17th Mediterranean Conference on Control & Automation
Makedonia Palace, Thessaloniki, Greece
June 24 - 26, 2009
Adaptive Temperature Control in a Tubular Chemical Reactor
P. Dostál, V. Bobál, and J. Vojtěšek
Abstract—The paper deals with adaptive control of a tubular
chemical reactor. As a part of the control design, preliminary
steady-state and dynamic analysis of the process is incorporated.
A nonlinear model of the process is approximated by a
continuous-time external linear model with parameters estimated
using a corresponding delta model. The controller design is based
on the polynomial approach. The adaptive control is tested on the
nonlinear model of the tubular chemical reactor with a
consecutive exothermic reaction.
I. INTRODUCTION
T UBULAR chemical reactor are units frequently used in
chemical and biochemical industry. From the system
theory point of view, tubular chemical reactors belong to a
class of nonlinear distributed parameter systems. Their
mathematical models are described by sets of nonlinear partial
differential equations (NPDEs). The methods of modelling
and simulation of such processes are described e.g. in [1] –
[5].
It is well known that the control of chemical reactors, and,
tubular reactors especially, often represents very complex
problem. The control problems are due to the process
nonlinearity, its distributed nature, and high sensitivity of the
state and output variables to input changes. In addition, the
dynamic characteristics may exhibit a varying sign of the gain
in various operating points, the time delay as well as non-
minimum phase behaviour. Evidently, the process with such
properties is hardly controllable by conventional control
methods, and, its effective control requires application some
of advanced methods. Some these methods with their
applications in the control of chemical reactors are described
e.g. in [6]. At all events, a previous analysis of the process
behaviour is obligatory.
One possible method to cope with this problem is using
adaptive strategies based on an appropriate choice of an
continuous-time external linear model (CT ELM) with
recursively estimated parameters. These parameters are
consequently used for parallel updating of controller‘s
parameters.
For the CT ELM parameters estimation, either the direct
method [7] or application of an external delta model with the
Manuscript received January 15, 2009. This work was supported by the
Ministry of Education of the Czech Republic under the grant MSM
7088352101.
Petr Dostál, Vladimír Bobál and Jiří Vojtěšek are with the Tomas Bata
University in Zlin, Faculty of Applied Informatics, Nad Stranemi 4511, 760
05 Zlín 5, Czech Republic (phone: +420 57 6035195; fax: +420 57 6035279;
e-mail: {dostalp, bobal, vojtesek}@ fai.utb.cz).
978-1-4244-4685-8/09/$25.00 ©2009 IEEE
same structure as the CT model can be used. The basics of
delta models have been described in e.g. [8], [9]. Although
delta models belong into discrete models, they do not have
such disadvantageous properties connected with shortening of
a sampling period as discrete z-models. In addition,
parameters of delta models can directly be estimated from
sampled signals. Moreover, it can be easily proved that these
parameters converge to parameters of CT models for a
sufficiently small sampling period (compared to the dynamics
of the controlled process), as shown in [10].
This paper deals with continuous-time adaptive control of a
tubular chemical reactor with a countercurrent cooling as a
non-linear single input – single output process. With respect to
practical possibilities of a measurement and control, the mean
reactant temperature and the coolant flow rate are chosen as
the controlled output and the control input. The parameters of
its CT ELM are estimated via corresponding delta model. The
resulting controllers are derived using polynomial approach
[11] and the pole assignment method (see, e.g. [12]). The
approach is tested on a mathematical model of the tubular
chemical reactor.
II. MODEL OF THE REACTOR
An ideal plug-flow tubular chemical reactor with a simple
k1 k2
exothermic consecutive reaction A → B → C in the liquid
phase and with the countercurrent cooling is considered as
