You are on page 1of 6

Multivariable Model-Based Control Strategies for

Level Control in a Quadruple Tank Process

Anca Maxim, Clara M.Ionescu, Cosmin Copot, Corneliu Lazar


Robin De Keyser Department of Automatic Control and Applied Informatics
Department of Electrical energy, Systems and Automation "Gheorghe Asachi" Technical University
Ghent University Iasi, Romania
Gent, Belgium clazar@ac.tuiasi.ro
Robain.DeKeyser@UGent.be

Abstract— This paper presents three model-based control However, when the performance achieved with
strategies applied to a multivariable process. First, a simple and decentralized techniques is not sufficient, decoupling control
rather naive approach is employed, i.e. treating the process as schemes are employed to eliminate the effect of the undesirable
two SISO (Single Input Single Output) loops and design PID cross-couplings. There are three types of basic decoupling
controllers. Obviously, this approach is effective, but does not techniques [8], ideal - where the inverse of the process transfer
take into account the interaction between the loops. Next, function is computed, simplified - in which the decoupler is
interaction is compensated by using dynamic decouplers and obtained directly from the ratio of the original transfer function
control performance is improved. Finally, a multivariable IMC matrix elements and inverted decoupling control, also named
(Internal Model Control) method is applied. All the results were
feedforward decoupling control.
validated on the laboratory setup with coupled quadruple tanks
from Quanser. This is an interesting and challenging testbed for In this paper, both control techniques (decentralized and
control, i.e. it poses non-minimum phase transmission zeros. Our decoupling control) were implemented using a PID controller.
experimental results show that the IMC outperforms the PID These methods have been tested on a real plant which consists
control at the cost of additional design complexity. All controllers of a quadruple tank from Quanser (www.Quanser.com). The
were successfully tested for setpoint trajectory and disturbance process is multivariable, highly coupled and presents a non-
rejection and tackled well the noise in the system. minimum phase (NMP) transmission zero. This setup is then
used to illustrate the effectiveness of the proposed algorithms
Keywords— multivariable control, transmission zeros, non-
minimum phase, decentralized control, decoupling control, internal
in real-life tests. The third part of this paper focuses on Internal
model control. Model Control (IMC) [9]. The IMC design procedure for
MIMO systems given in [1] was implemented and the results
were compared with decentralized and simplified decoupling
I. INTRODUCTION control.
Industrial applications of inter-connected systems are
The outline of this paper is as follows. The setup is briefly
manifold. To name a few: petro-chemical, paper-making and
given in section 2. The control algorithms are described in the
water treatment plants. Many of these applications present
third section and their results are presented in the fourth
difficult control issues, such as time delays, multivariable
section. A conclusion section summarizes the main outcome of
interaction, and non-minimum phase dynamics [1].
this work.
Control of such systems can de done either centralized, or
decentralized (SISO). The second choice is still a dominant II. DESCRIPTION OF THE MIMO SYSTEM
preference in industrial application, mainly because it exhibits
a list of advantages: flexibility in operation, simplified design
and tuning [2]. Despite of its practical benefits, this The schematic overview of the multivariable system
methodology cannot suppress the interactions of the MIMO represented by the quadruple tank water level is presented in
(Multi Input Multi Output) process, so one input affects all the figure 1. The control objective is to regulate the level of the
outputs, not only the corresponding one [3][4][5]. A method to water in the two lower tanks by manipulating the water flows.
restrict the interactions, based on concepts of reference The plant has two manipulated inputs, i.e. the voltages of the
conditioning technique is given in [6]. When the coupling of two pumps Vp1(t) and Vp2(t) (expressed in Volts), and two
the process is significant, choosing the 'right' control structure controlled outputs, i.e. the water levels of the two tanks below,
(pairing problem i.e. which input is used to control each of the L2(t), the level in Tank2, and L4(t), the level in Tank4 (both
outputs) can ease the task of controller design. In order to solve expressed in cm). There is a strong coupling effect between the
the pairing problem, the Relative Gain Array (RGA) [7] tool inputs and the outputs, (e.g. in Tank2 there are two inputs: the
can be used. flow from Pump1 through Out2, marked with dashed red line,
and the flow from Pump2 through Out1, denoted with green

978-1-4799-2228-4/13/$31.00 ©2013 IEEE

343
continuous line, that is the output flow from Tank1). Hence, Because a decentralized architecture is in essence a SISO
the controlled level in Tank2 is influenced by the two inputs, control scheme, there are many ways in which the input-output
and by adjusting the percentage of water flow from each input, pairing can be chosen. To check the input-output pairings, we
one can change the system for having minimum phase or non- apply the relative gain array (RGA) method and obtain:
minimum phase dynamics

