Professional Documents
Culture Documents
www.elsevier.com/locate/jprocont
Abstract
A sliding mode control methodology using output information is demonstrated in application to the HDA-plant, a plant for
production of benzene. This process is a highly integrated, non-linear large scale process with non-minimum phase and relative
degree zero characteristics. The non-linear control law is designed on the basis of a linear observer-based control system. The non-
linear control law uses the states of the linear observer. The performance in the sliding mode is determined by a linear stable sub-
manifold of the linear closed loop control system chosen via a robust pole selection scheme. The sliding mode control is optimized
to operate in a wide operating region. # 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Sliding-mode control; H1-Control; Chemical processes
13 possible actuators are the most eective for control also veried by investigating the frequency responses of
(Table 3). Further, the external disturbances of Table 2 the linearizations in the considered operation envelope.
aect the closed loop system performance. Chemical The next section will consider the necessary theore-
plants are often subject to delays: These are to be tical background for controller design which is
expected within the HDA-plant for the product purity employed for the HDA-plant in connection with an
measurement. Cao et al. [10] investigated delays of up to order reduced linear model.
10 min. The HDA-plant model is highly non-linear. One
characteristic of the responses of output measurements
of the open loop plant to small positive and negative 3. An H1-controller based approach
step changes of the inputs is that they are not inverted
to each other. Fig. 2 shows that the non-linear fast Suppose a linear, time-invariant system is given by:
dynamics seem to match the responses of the Aspen- :
Custom-Modeller-provided linearizations, while slow x AxBuBd d; x 2 Rn ; B 2 Rnm ; Bd 2 Rnmd
dynamics prove to be highly non-linear. This has been yCx DuDd d; C 2 Rpn
st: sp: A B Bd
;
C D Dd
Table 1
Measurements available for control
where u 2 Rm n 5 m are the available actuators, d the
Measurements Required rangea Scaling plant disturbances and y 2 Rp the output measurements.
1. Flash inlet temperature 98102 (F) O1 14 A mixed sensitivity H1 -control problem can be posed
2. Production rate 240280 (lb mol/h) O2 15
3. Product purity At least 99.97 (%) (mol-%) O3 0:01
1
lb mol=h Table 2
4. Hydrogen to aromatics 4.95.1 O4 0:2
1
lb mol=h Disturbances inuencing the control
ratio
5. Flash outlet vapour 458460 (psia) O5 15 Disturbance Nominal value Scaling
pressure
a
1. Pre-ash cooler: cooling euent 59 (F) I7 9
Substance quantity: 1 lb mol=453.593 mol, (lb mol)=pound- temperature
mole: Temperature: T (F) 1.8T ( C)+32, (F) Fahrenheit; pressure: 2. Purge down stream pressure 350 (psia) I 8 50
1 psia=6.895 kPa, (psia) pounds per square inch (absolute pressure).
G. Herrmann et al. / Journal of Process Control 13 (2003) 129138 131
1
Table 3 and S1 s I GsK1 sWy s . A sucient condi-
Most eective actuators for low control eort, high robustness and tion for this is that TyH;0 ;r0 0 1 4 1. Weights Wy have
good disturbance rejection
been used with poles at 0 which precludes the use of the
Actuator Unita Scaling standard H1 -synthesis procedure of Zhou et al. [11].
Safonovs pole shifting methodology has thus been used [12].
lb mol=h
1. Benzene column splitter: reux ratio I 1 0:265
lb mol=h
2. Compressor horse power (hp) I 2 64:494
3.1. A linear observer based closed loop control system
3. Pre-ash cooler duty (BTU/h) I 3=2658.677
4. Gas feed ow (lb mol/h) I 4=98.500 with auxiliary non-linear control input signal
5. Toluene feed ow (lb mol/h) I 5=55.875
6. Purge outlet valve opening I 6=9.183 Considering only the controlled plant input u and the
a
measured output yHy from Fig. 3, the plant augmented
Power: 1 hp=745.71W, (hp)=horse power; energy: 1
BTU=1054.185155 J, (BTU) British thermal unit. with the H1 -design weights is given by G:
:
st: sp: x G AG BG xG
G ; 1
yHy CG DG u
according to Fig. 3 choosing the sensitivity weight Wy as
part of the controller and including integrators to achieve and a linear controller for (1) with auxiliary inputs uNL
a steady state error of 0. Further, the actuator is weighted and uX is
2 3
by Wuu and Wur, a lter for unstructured uncertainty and
: x^G
a lter limiting the derivative of the actuator signals. Wp st: sp: AK BK BNL BX 6 6 uK 7;
7
K x^G 4
shapes the class of disturbances d. The H1 -optimization
yK CK DK 0 0 uNL 5
criterion minimizes the H1 normTyH ;0 ;r0 0 1 of the uX
transfer
T T T
function TyH ;0 ;r0 0 from the exogenous inputs
;
T T r to the controlled outputs yH where the controller K1 AK ; BK ; CK ; DK has been
T
yHy ; yHuu ; yTHur (Fig. 3). Robustness with respect to designed for uK yHy and u yK so that x^G are obser-
unstructured additive uncertainty, G Wuu; kk1 < 1 ver states and at least a weak pole separation principle
st: sp:
of the nominal plant, G (A, B, C, D) is achieved if ([13], p. 178) applies to this closed loop system. The
Wuu K1 Wy S1 4 1 where K1 is the designed control matrices BNL, BX and the input uX are determined later.
