You are on page 1of 9

Industrial & Engineering Chemistry Research Article

Table 1. Tuning Parameters for Different Cascade


Controllersa
GMC(m)−GMC(s) GMC(m)−PI(s)
cascade parameters product A product B product A product B
master controller (m) k1 1.1 1.0 1.0 0.6
k2 0.0 0.0 0.15 0.1
slave controller (s) kc1 2.4 2.4 19.0 18.0
kc2 0.0 0.02 0.125 0.125
PI(m)−GMC(s) PI(m)−PI(s)
cascade parameters product A product B product A product B
master controller (m) k1 37 33 37 33
k2 1.3 1.3 1.3 1.3
slave controller (s) kc1 2.0 2.0 2.0 2.0
kc2 0.0 0.0 0.0 0.0
a
Fine-tuned parameter values for different cascade controllers: k1 =
proportional gain for master control; k2 = integral gain for master
control; kc1 = proportional gain for slave control; kc2 = integral gain for
slave control.

Table 2. Overall Comparison of Control Performance for


Figure 1. Representation of Chylla Haase polymerization reactor: (a) Multiproduct Semibatch Polymerization Reactor Challenge
schematic representation with cascade control strategy, TC = tem- Control ProblemProducts A and B
perature controller, TT = temperature transmitter; (b) block diagram GMC GMC PI (master)/ PI
of cascade control strategy. (master)/ (master)/PI GMC (master)/
controller GMC (slave) (slave) (slave) PI (slave)
feedback method to handle sensitivity and input saturation with product A RMSOD 0.1084 0.3416 3.5889 3.5275
application of GMC based on exact model and data-driven (master control)
models. Further, with the same approach, Rani and Patwardhan21 product A RMSOD
(slave control)
3.040 5.8111 60.4306 25.124
have applied neural network based GMC for this system. product A 0.1369 0.2741 0.4225 0.2101
Cascade control approach has been the most commonly NRMSID
applied approach for this challenge problem. Clarke-Pringle and product B RMSOD 0.1248 0.5162 4.8121 4.4896
MacGregor22 have proposed a nonlinear adaptive temperature (master control)
control strategy, where some of the unknown process param- product B RMSOD 3.998 5.6026 67.9129 53.529
(slave control)
eters are estimated using an extended Kalman filter to find the product B 0.1631 0.2274 0.4084 0.3984
jacket temperature set point for the slave controller in a cascade NRMSID
structure. Binder et al.23 have presented an adaptive control vector
parametrization strategy with the help of wavelets for optimal by Chylla and Haase20 as a challenge problem. Achieving good
control for this system in order to find the set point trajectory for temperature control in these reactors is often difficult because
jacket temperature. Graichen et al.24 have applied feed forward physical properties of the contents, such as mass, heat capacity,
control with online parameter estimation using EKF based on the and heat transfer coefficient vary from run to run and within a
first-principles model to find the set point trajectory for jacket run. The system consists of a stirred tank reactor to prepare
temperature. Vasanthi et al.25 developed a cascade controller with a specialty emulsion polymers. The schematic representation of
self-tuning master control loop to obtain the desired control the polymerization reactor is illustrated in Figure 1a. During
performance for the Chylla and Haase polymerization process typical reactor operation, four or five batches of a particular
maintaining the reactor temperature within the tolerance interval of product would be made in succession. Afterward, the reactor
±0.6K from the set point. would be cleaned before changing to a new product. Data has
The focus of the cascade control approach for this problem has been provided for two different products (called products A
been on using different control approaches for the master loop and B), which are representative of the large variety of products
while retaining conventional PI controller in the slave loop. In the made. The operating recipe for product A is given as follows:
present study, the semibatch polymerization reactor challenge (1) in phase 1, place initial charge of solids and water into
control problem of Chylla and Haase20 is considered to evaluate the reactor at ambient temperature; (2) raise temperature of
different combinations of cascade control configurations for master the initial charge to the reaction temperature set point of
as well as slave loops. The configurations considered for com- 180 °F; (3) in phase 2, feed pure monomer into the reactor at
parison are master controller and slave controller by GMC 1.0 lb/min for 70 min; and (4) after the feed addition period is
[GMC−GMC], master controller by GMC and slave controller by complete, in phase 3, hold at the reaction temperature for
PI [GMC−PI], master controller by PI and slave controller by 60 min. For product B, the operating recipe is described as
GMC [PI−GMC] and both master controller and slave controller follows: (1) In phase 1, place initial charge of solids and water
by PI [PI−PI]. into the reactor at ambient temperature; (2) raise temperature
of the initial charge to the reaction temperature set point of
2. SEMIBATCH POLYMERIZATION REACTOR 176 °F; (3) in phase 2, feed pure monomer into the reactor at
An industrial case study for temperature control of a multi- 0.8 lb/min for 60 min; (4) after the feed addition period is
product semibatch polymerization reactor has been presented complete, in phase 3, hold at reaction temperature for 30 min;
B dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX
Industrial & Engineering Chemistry Research Article

