Académique Documents
Professionnel Documents
Culture Documents
pubs.acs.org/IECR
ABSTRACT: This work uses an open-loop analysis of a novel plate reactor to determine the optimum number of feed points and
their locations to achieve maximum conversion. Once the open-loop design is optimized, a control system is designed to achieve
maximum conversion of the reactants by regulating the reactor temperature using discrete proportional-integral-derivative (PID)
controllers. It was found that the optimum number of feed points is two, equidistantly spaced from each other, and also that a PID
controller design based on the internal model control method gives the best closed-loop performance.
r 2012 American Chemical Society 860 dx.doi.org/10.1021/ie2019245 | Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
Figure 2. Left: Schematic of a few rows of a reactor plate. The uid enters at the top; the small arrows indicate the ow direction. Y1 and Y2 mark the
locations of temperature sensors. Middle: A side view of the plate reactor. Right: A drawing of the ow inserts and the ow channels.1
Table 1 lists the parameters and variables associated with this concentrations form a system of nonlinear partial differential
process. Crucial issues in the design of the plate reactor are the equations (NLPDEs):2
number and locations of the distributed feed points that give the
optimal conversion. In the example, H2O2 is the component to
be distributed at the multiple feed points and the concentration Tr 2 Tr Tr 4h H
De vr Tr Tw r
of Na2S2O3 at the reactor exit is the criterion used to determine t z 2 z dr Fr cp Fr cp
the optimal conversion. Tw 2 Tw Tw 4h
Computational Model. In the study by Haugwitz2 a one- De vw Tr Tw
t z 2 z dw Fw cp
dimensional (1D) model was sufficient to represent the domi- CA 2 CA CA
nant characteristics of the plate reactor because the cross- Dm vr 2r 2
sectional area is much smaller than the channel length. In other t z 2 z
CB CB
2
CB
words, the cross-sectional gradients of temperature and concen- Dm vr 4r
tration are negligible when compared to the gradients along the t z 2 z
reactor length. The cooling plate component of the reactor is CC CC
2
CC
Dm vr r
approximated as a cooling jacket with a diameter (dw) that t z 2 z
surrounds the reactor. Because the flow rate of the reactants is
assumed to be instantaneous, pressure dynamics are neglected.2 Here r (mol/m3 3 s) represents the reaction rate defined in eq 1,
The 1D model is derived from appropriate mass and energy t is time, z is the spatial coordinate along the length of the reactor,
balances. The constitutive equations that describe the reactor Tr (C) is the reactor temperature, Tw (C) is the cooling water
and cooling water temperatures and the reactant and product temperature, CA (mol/m3) is the concentration of Na2S2O3, CB
861 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
Figure 5. Two-feed-point design. O, Tr; /, Tw; 3, Na2S2O3 conversion. Left: Entrance and 1/2 the reactor length. Right: Entrance and 1/3 the reactor
length.
Figure 6. Three-feed-point design: entrance, 1/3 the reactor length, and 2/3 the reactor length. O, Tr; /, Tw; 3, Na2S2O3 conversion. Left: Equal feed
rates. Right: Design to achieve the same maximum temperature, 63.5 C.
reactors in series. The heat of reaction is distributed between the (70 C) measured after the second feed point is less than 90 C
two reactors, and the designed cooling water capacity must by 22%. Comparing the designs (not shown) with one feed at the
permit Tr e 90 C. The final conversion of Na2S2O3 achieved entrance and the other at either 2/3 or 1/2 the length of the
with this design is 96.15%. The left panel of Figure 5 shows the reactor, the nal conversion in the former is found to be 5% less.
