state mass and flowrate. A low value for the reflux ratio design factor
energy balances describing the reactor and the distil- also improves the controllability. Over design of the
lation column (Lyman, 1995). heat exchanger surface area offers no dynamic benefit
The process is simulated for 15 h of simulation time and there is an economic penalty, so the best value is a
with a disturbance implemented at 1 h into the simula- small one as long as it is large enough to allow the
tion. Several different disturbance types and locations reactor to withstand the largest expected disturbance.
were tested. To illustrate the method, a single distur- The best choice for the number of reactors, zn, and D is
bance (a step in the fresh feed composition from 0.9 to not clear, and further testing must be done for these
1.0 mole fraction of component A) will be described here. factors.
3490 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996

Figure 3. Ternary process with two recycles.

Table 7. Process Design Factors

Another important point is that there are many
designs that have worse dynamic performance than case level
13 and are more expensive than case 1. Spending more factor 1 2 3
money on the plant does not always improve the B1 20 60 100
dynamics. You need to identify the proper changes to D2 10 40 70
make to improve the controllability. xd2 [0.99 0 0.01] [0.90 0 0.10] [0.80 0 0.20]
The designed experiment allows numerous factors to
be sorted efficiently into those that significantly affect distillate D2 (component A recycle, light recycle), the
the dynamic performance and those that do not. In composition of the A recycle stream xd2,C, and the
order to obtain physical understanding of the results, flowrate of the first column bottoms B1 (component B
further simulations must be performed in which the recycle, heavy recycle). The levels of each of these
design or control differences can be isolated. The shape factors are shown in Table 7 (xd2 values are liquid
of the disturbances tested did not give significant compositions for components A, B, and C, respectively).
differences in the conclusions (results not shown). The optimum steady-state design is the center point of
However, the disturbance location greatly influences the this designed experiment. The maximum and the
conclusions. It is very important to test those distur- minimum values are equidistant from the center point
bances that are the most likely to occur to arrive at in each factor, and one of the endpoints of each factor
useful conclusions. is the same as the maximum or minimum value
DSA could have been used to filter out two of the examined in the search for the optimum. Each case in
factors prior to the dynamic simulation step. This would the factorial design has a unique reactor volume,
have saved simulation effort. number of trays, vapor boilup, and reflux rates for both
columns as determined by the levels of D2, B1, and xd2.
4. Example 2sTernary Process 4.2. Design Results. The reactor volumes, operat-
ing cost, and NPV are shown in Table 8. Case 14 has
4.1. Process Description. This process and its
the largest NPV and is the steady-state economic
control were studied by Tyreus and Luyben, 1993. The
optimum design.
reaction is A + B f C where component A is the lightest
and component B is the heaviest. Two columns are used 4.3. Disturbance Sensitivity Analysis. The ma-
to separate the reactants remaining in the reactor nipulated variables that control product quality are V1
effluent stream and return them to the reactor in two and V2. The manipulated variable changes are calcu-
recycle streams. The process is shown in Figure 3. The lated and compared at two steady states for a load
first column separates unreacted component B (lowest disturbance in which the fresh feed of B changes from
volatility), and the second column separates unreacted pure B to 95% B and 5% A.
component A (highest volatility). The distance between the new (post disturbance) and
The design calculations are conducted using a proce- old (initial) steady-state values of V1 and V2
dure similar to that used by Tyreus and Luyben (1993). x(∆V1)2 + (∆V2)2 (lb mol/h)) is shown in Table 9 along
The details are given in Lyman, 1995. with the ISE results (calculated later in the dynamic
The design procedure was repeated for each design simulations). ANOVA results shown in Table 10 show
of a three-by-three factorial design (27 cases). The three that low values for the recycle flowrates, B1 and D2, give