⎡ −0.49 1.49 ⎤
Λ=⎢ ⎥ (2)
⎣ 1.49 −0.49⎦

This suggests that the pairing 1-1/2-2 is not suitable, since the
main diagonal has negative values. This outcome was
expected, as already mentioned in the description of the
process, since u2 is the dominant input for y1. Hence, the 1-2/2-
1 pairing from Fig. 2 will be further used in the remaining of
the paper. The transfer function matrix for the changed pairing
is given by:

Fig. 1. Schematic diagram of the quadruple tank process from Quanser. See
text for explanations. ⎡ 2.49 1.64 ⎤
⎢ 2 18.43s+1 ⎥
178.82s +26.74s+1
Based on the configuration depicted in Figure 1, in Tank2 Ĝ ( s ) = ⎢ ⎥ (3)
⎢ 1.28 2.56 ⎥
there is a greater flow coming from Pump2, via Tank1, than the ⎢
flow coming directly from Pump1. This exotic situation ⎣ 15.92s+1 172.22s +27.59s+1 ⎥⎦
2

originates from the fact that the outlet diameter Out1 is bigger
than the diameter Out2, while the outgoing orifices from each In the schematic representation from Fig. 1 of the MIMO
tank Doi, i=1...4 have all the same diameter. The same situation closed loops, w1,2 denote the setpoints for the outputs y1,2, C1,2
applies for Tank4. It follows the conclusion that the dominant are the controllers for each loop, and Gi,j with i,j=1,2 are the
flow in the tanks 2 and 4 comes from the manner in which the transfer functions from input j to the output i.
physical coupling is implemented via the choice of the setup
[10], [11].
The identified process from the previously described
testbed has a non-minimum phase transmission zero, i.e. a
feature most important for model based control design. The
model transfer function matrix is given in (1):

⎡ 1.64 2.49 ⎤
⎢ 18.43s + 1 2 ⎥
178.82 s + 26.74 s + 1
G (s) = ⎢ ⎥ (1) Fig. 2. Control Scheme with 1-2/2-1 pairing.
⎢ 2.56 1.28 ⎥
⎢ ⎥
⎣ 172.22s 2 + 27.59s + 1 15.92 s + 1 ⎦

with the transmission zeros z1=-0.26; z2= 0.07; z3= -0.06;


z4= -0.05

III. CONTROL STRATEGIES

A. Decentralized control
The simplest way to control a MIMO system is to design a
controller for every input-output pair, in a SISO approach.
However, this mechanism of directly exploiting SISO methods
in a MIMO setup is efficient only in presence of weak
interaction between the loops. If the system is highly coupled, a Fig.3. Condition number analysis.
change in the reference in one loop will also affect the output
in the other loop significantly. Consequently, the control results Before the controllers were designed for the decentralized
may become unacceptable and a decoupled strategy must be control, the condition number was computed, which is
adopted, which will integrate the strong interactions in the defined as the ratio between the gains in the strong and weak
control design. directions [5]. This was computed for both full block system

344
γ = σ (Gˆ ) σ (Gˆ ) and the diagonal system γ = σ (Gˆ d ) σ (Gˆ d ) , In Fig. 5 the decoupling control scheme, used to derive the
decouplers is presented.
where Gˆ d stands for the transfer matrix obtained from the main
diagonal of (3), replacing the off diagonal with 0. The analysis
was done in order to see if the diagonal matrix (decentralized
system) is easier or more difficult to control than the full
system. In Fig. 3 are depicted the results. These suggest that for
the same range of frequencies the process in decentralized form
is easier to control than the full process.
After this investigation, and with the input-output pairing
convenient chosen, the next step is to design two feedback
controllers. When a process model is available, the controller
design can be done using computer aided design (CAD) tools.
Fig.5. Decoupling control scheme with 1-2/2-1 pairing. MC denotes
monovariable controller

In the control scheme of Fig.5, there are two feedback


controllers C1 and C2 and two decoupling controllers D21 and
D12. The decouplers Dij, i, j=1...2, have the role of canceling
the interaction effect from input j on the output i.
The decoupling transfer function matrix is defined by:

D = [ I − Gh ]−1 (4)

with I the identity matrix and

⎡ 0 D21 ⎤
Gh = ⎢ (5)
⎣ D12 0 ⎥⎦

In order to compute the transfer function matrix Gh which is


part of the decoupling matrix D, we start from the premises that
Fig.4 The Nichols plot of the FRTool CAD interactive graphical user the resulting decoupled system has the transfer matrix:
interface. C denotes controller and P denotes process transfer function (in this
case, direct path transfer function G(1,2), from input 2 to output 1,without
taking into account the coupling effect). ˆ
GOL = GG (6)
MC

