See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228436191

Dynamic Matrix controller based on Sliding Mode Control.

Article · January 2008

CITATIONS READS

0 578

4 authors:

Oscar Camacho Luís Valverde

Escuela Politécnica Nacional University of the Andes (Venezuela)
192 PUBLICATIONS   1,067 CITATIONS    4 PUBLICATIONS   1 CITATION   

SEE PROFILE SEE PROFILE

Edinzo Iglesias Francklin Rivas

University of the Andes (Venezuela) University of the Andes (Venezuela)
22 PUBLICATIONS   133 CITATIONS    162 PUBLICATIONS   498 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Salones Inteligentes View project

Hybrid control schemes for chemical processes with dominant time-delay View project

All content following this page was uploaded by Francklin Rivas on 15 August 2014.

The user has requested enhancement of the downloaded file.

 AMERICAN CONFERENCE ON APPLIED MATHEMATICS (MATH '08), Harvard, Massachusetts, USA, March 24-26, 2008

Dynamic Matrix controller based on Sliding Mode Control.

OSCAR CAMACHO1 LUÍS VALVERDE 2. EDINZO IGLESIAS. 3. FRANCKLIN RIVAS4
1
Postgrado en Automatización e Instrumentación. Universidad de Los Andes. Facultad de
Ingeniería, Mérida 5101. Venezuela(ocamacho@ula.ve)
2
Postgrado en Automatización e Instrumentación. Facultad de Ingeniería. Mérida 5101.
Venezuela
3
Escuela de Ingeniería Química. Universidad de Los Andes. Facultad de Ingeniería. Mérida
5101. Venezuela, (iedinzo @ula.ve)
4
Laboratorio de Sistemas Inteligentes. Universidad de Los Andes. Facultad de Ingeniería. Mérida
5101. Venezuela, (rivas@ula.ve)

Abstract: - This work presents a system of tuning equations, based on the sliding surface response, to
predict the changes in the parameters of the process, which were used to improve and to tune the predictive
dynamic matrix controller parameters. The controller presents a fixed structure and its tuning parameter
equations were developed relating the characteristics of the sliding surface and the characteristic
parameters of the first order plus deadtime model. Simulations on a blended tank with variable level that
presents non linear behavior were considered.

Key-Words: - Dynamic Matrix Controller, Sliding Mode Control, tuning parameters, robustness.

1 Introduction
Time delays or dead times appear commonly in the
problem of control of different systems, such as
The main limitation of the original SPC is related
chemical and manufacturing processes. Time
to the class of systems for which it could be used,
delays can be originated by several situations like
since it is restricted to stable plants. Later on this
transportation lags, the effects of recycle loops on
research field, Model Predictive Control (MPC)
systems or by the approximation of higher order
[2] begins at the end of the 70. These kinds of
systems by lower dimension ones. Also time delay
controllers use a dynamical model of the process,
systems can be originated naturally as a
to predict the effect of the future controller actions
consequence of the modeling process, as in the
on the system output.
case of chemical processes[1].
MPC includes a series of algorithms among which
Several controllers have been developed for stable
the Dynamic Matrix Controller (DMC) is one of
processes. When the time delay is located at the
the most important ones. DMC were developed for
input (or output) of the system, a commonly used
Cutler and Ramaker [3], and it has been used in the
strategy is to eliminate the effect of the delayed
industrial world, mainly in the petrochemical
signal to obtain a free delay system. This method
industries. DMC is a linear control technique
works only in the case of sufficiently small delay.
where the process is represented by a first order
An alternative approach consists in the
plus deadtime (FOPDT) model. The model
approximation of the delay term by the
response to an unit step change is used to predict
consideration of a Taylor series expansion or the
the future response of the dependent variables and
use of Padé approximations via a rational transfer
formulates a series of control actions for all the
function. In the case of linear systems, the classical
independent variables. The actions are selected to
strategy is to use the well-known Smith predictor
minimize the error of the process on the time
compensator (SPC) [1], which provides a future
horizon.
estimation of the output signal that can be used in
the design of a control feedback
DMC presents some advantages, they can be
mentioned as follows: Intuitive and simple tuning,
it can be used for systems with complex dynamic,