reactor and cooling water temperatures and conversion of Three-Feed-Point Design. Consider three feed points
Na2S2O3 along the reactor length. equally spaced: entrance, 1/3 the length of the reactor, and
If, instead, the design is altered such that the second feed point 2/3 the length of the reactor. If the feed rate is the same at each
is located at 1/3 the reactor length, the conversion achieved is feed point, then the highest reactor temperature is achieved
98.95% (see right panel of Figure 5). Since the second feed point after the second feed point, as shown in the left panel of
is located nearer the rst one (cf. second feed point at 1/2 the Figure 6. The amount of heat generated in the first reactor
reactor length), there is not sucient cooling capacity and section is less compared to the one- and two-feed-point designs
residence time to remove the heat of reaction generated at the due to the smaller feed rate, shorter residence time, and
rst feed point. The reactor temperature (125 C) measured sufficient cooling capacity to remove the heat of reaction.
after the second feed location is found to exceed 90 C by Between the second and third feed points the heat of reaction
almost 40%. generated is such that the designed cooling capacity is insuffi-
Consider another two-feed-point design where the second cient; thus, Tr > 90 C. The behavior in the final third of the
feed point is located further from the entrance of the reactor, in reactor is similar to that in the first third of the reactor. The final
this case 2/3 the reactor length. For this design, the temperature conversion is 85.35%.
863 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
If the three-feed-point design is such that the location of the above dierent cases, it was found that when the maximum
feed points gives the same maximum reactor temperature in each reactor temperature of each section is close to the temperature
section, i.e., between any two consecutive feed points, then the constraint the achievable conversion is about 96% regardless of
required feed rates are 40.42, 18.33, and 41.25%, respectively. how many feed points are selected. However, more feed points
The highest temperature achieved is 63.5 C with a conversion of mean a more costly design. A two-feed-point reactor design, with
89.25% (right graph of Figure 6). the entrance and midpoint of the reactor length and with the
The number of feed points, their locations, and their feed rates same feed rate, appears to be the best choice from this iterative
could be selected to achieve the highest allowable (near 90 C) investigation.
reactor temperature in each reactor section with the aim of Optimizing the Feed-Point Locations. According to the
maximizing conversion (see Figure 7). If the second and third above analysis, when the highest reactor temperature achieved
feed points are located at 20 and 42% of the length of the reactor in each reactor section (as delineated by the number of feed
and the feed rates from feed point 1 to feed point 3 are 50, 16.67, points) is close to the reaction temperature constraint, a high
and 33.33%, the conversion achieved is 95.88%. conversion is possible. To optimize the location, assume equally
A reactor design with four feed points also was studied.9 spaced feed points. The accumulated difference between the
Similar to the case of a reactor design with three feed points, the maximum temperature in each reactor section and the constraint
conversion was nearly the same only when the maximum reactor is minimized by adjusting the feed rate at each feed point. The
temperature in the four reactor sections was 90 C. From the objective function, , to be minimized is given by
m
min
qB, j > 0
jTj, max T j
j1
Tj, max e T 90 C 4
Figure 8. Left: Conversion for dierent numbers of equally spaced feed points. Right: Accumulated dierence between Tj,max and T* for the dierent
numbers of feed points.
Figure 11. Closed-loop response of the T1C. 3, BLT; ), AMIGO; /, IMC. Left: Reactor temperature, T1. Right: Cooling water ow rate, qw1.
Figure 12. Closed-loop response of the T2C. 3, BLT; ), AMIGO; /, IMC. Left: Reactor temperature, T2. Right: Cooling water ow rate, qw2.
To apply the AMIGO design, the process dynamics for stable where gc(s) is the controller transfer function. The AMIGO
processes are approximated by a rst-order process with time design is based on the MIGO design method but with a
delay (FOPTD): constraint on the maximum value of the sensitivity function:
1 dy 1:2 e M max jSj e 2:0 9
Kut >0
dt A value of 1.4 is stated to be a good compromise between
where K is the process steady-state gain, is the process time performance and robustness in many cases.13,14
constant, and is the time delay. Applying the Laplace transform strom et al.5 showed that, for a unit step disturbance, the
operator (L ) with the proper limits, the above time-domain integrated error was equal to 1/KI, where KI (=Kc/I) is the
model can be converted to a Laplace transform model in the integral gain. By maximizing the integral gain (or minimizing the
complex variable s: integrated error) to make M = 1.4, the PID controller parameters
can be determined.15 The parameters obtained from the MIGO
K design procedure are correlated to the parameters (K, , and )
gs es 7 of the process transfer function.14 Approximating this correlation
s 1
with an algebraic relation gives the AMIGO tuning rules (see
where gs L yt=L ut. Table 2) for an FOPTD transfer function.