shown in Fig. 1.
qc, Tc out qc, Tc L
qr, Tr 0 qr, Tr out
ci 0 ci out
z
L
Fig. 1. Tubular chemical reactor.
Heat losses and heat conduction along the metal walls of
tubes are assumed to be negligible, but dynamics of the metal
walls of tubes are significant. All densities, heat capacities,
and heat transfer coefficients are assumed to be constant.
Under above assumptions, the reactor model can be described
1209
by five NPDEs in the form can be taken into account as the control variable, whereas
other inputs enter into the process can be accepted as
∂cA ∂c disturbances. The mean reactant temperature given by
+ vr A = − k1 c A (1)
∂t ∂z 1 L
Tm (t ) = ∫ Tr ( z , t ) d z (9)
∂c B ∂c B L 0
+ vr = k1 c A − k 2 c B (2) is considered as the controlled output.
∂t ∂z
∂T r ∂T r Qr 4U 1 TABLE I
+ vr = − (T r − T w ) (3) USED PARAMETER VALUES
∂t ∂z (ρc p ) r d 1 (ρc p ) r
L=8m n1 = 1200
∂T w 4 d1 = 0.02 m d2 = 0.024 m
= ⎡⎣ d 1U 1 (T r − T w ) +
∂t (d 22 − d 12 ) (ρc p ) w (4) d3 = 1 m
ρr = 985 kg/m3 cpr = 4.05 kJ/kg K
+ d 2 U 2 (Tc − T w ⎤⎦
ρw = 7800 kg/m3 cpw = 0.71 kJ/kg K
∂T c ∂T c 4 n1 d 2 U 2 ρc = 998 kg/m3 cpc = 4.18 kJ/kg K
− vc = (T w − Tc ) (5)
∂t ∂z (d 32 − n1 d 22 ) (ρc p ) c U1 = 2.8 kJ/m2s K U2 = 2.56 kJ/m2s K
16
k10 = 5.61⋅10 1/s k20 = 1.128⋅1018 1/s
with initial conditions
E1/R = 13477 K E2/R = 15290 K
c A ( z ,0) = c As ( z ) , c B ( z ,0) = c Bs ( z ) , Tr ( z ,0) = Trs ( z ) , 4
(-ΔHr1) = 5.8⋅10 kJ/kmol (-ΔHr2) = 1.8⋅104 kJ/kmol
T w ( z , 0) = Tws ( z )
, T c ( z ,0 ) = Tcs ( z )
and boundary conditions III. COMPUTATION MODEL
c A (0, t ) = c A 0 (t ) (kmol/m3), c B (0, t ) = c B 0 (t ) (kmol/m3), For computation of both steady-state and dynamic
Tr (0, t ) = Tr 0 (t ) (K), Tc ( L, t ) = Tc L (t ) (K). characteristics, the finite diferences method is employed. The
procedure is based on substitution of the space interval
Here, t is the time, z is the axial space variable, c are
z ∈< 0, L > by a set of discrete node points { z i } for i = 1, … ,
concentrations, T are temperatures, v are fluid velocities, d are
diameters, ρ are densities, cp are specific heat capacities, U are n ,and, subsequently, by approximation of derivatives with
heat transfer coefficients, n1 is the number of tubes and L is respect to the space variable in each node point by finite
the length of tubes. The subscript (⋅)r stands for the reactant differences. Then, NPDEs (1) – (5) are approximated by a set
mixture, (⋅)w for the metal walls of tubes, (⋅)c for the coolant, of NODEs in the form
and the superscript (⋅)s for for steady-state values. d c A (i )
= − ⎡⎣b 0 + k 1 (i ) ⎤⎦ c A (i ) + b 0 c A (i − 1) (10)
The reaction rates and heat of reactions are nonlinear dt
functions expressed as d c B (i )
⎛ −E j ⎞ = k 1 (i ) c A (i ) − ⎡⎣b 0 + k 2 (i ) ⎤⎦ c B (i ) +
k j = k j 0 exp ⎜ dt (11)
⎟ , j = 1, 2 (6)
⎝ RTr ⎠ + b 0 c B (i − 1)
Q r = (−ΔH r1 ) k 1 c A + (−ΔH r 2 ) k 2 c B (7) dT r (i )
where k0 are pre-exponential factors, E are activation energies, = b1 Q r (i ) − (b 0 + b 2 ) T r (i ) + b 0 T r (i − 1) +
dt (12)
( - ΔHr) are in the negative considered reaction entalpies, and + b 2 T w (i )
R is the gas constant.
The fluid velocities are calculated via the reactant and dT w (i )
= b3 ⎡⎣T r (i ) − T w (i ) ⎤⎦ + b 4 ⎡⎣T c (i ) − T w (i ) ⎤⎦ (13)
coolant flow rates as dt
4 qr 4 qc dT c (m)
vr = 2
, vc = 2 2
(8) = −(b5 + b 6 ) T c (m) + b5 T c (m + 1) +
π n1 d 1 π ( d 3 − n1 d 2 ) dt (14)
The parameter values with correspondent units used for + b 6 T w ( m)
simulations are given in Table 1.
for i = 1, ... , n and m = n − i + 1 , and, with initial conditions
From the system engineering point of view,
s s s
c A ( L, t ) = c A out , c B ( L, t ) = c B out , T r ( L, t ) = T r out and c A (i, 0) = c A (i ) , c B (i,0) = c B (i ) , T r (i, 0) = T r (i ) ,
T c (0, t ) = T c out are the output variables, and, q r (t ) , q c (t ) , T w (i, 0) = T ws (i ) and T c (i, 0) = T cs (i ) for i = 1, ... , n .
c A 0 (t ) , T r 0 (t ) and T c L (t ) are the input variables. Among The boundary conditions enter into Eqs. (10) – (14) for i = 1
.Now, nonlinear functions (6) and (7) take discrete forms
them, for the control purposes, mostly the coolant flow rate