factors varied were the flowrate of the second column lower values for the DSA measurement, and therefore
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3491
Table 8. Process Design Results for the Ternary Process
Vr capital cost oper cost NPV
case B1 D2 xd2 (lb mol) (106$) (103$) (106$)
1 1 1 1 7305 2.322 222.7 8.678
2 1 1 2 7962 2.384 218.2 8.63
3 1 1 3 8847 2.5 219.3 8.511
4 1 2 1 2971 1.604 290.5 9.169
5 1 2 2 3260 1.625 285.9 9.163
6 1 2 3 3655 1.693 284.6 9.099
7 1 3 1 2419 1.517 338.2 9.095
8 1 3 2 2658 1.517 338.2 9.095
9 1 3 3 2984 1.576 333.6 9.052
10 2 1 1 4302 1.787 258.2 9.094
11 2 1 2 4689 1.818 254 9.077
12 2 1 3 5210 1.892 255.2 8.998
13 2 2 1 1599 1.296 321.6 9.372
14 2 2 2 1754 1.296 317.8 9.385
15 2 2 3 1967 1.334 317.3 9.348
16 2 3 1 1221 1.245 370 9.26
17 2 3 2 1341 1.225 371 9.278 Figure 4. Load disturbance, example dynamic response.
18 2 3 3 1506 1.254 367 9.261
19 3 1 1 3965 1.707 281.7 9.095 Table 10. Results of ANOVA on DSA for a Load
20 3 1 2 4322 1.721 278.6 9.091 Disturbance for the Ternary Process
21 3 1 3 4802 1.788 279.9 9.02
22 3 2 1 1390 1.244 348.9 9.333 factor or sum of squares
23 3 2 2 1526 1.238 345.2 9.35 factor interaction x(∆V1)2 + (∆V2)2(104)
24 3 2 3 1711 1.27 345.1 9.319
25 3 3 1 1016 1.195 395.8 9.225 total 4.622
26 3 3 2 1116 1.169 397.1 9.246 B1 2.662
27 3 3 3 1253 1.192 394.2 9.232 D2 1.926
xd2 0.010
Table 9. Ternary Process DSA and ISE Results B1, D2 0.014
B1, xd2 0.005
load dist D2, xd2 0.005
case B1 D2 xd2 DSA ISE (108) B1, D2, xd2 0.0006
error 0
1 1 1 1 215.5 0.0796
2 1 1 2 208.0 0.1955 Table 11. Results of ANOVA on Integral Squared Error
3 1 1 3 209.2 0.1875 of xb2,A + xb2,B for a Load Disturbance Using CS 1
4 1 2 1 265.0 1.8628
factor or sum of squares
5 1 2 2 257.4 3.2947
6 1 2 3 255.1 1.5884 factor interaction (10+17)
7 1 3 1 286.1 5.2970 total 9.92
8 1 3 2 284.7 7.4279 B1 0.21
9 1 3 3 276.3 5.0183 D2 8.20
10 2 1 1 260.5 0.1579 xd2 0.50
11 2 1 2 255.4 0.2977 B1, D2 0.68
12 2 1 3 257.1 0.3130 D2, xd2 0.07
13 2 2 1 299.5 1.3730 B1, xd2 0.19
14 2 2 2 294.5 2.3596 all three 0.06
15 2 2 3 294.1 1.8556 error 0
16 2 3 1 320.8 4.3203
17 2 3 2 321.8 4.7103 1 as proposed by Tyreus and Luyben (1993). Another
18 2 3 3 316.9 3.8958
19 3 1 1 296.4 1.3367
control structure (CS 4 of Tyreus and Luyben, 1993) was
20 3 1 2 293.2 1.8055 also studied, and the results were essentially the same.
21 3 1 3 295.0 1.4740 At 1 h of simulated time the composition of Fob
22 3 2 1 335.4 1.7546 changes from pure B to 5% A and 95% B. The impact
23 3 2 2 332.1 2.8707 of this disturbance is illustrated in Figure 4. As the
24 3 2 3 332.9 1.8769 additional A is fed to the system, some of the A reacts
25 3 3 1 353.6 3.9791
26 3 3 2 356.1 5.2978 with B, causing a drop in the B inventory in the bottoms
27 3 3 3 354.5 3.7023 of column 1 and an increase in the A inventory in the
reflux drum of column 2. The flowrates of fresh feeds
better controllability is predicted for these designs. The adjust appropriately. The net result is a production rate
physical explanation for this effect is that low recycle increase.
flowrates give larger reactor volumes (Table 8) that The load disturbance is simulated on all of the
provide greater filtering of the disturbance in the designs, and the resulting integral squared errors were
column feed stream. Based on the DSA results, the xd2 analyzed using ANOVA to give the results shown in
factor could be eliminated. In this designed experiment Table 11. These results indicate that the flowrate of
this would reduce the experiment to two factors at three the A recycle stream (D2) is the main factor that
levels for each or 9 cases instead of the 27 cases with determines the controllability of the process to the load
all three factors. disturbance. Furthermore, a process design with a
Similar DSA results were found for production rate small D2 and a large reactor is more controllable.
changes. 4.5. Ternary Process Conclusions. The steady-
4.4. Dynamic Controllability. The controllability state design of the ternary system with two recycles
of the process is determined using dynamic simulations indicates that the highest NPV is obtained by using
with a load disturbance and a production rate change moderate recycle flowrates and reactor sizes. The use
(not shown) for each case. The control structure is CS of a load disturbance to evaluate the controllability of
3492 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996