ISSN: 1790-5117 379 ISBN: 978-960-6766-47-3

 AMERICAN CONFERENCE ON APPLIED MATHEMATICS (MATH '08), Harvard, Massachusetts, USA, March 24-26, 2008

the multivariable case is easily implemented, it is In the PDMCr structure, equation (1) can be
favorable for systems with long delays and the expressed as function of the characteristic
inclusion of restrictions is simple for controller's parameters of the process. For the case of an
design. DMC also presents inconvenience such as FOPDT model are gain, time constant, dead time
the necessity of an appropriate process model. besides the suppression factor, therefore, the new
model can be represented as follows:
Sliding Mode Control (SMC) is a type of Variable ∆ M = f (K p , τ , t 0 , λ )⋅ E (2)
Structure Control (VSC) that was developed in the Equation (2) has the advantage of including the
Soviet Union [4, 5]. The techniques of SMC have effect of the process parameters in the control law.
been employees in diverse systems; robustness is Therefore, Eq. (2) can be adjusted taking into
one of its principal advantages in the control of account changes in process parameters. As
non linear and time variant systems and systems mentioned above, this control law was proposed in
with uncertainties. However, also presents 2006 by Iglesias and Smith as:
disadvantages such as when the delay is too big the
performance of the system decreases.
∆M =
1 
K p  
(
ZCHxn M  DadjT ⋅ Dadj + λ2P I )
−1
DadjT 

CHx( PH −n ) 
⋅E
This paper tries to take advantages of both
techniques, from the DMC the predictive (3)
advantages and from the SMC the robustness
attributes. Where
 1 − exp(− ki ⋅ Ts τ nuevo) 
Dadj = D ⋅   (4)
The paper is divided as follows: Section 2 presents  (
 1 − exp − ki ⋅ Ts τ prev 
)
a brief description of the DMC; section 3 shows Where:
the methodology, section 4 presents the results and ki is the i-th term in the D row, Ts is the sampling
finally in section 5 the conclusions are presented.
time τ nuevo is the new time constant of the
,
2 The Parametric DMC. process, τ prev is the old time constant D is the
The conventional DMC was designed as a strategy transient information after dead time. CH is the
to work with linear systems, or processes with control horizon, PH is the prediction horizon,
small deviations from its operating point. Z is a zeros vector, I is the identity matrix, n
Unfortunately, the industrial processes are represents the periods of sampling.
complex with nonlinear characteristics. Several
articles report that when the DMC is used in non
linear processes the process response can go from 3 Methodology.
very slow to oscillatory [6], [7], [8] and [9]. For To develop the new tuning equations a factorial
this reason, several proposals to modify the experiment was designed, and an analysis of
algorithm DMC and to improve its performance variance (ANOVA) was performed to determine
for non linear processes have been proposed. One the variables that have a significant effect on the
of these approaches was developed in 2006 by optimal suppression factor λ. The experiment
Iglesias and Smith [10], where they proposed the consisted in modeling a general process as a first-
structure of the Parametric DMC called PDMCr. order-plus dead time (FOPDT) and determine,
The new structure was designed to include variable using constrained optimization, the best λ value to
terms whose values can change as necessary minimize a cost function. The FOPDT model
adaptation of process variations. contains three parameters, process gain, Kp, time
constant, τ, and dead time to.
The control law proposed by Cutler and Ramaker
for the conventional DMC is expressed as follows: Therefore, the proposed tuning equations set, were
∆ M = (A T A + λ 2 I ) A T E
−1 obtained from experiments based on FOPDT
(1)
models, where each one of the parameters were
Where: varied in 10%, 30%, 50% as are shown in Table
A: Is the dynamic matrix 1. A total of 36 simulations were performed.
∆ M : Is the output vector
λ : Is the suppression factor
I : Is the identity matrix

ISSN: 1790-5117 380 ISBN: 978-960-6766-47-3

 AMERICAN CONFERENCE ON APPLIED MATHEMATICS (MATH '08), Harvard, Massachusetts, USA, March 24-26, 2008

Table1. Models used for the experiment

2

∆Kp 0

Kp τ t0 τ 10% 30% 50%

-2

-4

Maximum Peak
0.5 5 0.5 0.05 0.15 0.25 -6

-8
Minimum
1.5 10 1.0 0.15 0.45 0.75

S(t)
-10 Peak

-12
2.5 15 1.5 0.25 0.75 1.25 -14

Kp τ t0 τ ∆τ -16

-18

0.5 1.5 2.5 0.5 1.5 2.5 -20

0 50 tmin tmax 100 150
time(min)