1210
 ⎛ −E j ⎞ 3.0 s
cA
k j (i ) = k j 0 exp ⎜
s

⎜ RT (i ) ⎟⎟
, j = 1, 2 (15) 1 - qc = 0.2

Concentrations (kmol/m )
2.5
s

3
⎝ r ⎠ 1 2
2 - qc = 0.25
s
3 - qc = 0.3
Q r (i ) = (−ΔH r1 ) k 1 (i ) c A (i ) + (−ΔH r 2 ) k 2 (i ) c B (i ) (16) 2.0 3

for i = 1, … , n. 1.5
The parameters b in Eqs. (10) – (14) are calculated as 1.0
v 1 4U 1 3
b 0 = r , b1 = , b2 =
s
, 0.5 cB
1 2
h (ρc p ) r d 1 (ρc p ) r
0.0
4 d 1U 1 4d 2 U 2
b3 = 2 2
, b4 = 2 2
, (17) 0 1 2 3 4 5 6 7 8
(d 2 − d 1 ) (ρc p ) w (d 2 − d 1 ) (ρc p ) w z (m)
vc 4 n1 d 2 U 2 Fig. 2. Concentration profiles for various coolant flow rates.
b5 = , b6 =
h (d 32 − n1 d 22 ) (ρc p ) c
s
1 - qc = 0.2
and, the expression (9) is rewriten to the discrete form 360 1

Reactant temperature (K)

s
2 - qc = 0.25
n
1
∑ T (z , t)
2 s
3 - qc = 0.3
Tm (t ) = r i (18) 350
n i =1
340
A steady-state model can simply be derived equating the
3
time derivatives in (10) – (14) to zero. Then, after some 330
algebraic modifications, the steady-state model takes the form
of difference equations. Computation of the steady-state 320
characteristics is necessary not only for a steady-state analysis
s
0 1 2 3 4 5 6 7 8
but the steady state values y (i ) also constitute initial z (m)
conditions in ODEs (10) – (14) (here, y presents some of Fig. 3. Reactant temperature profiles for various coolant flow rates.
variables in the set (10) – (14)).
2.4
Output concentrations (kmol/m )
3

IV. STEADY-STATE AND DYNAMIC CHARACTERISTICS

2.0
Typical concentration and reactant temperature profiles
s s 1.6
along the reactor tubes computed for c A 0 = 2.85 , c B 0 = 0 ,
1.2
s s s
Tr 0 = 323 , Tc 0 = 293 and qr = 0.15 for various values of 0.8
the coolant flow rates are shown in Figs. 2, 3. The presence of 0.4 s
cA out
a peak on the reactant B profile is given by considered 0.0
s
cB out
reaction type. The existence of a maximum on the reactant
temperature profiles is a common property of many tubular 0.20 0.24 0.28 0.32 0.36 0.40
3
reactors with exothermic reactions. Coolant flow rate (m /s)
The dependences of output concentrations and the reactant Fig. 4. Dependence of output concentrations on the coolant flow rate.
mean temperature on the coolant flow rate are shown in Figs.
4, 5. The courses document strong sensitivity of outputs to the
345
flow rate which will be considered as the control input.
Mean temperature (K)