Table 12. Product Stream Compositions Table 13. Design Constants for the Ternary Process
composition flowrate parameter value units
stream component (mol fr) (lb mol/h)
z0a 1/0/0/0 mol fr
D2 A 0.010 0.5076 z0b 0/1/0/0 mol fr
B 0.005 0.2538 R 9/1.5/4/1
C 0.985 50.0 U 200 Btu/h/ft2/°F
D 0 0 Ah 1000 ft2
T 140 °F
total 1.0 50.761 Tf0 140 °F
B3 A 0 0 Td3 140 °F
B 0.005 0.251 Tcw,in 100 °F
C 0 0 Hvap 250 Btu/lb
D 0.995 50.0 λ -10000 Btu/lb mol
total 1.0 50.251 M 50 lb/lb mol
Cp 0.75 Btu/lb/°F
these designs shows that those processes that have a Mcw 18 lb/lb mol
large reactor and a small light recycle are more easily Cp,cw 1.0 Btu/lb/°F
controlled. The production rate change (not included kf0 0.2968 × 1011 h-1
here) shows that a large reactor with both recycles small kr0 0.1930 × 107 h-1
gives better controllability. These highly controllable Ef 30000 Btu/lb mol
Er 20000 Btu/lb mol
processes have low reactant concentrations in the reac-
tor; particularly for the light reactant. This can be Table 14. Design Factors for the Ternary Process
shown by locating the six lowest ISE designs on Table
factor level value units
9. The best designs have small values of zA. The best
process using dynamic controllability considerations is Vr 1 4000 lb mol
2 5000
case 1, while the economic optimum process is case 14.
3 6000
The cost of the improved controllability is a decrease of D3 1 45 lb mol/h
$707 000 in the NPV. 2 55
3 65
5. Example 3sQuaternary Process xd3 1 0/0.985/0.010/0.005 mol fr
2 0/0.850/0.100/0.050
The third process studied consists of a reactor, a 3 0/0.700/0.200/0.100
stripper, and two distillation columns with four com-
ponents from the work of Yi, 1994. light recycle, and intermediate purity light recycle) at
5.1. Process Description. The reaction is A + B $3 310 000. For this wide range of designs the NPV is
h C + D that occurs in the liquid phase and is fairly flat. The lowest NPV is case 25 at $2 850 000 (a
exothermic. The relative volatilities are assumed to be difference of only $460 000).
constant: RA ) 9, RB ) 1.5, RC ) 4, RD ) 1. The process 5.3. Disturbance Sensitivity Analysis. The con-
is shown in Figure 5 with one control structure in place. trol structure is shown in Figure 5.
The lightest component A is removed from the reactor The load disturbance is a shift in the fresh feed
effluent using a stripping column. The bottoms product composition of A from 1/0/0/0 to 0.95/0.05/0/0 in mole
of the stripper is fed to a second column for removal fractions of A/B/C/D. The manipulated variables that
and purification of product C, and the bottoms of column control product quality are the three vapor boilup
2 is fed to column 3 for removal and purification of streams. DSA tracks the distance between the two
product D. The distillate stream from column 3 con- steady states in terms of the three vapor boilup rates
tains mostly B and is recycled to the reactor. There are (x(∆V1)2 + (∆V2)2 + (∆V3)2).
four unit operations and two recycle streams. The steady-state distance measurements are shown
5.2. Design Results. As a basis for the process in Table 18. By comparing values in Table 18, one can
design, the reaction rate is fixed at 50 lb mol/h. The find the directions of the factor levels that are predicted
compositions of the product streams are assumed to be to give improved controllability. The higher the purity
as shown in Table 12. Fixing the reaction rate at 50 lb of xd3,B, the better is the DSA. Larger D3’s and Vr’s give
mol/h also fixes the flowrate of both C and D in the lower DSA values. Each of these factor directions tends
product streams at 50 lb mol/h. Since the product to reduce the magnitude of the disturbance that is
streams are not pure C and D, the total flowrates are passed to the columns. This is explained in the follow-
slightly larger than 50 lb mol/h, and hence the flowrates ing section. The factor effects show that none of the
of the fresh feeds are slightly larger also. The fresh main factors can be eliminated from the dynamic testing
feeds F0a and F0b are assumed to be pure A and B, of the load disturbance. These results are compared
respectively. A further assumption is that the recycle with the dynamic simulation results in the next section.
stream from the third column contains no A component The DSA results predict that case 25 is the best design.
(xd3,A ) 0). This is a good assumption since any A that 5.4. Dynamic Controllability. The dynamic con-
makes it to the bottoms stream of the stripping column trollability of the quaternary process is measured using
will go out with the D2 product. The reactor tempera- the integral squared error of the sum of deviations in
ture is 140 °F, and the reaction rate constants for the each of the impurity levels in both product streams. The
forward and reverse reactions are 0.35 and 0.10 h-1, deviations that are added together are deviations in
respectively. All of the design constants and parameters variables: xd2,A, xd2,B, xd2,D, xb3,A, xb3,B, and xb3,C. The
are shown in Table 13. ISE is calculated for each of the 27 cases using control
The parameters Vr, D3, and xd3 are selected to fix the structure 1 with single-ended control on column 3 for a
plant designs. These factors and their levels are shown load disturbance consisting of a shift in the fresh feed
in Table 14. composition of A from 1/0/0/0 to 0.95/0.05/0/0 in mole
The design procedure is detailed in Lyman, 1995. The fractions of A/B/C/D.
resulting designs are shown in Tables 15-17. The An example of the dynamic responses is shown in
highest NPV is provided by case 2 (small reactor, small Figure 6 using the results from case 1. The reactor
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3493
Table 15. Design Results for the Quaternary Process (Cases 1-9)
case
1 2 3 4 5 6 7 8 9
Vr (lb mol) 4000 4000 4000 4000 4000 4000 4000 4000 4000
D3 (lb mol/h) 45 45 45 55 55 55 65 65 65
xd3,B (mol fr) 0.985 0.850 0.700 0.985 0.850 0.700 0.985 0.850 0.700
xd3,C (mol fr) 0.010 0.100 0.200 0.010 0.100 0.200 0.010 0.100 0.200
xd3,D (mol fr) 0.005 0.050 0.100 0.005 0.050 0.100 0.005 0.050 0.100
Nt1 (trays) 11 11 13 10 11 12 10 10 11
Nt2 (trays) 19 15 13 20 15 13 20 15 13
Nt3 (trays) 52 40 35 52 40 35 52 40 35
zA (mol fr) 0.296 0.372 0.515 0.246 0.304 0.405 0.209 0.265 0.351
zB (mol fr) 0.182 0.138 0.089 0.220 0.176 0.124 0.259 0.203 0.148
zC (mol fr) 0.333 0.316 0.253 0.345 0.338 0.307 0.345 0.352 0.333
zD (mol fr) 0.190 0.174 0.143 0.188 0.183 0.165 0.187 0.180 0.169
F (lb mol/h) 300 340 430 305 328 380 307 338 380
V1 (lb mol/h) 154 194 284 149 172 224 141 172 214
V2 (lb mol/h) 139 130 128 142 138 130 151 142 132
V3 (lb mol/h) 269 261 240 304 284 256 341 316 279
R2 (lb mol/h) 89 80 77 91 87 79 100 91 81
R3 (lb mol/h) 224 216 195 249 229 201 276 251 214
B1 (lb mol/h) 146 146 146 156 156 156 166 166 166
B2 (lb mol/h) 95 95 95 105 105 105 115 115 115
cap cost (1000 $) 1646 1548 1531 1675 1563 1518 1708 1580 1523
oper cost (1000 $) 308 320 357 326 325 334 346 345 342
NPV (1000 $) 3254 3310 3204 3166 3279 3294 3063 3196 3263