1.0 3.0 5.0 1.0 3.0 5.0

Figure 2. Characteristic values of the sliding
1.5 4.5 7.5 1.5 4.5 7.5 surface response
Kp τ t0 τ ∆t 0 τ 3.2 Analysis of Variance.
Once completed the 729 simulations, taking into
0.05 0.15 0.25 0.05 0.15 0.25
account the characteristic values of the sliding
0.10 0.30 0.50 0.10 0.30 0.50 surface response as the output and as independent
variables the FOPDT characteristics parameters a
0.15 0.45 0.75 0.15 0.45 0.75
variance analysis (ANOVA) were prepared.
The variance analysis and the regression
The purpose of this conducted test is to observe if techniques are used to find out the non linear
changes in the process parameters can induce models that relate the sliding surface response
changes in the sliding surface. Fig. 1 shows one of characteristics and the FOPDT process parameters.
the tests.
Table 2. ANOVA considering the minimum pick.
0.65
Source Sum Sq d.f Mean Sq. F Prob>F
0.6

Kp 0.00579 2 0.0029 33.34 0

cset(t)

0.55

0.5 τ 0.00499 1 0.00499 57.39 0

0.45
0 5 10 15 20 25 30 35 40 45 50 t0/τ 0.00263 2 0.00132 15.13 0
time(min)
0.65 ∆Kp 0.00294 5 0.00059 6.76 0
0.6 c(t) ∆τ 0.00195 6 0.00032 3.73 0.0025
cset(t) y c(t)

cset(t)
0.55
∆t0/τ 0.00132 7 0.00019 2.17 0.0453
0.5
Error 0.00713 82 0.00009
0.45
0 5 10 15 20 25 30 35 40 45 50
time(min)
Total 0.10155 107

Figure 1. A test used in the experiment

.
3.1 Obtaining the Characteristic Values of Table 3. ANOVA considering the maximum pick
the Sliding Surface response.
Source Sum Sq d.f Mean Sq. F Prob>F
Using the values presented in Table 1, all
simulations were carried out. From the Kp 1582 2 791 41.47 0
simulations, it was observed that the S(t) τ 4437 1 4436.95 232.63 0
characteristic values response affected by changes t0/τ 293.2 2 146.62 7.69 0.0009
in the process parameters are: minimum time, ∆Kp 1062.9 5 212.59 11.15 0
maximum time, minimum pick, maximum pick ∆τ 331.7 6 55.28 2.9 0.013
and the period (see Fig. 2). ∆t0/τ 1130.4 7 161.49 8.47 0
Error 1564 82 19.07
Total 55956.7 107

ISSN: 1790-5117 381 ISBN: 978-960-6766-47-3

 AMERICAN CONFERENCE ON APPLIED MATHEMATICS (MATH '08), Harvard, Massachusetts, USA, March 24-26, 2008

Table 4. ANOVA considering the minimum time Y2: Model that takes into account the minimum
time
Source Sum Sq d.f Mean Sq. F Prob>F Y2 = 54.127 + 12.104 ⋅ x1 − 18.51 ⋅ x3 − 1.4772 ⋅ x4
Kp 317.2 2 158.623 40.99 0 − 0.64188 ⋅ x5 − 8.4622 ⋅ x6 + 1.783 ⋅ (x1 ⋅ x3 )
τ 781.6 1 781.647 202.01 0 − 22.511 ⋅ (x1 ⋅ x4 ) + 0.079645 ⋅ (x1 ⋅ x5 )
t0/τ 82.1 2 41.03 10.6 0.0001 + 4.3791 ⋅ (x1 ⋅ x6 ) + 10.11 ⋅ (x3 ⋅ x4 ) − 0.85161 ⋅ (x3 ⋅ x5 ) (6)
∆Kp 92.5 5 18.494 4.78 0.0007 + 16.291 ⋅ (x3 ⋅ x6 ) + 2.0966 ⋅ (x4 ⋅ x5 ) + 3.2492 ⋅ (x4 ⋅ x6 )
∆τ 121.2 6 20.205 5.22 0.0001 ( ) ( )
− 0.6627 ⋅ (x5 ⋅ x6 ) − 0.32554 ⋅ x1 − 11.749 ⋅ x3
2 2