Dynamic charakteristics were computed in the 340

neighbourhood of the chosen point qcs = 0.27 m3 / s and 335
Tms = 334.44 K . The input and both outputs were considered 330
as deviations from steady values. This form is frequently used 325
in the control. The deviations are denoted as follows:
320
Δqc = qc (t ) − qcs , ΔcB (t ) = cB (t ) − cBs , ΔTm (t ) = Tm (t ) − Tms
The responses in Figs. 6 and 7 show a better applicability of 0.20 0.24 0.28 0.32 0.36 0.40
the reactant mean temperature as the controlled output in Coolant flow rate (m /s)
3

comparison with the output component B concentration Fig. 5. Dependence of the reactant mean temperature on the coolant flow
having in this regard very unfavourable properties. rates.

1211
 0.4 where t ′ is the discrete time.
0.2 When the sampling period is shortened, the delta operator
3
0.0 approaches the derivative operator, and, the estimated
ΔcB out (kmol/m )

parameters a′, b′ reach the parameters a, b of the CT model

3

-0.2 4
-0.4 2
(20) as shown in [10].
-0.6
-0.8 1 1 - Δqc = -0.04 3 - Δqc = 0.02 VI. DELTA MODEL PARAMETER ESTIMATION
-1.0 2 - Δqc = -0.02 4 - Δqc = 0.04
-1.2 Substituting t ′ = k − 2 , equation (5) may be rewriten to the
-1.4 form
0 25 50 75 100 125 150 175 200 δ 2 y (k − 2) + a1′ δ y (k − 2) + a′0 y (k − 2) =
Time (s) (24)
= b1′ δu (k − 2) + b′0 u (k − 2)
Fig. 6. Component B output concentration step responses.
In the paper, the recursive identification method with
exponential and directional forgetting according to [13] was
6 used.
4 1 Establishing the regression vector
Φ δ (k − 1) = ( −δ y (k − 2) − y (k − 2) δu (k − 2) u (k − 2) )
T
2 2
0 where
ΔTm (K)

1 - Δqc = - 0.04 3 - Δqc = 0.02

-2 3 y (k − 1) − y ( k − 2)
2 - Δqc = - 0.02 4 - Δqc = 0.04
-4 δ y (k − 2) = ,
T0
-6 4 u (k − 1) − u (k − 2)
-8 δu (k − 2) = ,
T0
-10
0 25 50 75 100 125 150 175 200
the vector of delta model parameters
Time (s) Θ δT (k ) = ( a1′ a′0 b1′ b′0 ) (25)
Fig. 7. Reactant mean temperature step responses. is recursively estimated from the equation
δ 2 y (k − 2) = ΘδT (k ) Φδ (k − 1) + ε(k ) (26)
V. CT AND DELTA EXTERNAL LINEAR MODEL
where
For the control purposes, the controlled output and the control y (k ) − 2 y (k − 1) + y (k − 2)
input are defined as δ 2 y (k − 2) = . (27)
T02
q (t ) − q s
y (t ) = ΔT m (t ) = T m (t ) − T ms , u (t ) = 10 c s c . (19)
qc VII. CONTROLLER DESIGN
These expressions enable to obtain variables of approximately The control system is depicted in Fig. 9. In the scheme, w
the same magnitude. is the reference signal, v denotes the load disturbance, e is the
The second order CT ELM has been chosen in the form of tracking error, u is the control input, and y is the controlled
the second order linear differential equation output.
y (t ) + a1 y (t ) + a 0 y (t ) = b1 u (t ) + b 0 u (t )
 (20)
or, in the transfer function representation as v
b s + b0
G( s) = 2 1 . (21) w e u y
s + a1 s + a 0 Q G
-
Here, with regard to step responses in Fig. 7, small values b1
can be expected but a consideration of b1 will not significantly
affect a subsequent control.
Establishing the δ operator Fig. 8. Control system.
q −1
δ= (22)
T0 Further, G represents the ELM and Q is the feedback
where q is the forward shift operator and T0 is the sampling controller. Both w and v are considered to be step functions.
period, the delta ELM corresponding to (20) takes the form The transfer functions of the controlled system and the
2 controller are in general forms
δ y (t ′) + a′1 δ y (t ′) + a′0 y (t ′) = b1′ δu (t ′) + b′0 u (t ′) (23)