Table 16. Design Constants for the Quaternary Process (Cases 10-18)
case
10 11 12 13 14 15 16 17 18
Vr (lb mol) 5000 5000 5000 5000 5000 5000 5000 5000 5000
D3 (lb mol/h) 45 45 45 55 55 55 65 65 65
xd3,B (mol fr) 0.985 0.850 0.700 0.985 0.850 0.700 0.985 0.850 0.700
xd3,C (mol fr) 0.010 0.100 0.200 0.010 0.100 0.200 0.010 0.100 0.200
xd3,D (mol fr) 0.005 0.050 0.100 0.005 0.050 0.100 0.005 0.050 0.100
Nt1 (trays) 11 11 12 10 10 11 9 10 11
Nt2 (trays) 19 15 13 20 15 13 20 15 13
Nt3 (trays) 52 40 35 52 40 35 52 40 35
zA (mol fr) 0.254 0.314 0.408 0.212 0.265 0.344 0.184 0.228 0.297
zB (mol fr) 0.192 0.151 0.108 0.230 0.182 0.134 0.262 0.213 0.159
zC (mol fr) 0.353 0.346 0.313 0.361 0.365 0.344 0.365 0.370 0.361
zD (mol fr) 0.201 0.189 0.171 0.197 0.188 0.178 0.189 0.189 0.182
F (lb mol/h) 284 313 359 292 320 354 306 322 352
V1 (lb mol/h) 138 167 213 136 164 198 140 156 186
V2 (lb mol/h) 139 130 128 142 138 130 151 142 132
V3 (lb mol/h) 269 261 240 304 284 256 341 316 279
R2 (lb mol/h) 89 80 77 91 87 79 100 91 81
R3 (lb mol/h) 224 216 195 249 229 201 276 251 214
B1 (lb mol/h) 146 146 146 156 156 156 166 166 166
B2 (lb mol/h) 95 95 95 105 105 105 115 115 115
cap cost (1000 $) 1764 1663 1626 1794 1677 1626 1825 1698 1638
oper cost (1000 $) 299 305 318 318 321 320 346 336 327
NPV (1000 $) 3166 3245 3240 3071 3180 3234 2948 3107 3199
composition controller increases F0a to maintain zA. The The largest effect upon the ISE is due to a two-factor
excess B that enters with F0a in this disturbance collects interaction between D3 and xd3. The ISE values for each
in the third column reflux drum and the level controller combination of D3 and xd3 are averaged together and
reduces F0b. V1 increases to bring xb1,A back to setpoint. plotted in Figure 8. Small D3 and more impure xd3
The impurities in D2 and B3 are controlled by V2 and result in especially large ISE’s and therefore poor
V3. This case does not use dual-ended composition controllability. The main contributor to the ISE is the
control; hence, xd3,B can vary. deviation of xd2,A. This can be explained using steady-
The ISE values that result from the dynamic simula- state arguments as follows. Small D3 and more impure
tions are shown in Table 18. The best process using xd3 result in a low flowrate of component B returned to
dynamic considerations alone is case 26, although the the reactor in the recycle stream. This must be com-
four best cases (cases 26, 17, 27, and 14) have very close pensated for by a large zA to achieve the same produc-
ISE’s. tion rate. A large vapor boilup rate in the stripper is
An ANOVA analysis of this vector gives the design required to remove the A from the reactor effluent. The
factor and factor interaction effects. These are shown flowrate of reactor effluent is correspondingly large. The
plotted along side the DSA results in Figure 7. From load disturbance results in a reactor effluent flowrate
these results, all three factors are seen to be equally change, while the reactor composition of A is fixed. The
important, and there is a significant two-factor interac- net result is that small D3 and more impure xd3 designs
tion between D3 and xd3. Examination of the ISE values have a larger reactor effluent flowrate (F) that carries
shows that larger reactor volumes, larger D3 flowrates, a larger disturbance to the stripper expressed in lb mol
and intermediate xd3 compositions give better control- A/h. This value is calculated and averaged together for
lability. Specific case comparisons can be used to each combination of D3 and xd3 in the same manner
understand these findings. described above. The combination of D3 and xd3 that
3494 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996