In this paper, the CAD is based on the Frequency Response


tool (FRtool) for Matlab® as described in [12]. However, the where subscript OL comes from Open Loop.
reader may use the Root Locus approach (RLtool) in Matlab® With GMC is denoted the transfer function matrix of the
or any other model-based PID design method in order to feedback controllers and decoupling controllers and has the
produce a well-tuned PID. Fig. 4 shows the tuning of the PID form:
controller on the process P(s) with the specifications: overshoot
%OS<5% and settling time Ts<1000seconds. The
corresponding PID parameters are given in Table 1. Notice that ⎡C 0⎤
P(s) in case of decentralized control is the direct path transfer GMC = DC = D ⎢ 2 (7)
function G(1,2) for the first loop, and G(2,1) for the second ⎣0 C1 ⎥⎦
loop. The corresponding designed controllers are C1(FR2)
respectively C2 (FR4) from Fig. 2. In this way, interaction is where C is the transfer function matrix of the feedback
not taken into account. controllers.
From (6) and (7) through matrix algebra one has that:
B. Decoupling control
This control strategy takes into account the interactions
between the loops and compensates them through additional ⎡C 0 ⎤ −1 ˆ
Gh = I − ⎢ 2 G G (8)
decoupling controllers. Ideally the result will be two separate
⎣0 C1 ⎥⎦ OL
SISO systems, where the setpoint changes affect only the
desired controlled variables.

345
By definition, the decouplers D12 and D21 cause a zero
cross-interaction in the open loop system. Thus, the inverse of ⎡ 1 ⎤
⎢ − z s +1 ⎥
the resulting open loop transfer function matrix becomes: ⎢ 1 0 ⎥
⎢ s +1 ⎥
Gbad ( s ) = ⎢⎢ z
ˆ ⎥
⎥ (12)
⎡ 1 ⎤ 1
⎢C G 0 ⎥ ⎢ − s + 1⎥
−1
GOL =⎢
2 12 ⎥ (9) ⎢ 0 z ⎥
⎢ 1 ⎥ ⎢ 1 ⎥
⎢ 0 ⎥ ⎢⎣ s +1 ⎥
⎣ C1G21 ⎦ z ⎦

From (8) and (9) the decoupling transfer function matrix is:
The next step is to compute the controller by inverting only
the 'good part', Gˆ good
−1
( s ) . Because the resulting transfer matrix
⎡ G11 ⎤
⎢ 1 − is improper a filter F ( s) is added to make it (semi)proper:
G12 ⎥
D=⎢ ⎥ (10)
⎢ G22 ⎥
⎢− 1 ⎥
Q ( s ) = Gˆ good
⎣ G21 ⎦
−1
( s) F ( s) (13)

C. Internal Model Control The choice of the filter is the 'extended' filter, optimal for
ramp setpoint and useful for step disturbance at the input of the
The IMC (Internal Model Control) principle given in Fig.6 process:
requires a model of the process in order to compute a
controller. The main idea is to design a controller/compensator
which is based on the inverse of the model of the process. This 1 + (nλ − p1 ) s
approach works in the case of a traditional process. However, Fe ( s ) = (14)
(1 + λ s ) n
in our case, the non-minimum phase transmission zero will
become an unstable pole for the controller and care must be
taken in the controller design phase. with n such that Q ( s ) is (semi)-proper, ߣ a tuning parameter
and p1 is the coefficient of the numerator of the transfer
function from the diagonal of (12)
Gˆ bad _ num ( s ) = 1 + p1 s + p2 s 2 + …

IV. RESULTS
In the decentralized/decoupling scheme, the feedback
controllers were designed via FRTool using the specifications
given in section III.
Fig.6. Generic IMC control scheme.
In IMC controller, starting from the model transfer function
The solution of this problem is to factorize the model defined in (3), after the factorization (11) we can solve for the
transfer matrix into an invertible part Gˆ good ( s ) and a non- 'good matrix' form Gˆ good ( s ) = Gˆ bad
−1
( s )Gˆ ( s ) and find that
invertible part Gˆ bad ( s ) :

⎡ −13.71s + 1 ⎤
0
Gˆ ( s ) = Gˆ bad ( s ) Gˆ good ( s ) (11) ⎢ 13.71s + 1 ⎥ 13.71s + 1
Gˆ = ⎢ ⎥ Gˆ (15)
⎢ −13.71s + 1 ⎥ 
−13.71
s + 1