Figure 2. Comparison of performance of cascade controller combinationsProduct A for summer batch 1: (a) reactor temperature profiles, (b)
input profiles of master control, (c) mean jacket temperature profiles, and (d) input profiles of slave control.

(5) in phase 4, feed pure monomer into the reactor at and Haase20). The model equations considered for semibatch
0.8 lb/min for 40 min; and (6) after the second feed addition polymerization reactor are as follows:
period is complete, in phase 5, hold at reaction temperature for Material balance in terms of moles of monomer is given as
45 min. The recipe for product B consists of two feed periods;
however, only the first one is considered in the present study. dn m
= FM − RP
During the second feed, the heat transfer coefficient becomes dt (1)
zero and no control is possible. This observation has also been
where nm = number of moles of monomer in the reactor
made by Clarke-Pringle and MacGregor.22 Thus, the control
(lb·mol), FM = molar flow rate of monomer into the reactor
performance is evaluated only in the first three phases for both
((lb·mol)/min), and RP = rate of polymerization ((lb·mol)/min).
the products in the present study.
The energy balance around the reactor is given as
The detailed model (consisting of four states, i.e., moles of
monomer; reactor temperature; outlet jacket temperature; and
inlet jacket temperature, described by four differential equations
∑ miCpi dT = mMC pM(Tamb − T ) + R p(−ΔHP) − UA(T − Tj)
i dt
and several algebraic equations) and data as well as problem
description have been defined in the original problem (Chylla − (UA)loss (T − Tamb) (2)

C dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX


Industrial & Engineering Chemistry Research Article

Figure 3. Comparison of performance of cascade controller combinationsProduct A for summer batch 5: (a) reactor temperature profiles, (b)
input profiles of master control, (c) mean jacket temperature profiles, and (d) input profiles of slave control.

where T = reactor temperature (°F), Tamb = ambient dTjout(t )


temperature (°F), U = overall heat transfer coefficient (Btu/ mcC pc = mcC pc[Tjin(t − θ1) − Tjout(t )] + UA(T − Tj)
dt
(ft2·min·°F)), mi = mass of component i in the reactor (lb), i = (3)
M for monomer, W for water, and S for solids, mM = mass flow where mc = mass of coolant in jacket (lb), Cpc = heat capacity of
rate of monomer (lb/min), (UA)loss is heat loss to environment coolant in jacket (Btu/(lb·°F)), θ1 = transport delay in jacket
per unit temperature (Btu/(min·°F)), Tj = average jacket (min), and θ2 = transport delay in recirculation loop (min).
temperature (°F), −ΔHP is heat of polymerization (Btu/(lb· The inlet jacket temperature is the delayed exit jacket tem-
mol)), A = jacket heat transfer area (ft2), c(t) = slave controller perature given by the equation
output (0−100%), Cp = specific heat of component i (Btu/
dTjin(t ) dTjout(t − θ2) Kp
(lb·°F)), and MWM = molecular weight of the monomer mix = + [Tjout(t − θ2) − Tjin(t )] +
(lb/(lb·mol)). dt dt τp
The exit temperature of the jacket is given by (4)

D dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX


Industrial & Engineering Chemistry Research Article

Table 3. Effect of Other Disturbances on Control Performance for Multiproduct Semibatch Polymerization Reactor Challenge
Control Problem Using GMC−GMCProduct A and Product B
effect of unmeasured disturbances product A product B
RMSOD RMSOD
(master RMSOD (master RMSOD
disturbances variations control) (slave control) NRMSID control) (slave control) NRMSID
base case i = 1; no load disturbance; no 0.0961 2.5436 0.1376 0.1016 3.0404 0.1607
noise; θ1d = 3; θ2d = 2
impurity factor in rate i = 0.9 (−10%) 0.179 2.506 0.137 0.188 2.3158 0.148
expression i = 1.1(+10%) 0.1625 2.3432 0.1354 0.205 3.9511 0.1656
i = 0.8 (−20%) 0.3618 2.3854 0.1190 0.4053 1.8320 0.1383
i = 1.2(+20%) 0.2569 2.3994 0.1187 0.3246 5.0881 0.1715
load disturbance in −10% flow disturbance 0.0888 2.1932 0.1241 0.0999 2.5004 0.147
monomer feed flow rate +10% flow disturbance 0.1078 2.893 0.1414 0.1138 3.8641 0.1671
−20% flow disturbance 0.0838 1.8534 0.1105 0.1044 2.1396 0.1414
+20% flow disturbance 0.1242 3.2910 0.1269 0.1382 5.1297 0.1781
delay times in recirculation θ1d = 4; θ2d = 2 0.1421 2.7801 0.0890 0.1998 3.2374 0.1833
loop θ1d = 3; θ2d = 1 0.2375 5.4173 0.3812 0.2981 3.5123 0.1990
θ1d = 3; θ2d = 3 0.1014 2.4632 0.0922 0.1920 3.9758 0.2242
effect of measured disturbances product A product B
RMSOD (master RMSOD (slave RMSOD (master RMSOD (slave
disturbances variations control) control) NRMSID control) control) NRMSID
i = 0.9(−10%)
impurity factor in rate expression 0.0901 2.329 0.1279 0.0924 2.2835 0.1496
i = 1.1(+10%) 0.0981 2.6183 0.1392 0.1163 3.9634 0.1635
i = 0.8(−20%) 0.0817 2.0688 0.1147 0.0878 1.7729 0.1371
i = 1.2(+20%) 0.0998 2.6894 0.1204 0.1373 5.1375 0.1704
load disturbance in monomer −10% flow 0.0845 2.1646 0.1251 0.0913 2.3562 0.1462
feed flow rate disturbance
+10% flow 0.1052 2.8072 0.1338 0.1209 4.1402 0.1684
disturbance
−20% flow 0.0749 1.8604 0.1109 0.0852 1.9428 0.1405
disturbance
+20% flow 0.1233 3.3302 0.1267 0.1499 5.6913 0.1798
disturbance
presence of measurement noise in reactor temperature product A product B
white noisenormal distribution with standard deviation of 0.02 0.0987 2.675 0.1493 0.1053 3.1221 0.1668
white noisenormal distribution with standard deviation of 0.5 0.5868 11.4121 0.3694 0.5371 11.8458 0.3631

where 1000 solids(t )


a= ; f (t ) =
T + 460 batch weight(t )
T jin + T jout mM
Tj = and FM = solids(t ) = solids(t = 0) − X (t )·total monomer added(t )
2 MWM (5)
batch weight(t ) = solids(t ) + total monomer added(t )(1 − X (t ))
+ initial water
R p = i(knm) (6) h = 143.4 exp(−5.13 × 10−3μwall )
T + Tj
⎛ 6400 ⎞ 0.4 μwall = μ(Twall); Twall =
2
k = k 0 exp⎜ ⎟μ
⎝ T + 460 ⎠ (7) 1
= {0.000, 0.001, 0.002, 0.003, 0.004}
hf
where Kp = heating/cooling process gain (°F/%), k = first order for batch numbers {1, 2, 3, 4, 5}
kinetic constant (1/min), k0 = pre-exponential factor (1/min), (10)