1212
 b( s ) q( s )
G(s) = , Q( s) = (28)
a( s) p( s) Computation of b, a
Parameter estimation
where q and p are polynomials in s, and, deg q ≤ deg p . controller parameters

The controller design described in this section stems from

q, p v
the polynomial approach. General conditions required to T0 T0
govern the control system properties are formulated as strong y
w u Controlled
stability (in addition to the control system stability, also the Controller
process
-
stability of controllers is required), internal properness,
asymptotic tracking of a step reference and step load
disturbance attenuation.
It is well known from the algebraic control theory that a Fig. 9. Adaptive control scheme.

controller which satisfies above requirements is in the

polynomial ring given by a solution of the polynomial VIII. CONTROL SIMULATION
(Diophantine) equation Also the control simulations were performed in a
neighbourhood of the operating point qcs = 0.27 m3 / s and
a ( s ) p ( s ) + b( s ) q ( s ) = d ( s ) (29)
with a stable polynomial d(s) on the right side. Tms = 334.44 K . For the start (the adaptation phase), a P
For step input signals w and v, the polynomial p is in the form controller with a small gain was used in all simulations.
p( s ) = s p ( s ) . (30) The effect of the pole α on the control responses is
The degrees of unknown polynomials in (29) and (30) are transparent from Figs. 10 and 11. Here, three values of α were
deg q = deg a , deg p = deg a − 1 , deg d = 2 deg a . (31) selected. The control results show sensitivity of the controlled
output and control input to α. Obviously, careless selection of
Then, for the ELM (21), the controller transfer function takes
this parameter can lead to controlled output responses of a
the form
poor quality or even to unstability. Further, a increasing α
q( s) q s 2 + q1 s + q0
Q( s) = = 2 (32) leads to higher values and changes of the control input which
s p ( s ) s ( s + p0 ) has been constrained on the interval −1.5 ≤ u (t ) ≤ 1.5 .
The polynomial d with roots determining the closed-loop As an influence illustration of a additive random
poles is chosen as disturbance on the control, the control simulated in a presence
d ( s ) = n( s ) ( s + α ) 2 (33) of the random signal v(t ) = c A0 (t ) − c As is shown in Figs. 12.
where n is a stable polynomial obtained by spectral The simulation result document an usability of the method
factorization also in this case.
a ∗ ( s ) a ( s ) = n∗ ( s ) n ( s ) (34) A presence of the integrating part in the controller enables
rejection of various step disturbances entering into the
and α is the selectable parameter.
process. Here, step disturbances Δc A0 = 0.15 kmol / m3 ,
Note that a choice of d in the form (33) provides the control
of a good quality for aperiodic controlled processes. Δqr = −0.03 m3 / s and ΔTr 0 = 2 K at times t v = 220s ,
The coefficients of n then are expressed as t v = 440 s and t v = 640 s were considered. The controller
n0 = a02 , n1 = a12 + 2 n0 − 2 a0 (35)
and, the controller parameters can be obtained from solution
of the matrix equation 4

⎛1 b1 0 0 ⎞ ⎛ p0 ⎞ ⎛ d3 − a1 ⎞ 2
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ a1 b0 b1 0 ⎟ ⎜ q2 ⎟ ⎜ d 2 − a0 ⎟ 0
Ω = (36) w
y (K)

⎜ a0 0 b0 b1 ⎟ ⎜ q1 ⎟ ⎜ d1 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ -2
⎝0 0 0 b0 ⎠ ⎝ q0 ⎠ ⎝ d 0 ⎠
where -4
α = 0.05
α = 0.075
d3 = n1 + 2α, d 2 = 2α n1 + n0 + α 2 -6 α = 0.1
(37)
d1 = 2α n0 + α 2 n1 , d0 = α 2 n0 0 100 200 300 400 500 600 700 800
Evidently, the controller parameters can be adjusted by Time (s)
selectable parameter α. Fig. 10. Controlled output responses for various closed-loop poles.
The adaptive control system is shown in Fig. 9.