Dene the sensitivity function: Internal Model Control. An internal model controller (IMC)
design is based on a realizable model of the process.6,16 It is well-
Ss 1 gs gc s1 8 known that the ideal feedback controller is the inverse of the
866 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
Table 4. Performance Criteria of the T1C and T2C con- Table 5. Closed-Loop Dominant Poles of the Controller
troller designs Designs
BLT AMIGO IMC BLT AMIGO IMC
rise time (s) T1C 21.2 2.3 6.0 real value T1C 0.7019 0.1705 0.5231
T2C 30.7 2.7 6.9 T2C 0.6852 0.2215 0.6338
settling time (s) T1C 4.2 3.5 2.3 imaginary value T1C (0.0369
T2C 6.6 2.2 2.8 T2C (0.0438
set-point dev (%) T1C 0.82 0.64 0.63
T2C 0.84 0.66 0.65
reactor section as shown in Figure 10. This gure is similar to that
decay ratio T1C 0.005
found in ref 2 with two feed points and one cooling water stream
T2C 0006 actuator. Two discrete PID controllers are designed to regulate
the CVs to the same set point of 90 C. The MVs are the
respective cooling water ow rates, qw1 and qw2, in each section. If
process; by providing a model of the process and the desired
the closed-loop dynamics of the T1:qw1 controller (T1C) was
control objectives, it is straightforward to obtain the appropriate
designed to be faster than that of the T2:qw2 (T2C) controller, it
controller structure and parameters. However, it is not always
was found that the rst controller will have almost no impact on
feasible to obtain the exact inverse of the process model when the
the second one. This choice of MVs and CVs mitigates cross-
model has a time delay. In that case, only the inverse of the couplings.
realizable part is used.
The FOPTD models are identied from a series of step
Procedurally, factor the transfer function model ~g (s) = ~g +(s)
response tests applied to the nonlinear process model introduced
~g (s) such that ~g (s) contains only the realizable parts and
above. The time domain data that relate the eect of the MVs on
~g +(s) contains everything else. Dene the IMC controller by
the CVs can be transformed to a Laplace domain model in the
c(s) ~g 1(s) f(s), where f(s) is a lter in the form f(s) =
complex variable s. Using matrix notation, these relations can be
(s + 1)n. The order n of the lter is chosen to guarantee that
written compactly as
the IMC controller is at least proper. For example, in the case of " # " #" #
an FOPTD model of the process, assigning n = 1 gives a T1 g11 s g12 s qw1
realizable design. The parameter can be adjusted to deter-
T2 g21 s g22 s qw2
mine the speed of the closed-loop response. As decreases,
the closed-loop response becomes faster and may become
where
unstable. Conversely, as increases, the closed-loop response 2 3
is stable but sluggish. To compare the IMC and the AMIGO " # 0:55e0:3s
designs, is chosen to satisfy eq 9. Table 2 lists the three PID g11 g12 6 2:94s 1 0 7
Gs 6
4
7
controller designs. g21 g22 0:62e0:44s 5
0
2:75s 1
DECENTRALIZED CONTROL The use of computers to automate control is a common
For the simultaneous objectives of safe operation and max- practice in many industries. Thus, all sensor signals are sampled
imum conversion, the reactor temperature should be regulated at at discrete points in time, t0, t1, ..., tj, and all actuator commands
90 C. The obvious variables to be controlled (CVs) are T1 and are implemented as discrete command signals. However, the
T2, the maximum reactor temperatures measured in each reactor process remains a continuous operation. To design a digital
section of the two-feed-point design. controller, it is necessary to discretize the operation of the
For convenience and accurate control, Haugwitz et al. select process. Here, we apply the z-transform of G(s) by choosing a
the feed temperature of Na2S2O3 (TA,feed) and the cooling water proper sample time t. In practice it is recommended that the
inlet temperature Tcool as the manipulated variables (MVs).2 sample time be selected such that t e 0.1dom.10,19 In the plate
That study pairs T1 with TA,feed and T2 with Tcool. Simulated reactor design, dom 3 s; thus t = 0.3 s is selected as the
results show that these pairings have slow closed-loop response sample time. The resulting discrete transfer functions are
2 3
times and there is some observable degree of interaction between 0:053 1
the CVs. An RGA analysis cannot conrm that there will not be z 0
6 z 0:903 7
strong interactions. Further, the physics of the system shows that Gz 6 4 0:035z 0:029 2 5
7
Tcool introduced at the entrance to the reactor clearly aects what 0 z
z 0:897
happens downstream from the entrance.