+ 27.163 ⋅ (x ) + 0.36453 ⋅ (x ) − 5.8308 ⋅ (x )

∆t0/τ 478.7 7 68.391 17.67 0 2 2 2
4 5 6
Error 317.3 82 3.869
Total 10991.6 107 Y3: Model that takes into account the period

Y3 = 7.8002 + 1.7345 ⋅ x2 − 8.155 ⋅ x3 − 22.957 ⋅ x4

Table 5. ANOVA considering the period + 3.4038 ⋅ x5 − 29.598 ⋅ x6 − 1.8113 ⋅ (x2 ⋅ x3 )
− 1.2534 ⋅ ( x2 ⋅ x4 ) + 0.21358 ⋅ (x2 ⋅ x5 )
Source Sum Sq d.f Mean Sq. F Prob>F + 3.6559 ⋅ ( x2 ⋅ x6 ) + 13.709 ⋅ (x3 ⋅ x4 ) − 0.48949 ⋅ (x3 ⋅ x5 ) (7 )
Kp 482.6 2 241.28 25.45 0 + 38.242 ⋅ ( x3 ⋅ x6 ) + 1.0353 ⋅ (x4 ⋅ x5 ) − 8.9074 ⋅ (x4 ⋅ x6 )
τ 1494 1 1494.02 157.61 0 ( )
− 0.083948 ⋅ (x5 ⋅ x6 ) + 0.11496 ⋅ x2 + 8.5389 ⋅ x3
2
( ) 2

t0/τ 74.5 2 37.26 3.93 0.0234

( ) ( )
+ 15.8 ⋅ x4 − 0.64706 ⋅ x5 − 35.765 ⋅ x6
2 2
( )
2

∆Kp 659 5 131.8 13.9 0

∆τ 461.8 6 76.96 8.12 0 Where:
∆t0/τ 380.8 7 54.4 5.74 0 x1 : is the process gain.
Error 777.3 82 9.48
x2 : is the time constant
Total 19238.9 107
x3 : is the controllability relationship
x4 : is the gain variation (∆Kp )
To observe that values are significant on each x5 : is the constant time variation (∆τ)
output variable a limit is chosen for the value of P x6 : is the controllability relationship variation
in the statistical F. this limit should be smaller than ∆(t 0 / τ)
0.05.

3.3 Models of S (t) 4. Results

In the design of the different models were realized In this section a mixing tank [5], Fig. 3, is used to
non linear regressions to adjust the parameters of compare the proposed controller against the SMC
the surface as a function of the changes in the and the DMC.
process. Each pattern represents to an output
variable of the sliding surface. To verify the good
adjustment R2 is used to measure the global quality
of the pattern. Therefore, the models obtained are:

Y1: Model that considers the minimum peak

Y1 = −0.12361+ 0.05242⋅ x1 + 0.062124⋅ x3 + 0.030563⋅ x4
− 0.0050835⋅ x5 − 0.093941⋅ x6 − 0.035098⋅ (x1 ⋅ x3 )
− 0.077549⋅ (x1 ⋅ x4 ) + 0.00016054⋅ (x1 ⋅ x5 )
+ 0.011387⋅ (x1 ⋅ x6 ) − 0.010125⋅ (x3 ⋅ x4 )
+ 0.0011869⋅ (x3 ⋅ x5 ) + 0.087821⋅ (x3 ⋅ x6 )
+ 0.023212⋅ (x4 ⋅ x5 ) + 0.027632⋅ (x4 ⋅ x6 ) (5)
− 0.0076065⋅ (x5 ⋅ x6 ) − 0.0026925⋅ ( ) x12
( ) + 0.03492⋅ (x )
− 0.0071276⋅ x32 4
2

+ 0.00062833⋅ (x ) − 0.021659⋅ (x ) Figure 3. Mixing tank

2 2
5 6

ISSN: 1790-5117 382 ISBN: 978-960-6766-47-3

 AMERICAN CONFERENCE ON APPLIED MATHEMATICS (MATH '08), Harvard, Massachusetts, USA, March 24-26, 2008

Figure 4. Comparison among the different

In Table 6 can be observed the characteristic approaches
FOPDT model parameters, they are used to tune
the controllers. They are obtained by identification
[1]. DMC+SMC Controller
0.86 Conventional DMC Controller
SMC Controller
0.84 cset(t)
Table 6. FOPDT characteristic parameters.