0
Many methods for factorization are available in the ⎢⎣ 13.71s + 1
⎥⎦ Gˆ good
literature [14]. In this paper, all-pass factorization [1] is used. 
Gˆ bad
This method consists in placing the RHP (Right Half Plane)
transmission zero, denoted z, on the diagonal of the non- Thus the controller is
invertible part together with a pole at the reflection of the RHP
zero. Also the 'bad part', must be normalized with the static
gain=1 in order to ensure no steady state error. −1
⎡ 13.71s + 1 ˆ ⎤
Q( s ) = ⎢ G⎥ Fe ( s ) (16)
⎣ −13.71s + 1 ⎦

346
where Fe has the form: In Fig. 8 are presented the results for disturbance rejection
performance for the same PID controllers. The disturbance
consists in eliminating the water from Tank1 directly in the
reservoir, without going to Tank2 first. As a result, the level in
⎡ 25.71s + 1 ⎤
⎢ 0⎥ Tank2 drops and the controller changes the pump voltage to
36s 2 + 12 s + 1
Fe ( s ) = ⎢ ⎥ (17) recover the reference value of 11cm. Due to the coupling
⎢ 25.71s + 1 ⎥ between the pumps (recall Fig 1), the level in Tank4 will be
⎢ 0
⎣ 36s 2 + 12s + 1 ⎥⎦ also disturbed.
Next we test decoupled control. Fig. 9 and Fig. 10 depict
All the controllers were tested in two experiments the results obtained with the PID controllers 're-tuned' for the
consisting in setpoint tracking and disturbance rejection. In the 'augmented process' (i.e. plant + decoupling controllers). As
setpoint tracking test, the reference for water level in Tank2 is expected, the additional compensators cancel (in part) the
changed from 10cm to 11cm, while keeping the reference for strong interaction between the two loops. This can be observed
water level in Tank4 constant to 10cm. Next, the reference for by the weak reaction in Tank2 (FR2) at the setpoint change for
Tank4 as been changed to 11cm, while keeping the water level Tank4 (FR4), or, respectively, in the disturbance rejection
in Tank2 constant to 11cm. These values were chosen because experiment.
the quadruple tank process was linearized around the operating
point of 10cm, and a higher variation for the level value gave An alternative method to further refine the results obtained
disappointing results, obviously due to nonlinearity. with the decoupled control scheme may be to use IMC
procedure. In this paper, this direction was followed and the
Fig. 7 depicts the results of the closed loop performance for results are presented in Fig.11 and Fig.12.
the setpoint tracking with PID controllers tuned with FRTool in
decentralized control strategy. The controllers for Tank2 and 11.5

Tank4 are denoted by FR2 and FR4, respectively. The strong 11


Ref2

interaction between the loops can be observed when the

Level [cm]
FR2
setpoint changes, fact which motivated the next step i.e. that of 10.5

FR4
decoupled control design. 10

Ref4
9.5
700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800
11.5 Time[s]
Ref2
11
Level [cm]

6.2
FR2
10.5 6
FR2
Voltage [V]

FR4
5.8
10 5.6
Ref4 5.4
9.5
700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 5.2
Time[s] FR4
5
700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800
Time[s]
6.5 FR4
Voltage [V]

6
Fig.9. Real life test for setpoint tracking in decoupled control
5.5
FR2
5

4.5
700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
11
Time[s]
10
Level [cm]

FR2
9 FR4
Fig.7. Real life test for setpoint tracking in decentralized control.
8

7
12
1950 2000 2050 2100 2150 2200 2250 2300 2350
FR2
Time[s]
Level [cm]

11

9
10
FR4 8
Level [cm]

7 FR4
9
1950 2000 2050 2100 2150 2200 2250 2300 2350 2400 6
Time[s] 5
FR2
4
8
3
1950 2000 2050 2100 2150 2200 2250 2300 2350
Voltage [V]

7 FR4 Time[s]
6

5
FR2

4 Fig.10. Real life test for disturbance rejection in decoupled control


1950 2000 2050 2100 2150 2200 2250 2300 2350 2400
Time[s]

Fig.8. Real life test for disturbance rejection in decentralized control.