Figure 5. Quaternary process with control structure 1.

Table 17. Design Constants for the Quaternary Process (Cases 19-27)
case
19 20 21 22 23 24 25 26 27
Vr (lb mol) 6000 6000 6000 6000 6000 6000 6000 6000 6000
D3 (lb mol/h) 45 45 45 55 55 55 65 65 65
xd3,B (mol fr) 0.985 0.850 0.700 0.985 0.850 0.700 0.985 0.850 0.700
xd3,C (mol fr) 0.010 0.100 0.200 0.010 0.100 0.200 0.010 0.100 0.200
xd3,D (mol fr) 0.005 0.050 0.100 0.005 0.050 0.100 0.005 0.050 0.100
Nt1 (trays) 10 10 11 10 10 11 9 10 11
Nt2 (trays) 19 15 13 20 15 13 20 15 13
Nt3 (trays) 52 40 35 52 40 35 52 40 35
zA (mol fr) 0.232 0.287 0.365 0.191 0.238 0.306 0.165 0.205 0.266
zB (mol fr) 0.195 0.154 0.113 0.237 0.189 0.142 0.269 0.220 0.166
zC (mol fr) 0.370 0.368 0.342 0.370 0.378 0.365 0.372 0.381 0.377
zD (mol fr) 0.203 0.192 0.180 0.202 0.195 0.188 0.194 0.195 0.190
F (lb mol/h) 283 310 343 284 308 336 298 313 338
V1 (lb mol/h) 137 164 197 128 152 180 132 147 172
V2 (lb mol/h) 139 130 128 142 138 130 151 142 132
V3 (lb mol/h) 269 261 240 304 284 256 341 316 279
R2 (lb mol/h) 89 80 77 91 87 79 100 91 81
R3 (lb mol/h) 224 216 195 249 229 201 276 251 214
B1 (lb mol/h) 146 146 146 156 156 156 166 166 166
B2 (lb mol/h) 95 95 95 105 105 105 115 115 115
cap cost (1000 $) 1872 1770 1728 1906 1788 1735 1937 1810 1748
oper cost (1000 $) 299 304 309 314 315 310 342 331 319
NPV (1000 $) 3060 3144 3167 2974 3089 3158 2850 3013 3114