Because each plate reactor has its own cooling jacket, two Although the forms of the two transfer functions, g11 and g22,
cooling water streams can be used to control the peak tempera- are similar in the s-domain, their discrete counterparts in
tures in the reactor and to decouple the interactions. Figure 9 the z-domain are dissimilar. This is because their time delays in
shows the reactor and cooling water temperatures and the the s-domain are dierent: g22 has one additional time delay when
conversion rate of Na2S2O3. Comparing this gure with the left compared to g11. The parameter values of the three PID
panel of Figure 5, the values and locations of T1 and T2 do not controllers are listed in Table 3.
change. The conversion rate of Na2S2O3 remains unchanged. Comparison of the Three Digital Controller Designs. We
In the present design, two actuators (cooling water stream consider two constant disturbances (d1 and d2) in the energy
valves) are available to regulate the reactor temperature in each content of the feeds at both feed points.2 These disturbances will
867 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
Figure 13. Closed-loop response of the T1C. 3, BLT; ), AMIGO; /, IMC. Left: Reactor temperature T1. Right: qw1.
Figure 14. Closed-loop response of the T2C. 3, BLT; ), AMIGO; /, IMC. Left: Reactor temperature T2. Right: qw2.
be implemented as step functions: smaller. Further, since there are no time delays in the disturbance
transfer function, the disturbances will have a faster effect on T1
dTr 1 d1
f 1Tr , Tw , CA , CB , CC , u and T2 when compared to the effect of the actuators. The discrete
dt Fr cp Vr versions of eq 11 are
10
dTr 151 d2
f 151Tr , Tw , CA , CB , CC , u 0:0041 0:0037
dt Fr c p V r gd1 z gd2 z 12
z 0:657 z 0:684
( (
0 t e 70 s 0 t e 160 s Simulated results show that, together, the effect of both dis-
d1 d2
100 J=s t > 70 s 100 J=s t > 160 s turbances results in a 2.8% increase in T1 and a 3.2% increase
in T2.
where f[k] = dTr[k]/(dt) in eq 3. The Laplace transfer functions To compare the closed-loop control performance among the
of these two disturbances are three digital PID designs (BLT, AMIGO, and IMC), the follow-
0:0119 0:0117 ing criteria are used: rise time, settling time, ratio of the maximum
gd1 s gd2 s 11 deviation to its set-point value, and the decay ratio.
0:714s 1 0:789s 1
Figures 11 and 12 show the closed-loop results for the T1C
Comparing these to the process transfer functions in the and T2C controllers, respectively, with the dierent PID feed-
s-domain, both the process time constants and the process back controller design methods. Table 4 lists values of the per-
steady-state gains of the disturbance transfer functions are formance criteria.
868 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
Figure 15. Bode graphs of T1C. Top: BLT. Middle: AMIGO. Bottom: IMC.
From Table 4, the BLT design gives the longest rise and T1C: The set-point deviations in the case of the AMIGO and
settling times and the greatest set-point deviation. The value of IMC designs are nearly the same: 0.64 and 0.63%, respectively.
the detuning factor makes the closed-loop response sluggish The AMIGO performance has a longer settling time, and there
when compared to the other controller designs. are some oscillations in the closed-loop response. The IMC
869 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
performance shows a larger rise time when compared to the T1C and T2C controllers increase by 16.7 and 20%, respec-
AMIGO performance. A conclusion that can be reached is that tively, when compared to the noise-free case.
the IMC controller design gives the best closed-loop perfor-
mance to regulate T1.