μ = product viscosity (cP), and i = impurity factor. The overall heat transfer coefficient is then given by the
The viscosity relation for product A is given as equation
μ = 0.032e(16.4f )10[2.3(a − 1.563)] (8) 1
U= 1 1
h
+ hf (11)
and for product B, it is given as

μ = 0.032e(119.1f )10[2.3(a − 1.563)] (9)


where h = film heat transfer coefficient (Btu/(ft2·min·°F)) and
f = fraction of solids.
where The process gain equation is expressed as
E dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX
Industrial & Engineering Chemistry Research Article

Figure 4. Comparison of performance of cascade controller combinationsProduct A for winter batch 1: (a) reactor temperature profiles, (b) input
profiles of master control, (c) mean jacket temperature profiles, and (d) input profiles of slave control.

⎧ 0.8 × (30)(−c(t )/50)(T − T in) 0 ≤ c(t ) < 50 ⎫ For this challenge problem, the information available to the
⎪ inlet j ⎪
⎪ ⎪ control algorithm is the product being made, and measure-
K p = ⎨0 c(t ) = 50 ⎬
⎪ ⎪ ments of the reactor temperature, temperature of the inlet
⎪ 0.51 × (350 − Tj )(30)
in (c(t )/50 − 2)
50 < c(t ) ≤ 100 ⎪
⎩ ⎭ (12) water to the jacket, temperature of the exit water from the
jacket, and the monomer feed flow rate. The controller must
where c(t) is the valve opening. work over a series of five consecutive batches for both the
The reaction temperature determines the chemical compo-
sample products under both summer and winter operation.
sition and particle size distribution of the emulsion polymer.
In the present study, two basic controllers considered in
Precise control of the reaction temperature is required
throughout the batch to produce an acceptable product. different cascade combinations are GMC and PI controllers.
Reaction temperature is controlled by manipulating the It is assumed that the process model is available for formula-
temperature of water which is recirculated through the jacket tion of GMC. Before presenting different combinations of
of the reactor. The deviations in reactor temperature of less cascade controller configurations, a brief overview of GMC is
than 1 °F are considered as satisfactory (Chylla and Haase20). presented.
F dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX
Industrial & Engineering Chemistry Research Article