347
11.5
Ref2
11

Level [cm] IMC2


10.5
IMC4 V. CONCLUSIONS
10

9.5 Ref4 This paper presented a successful implementation and real-


9
400 500 600 700 800 900 1000
life test of three model-based control strategies applied to a
Time[s] multivariable system with NMP transmission zero and
7
significant coupling. Firstly, a decentralized control is
6 IMC2 implemented, without taking into account the interaction
Voltage [V]

IMC4
5 between the loops. As expected, the results suggest that a
4 decoupling approach may give a better performance in terms of
3
400 500 600 700 800 900 1000
coupling. Both decentralized and decoupling controls achieve
Time[s] poor performance for disturbance rejection tests. This
motivates the application of the more advanced IMC control.
Fig.11. Real life test for setpoint tracking in IMC control with Fe filter. The experimental results confirm the necessity of more
complex control algorithms when stringent performance is
Fig. 11 depicts the results of the closed loop performance envisaged and coupling, as well as RHP zeros need to be
for setpoint tracking with the IMC controllers, designed with tackled efficiently. Further extension of this work could be
the filter from (14) and diminished interaction between the testing other procedures (e.g. model based predictive control)
loops can be noticed. to improve the results by adding constraints and maximize the
closed loop bandwidth and nonlinear control.
12

REFERENCES
Level [cm]

10

IMC2
8
IMC4

6 [1] W. Bequette, Process Control: Modeling, Design and Simulation,


1200 1250 1300 1350 1400 1450 1500 1550 Prentice Hall Professional, 2003.
Time[s]
[2] P. Campo and M. Morari, "Achievable closed-loop properties of systems
12 under decentralized control: conditions involving the steady-state gain",
10 IEEE Transactions on Automatic Control, vol. 39(5), pp. 392-394, 1994.
Voltage [V]

IMC4
8
[3] K. Åström, K. Johansson and Q. Wang, "Design of decoupled PID
6

4
IMC2 controllers for MIMO systems", Proceeding of the American Control
2 Conference, Arlington, VA., pp. 2015-2020, 2001.
1200 1250 1300 1350 1400 1450 1500 1550 [4] J. Freudenberg and R. Middleton, "Design rules for multivariable
Time[s]
feedback systems", Proceedings of the Conference on Decision and
Control, Kobe, Japan, pp. 1980-1985, 1996.
Fig.12. Real life test for disturbance rejection in IMC control with Fe filter. [5] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control:
Analysis and Design , Wiley, New York , 1996
Fig. 12 describes the results obtained for the disturbance [6] F. Garrelli, R.J. Mantz and H. De Battista, "Limiting interaction in
decentralized control of MIMO systems", Journal of Process Control,
rejection experiment. The setup for all the experiments vol. 16,(5), pp. 437-483, 2006.
performed for this paper was identical. The only difference was [7] E. Bristol, "On a new measure of interaction for multivariable process
in disturbance rejection test, where there was impossible to control", IEEE Transactions on Automatic Control, vol. 11, pp. 133–
perform identical experiments (i.e. by manually open the 134, 1966.
disturbance tap from Tank3). Considering this, it can be [8] W.J. Cai, W. Ni, M.J. He and C.Y. Ni, "Normalized Decoupling-A New
Approach for MIMO Process Control System Design", Ind. Eng. Chem.
observed that the IMC controller outperforms the PID Res., vol. 47 (19), pp. 7347-6356, 2008.
controller. [9] C.E. Garcia and M. Morari, "Internal Model Control-1.A unifying
review and some new results",Ind. Eng. Chem. Process Des. vol. 21. pp.
The FRTOOL-PID controller parameters are presented in 308-323, 1982.
Table I. The IMC controller was developed, with the extended [10] K.J. Johansson, "The Quadruple-Tank Process. A Multivariable
filter described in section III, and the values for the tuning Laboratory Process with an Adjustable Zero", IEEE Trans. Cont.
parameters n=2 and λ = 6 . Syst.Tech., vol. 8(3), pp. 456-465, 2000
[11] K.J. Johansson, A. Horch, O. Wijk and A. Hansson, Teaching
Multivariable Control Using the Quadruple-Tank Process, Proceedings
TABLE I. PID CONTROLLER PARAMETERS of the 38th IEEE Conference on Decision and Control, pp. 807-812,
1999
PID CONTROLLER PARAMETERS [12] R. De Keyser and C.M. Ionescu, "FRtool: a frequency response tool for
Controller CACSD in Matlab", IEEE Conf. on Computer Aided Control Systems
Kp Ti Td Design, Munich, pp. 2275-2280, 2006
FR2 (Fig 7/8) 0.035 4 1 [13] M. Morari and E. Zafirou, Robust Process Control, Prentice-Hall,
Englewood Cliffs, New Jersey, pp. 293-324, 1989.
FR4 (Fig 7/8) 0.05 4 1
FR2 (Fig 9/10) 0.02 0.68 2.73
FR4 (Fig 9/10) 0.02 0.68 2.73

348

You might also like