T2C: The closed-loop response with the BLT design shows a CONCLUSION
sluggish response. The AMIGO performance gives shorter rise A formal analysis of a novel plate reactor system was carried
and settling times when compared to the IMC performance; out to determine the optimal number of distributed feed points
however, there are some oscillations. It is not dicult to conclude for one of the feed components. It was shown that a reactor
that the IMC controller design gives the best closed-loop design with two feed points equally spaced along the reactor
performance to regulate T2C. length was optimal assuming the same component feed rate.
The characteristics of the closed-loop responses of these Regulation of this two-feed-point design was investigated using
controller designs can be explained by their dominant closed- three digital PID controller designs. The performance of the
loop poles. Table 5 lists the dominant poles of the closed-loop controllers to regulate the reactor temperature in two sections of
system in the z-domain. Poles near the unit circle of the z-plane the reactor was compared when disturbances and measurement
result in sluggish responses, while poles that are complex give noise were present.
oscillatory responses. Only the AMIGO design has closed-loop The main contributions of this work are as follows. The
poles that are complex. The dominant poles with the BLT design location and number of feed points to maximize conversion
are points that are closest to the unit circle boundary, which within the thermal constraint was found using optimization
results in the most sluggish response. theory. Prior studies provided no such rigorous justication.
Controller Robustness. To study the robustness of the Discrete digital PID controllers were designed and compared for
controllers, the disturbances are decreased rather than increased this two-feed-point design. Such controllers are practical con-
at the feed points. sidering that modern chemical plants routinely apply digital
control technologies to regulate their plants. It was concluded
( (
0 t e 70 s 0 t e 160 s that the internal model controller design method was the most
d1 d2 eective in regulating the process.
100 J=s t > 70 s 100 J=s t > 160 s
The plate reactor model was developed as a one-dimensional
plug ow reactor. If this assumption is not valid (i.e., cross-
The two disturbances together cause a 2.1% decrease in T1 and sectional gradients of temperature and concentration cannot be
a 2.5% decrease in T2. Figures 13 and 14 show the closed-loop neglected), then it is reasonable to consider using a method such
results for the T1C and T2C feedback digital controller as computational uid dynamics to provide a more realistic
designs. The performances of the two controllers are similar. model to determine the optimal number and location of the
The AMIGO regulator exhibits oscillation, and as before the feed points, and to design and validate appropriate regulators.
IMC design gives the best closed-loop performance.
Closed-Loop Stability. The gain and phase margins are AUTHOR INFORMATION
measures of stability of the closed-loop system. Both measures
can be quantified from a Bode graph of the closed-loop response. Corresponding Author
Figure 15 shows the Bode graph of the three controller designs *E-mail: karlene.hoo@ttu.edu. Tel.: (806) 742-4079. Fax: (806)
for the T1C controller. 742-3552.
From these graphs, the gain margins of the IMC and the
AMIGO designs are similar, but in the former, the phase margin
is 30.7% greater than the latter. This means the IMC design is ACKNOWLEDGMENT
more stable. The BLT design has the largest gain and phase
Y.C. was supported by the TTU Process Control and Opti-
margins, which indicates a very stable feedback system. However, mization Consortium.
