Computers

&Chemical
Engineering
ELSEVIER Computers and Chemical Engineering 24 (2000) 1503-1509
www.elsevier.com/locate/compchemeng

Model based control of a four-tank system

Edward P. Gatzke, Edward S. Meadows, Chung Wang, Francis J. Doyle III *
Department of Chemical Engineering, University of Delaware, Newark, DE 19716, USA

Abstract

A multi-disciplinary laboratory for control education has been developed at the University of Delaware to expose students to
realistic process system applications and advanced control methods. One of the experiments is level control of a four-tank system.
This paper describes two model-based methods students can implement for control of this interacting four-tank system. Sub-space
identification is used to develop an empirical state space model of the experimental apparatus. This model is then used for model
based control using internal model control (IMC). This represents an application of inner-outer factorization for non-minimum
phase multivariable IMC design. Modeling is also performed using step tests and Aspen software for use with dynamic matrix
control (DMC). 0 2000 Elsevier Science Ltd. All rights reserved.

Keywords: Predictive control; Internal model control; Process control education; Experimental apparatus

1. Introduction multidisciplinary group facilitates peer learning in that

students familiar with the principles of a given experi-
Process control courses for chemical engineers often ment must help other group members. The students
emphasize complex theoretical and mathematical issues enrolled in the process control laboratory have all
while devoting limited time to implementation of appli- taken the basic process control course taught in their
cation of control methods. In order to reinforce and individual departments. The basic knowledge of process
demonstrate the concepts presented in a lecture, practi- control allows the experimental laboratory course to
cal laboratory applications can be developed for stu- cover advanced applications of process control.
dents. The experience of working on a laboratory A four-tank level control system has been con-
experiment in an academic setting exposes students to structed for use in the control laboratory. This paper
realistic industrial problems. Laboratory work can ex- details the use of model identification and model-based
pose students to process details that are often neglected control methods for control of this interacting four-
in computer simulation, including measurement noise, tank system. The four-tank system is based on the
measurement bias, process nonlinearity, equipment fail- system presented by Johansson and Nunes, 1998. A
ure, actuator constraints and external disturbances. schematic of the process is shown in Fig. 1. The exper-
A multi-disciplinary process control laboratory has iment has two inputs (pump speeds) which can be
been developed at the University of Delaware. Experi- manipulated to control the two outputs (tank levels).
ments currently include an inverted pendulum, an elec- The system exhibits interacting multivariable dynamics
tric servo motor, a gyroscope, a distillation column, a because each of the pumps affects both of the outputs.
spring mass damper system, a virtual boiler and a The system has an adjustable multivariable zero that
four-tank system. The undergraduate experimental con- can be set to a right-half or left-half plane value by
trol course is offered every year in the spring semester. changing the valve settings of the experiment. Unmea-
Students from mechanical, electrical and chemical engi- sured disturbances can be applied by pumping water
neering disciplines participate in this class. The students out of the top tanks and into the lower reservoir. This
are grouped with students from other majors. This exposes students to disturbance rejection as well as
reference tracking.
* Corresponding author. A full nonlinear mass balance model of the system is
E-mail address: [email protected] (F.J. Doyle III). given in Table 1. Here, Bernoulli’s law is used for flows

0098-1354/00/$- see front matter 0 2000 Elsevier Science Ltd. All rights reserved.
PII: SOO98-1354(00)00555-X
 1504 E.P. Gatzke et al. /Computers and Chemical Engineering 24 (2000) 1503-1509