gives the largest deviation in the flowrate of A to the to xd2,A) and hence will give a lower ISE value. The
stripper corresponds to the combination that gives the manipulated variables move less for a larger reactor,
poorest controllability. as shown in the plots on the right side of Figure 9.
The larger the reactor inventory, the smaller is the The effect of different recycle flowrates can be seen
steady-state change in F. Also, the dynamic transition
in Figure 10. In this figure cases 2, 5, and 8 have D3
gets easier. This is shown by comparing three cases in
flowrates of 45, 65, and 75 lb mol/h, respectively. These
which only the reactor inventory differs. Cases 3, 12,
and 21 have reactor inventories of 4000, 5000, and 6000 cases were selected such that the other factors are the
lb mol/h, respectively. The other two factor levels are same for all three. The larger the recycle flowrate, the
the same for all three. The change in FzA (the distur- better is the controllability. As the recycle flowrate is
bance imposed upon the separation section) when mov- increased, there is more component B in the reactor and
ing between these two steady states is smaller for the less A is needed to achieve the same production rate
larger reactor processes. For cases 3, 12, and 21 the (the reactor temperature and volume are held constant).
steady-state changes in FzA are 4.51, 2.97, and 2.49 lb Since the compositions of component A in the reactor
mol A/h, respectively. The larger reactors also act to are held constant in each design, the steady-state
reduce the dynamic deviations in zA as shown in Figure deviation in the stripper feed can be expressed as zA∆F.
9. This results in a smaller deviation in xb1,A (equivalent With the larger recycle flowrates, zA is lower and the
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3495
Table 18. DSA Results for the Quaternary Process

case Vr (lb mol) D3 (lb mol/h) xd3,B (mol fr) x(∆V1)2 + (∆V2)2 + (∆V3)2 ISE results (109)
1 4000 45 0.985 4.17 0.1857
2 4000 45 0.850 5.03 0.1620
3 4000 45 0.700 6.84 0.5308
4 4000 55 0.985 3.78 0.0906
5 4000 55 0.850 4.20 0.1198
6 4000 55 0.700 5.14 0.2072
7 4000 65 0.985 3.42 0.1683
8 4000 65 0.850 4.03 0.0637
9 4000 65 0.700 4.78 0.1018
10 5000 45 0.985 3.88 0.1117
11 5000 45 0.850 4.52 0.1270
12 5000 45 0.700 5.42 0.2572
13 5000 55 0.985 3.59 0.1509
14 5000 55 0.850 4.09 0.0579
15 5000 55 0.700 4.70 0.1122
16 5000 65 0.985 3.43 0.1281
17 5000 65 0.850 3.79 0.0579
18 6000 65 0.700 4.36 0.0823
19 6000 45 0.985 3.89 0.0967
20 6000 45 0.850 4.50 0.0780
21 6000 45 0.700 5.14 0.1391
22 6000 55 0.985 3.48 0.0900
23 6000 55 0.850 3.93 0.0848
24 6000 55 0.700 4.41 0.0860
25 6000 65 0.985 3.33 0.1789
26 6000 65 0.850 3.66 0.0567
27 6000 65 0.700 4.15 0.0579

Figure 6. Example dynamic simulation results using case 1 with

a load disturbance. Figure 8. Effect of D3 and xd3 interaction on ISE for CS 1 with
load disturbance.

Figure 7. Factor and interaction effects for CS 1 with load

disturbance. Figure 9. Effect of reactor volume (Vr) on dynamic controllability.

change in FzA is lower. This change is 2.54, 1.81, and in their recycle compositions but have the same values
1.49 lb mol A/h for cases 2, 5, and 8 respectively. for the other two factors. This figure shows that the
The dynamic responses of two impurity concentra- most impure recycle (case 3) has the largest deviation
tions are plotted in Figure 11 for three cases that differ in xb1,A while the other two are nearly identical. This
3496 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996