3. GENERIC MODEL CONTROL OVERVIEW (r * − Lf h(x))


u=
In this section, GMC is presented as a typical nonlinear control Lg h(x)
algorithm. Consider a control-affine single-input single-output t
(SISO) system described by the following equations: (k1(y sp − y) + k 2 ∫ (y sp − y) dτ − Lf h(x))
0
=
ẋ = f(x(t ), t ) + g(x(t ), t ) ·u(t ) x(0) = x 0 Lg h(x) (18)
y = h(x(t )) (13) The closed-loop transfer function resulting from application of the
above control law is expressed as
where x represents the state variable vector of dimension n, u is
the manipulated input, t denotes the present time, x0 represents the Y (s) k s + k2
= 2 1
initial state vector, y represents the output variable, f(·) and g(·) are Y sp(s) s + k1s + k 2 (19)
vector functions, and h(·) is a scalar function. For systems where sp
where Y(s) and Y (s) represent the transfer functions of the output
relative degree is well-defined, the relative degree r of the output y variable and the set point, respectively.
with respect to the manipulated input u is defined as the number of
times the output is to be differentiated with respect to time so as to 4. CASCADE CONTROL CONFIGURATIONS FOR THE
explicitly depend on the manipulated input. Equation 13 can be CHYLLA−HAASE PROBLEM
successively differentiated and expressed in the following form:
The block diagram of cascade control strategy is shown in
dy Figure1b. In the present study, different cascade configurations
z1̇ = y(1) = = Lf h(x) = z 2 are developed for controlling the reaction temperature at the
dt
desired set point, i.e.,
d2y
z 2̇ = y(2) = = Lf 2h(x) = z 3 (a) Master controller and slave controller as GMC [GMC−
dt 2 GMC],
: (b) Master controller as GMC and slave controller as PI
[GMC−PI],
dr − 1y
zṙ − 1 = y(r − 1) = r−1
= Lf r − 1h(x) = zr (c) Master controller as PI and slave controller as GMC
dt [PI−GMC], and
(r ) (d) Master controller and slave controller as PI [PI−PI].
zṙ = y = Lf h(x) + Lg Lf r − 1h(x) ·u
r
4.1. Master Controller and Slave Controller As GMC
= b(z) + a(z) ·u (14) [GMC−GMC]. In this control scheme, the master controller,
i.e., GMC, regulates the reaction temperature T by manipulat-
to represent the first r derivatives of the output variable by defining ing the mean cooling jacket temperature Tjset. The slave
the transformed states zi, where the definitions of different orders of controller, i.e., GMC manipulates the valve opening c(t) to
Lie derivatives of the scalar function h(x) with respect to vector control the mean jacket temperature Tj. The controller equations
functions f(x) and g(x) are discussed in detail by Henson and are given as follows:
Seborg.26 The remaining states can then be represented by the Master Loop. In order to derive GMC law, it is necessary to
following transformed equations, define a reference rate trajectory for the output derivative and
zṙ + 1 = qr + 1(z) + pr + 1 (z) ·u to force the process output rate to match the reference rate
trajectory.
: The reference trajectory is given according to eq 16 as
zṅ = qn(z) + pn (z) ·u (15) t
r * = k1(Tset − T ) + k 2 ∫0 (Tset − T ) dτ
(20)
where pi(·) and qi(·) are scalar functions of the transformed variable
vector z, zr+1 to zn denote the internal states, and eq 15 represents The process output rate for the reaction temperature is
internal dynamics. defined by eq 2. The manipulated variable, mean cooling jacket
For relative degree one systems, Lee and Sullivan12 have temperature, Tjset, is calculated by equating reference trajectory
proposed GMC as a mechanistic model based control approach r* to reaction temperature rate as
where the control law is derived by defining a reference rate
⎡ t
trajectory for the output derivative and forcing the process output 1 ⎢
rate to match the reference rate trajectory defined as
Tjset =
UA ⎢⎣
(k1(Tset − T ) + k 2 ∫0 (Tset − T ) dτ )

t ⎛
r * = k1(y sp − y) + k 2 ∫0 (y sp − y) dτ
(16) × ⎜⎜(∑ miCp, i) − (mM0C PM(Tamb + T ) + RP0(−ΔHP)
⎝ i
where r* is the desired “rate of change” of process output, ysp is the ⎞⎤
set point, and k1 and k2 are the controller tuning parameters. The − UA × T − (UA)loss (T − Tamb))⎟⎟⎥
reference rate is proportional to the distance from the set point and ⎠⎥⎦ (21)
includes integral action to eliminate offset. For systems of relative
degree one, eq 14 reduces to Slave Loop. In this case, the reference trajectory is given
according to eq 16 as
dy
y(1) = = Lf h(x) + Lg h(x) ·u r1* = kc1( Tjset − 0.5(Tjin + Tjout))
dt (17)
t
The explicit form of the GMC law can be derived from eqs 16 and
17 as
+ kc2 ∫0 ( Tjset − 0.5(Tjin + Tjout)) dτ
(22)

G dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX


Industrial & Engineering Chemistry Research Article

Figure 5. Comparison of performance of cascade controller combinationsProduct A for winter batch 5: (a) reactor temperature profiles, (b) input
profiles of master control, (c) mean jacket temperature profiles, and (d) input profiles of slave control.