I
out of the tanks, hi is the level of water in tank i; u, and
+ + _ u2 are the manipulated inputs (pump speeds); and d,
Tank 3 Tank 4 and d2 are external disturbances representing flow out
of tanks three and four. The linearized model is given
in Table 2 and the estimated model parameters for the
experimental setup are given in Table 3. Ai is the area
of tank i and ai is the area of the pipe flowing out of
tank i. The ratio of water diverted to tank one rather
Tank 1 Tank 2
than tank three is yr and y2 is the corresponding ratio
diverted from tank two to tank four. It can be shown
that for the linear system, a multivariable right half
plane zero will be present when yr + y2 < 1.
Sub-space identification is used to develop an empiri-
cal linear state space model of the experimental appara-
Fig. 1. Schematic of the interacting four-tank process. The two tus. This model is then used for formulation and
manipulated variables are the pump speeds. The two controlled application of model-based control methods. The iden-
variables are the levels of tanks one and two. Unmeasured flow
disturbances can affect tanks three and four.
tified state-space model is used explicitly in the control
algorithms. The methods considered in this work in-
clude multivariable internal model control (IMC) and
dynamic matrix control (DMC). Non-minimum phase
Table 1 behavior of the system requires that inner-outer factor-
Non-linear model equations for the four-tank system
ization is used for the multivariable IMC design. The
right-half plane pole also creates performance limita-
tions for the closed-loop system. Because the system
has an adjustable zero, the system can be adjusted to
exhibit minimum phase behavior and demonstrate im-
proved control response.
The physical system is constructed so that it appears
to emulate elements of an industrial unit operation.
Four five-gallon tanks are used in the simulation. The
addition of submersible pumps in the tanks can simu-
Table 2 late a leak disturbance in the tanks. An industrial
Linearized model equations for the four-tank system
distributed control system (DCS) from ABB-Bailey is
used for control of the system using the freelance
software package. The operator interface relies on a PC
communicating with the Bailey Process Station using
TCP/IP on a private LAN. An interface has been
developed using dynamic data exchange (DDE) that
i
dt 1 allows controllers developed using MATLAB and Sim-
0 0 - 0
T3 ulink to be used for simulation runs. The apparatus has
1 the flavor of an industrial system while retaining the
0 0 0
T4 flexibility needed to quickly implement and test ad-
J vanced control strategies.

0
0 0 2. Sub-space process modeling
Y&s 0 0
A3 kdI d Most advanced process control methodologies re-
U+
( 1- Ydkz - 0
0 ‘43 quire the development of accurate models of the sys-
A3 0 k d2 tem. For identification, the process is typically forced
(1-*/,)k, A, by known inputs and the resulting output responses are
0
4 used for development of a linear model relating system
inputs to outputs. Some process data should be used
for validation of new process models. In industrial
applications, one must consider tradeoffs between the
need for accurate process models, the disadvantage of
 E.P. Gatzke et al. /Computers and Chemical Engineering 24 (2000) 1503-1509 1505

Table 3 The resulting input and output data are used to

Model parameters of the experimental four-tank system create the linear model. For this task, a sub-space
identification procedure is used. This procedure is de-
ah a2 2.3 cm* k, 5.51 cm3/s
2.3 cm2 6.58 cm3/s scribed in Van Overshee and De Moor, 1994. The
6, a4 k2
A,, A,, A,, A, 730 cm2 g 981 cm/s2 method develops an empirical linear state space
QJ) 60% YI 0.333 model from input and output data. For this type of
60% 0.307 modeling, no prior process knowledge is used; no as-
CO)
u269

i-1 53.8 s 14.1 cm

48.0 s 11.2 cm
sumptions are made about the state relationships or
T2 h(O)
T3 38.5 s h,(O) 7.2 cm number of process states. Only the number of states
T4 31.1 s h&‘) 4.7 cm used in the resulting process model must be deter-
mined.
It should be noted again that the sub-space identifi-
cation method is an empirical method. In many cases,
a model based on fundamental process knowledge
(first principles modeling) may be more desirable.
This may be true for a process with significant non-
linear characteristics and varied operating regimes.
For process that operates around a single steady state
or have little nonlinear character, empirical linear
modeling can effectively be used for process control.
Although there are fundamental models available
for the four-tank system, the sub-space identification
method was used to show that empirical modeling
could effectively be applied to real systems. Sub-space
identification of a four-tank process has been demon-
strated in Johansson and Nunes, 1998. For the cur-
rent study, the normal input levels were 60% for both
Fig. 2. Model validation for sub-space identification on the four-tank pumps. A binary sequence was used that switched
system.
between levels of 40 and 80%. Switching could occur
every 35 s. The resulting output levels never exceeded
+ 8 or - 10 cm in relation to the nominal operating
point. The N4SID algorithm from the Matlab Iden-
tification Toolbox was used to calculate the linear
models. Four states are needed to effectively capture
the process character. This result agrees with knowl-
edge of the process and the existing fundamental
Fig. 3. Internal model control (IMC) block diagram showing ‘Inner- model. Fig. 2 shows a comparison of the process and
Outer’ factorization of plant G(s) into the invertible portion M(s) - ’ the model, as well as process residuals.
and the non-invertible portion N(s). F(s) is a filter used to adjust the
aggressiveness of the control loop.
3. Internal model control
large excursions from normal operating conditions
during modeling, the problem of extended periods of Internal model control (IMC) is a very effective
process down time and the desire for ‘plant friendly’ method of utilizing a process model for feedback con-
input sequences. trol. IMC directly uses the process model and re-
To effectively balance the process modeling trade- quires very limited on-line computation. For a full
offs, one may employ a pseudo-binary random binary discussion of IMC, see the monograph by Morari
sequence (PRBS) of inputs. This type of input se- and Zafiriou, 1989. IMC uses a process model and
quence can effectively excite a multivariable process ‘inverts’ parts of the model for use as a controller for
for use in identification and provide an accurate pro- the process (Fig. 3). Some portions of a linear pro-
cess model while avoiding large process excursions cess model cannot be inverted. These non-invertible
and periods of off specification production. The factors include time delays and right-half-plane (RHP)
PRBS of inputs can also be considered a ‘plant zeros. In addition, a process model that is not semi-
friendly’ sequence because only two different levels of proper cannot be inverted. A linear filter can be
input are used. added to make the process model invertible. The filter
1506 E.P. Gatzke et al. /Computers and Chemical Engineering 24 (2000) 1503-1509