to a lesser extent recycle purity. The binary process

exhibited improved controllability with a large reactor
and small recycle flowrate. A potential benefit was
found in the use of dual composition control. The
ternary process controllability is improved with a large
reactor and a small light recycle flowrate. A small
heavy recycle flowrate was also shown to improve the
controllability during a production rate change. The
quaternary process gives better controllability with
designs that feature a large reactor, large heavy recycle
flowrate, and a more pure heavy recycle stream. The
use of dual composition control improves the control-
lability during a production rate change (not shown).
The quaternary process results are consistent with
those from the simpler processes. For fixed Vr and xd3,
a small light recycle stream coincides with a large heavy
recycle stream (D3) to achieve the same production rate.
Figure 10. Effect of recycle flowrate (D3) on dynamic control-
lability. For fixed Vr and D3, a smaller ynt1,A results from a more
pure heavy recycle stream (large xd3,B). Note that less
component A is required to drive the reaction with a
greater amount of component B in the reactor.
The observation of the interaction of design and
control in these recycle processes allow a set of design
heuristics to be identified. These heuristics can be used
to provide guidance to the process designer. The first
three heuristics are based upon designing processes in
which changes in the reactant flowrates to the separa-
tion section (Fzj where j is for the reactants) are small.
1. The recycle process should be built with a large
reactor. This will drive the product of the reactant
concentrations (zAzB for the ternary process) down for
a given production rate and reaction rate constant. The
reaction rate constant (reaction temperature) could be
used in the same way for a given production rate and
reactor holdup, although this may give undesirable side
Figure 11. Effect of recycle composition (xd3) on dynamic control-
reactions or cause other problems. The larger-sized
lability. reactor, while providing improved controllability, entails
greater capital costs and introduces safety and envi-
corresponds with the ISE values for these three cases. ronmental concerns corresponding to the greater mate-
This can be explained by comparing the absolute steady- rial storage.
state deviation in the impurity flow in the reactor 2. Light recycle streams (recycle of the more volatile
effluent stream. ∆(FzA) for cases 1, 2, and 3 are 1.85,
reactant) should be kept small. Often there exists a
2.54, and 4.51 lb mol A/h. The process with the largest
tradeoff in the recycle flowrates. Once k and Vr have
deviation here has the largest ISE. The higher purity
been fixed, the product of the reactant concentrations
recycle sends more component B to the reactor, giving
a lower steady-state value of zA. This tends to reduce is fixed for a given production rate. Since the heuristic
the deviation in the reactor effluent stream and lower recommends a smaller flowrate of light recycle, the
the ISE. heavy recycle flowrate must necessarily be larger to give
5.5. Quaternary Process Conclusions. From a the same product of reactant concentrations (zAzB for
steady-state economic point of view, the plant that the ternary process).
should be built is represented by case 2. It has a small 3. Light recycle streams should be impure in the light
reactor (4000 lb mol), a small D3 recycle flowrate (45 lb component. The direction for the heavy recycle compo-
mol/h), and an intermediate recycle composition (0/ sition is less clear. A more pure heavy recycle will tend
0.850/0.100/0.005). Although this plant has the highest to increase the composition of the heavy component and
NPV ($3.310 × 106), it does not give the best control- reduce the composition of the light component in the
lability to disturbances. The best process for the load reactor. This may be beneficial if the light component
disturbance is case 26. composition is sufficiently reduced. A less pure heavy
Controllability with respect to the load disturbance recycle will tend to reduce the composition of the heavy
is strongly affected by all three design factors. The DSA component in the reactor, which could also have a
indicates this without the need for the dynamic simula- beneficial impact on the controllability. The direction
tions. In order to improve the controllability to the load in which to move the heavy recycle purity depends upon
disturbance, the reactor size should be large, as should the initial steady state.
the heavy recycle flowrate. Highly impure heavy recycle 4. There is a control benefit possible by using dual
streams are to be avoided. composition control or reflux/feed ratio control (Lyman,
1995; Lyman and Luyben, 1996).
6. Heuristics
These rules have been found to apply well to these
For the processes studied here, the most important three processes. However, further testing of the rules