1
Fig. 2. Responses to small positive/negative step-changes ( 1%) of the benzene column reux ratio for non-linear plant and linearization (, non-
linear responses to pos. step; ...., non-linear responses to neg. step; . . , linear responses to pos. step; - - - -, linear responses to neg. step).
132 G. Herrmann et al. / Journal of Process Control 13 (2003) 129138
Fig. 3. Augmented plant for H1 -optimization considering additive uncertainty Wuu ; kk1 4 1.
For the closed loop system investigated here, assume not aect the actual closed loop stability and this per-
the auxiliary signal uNL (Fig. 4) is added to the linear mits them to be neglected when considering stability.
controller output signal yK, so that u uNL yK ; uK The introduction of integral action ensures zero steady
yHy . For eG xG x^G , the closed loop can be derived state error for a given demand r=const. Having estab-
from a transfer matrix formula for a feedback connec- lished a stable linear closed loop system with pole
tion ([11], p. 66) separation of observer and feedback, a sliding-mode
2 3
state-feedback control can be constructed.
: xG
xG A A1 B 0 6 eG 7
6 7
: 2
eG A2 A3 B1 BX 4 uNL 5 3.2. Derivation of a sliding mode hyperplane from a
uX stable plant representation
Then there is a transformation matrix T and a matrix where the stable matrices and and A12 are dened
with eigenvalues l1 ; . . . ; lr ; lr1 ; . . . ; lNm such that in (5). By construction, the hyperplane dening the
sliding surface is given by
I 0
T ; S z1; : 0
S1 S2
1 A12 5 Suppose the sub-state can be forced to remain on the
TAT ;
0 sliding mode hyperplane in nite time then uncertainty and
disturbances within the range space of B2 can be com-
S2 2 Rmm ; 2 Rmm ;
pletely rejected and the control system dynamics are gov-
:
where erned by the reduced order dynamics z 1 z1 . Dene a
Lyapunov matrix P P T > 0; P 2 Rmm , where
S1 S2 0 Im U T ; V~ U^ V;
" # P T P Im and suppose uncertainty and dis-
^1 def turbances are in the range space of B2 and are parametric
^ ; V~ v1 . . . vr vr1... vNm
0 or constant bounded, then the non-linear control signal
S2 B2 1 P
uNL 1 zT1 T 2 ;
kPk NL 6
and U 2 RNN ; V 2 RNmNm orthogonal and ^ 1 > 0;
^ 1 2 RNmNm diagonal resulting from a singular value NL > 0; 1 ; 2 5 0
decomposition, if and only if INm 0 V~ has full rank.
can achieve sliding, 0, for NL 0 in nite time and
Proof. For brevity, the proof is sketched. In Herrmann large enough 1 ; 2 [17]. However in practice, the choice
[18], the full proof is provided. Suppose NL > 0 is used to prevent chattering, high frequency
134 G. Herrmann et al. / Journal of Process Control 13 (2003) 129138
switching of the non-linear control ([14] and [17], p. 15). 4. Controller design for the HDA-plant
This implies that ultimately pseudo-sliding motion,
kk 4 NL , can be assured only where : R ) R is The control design takes the two disturbances, six
an increasing continuous function satisfying 0 0. actuators and the ve output measurements into
The application of an observer also reduces chattering account. Thus, a model linearization GHDA j! of 270
of the control: such control structures have been used in states, eight inputs and ve outputs for the operation
practice by Bhatti [15] and have been reported by Utkin point of the nominal values of Tables 1 and 2 (middle of
[16] as practically important. However, strict mathema- required value range) and a production rate of 265 lb
tical proofs of robust stability for this linear observer/ mol/h is considered. For model order reduction and
non-linear controller set-up have not been shown so far. controller design, input and output scalings
A question remains in the case of the availability of I sc diagI 1 ; I 2 ; . . . ; I 8 , Osc diagO1 ; O2 ; . . . ; O5
several choices for the sliding mode plane as to which (Tables 13) have been used. From an evaluation of the
choice is preferable. For high order plants, the number Hankel singular values of the balanced realization of
of choices (max. Nm N!