The process output rate for the mean jacket temperature Tj τp ⎛ ⎛ dT out dTjout(t − θ2) ⎞⎞
Kp = ⎜r1* − 1 ⎜ j − − Tjout(t − θ2) − Tjin(t )⎟⎟⎟⎟
is defined by taking the mean of eqs 3 and 4. The manip- ⎜ ⎜
2 ⎝ dt
0.5 ⎝ dt ⎠⎠
ulated variable, valve opening c(t), is calculated by equating (23)
reference trajectory r1* to the mean jacket temperature rate in After determining the heating/cooling process gain, the valve
two steps in eqs 23 and 24 as opening c(t) can be calculated based on eq 12 and given as

⎧− 50 × log(K /(0.8 × (T − T in)))/log(30) K p < (0.8 × (Tinlet − Tjin))/30 ⎫


⎪ p inlet j

⎪ ⎪
c(t ) = ⎨ 50 (0.8 × (Tinlet − Tjin))/30 < K p < (0.15 × (350 − Tjin))/30 ⎬
⎪ ⎪
⎪100 − 50 × log((0.15 × (350 − Tjin))/K p)/log(30) (0.15 × (350 − Tjin))/30 < K p ⎪
⎩ ⎭ (24)

H dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX


Industrial & Engineering Chemistry Research Article

Figure 6. Comparison of performance of cascade controller combinationsProduct B for summer batch 1: (a) reactor temperature profiles, (b)
input profiles of master control, (c) mean jacket temperature profiles, and (d) input profiles of slave control.

The lower and higher limits for Kp are defined based on control the mean jacket temperature Tj. The master control
substitution of c(t) value as 50 in the two expressions of eq 12 equation is given by eq 21, and the slave control equation is
for c(t) < 50 and c(t) > 50, respectively. Further, the values of given by
c(t) obtained are bounded between 0 and 100 using the
saturation function. c(t ) = kc1( Tjset − 0.5(T jin + T jout))
The procedure followed for derivation of control laws is t
similar in all the cases and is therefore not included in detail in + kc2 ∫0 ( Tjset − 0.5(T jin + T jout)) dτ
(25)
the subsections below.
4.2. Master Controller as GMC and Slave Controller 4.3. Master Controller as PI and Slave Controller as
as PI [GMC−PI]. In this control scheme, the master controller, GMC [PI−GMC]. In this control scheme, the master controller,
i.e., GMC, regulates the reaction temperature T by manipulat- i.e., PI, regulates the reaction temperature T by manipulating
ing the mean cooling jacket temperature Tjset. The slave the mean cooling jacket temperature Tjset. The slave controller,
controller, i.e., PI manipulates the valve opening c(t) to i.e., GMC manipulates the valve opening c(t) to control the
I dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX
Industrial & Engineering Chemistry Research Article

Figure 7. Comparison of performance of cascade controller combinationsProduct B for summer batch 5: (a) reactor temperature profiles, (b)
input profiles of master control, (c) mean jacket temperature profiles, and (d) input profiles of slave control.

mean jacket temperature Tj. The slave control is given by eq 23, 5. RESULTS AND DISCUSSION
and the master controller equation is given by In the present study, a challenging industrial benchmark
t multiproduct semibatch polymerization reactor is considered to
Tjset = k1(Tset − T ) + k 2 ∫0 (Tset − T ) dτ
(26)
control the reaction temperature at desired set point using
different cascade controller configurations. Chylla and Haase20
have specified that solutions can be provided for the challenge
4.4. Master Controller and Slave Controller as PI problem by defining the control problem as design of a
[PI−PI]. In this control scheme, the master controller, i.e., PI, controller capable of maintaining the desired reaction temper-
regulates the reaction temperature T by manipulating the ature throughout the batch in the presence of typical distur-
mean cooling jacket temperature Tjset. The slave controller, bances, such as change in fouling factor from batch to batch,
i.e., PI, manipulates the valve opening c(t) to control the ambient temperature change due to seasonal variations, changes
mean jacket temperature Tj. The master controller equa- in the product, and variation within each batch such as heat
tion is given by eq 26, and the slave control equation is given transfer change due to increasing viscosity of product. Con-
by eq 25. ditions often change from batch to batch due to increased
J dx.doi.org/10.1021/ie501515y | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

You might also like