T ran s m itte r O ut pu t, fra c t ió n T O (t )

Parameter Value 0.82

KP -0.78503 Fraction TO/Fraction CO 0.8

τ 2.0906 min
0.78
t0 3.5663 min
0.76

Figure 4 shows how the process response is 0.74

affected when the three controllers are used. As
can be observed the proposed approach presents a 0.72

better performance than the other two, the DMC 0.7

presents a more oscillatory response and the SMC
is slower. 0.68
3200 3300 3400 3500 3600 3700 3800 3900 4000
time (min)
Figure 5, also plots the same results as Figure 3.
The previous two charts have shown the Figure 5 Comparison among the different
advantages of this mixed scheme. approaches,
.

4.1 Performance.
DMC+SMC Controller Table 7 shows two performance indexes for the
Conventional DMC Controller
0.8
SMC Controller different controllers used in this work. IAE and
cset(t) ISE are used as the performance indexes.
T ran s m itte r O utpu t, frac tión T O (t)

0.75 Table 7. Performance Indexes

Index SMC DMC DMC+SMC

0.7
IAE 31.8981 75.1856 23.3072
ISE 1.5875 4.8948 1.1888
0.65

As can be observed in Table 7, both indexes show

0.6 than the proposed approach, DMC+SMC, presents
smaller indexes values than the others two.
800 850 900 950 1000 1050 1100 1150 1200
time (min)
5. Conclusions
In this work a system of tuning equations has been
designed based on the response of the sliding
surface. The equations were used to improve the
performance of the proposed approach. The results
showed in all the cases improvements in the
performance indexes with respect to the SMC and
the DMC. In spite of robustness indexes were not
presented in this paper, the results shown that the
proposed approach is less oscillatory than the
DMC and faster than the SMC.

ISSN: 1790-5117 383 ISBN: 978-960-6766-47-3

 AMERICAN CONFERENCE ON APPLIED MATHEMATICS (MATH '08), Harvard, Massachusetts, USA, March 24-26, 2008

6. – References

[1]Smith, C. A., and A. B. Corripio, (1997).

Principles and Practice of Automatic Process
Control, John Wiley & Sons, Inc., New York.
[2]Camacho, E. F. and Bordons, C (1998) Model
Predictive Control. Springer. London
[3]Cutler, C.R & Ramaker, B.C. (1979). Dynamic
Matrix Control-to Computer Control Algorithm.
Proceeding of the 86th National Meeting of the
American Institute of Chemical Engineering,
Houston, TX. WP5-B.
[4]Utkin, V. I., (1977), “Variable Structure
Systems with Sliding Modes”, Transactions of
IEEE on Automatic Control, AC – 22, pp. 212
– 222.
[5]Camacho, O. (1996). TO new Approach to
Design and Tune Sliding Mode Controllers for
Chemical Process, University of South Florida.
[6]Sanjuan, M. E. (2002). Design and
Implementation of Fuzzy Supervisors to Online
Perform Compensation of Nonlinearities in I
PILFER PID Control Loops, University of South
Florida.
[7]Peterson, T., Hernandez, E., Arkun, Y. and F.
Schork (1992). TO Nonlinear DMC Algorithm
and Its Application to Semi Batch
Polymerization Reactor, Chemical Engineering
Science, Vol. 47, N°4. pp. 737-753.
[8]Chang. C., Wang, S. and S. Yu (1992).
Improved DMC Design for Nonlinear Process
Control. AIChE Journal. Vol. 38, N°4, pp. 607-
610.
[9]Georgiou, A., Georgakis and W. Luyben
(1988). Nonlinear Dynamic Matrix Control for
High-Purity Distillation Columns. AIChE
Journal. Vol. 34, pp. 1287.
[10]Iglesias, E. & Smith, C. (2006), Using Fuzzy
Logic to Enhance Control Performance of
Sliding Mode Control and Dynamic Matrix
Control, University of South Florida, pp.68-77.

ISSN: 1790-5117 384 ISBN: 978-960-6766-47-3

View publication stats