a system with a large gain margin or a large phase margin also is
very sluggish,20,21 which is not attractive for achieving a fast but
stable closed-loop performance. The Bode graphs for T2C NOMENCLATURE
controllers (not shown) are similar to those for T1C controllers. AMIGO = approximate M-constrained integral gain optimization
Measurement Noise. Measurement noise always exists. As- BLT = biggest log modulus
sume the measurement noise is Gaussian with a mean value of CSTR = continuous stirred-tank reactor
zero and a variance that is 10% the signal variance. That is, the CV = control variable
SNR (signal-to-noise ratio) is 10:1. Assume that the size of the FOPTD = rst order plus time delay
disturbances are d1 = d2 = 100 J/s. The closed-loop performances IMC = internal model control
of the BLT and IMC controller designs exhibit no change when MIGO = M-constrained integral gain optimization
compared to the noise-free cases. Because the BLT design makes MV = manipulated variable
the closed-loop response sluggish, it appears to be insensitive to NLODE = nonlinear ordinary dierential equation
the presence of noise. In the case of the IMC design, there is an NLPDE = nonlinear partial dierential equation
explicit filtering function; thus the noise is attenuated. However, PFR = plug ow reactor
measurement noise has an obvious effect on the AMIGO PID = proportional integral derivative
controller performance (not shown). Such results also can be RGA = relative gain array
predicted from the gain and phase margins of these three SISO = single input single output
controller designs. For the AMIGO regulator, the settling time SNR = signal-to-noise ratio
of the T2C controller increases by 5.7%. The decay ratios of ZN = ZieglerNichols
870 dx.doi.org/10.1021/ie2019245 |Ind. Eng. Chem. Res. 2012, 51, 860871
Industrial & Engineering Chemistry Research ARTICLE
REFERENCES
(1) Haugwitz, S.; Hagander, P.; Nore, T. Control Eng. Pract. 2007,
15, 779792.
(2) Haugwitz, S. Modeling, Control and Optimization of a Plate
Reactor. Ph.D. Thesis, Lund Institute of Technology, Lund, Sweden,
2007.
(3) Mann, U. Principles of Chemical Reactor Analysis and Design;
Plains Publishing Co.: Lubbock, TX, 1999.
(4) Luyben, W. L. Ind. Eng. Chem. Process Des. Dev. 1986, 25,
654660.
(5) strom, K. J.; Hagglund, T. J. Process Control 2004, 14, 635650.
(6) Garcia, C. E.; Morari, M. Ind. Eng. Chem. Process Des. Dev. 1982,
21, 308303.
(7) Janickis, V.; Ivanauskas, R.; Stokien_e, R.; Kreiv_enien_e, N. Mater.
Sci. 2006, 12, 311317.
(8) Beers, K. J. Numerical Methods for Chemical Engineering Applica-
tions in MatLab; Oxford University Press: New York, NY, 2007.
(9) Chen, Y. Comparison of Discrete Controller Designs Applied to the
Regulation of a Plug Flow Reactor; Technical Report; Texas Tech
University: Lubbock, TX, 2010.
(10) Seborg, D. E.; Edgar, T. F.; Mellichamp, D. A. Process Dynamics
and Control, 2nd ed.; Wiley: New York, NY, 2004.
(11) Ziegler, J.; Nichols, N. Trans. ASME 1942, 64, 759768.
(12) Hagglund, T.; strom, K. J. Asian J. Control 2002, 4, 364380.
(13) strom, K. J.; Panagopoulos, H.; Hagglund, T. Automatica
1998, 34, 585601.
(14) Hagglund, T.; strom, K. J. Asian J. Control 2004, 6, 469482.
(15) strom, K. J.; Hagglund, T. PID Controllers: Theory, Design and
Tuning; ISA: Research Triangle Park, NC, 1995.
(16) Rivera, D. E.; Morari, M.; Skogestad, S. Ind. Eng. Chem. Process
Des. Dev. 1986, 25, 252265.
(17) Ogunnaike, B. A.; Ray, W. H. Process Dynamics, Modeling, and
Control; Oxford University Press: New York, NY, 1994.
(18) Eriksson, L. M.; Johansson, M. Simple PID tuning rules for
varying time-delay systems. In IEEE Conference on Decision and Control;
IEEE: New Orleans, LA, 2007.
(19) Kalman, R. E.; Bertram, J. AIEE Trans. 1958, 77 (Part 2), 602.
(20) Nagrath, L. J.; Gopal, M. Control System Engineering; New Age
International (P) Ltd.: New Delhi, India, 2006.
(21) Hong, Y.; Yang, O. W. W. Telecommun. Syst. 2007, 36, 161171.