dynamic factors are reactor size, recycle flowrate, and on different types of processes may be required.
 Ind. Eng. Chem. Res., Vol. 35, No. 10, 1996 3497
Nomenclature z0b,j ) composition of fresh feed of B [mole fraction of
component j]
Ah ) heat exchanger area [ft2] zin ) composition of feed to reactor [mole fraction of
A0 ) reaction preexponential factor [h-1] component A]
A0f ) forward reaction preexponential factor [h-1] zn ) composition of feed to column [mole fraction of
Bm ) column m bottoms flowrate [lb mol/h] component A]
Cp ) reaction mixture heat capacity [Btu/lb/°F]
Cp,cw ) cooling water heat capacity [Btu/lb/°F] Greek Symbols
Dm ) column m distillate flowrate [lb mol/h] Rj ) relative volatility of component j
Ea ) activation energy [Btu/lb mol] λ ) overall heat of reaction [Btu/lb mol]
Ef ) forward reaction activation energy [Btu/lb mol]
Er ) reverse reaction activation energy [Btu/lb mol]
F ) column feed flowrate [lb mol/h] Literature Cited
Fcw,n ) cooling water flowrate [lb mol/h] Astrom, K.; Hagglund, T. Automatic tuning of simple regulators
F0 ) fresh feed flowrate [lb mol/h] with specifications on phase and amplitude margins. Auto-
F0a ) fresh feed flowrate of A [lb mol/h] matica 1984, 20 (5), 645.
F0b ) fresh feed flowrate of B [lb mol/h] Box, G.; Hunter, J. The 2k-p fractional factorial designs. Part I.
hw ) weir height [in.] Technometrics 1961a, 3(3), 311-351.
Hvap ) heat of vaporization [Btu/lb] Box, G.; Hunter, J. The 2k-p fractional factorial designs. Part II.
Technometrics 1961b, 3 (4), 449-458.
k ) reaction rate constant [h-1]
Douglas, J. Conceptual Design of Chemical Processes; McGraw-
kf ) forward reaction rate constant [h-1] Hill Book Co.: New York, 1988.
kr ) reverse reaction rate constant [h-1] Luyben, W. Steady-state energy conservation aspects of distillation
M ) reaction mixture molecular weight [lb/lb mol] column control system design. Ind. Eng. Chem. Fundam. 1975,
Mbm ) column m bottoms holdup [lb mol] 14, 321.
Mcw ) cooling water molecular weight [lb/lb mol] Luyben, W. Dynamics and control of recycle systems. 1. Simple
Mdm ) column m reflux drum holdup [lb mol] open-loop and closed-loop systems. Ind. Eng. Chem. Res. 1993a,
n ) number of equal volume reactors 32, 466-475.
Nf,m ) column m, location of feed tray Luyben, W. Dynamics and control of recycle systems. 2. Com-
Nt,m ) column m, total number of trays parison of alternative process designs. Ind. Eng. Chem. Res.
1993b, 32, 476-486.
Qn ) reactor n heat transfer rate [Btu/h] Luyben, W. Dynamics and control of recycle systems. 3. Alterna-
Rm ) column m reflux flowrate [lb mol/h] tive process designs in a ternary system. Ind. Eng. Chem. Res.
RR ) reflux ratio 1993c, 32, 1142-1153.
RR factor ) ratio of actual to minimum reflux ratio Luyben, W.; Wenzel, L. Chemical Process Analysis; Prentice-Hall
t ) time [h] Inc.: Englewood Cliffs, NJ, 1988.
T ) reaction mixture temperature [°F] Lyman, P. A Method for Assessing the Effect of Design Parameters
Tcw,n ) reactor n cooling water temperature [°F] on Controllability. Ph.D. Thesis, Lehigh University, 1995.
Tcw,in ) inlet cooling water temperature [°F] Lyman, P.; Luyben, W. Production rate changes in a ternary two-
Tdm ) column m distillate temperature [°F] recycle process. Ind. Eng. Chem. Res. 1996, 35, 2198-2203.
Montgomery, D. Design and Analysis of Experiments, 2nd ed.;
T0 ) fresh feed temperature [°F] Wiley: New York, 1984.
U ) overall heat-transfer coefficient [Btu/h/ft2/°F] Papadourakis, A. Stability and Dynamic Performance of Plants
Vm ) column m vapor boilup rate [lb mol/h] with Recycle. Ph.D. Thesis, University of Massachusetts, 1985.
Vr ) reactor liquid holdup [lb mol] Peters, M.; Timmerhaus, K. Plant Design and Economics for
Vrn ) reactor n liquid holdup [lb mol] Chemical Engineers, 3rd ed.; McGraw Hill: New York, 1980.
xbm,j ) column m bottoms composition [mole fraction of Tyreus, B.; Luyben, W. Tuning PI controllers for integrator/
component j] deadtime processes. Ind. Eng. Chem. Res. 1992, 31, 2625-2628.
xdm,j ) column m distillate composition [mole fraction of Tyreus, B.; Luyben, W. Dynamics and control of recycle systems.
component j] 4. Ternary systems with one or two recycle streams. Ind. Eng.
Chem. Res. 1993, 32, 1154-1162.
xb ) column bottoms composition [mole fraction of compo-
Yi, C.-K. Design and Control of Coupled Reactor/Column Systems.
nent A] Ph.D. Dissertation, Lehigh University, 1994.
xd ) column distillate composition [mole fraction of com-
ponent A] Received for review November 10, 1995
xf ) column feed composition [mole fraction of component Revised manuscript received May 14, 1996
A] Accepted May 15, 1996X
yf ) vapor composition in equilibrium with feed [mole
fraction of component A] IE950682C
zj ) composition of reaction mixture [mole fraction of
component j]
z0a,j ) composition of fresh feed of A [mole fraction of X Abstract published in Advance ACS Abstracts, September

component j] 15, 1996.