!m!) prevents testing of each pos- Osc GHDA jwI sc , a 1020th order model was seen to be
sibility via simulation. A criterion involving perfor- most appropriate. Balanced truncation has been pre-
mance or robustness, for assessment of the sliding mode ferred to achieve good high frequency model matching
plane prior to simulation, is preferable. It has been because of the uncertainty associated with the linear-
argued by Edwards and Spurgeon [17], using ideas of ization at low frequency. A 17th order model G~ HDA has
robust eigenstructure assignment, if the condition num- been obtained.
ber of the matrix of right hand eigenvectors INm 0V of The H1 controller from Section 3 employs the
is minimized then the sliding mode poles become derived order reduced model and the weights:
insensitive to uncertainties or perturbations of the sys-
tem matrix A. The minimization of the condition num- Wur s diag1:9!ur ; 2:3!ur ; !ur ; 0:1!ur ; !ur ; !ur s;
ber provides a robustness measure for the system in 100 000s 1
sliding motion, but does not necessarily imply good !ur
s 100 000
performance. It is therefore better to evaluate pole
model linearizations Gj are obtained to evaluate the norm by H1 -optimization is not possible without per-
degree of uncertainty. This procedure keeps the set- formance degradation. Thus, this trade-o leads to an
points in reasonable proximity of the nominal set-point indirect reduction of the eective operation area of the
and does not introduce an unnecessarily high, spurious closed loop controller. This procedure is common in
uncertainty which is introduced by testing
all available
applied H1 -control (e.g. [19]) as the primary interest is
to achieve practically robust performance in the oper-
combinations. An envelope of maxj Gj G~ HDA has
ation area of interest.
been tted with a transfer function. The resulting The closed loop performance and robustness of the
transfer function was used to construct Wuu. Safonovs linear control for the non-linear HDA-plant is improved
pole shifting methodology [12] is used for the H1 - by the introduction of a sliding motion using the linear
design, shifting the augmented plant poles by a value of observer derived from the designed H1 -control. For
p~=0.0001 for design. The choice of plant model order this reason, the linear control is modied to introduce
and the weights results in a control system order of 48 pole separation (Section 3.1). This is found to be fea-
states and an additional ve states for the sensitivity sible since A2 from (2) is small compared to A3
weight Wy and integrators which are included in the
closed loop. The numerical H1 -optimization with the kA2 k 0:156526 kA3 k 1105 ; kAk 2:35104
chosen design
weights gives a closed loop norm of
Ty ;0 ;r0 4 2:3. For the specied controller design and the matrix A3 A2 is stable. The sliding mode
H 1
weights, the resulting
controller has the property that plane is chosen so that the condition number of
Wuu K1 Wy S1 5 1 (Fig. 5b) for parts of the frequency INm 0V is small. The pole combination with condi-
range; hence, robustness to plant non-linearities has tion number value of 222.1 yields good performance.
been sacriced
for improved
controller performance. The condition number of the eigenvector matrix for
Since Wuu K1 Wy S1 is constant in large parts of the some other pole choices can reach values of the order 1017
considered frequency range, a further reduction of this indicating poor robustness to unmatched uncertainties. It
has been found that other ad-hoc pole selection schemes 4.1. Controller performance
for the sliding mode were less successful. The para-
meters for the non-linear control element from (6) have The linear H1 -controller was designed so that the
been chosen with 1 100; 2 0, which allowed prac- settling times of the linear responses were less than 5 h.
tically
fast attainment of the pseudo-sliding mode The non-linear simulations show fast rise times for the
^
1 while NL 1105 has been chosen to prevent production rate (Fig. 6), but long settling times. This
results from the fact that low frequency uncertainty is
chattering of the control and numerical inaccuracy of
particularly high (Section 2) which only allows the
the simulation.
designer to obtain fast initial dynamics. Interaction with
the remaining four output measurements for the H1 -
controller is relatively slow (Fig. 7). Systematic tests
with controllers employing various design weights have
shown that increasing the bandwidth of the sensitivity
function S1 of the linear closed loop H1 -controller by
adjusting the weight Wy does not result in a faster
response with improved interaction of the non-linear
simulated closed loop system while a decrease of the
required bandwidth would decrease rise times and sub-
sequently settling. This has been particularly detected
for the purity measurement response. Subsequently, a
sliding mode controller is designed based on the H1 -
controller. This reduces the respective interaction and
gives faster settling times for the four output measure-
ments from Fig. 7, while for the production rate the
Fig. 6. Production rate tracking response of H1 and linear observer tracking response characteristics are very similar to the
based sliding mode control. linear control. This has been also veried employing
Fig. 7. Tracking errors of linear observer-based control due to production demands (see Fig. 6).