parameters then are available for adjusting the aggres- O=(A-BR-‘Q=C)=X+X(A-BR-‘Q=C)

siveness of the IMC controller.
- X(BR - ‘)(BR - ‘)=X
The following ‘inner-outer’ factorization for the sta-
ble process G(s) follows the procedure described in The controller for the IMC formulation is the inverse
Morari and Zafiriou, 1989. The linear process transfer of the invertible portion of the process model. For
function can be written: offset free steady state reference tracking in all chan-
nels, the product of the controller inverse gain and the
G(s) = C(sZ - A) - ‘B + D = N(s)M(s) - ’
process model gain must be identity. The current pro-
where N(s) and M(s) are stable. Additionally, Z= cess model factorization does not guarantee this. It can
N(i~)~N(icu). The non-invertible portion of the pro- conveniently be achieved by scaling the non-invertible
cess, N(s), is given by: portion of the plant by N(0) - ‘. Now, the non-invert-
ible process model is N(s)N(O) - ’ and the invertible
N(s)=(C-QF)(sZ-(A-BR-‘F))-‘BR-‘+Q process model is N(O)M- l(s). The IMC controller be-
comes M(s)N(O) - ‘. SISO IMC systems incorporate a
For N(s), the state space matrices are given as:
scalar filter for strictly proper process models so that
A,,,= A- BR-‘F BN= BR-’ the resulting controller is semi-proper. In this multivari-
able case, the process model is already semi-proper.
C,= C- QF Pv=Q Each of the error signals sent to the MIMO 2 x 2 IMC
controller can be filtered with a first-order linear filter,
The invertible part of the process, M(s) - ‘, is:
F&) = l/np + 1. This filter allows for adjustment of the
M(s)-‘=F(sZ-A)-‘B+R closed-loop dynamics.
Fig. 4 shows closed-loop performance of the IMC
For M(s), the state space matrices are given as: system for reference changes. In this example, li are
selected to 7 s.
A,=A B,=B The process has zeros at values of 0.0324 and -
C,=F D,=R 0.088. The input direction corresponding to the RHP
(as described in Skogestad & Postlethwaite, 1996) is:
The following relation is given for R and Q: [ - 0.736 0.677]= and the output direction is [ -
0.6930.7211T. Multivariable performance limitations re-
D=QR lated to RHP systems are described in Johansson and
Rantzer, 1997. In the nominal case, the process is
For this example, the orthogonal matrix Q is selected
identical to the process model and the filter time con-
as identity. To make the process semi-proper, D is set
stants are set to 0. The complementary sensitivity func-
to EZ, with E = 0.0001. For the factorization method, F
tion, T(s), for the nominal IMC system reduces to
is calculated as
N(s)N(O) - l. Ideally, T(s) is identity at all frequencies.
F = Q’C + (BR - ‘)=X The presence of a RHP zero creates a performance
limitation for the system. In Morari and Zafiriou, 1989,
where X is the solution of the algebraic Riccati it is shown that T(s) for a non-minimum phase system
equation: with a single zero can be arbitrarily selected so that
only a single row deviates from identity. This implies
that the performance degradation caused by the RHP
zero can be driven into a single output channel. The
complementary sensitivity function for the 2 x 2 tank
system with ideal performance in the first measurement
is:

1
1 0
T’(s)=[ pls/s+(- -s++/s+i

and the complementary sensitivity function for the sys-

tem with ideal performance in the second measurement
is:

1
Fig. 4. Closed-loop reference tracking using IMC controller with
-s++/s+s BzsIs+~
T*(s)= 0
1
A=7 s. [
 E.P. Gatzke et al. /Computers and Chemical Engineering 24 (2000) 1503-1509 1507

4. Dynamic matrix control

Dynamic matrix control (DMC) refers to the com-

mercial implementation of model predictive control
(MPC) provided by Aspen Technology. Aspen is a
member of the University of Delaware Process Moni-
toring and Control Consortium and has provided DM-
CPlus for student use in University of Delaware
teaching laboratories.
Like other MPC algorithms, DMC uses an optimiza-
tion criterion to choose control moves. Using an ex-
plicit process model for prediction of process outputs,
DMC chooses optimal sequences of control moves
based on a trade-off between speed of setpoint tracking
Fig. 5. Screen shot of the DMC four-tank system modeling interface.
and avoidance of rapid changes in control inputs. Some
specific features of the DMC algorithm include the
following:
applicability to multivariable and non-square
systems;
a step response model;
quadratic optimization criterion;
explicit trade-off between setpoint tracking and ag-
gressive of control action through an ‘input suppres-
sion factor’;
continuous adjustment of setpoints via linear pro-
gramming to target the most profitable steady-state
operating condition.

DMC has been presented for control of a similar

process (Dai & Astrdm, 1999. In the DMCPlus imple-
mentation presented here, students in the control teach-
Fig. 6. Screen shot of the four-tank system DMC interface showing a
ing laboratory learn an advanced multivariable control
closed-loop run. package that is not only industrially relevant, but is
identical to that found in industry. Fig. 5 shows the
DMCPlus modeling interface. Fig. 6 shows a screen
where /3i and lj2 are functions of the terms in the shot of the DMCPlus interface to the four-tank system.
output zero direction. Significant interaction can occur The DMCPlus system permits students to explore sev-
when the values of pi are large. For the given system, eral interesting aspects of the four-tank system:
/3r is 1.9 and p2 is 2.1. This implies that choosing either controller tuning: Aggressive tuning of a DMC or
input for ideal tracking in the nominal case will have other MPC controller has the effect of computing a
essentially the same amount of interaction. This result controller that is the inverse of the process. Since the
is expected due to the symmetric nature of the system. four-tank system can be adjusted to have a right
For the developed controller formulation, a first-order half-plane zero, aggressive controller tuning can lead
realization of T(s) is given as: to a controller that has a right half-plane pole and is
therefore unstable. Investigations on the limits of

1
0.039s + l/s + 0.32 1s + l/s + 0.32 controller tuning and ability to track setpoint
T(s) = changes offer interesting opportunities for student
Is/s + 0.32 - 0.039s + l/s + 0.32 experiments;
process identz&ztion: One of the first steps in imple-

1
1.2s + 1/32x + 1 32s/32s + 1 menting a DMC controller is the identification of a
= [ 32.~132~+ 1 - 1.2s + 1/32s + 1 step response model for the process. The associated
software tools can be used to store and implement
This also demonstrates that the system and controller different process models for different operating con-
formulation is symmetric. Both output channels should ditions. In this way, students can also investigate the
demonstrate equal performance limitations and effects of model mismatch on closed-loop perfor-
interactions. mance; and
1508 E.P. Gatzke et al. /Computers and Chemical Engineering 24 (2000) 1503-1509

Table 4 measured tank levels. This is equivalent to achiev-

Student evaluations of the multi-disciplinary laboratory experimentsa
ing a setpoint for tank levels with minimum
steady-state control action.
Ql Ql (SD.) Q2 Q2 (SD.)
(average) (average)