1213
 1.5 IX. CONCLUSIONS
1.0 In this paper, one approach to the continuous-time adaptive
control of the mean reactant temperature in a tubular chemical
0.5
reactor was proposed. The control strategy is based on a
u (1)

0.0 preliminary analysis of the steady-state and dynamic analysis

-0.5 of the process behaviour and on the assumption of the
temperature measurement along the reactor. The proposed
-1.0 α = 0.05
α = 0.075 algorithm employs an alternative continuous-time external
-1.5 α = 0.1
linear model with parameters obtained through recursive
0 100 200 300 400 500 600 700 800
parameter estimation of a corresponding delta model. The
Time (s) resulting continuous-time controller is derived using the
polynomial approach and given by a solution of a polynomial
Fig. 11. Control input responses for various closed-loop poles.
Diophantine equation. Tuning of its parameters is possible via
closed-loop pole assignment. The presented method has been
0.04
tested by computer simulation on the nonlinear model of the
v (kmol/m )
3

0.02
tubular chemical reactor with a consecutive exothermic
0.00
reaction. Results demonstrate an applicability of the presented
-0.02
control strategy. It can be remarked that similar results have
-0.04 been obtained using the output temperature as the controlled
4
output.
2
REFERENCES
0
y (K)

w [1] W. Luyben, Process modelling, simulation and control for chemical

-2 engineers. New York: McGraw-Hill, 1989.
-4 [2] J. Ingham, I.J. Dunn, E. Heinzle, and J. E. Přenosil, Chemical
Engineering Dynamics. Modelling with PC Simulation. Weinheim: VCH
-6 Verlagsgesellschaft, 1994.
[3] P. Dostál, V. Bobál, and J. Vojtěšek, "Simulation of steady-state and
0 100 200 300 400 500 600 700 800 dynamic behaviour of a tubular chemical reactor," in Proc. 22nd
Time (s) European Conference on Modelling and Simulation, Nicosia, Cyprus,
2008, pp. 487-492.
Fig. 12. Controlled output in the presence of unmeasured random
[4] F.L. Severance, System Modeling and Simulation. Chichester: Wiley,
disturbance in c A0 (α = 0.075). 2001.
[5] B.V. Babu, Process Plant Simulation. New Delhi: Oxford University
Press, 2004.
parameters were estimated only in the first (tracking) interval [6] J.P. Corriou, Process Control. Theory and Applications. London:
t < 150 s. The experiences of authors of this paper proved that Springer_Verlag, 2004.
an utilization of recursive identification using the delta model [7] G.P. Rao, and H. Unbehauen, "Identification of continuous-time
systems," IEE Proc.-Control Theory Appl., vol. 152, pp. 185-220, 2005.
in the phase of a constant reference and in a presence of step [8] R.H. Middleton, and G.C. Goodwin, Digital Control and Estimation - A
disturbances decreases the control quality. From this reason, Unified Approach. Englewood Cliffs: Prentice Hall, 1990.
during interval t ≥ 150 s, fixed parameters were used. The [9] S. Mukhopadhyay, A.G. Patra, and G.P. Rao, "New class of discrete-
time models for continuos-time systems," International Journal of
controlled output responses are shown in Fig. 13.
Control, vol. 55, pp. 1161-1187, 1992.
[10] D.L. Stericker, and N.K. Sinha, "Identification of continuous-time
systems from samples of input-output data using the δ-operator,"
7 Control-Theory and Advanced Technology, vol. 9, pp. 113-125, 1993.
6 [11] V. Kučera, "Diophantine equations in control – A survey," Automatica,
vol. 29, pp. 1361-1375, 1993.
5 [12] P. Dostál, F. Gazdoš, V. Bobál, and J. Vojtěšek," Adaptive control of a
w continuous stirred tank reactor by two feedback controllers", in Proc. 9th
4
y (K)

IFAC Workshop Adaptation and Learning in Control and Signal

3 Processing ALCOSP'2007, Saint Petersburg, Russia, 2007, P5-1 – P5-6
[13] V. Bobál, J. Böhm, J. Fessl, and J. Macháček, Digital Self-tuning
2 Controllers. Berlin: Springer Verlag, 2005.
1
0
0 100 200 300 400 500 600 700 800 900
Time (s)
Fig. 13. Controlled output in the presence of step disturbances (α = 0.1).

1214