G. Herrmann et al. / Journal of Process Control 13 (2003) 129138 137
Table 4
Mean absolute error evaluating the tracking error for several demand changes between 255 and 280 lb mol
h in the interval 0 h; 54:5 h
Acknowledgements
Fig. 8. Non-linear responses of the two H1 -based controllers to a
production
rate ramp demand change during 0.5 and 1 h from 265 to The authors would like to acknowledge the support
240 lb mol
h ( , linear; , sliding mode controller).
for G. Herrmann from the European Commission
(TMR-grant, project number: FMBICT983463).
1
output data sampled at 3min and analyzed via the mean
absolute error (Table 4) of the tracking errors (yr):
1 X Ns
yj ti rj ti ; j 1 . . . 5;
References
ej 8
Ns i1 [1] J.M. Douglas, Conceptual Design of Chemical Processes,
McGraw Hill, 1988.
The mean absolute error reduces by at least 30% up [2] J.J. McKetta, W.A. Cunningham, Encyclopedia of Chemical
Processing and Design, Marcel Dekker, New York, 1977.
to 90% for the four interaction outputs which is a sig- [3] R. Smith, Chemical Process Design, McGraw Hill, New York, 1995.
nicant improvement for general plant operation and [4] C. Brognaux, A Case Study in Operability Analysis: The HDA
product quality. The non-linear controller also operates plant, Masters Thesis, Imperial College, London, Prince Consort
in a wider envelope than the H1 -controller. The sliding Rd., London SW7 2BY, 1992.
mode controller remains stable for production rate [5] Y. Cao, D. Rossiter, D. Owens, Input selection for disturbance
rejection under manipulated variable constraints, Computers and
demands below 240 lb hmol , while the H1 -controller
Chemical Engineering 21 (SS) (1997) S403S408.
becomes oscillatory for a demand of 240 lb mol h and [6] W.L. Luyben, M.L. Luyben, B.D. Tyreus, Plantwide Process
unstable below this value (Fig. 8). Control, McGraw-Hill, New York, 1998.
The sliding mode controller and the linear H1 -con- [7] Y. Cao, D. Rossiter, D. Edwards, J. Knechtel, D. Owens, Mod-
elling issues for control structure selection in a chemical process,
troller are both robust to product purity measurement
Computers and Chemical Engineering 22 (SS) (1998) S411S418.
delays of more than 20 min. The sliding mode control [8] Y. Cao, D. Rossiter, An input pre-screening technique for control
generally shows better closed loop responses to small structure selection, Computers and Chemical Engineering 21 (6)
cooling-euent temperature disturbances than the H1 - (1997) 563569.
control, while for purge-down stream pressure dis- [9] Cao, Y., Rossiter, D., Globally optimal control structure selec-
turbances both controllers have similar performance. tion using the branch-and-bound method, in: Proceedings of the
5th IFAC Symposium on Dynamics and Control of Process Sys-
tems, Corfu, Greece, 1998.
[10] Y. Cao, D. Rossiter, D. Owens, Controllability improvement
5. Conclusion using secondary measurements, in: Proceedings of the Conf. on
Advanced Process Control 5, 1998.
An observer based sliding-mode control scheme has been [11] K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control,
Prentice Hall, NJ, 1996.
presented and applied to the highly non-linear, benzene [12] R.Y. Chiang, M.G. Safonov, Robust control toolbox, in Matlab
producing HDA-plant. The plant has a particular Toolbox Users Guide, Mathworks Inc., 1992.
highly non-linear steady state characteristic. A linear [13] P. Colaneri, J.C. Geromel, A. Locatelli, Control Theory and
138 G. Herrmann et al. / Journal of Process Control 13 (2003) 129138
DesignAn RH2 and RH1 Viewpoint, Academic Press, San [17] C. Edwards, S.K. Spurgeon. Sliding Mode Control: Theory and
Diego, 1997. Application, Taylor and Francis Ltd, London, 1998.
[14] C. Edwards, S.K. Spurgeon, Robust nonlinear control of heating [18] G. Herrmann, Discretization of Non-linear Controls with Appli-
plants, in: Proceedings of IEE, Vol. D/144(4), 1994, 227234. cation to Robust, Sliding-mode-based Control Systems, PhD
[15] A.I. Bhatti, Advanced Sliding Mode Controllers for Industrial Thesis, University of Leicester, 2000.
Applications, PhD Thesis, University of Leicester, 1997. [19] D.J. Walker, H1 -control of gasication plant, Proceedings of the
[16] V.I. Utkin, Sliding Modes in Control Optimization, Springer- Institution of Mechanical Engineers Part I, Journal of Systems
Verlag, New York, 1992. and Control Engineering 215 (2001) 235243.