Four-tank 4.0 0.93 4.4 0.91

Spring 3.5 0.92 3.3 1.1
5. Conclusions
mass
Pendulum 2.9 1.3 3.2 1.1 This paper has described advanced modeling and
Servo 2.9 0.99 3.2 1.1 control techniques that can be used by students in a
motor hands-on experimental control laboratory. The stu-
Virtual 2.4 1.2 3.1 1.3
dents will be introduced to sub-space identification
boiler
for plant-friendly process identification. IMC is ap-
a There were 15 students taking the class. Question 1 was: “rate the plied to the non-minimum phase process by mathe-
overall educational value of each experiment using a scale of l-5 matically factoring the model into invertible and
(1 = poor, 5 = excellent)“. Question 2 was: “rate the practical value of non-invertible sections. DMC is used to control the
each experiment using a scale of l-5 (1 = poor, 5 = excellent)“.
system and explicitly implement process constraints.
Students are introduced to advanced modeling and
control techniques in an application based environ-
l gain scheduled DMC: DMCPlus contains features ment. They are able to connect classroom theory with
to permit process gains to vary with operating con- concrete laboratory experiences.
ditions, thereby incorporating a particular form of The first offering of the laboratory class used the
nonlinear model into the DMC algorithm. In the four-tank system for SISO and multivariable decou-
fourtank system, this permits models to be used pling. The advanced identification and control meth-
over a wider operating range, and students can ods presented in this paper are being developed for
compare the differences between gain-scheduled use in upcoming course offerings. The 15 students
DMC to singlemodel DMC. from the first offering of the laboratory class were
asked to evaluate the laboratory course. These stu-
4.1. Implementation details dents were asked to rate the overall educational value
as well as the practical value of each laboratory ex-
DMC implementation on the four-tank system has periment on a scale of 1-5, 5 being the highest.
some significant differences from typical industrial Table 4 shows the quantitative ratings of the separate
implementations: experiments used in the first offering of the course.
an industrial implementation of DMC is imple- Note that the four-tank system received the highest
mented in a cascade configuration that is built rating of the five experiments. The students were also
upon a system of well-tuned, single-input/single- asked to give a free response evaluation of each ex-
output (SISO) regulatory controllers. Industrial periment. Some representative student quotes from
DMC then chooses setpoints to be implemented in the free response section of the four-tank evaluation
these regulatory controllers to effect process out- include:
puts. With the simpler experimental process, it is
not possible to implement DMC in a cascade for- Very good lab, extremely educational.
mulation since the necessary instruments to mea- This was the most intuitive control problem for
sure flow are not present. Therefore, the DMC me, very easy to see the direct physical results of our
control action. My personal favorite.
controller directly manipulates pump speed to
Real problem using industrial interfaces. Excellent
achieve optimal dynamic performance; and
practical problem, but it takes too long due to the
industrial implementation of DMC depends upon
time constant.
economic criteria to determine steady state set-
points, based upon prices of input and output
streams in a process. Since the laboratory system
has no economic criteria assigned to its operations, References
we have to substitute equivalent criteria as DMC
tuning parameters. In these experiments, we as- Dai, L., & Astrbm, K. J.(1999). Dynamic matrix control of a
signed a negative value (a cost) to pump speed and quadruple tank process. In Proceedings of the fourteenth ZFAC,
(pp. 295-300). Beijing, China.
asked for the DMC controller to choose operating Johansson, K. H., & Rantzer, A. (1997). Multi-loop control of
points that would maximize profits (minimize minimum phase systems. In Proceedings of American Control
costs) subject to a minimum-level constraint on Conference, (pp. 3385-3389). Albuquerque, NM.
 E.P. Gatzke et al. /Computers and Chemical Engineering 24 (2000) 1503-1509 1509

Johansson, K. H., & Nunes, J. L. R. (1998). A multivariable labora- Skogestad, S., & Postlethwaite, I. (1996). M&variable feedback
tory process with an adjustable zero. In Proceeding of American control. New York: Wiley.
Control Conference, (pp. 2045-2049). Philadelphia, PA. Van Overshee, P., & De Moor, B. (1994). N4SID: subspace al-
Morari, M., & Zafiriou, E. (1989). Robust process control. Englewood gorithms for the identification of combined deterministic-stochas-
Cliffs, NJ: Prentice-Hall. tic systems. Automatica, 30(l), 75-93.