Copyrigh' © [FAC Con'rol Science and Technology

(8,h Triennial World Congress) Kyo<o, Japan, 1981

ABOUT A HIERARCHICAL CONTROL FOR

COMPLEX DYNAMIC SYSTEMS

Z. Binder, A. Franco, R. Perret and D. Rey

Laboratoire d'Automatique de Grenoble, ENSIEG - INPG, B.P. 46,
38402 - Saint-Martin-d'Heres, France

Abstract. The stability and the way of using a hierarchical representation

of complex systems is presented. The treatment and the respect of mutual
constraints existing between subsystems is considered. The whole methodology
is based on tracking considerations. It could be applied to systems coupled
serially or according to loops . Finally experimental results performed on a
pilot plant are presented.
Keywords. Large scale systems; control system synthesis stability multi-
variable control systems; chemical process control.

INTRODUCTION X being the state vector.

E being the input vector.
We shall consider in this paper the case of S being the output vector.
high dimensional system which are met for ins-
b) Tracking criteria to be minimized is (2)
tance in the technological field. We think,
as examples, of petrochemical processes or of 1
networks (electrical, vapour ... ). The modeli-
J = Z[(Sr-Sd)tQs(Sr-Sd)+(Er-Ed)tR(Er-Ed)]
sation of such systems leads to linear high 1 00
dimensional representations which cannot be + -2 J [S-Sr) Qs(S-Sr)+(E-Er) R(E-Er)]dt
o t t
treated easily on line under their global form.
As a matter of fact, we have proposed a metho- The first term of this criteria is related to
dology for the hierarchical control of com- the static optimums Sr and Er, which corres-
plex systems, made of the connexion of sepa- pond to the wished objective values Sd and Ed.
rate stable units, which is the general case Sr and Er must satisfy equation (I). While ma-
for technological processes. We have also made king the assumptions that Sd and Ed are step
the assumption of linearity. functions, as soon as these last values are
known, Sr and Er can be computed by variatio-
Such a method has already been proposed and nal Lagrange method for instance. The second
developed under a heuristic form in our labo-
term of the criteria is related to the dyna-
ratory. This paper legitimates its validity mic optimum values of Sand E. The extensions
by showing sufficient conditions of stability to infinity of the integral gives the classic
of the global hierarchical system obtained,
result of a control by a constant state feed-
and by presenting rules for using it. This me- back. Moreover the amplitude of this feedback
thodology is illustrated by its application to
will be independent of initial conditions.
a complex pilot distillation process.
Practically when the integral is extended to
a limited time, it is known that the feedback
TRACKING CONTROL amplitude converges quickly towards a cons-
tant value identical to that determined for
Generally, in process control, we wish that an infinite time.
the evolution of the process input and output
variables be as near as possible of the evolu- In the following studies, rather than equa-
tion of an ideal reference process, which tions (2), we shall use equations (3), while
leads to minimize an error criteria concer- introducing the state vector. (3)
1
ning input and output vectors, considered as J Z[(Xr-Xd)tQ(Xr-Xd)+(Er-Ed)tR(Er-Ed)]
control objectives. We recall this classic
1
problem, according to the following presenta- + -2 Joo[X-Xr) Q(X-Xr)+(E-Er) R(E-Er)ldt
tion. o t t

a) System equations are with Q = C Qs C

t
x AX + BE
( I)
We shall always make the assumption of obser-
S CX vability. With these conditions, the solutions
of the tracking problem may be written as

1561
1562 Z. Binder e t aL

follows Some variables may be unconstrained so NI and

a) Static state MZ are pseudo permutation matrix of dimen-
-I sions (m.xI) and (m.e2), xI and eZ being the
Er Ed - R B IjJ dimensions of XI and E2. Such matrix has the
t
(4) property of giving a permutation of a subset
-I
Xr Xd - Q A IjJ of the components of the affected vector. NI
t
and M2 have the following properties :
IjJ is the vector made by Lagrange parameters. Im Im is a unit
This vector is solution of equations (4) and matrix (m.m)
(5) M2 M2 Nl . NI
12 I]
t t
A Xr + B Er 0 (5)
1 2 (eZ.e2) has zero components everywhere, but
b) Dynamic state ones for diagonal terms which correspond to
the selected comp~nents of EZ' The same pro-
Quadratic optimization gives the solution perty exists for II(xI.xI) while considering
E Er - K (X-Xr) (6) the selected components of X]' Besides we
call
with K R- I B H (7)
t PI = Nlt ·M 2
H being a solution of Riccati equation P lt = MZt·N I
AtH + HA + Q - H BR-IB H = 0 (8)
t Relation (10) is generally not satisfied be-
An essential property of this solution is cause models Uo u l Uz are just aggregated sim-
that the control so obtained by state feed- plified models, so an error f appears accor-
back is stable if the pair (A,D) is observa- ding to equation (11) :
ble with DtD = Q. ( I I)
If Q is a positive diagonal matrix of rank n
D = D = QIl2 is of rank n, and the pair (A,D) We shall try to compensate this error by modi-
is ob~ervable. Finally the tracking control fying the objectives SrI and ErZ such as the
is always stable if Q is a diagonal matrix of new objectives S+rl' x+r] and E+rZ be (fig.
rank n. Notice that this last condition is not 2)
necessary, but only sufficient. But its impor-
so (12)
tance come from the fact that such a diagonal
form is generally used for the state pondera- X+r] Xr] + g N]tf = Xrl+gP]E2-g I]X]
tion Q matrix. A corresponding Lyapunov func- + -
E r = Er -(]-g)M Z f=Er Z-(]-g)I 2E2+(]-g)P I X]
tion can be written : 2 2 t t
(X-Xr)t H(X-Xr) (9) with 0 ~ g ~ I
Ponderation matrix Q and R, may respectively This compensation is legitimated by the fact
be chosen as regards to quality aspects for that, if the tracking would be perfect, com-
Q and control cost for R. pensation itself would be perfect with :
+ +
f MZ E r2 - N] X r]
APPLICATIONS TO CASCADE SYSTEMS f MZ ErZ- (I-g)f - N]Xr]- gf

Let us consider first the simple case of a f M2 Er2 - NI Xr l - f

process made of two units V I ,V 2 serially cou- f - f and f = 0
pled (fig. I). The control level is made of
two stages. The lower level is made of two e- because the objectives coming from the higher
lements u I ' u2' giving respectively reference level respect the constraints. Finally the er-
inputs (Eru) and outputs (Sru) for units VI ror f is equal to zero. Input vectors of u I
and U2' These reference variables are elabo- and Uz may now be written as follows :
rated by models of VI in ul, of V2 in u2' The a) El = Erl - K](XI-x+r) (J 3)
inputs and outputs vectors of the models uI
u2 are El SI and E~ S2' Reference values Er], E I = -K](I+gI])X I + gKIP]E Z
SrI' Er2' Sr2 are ~ssued by the highest level
of the control hierarchy uQ. Inputs E] E2 are (]4)
obtained by implementing tracking algorithms
previously described. The level Uo represents
globally the process. At this level references
Er Sr for the whole process are given by the From (13) and (]4) . we can compute (15) and
operator and the associate global model ela- (16)
borates the reference for the inferior stage.
The problem for building such a hierarchy is
that of respecting interaction constraints,
\E,/ ['
(E2~ (]-g)Y
g: J X]

X2
+

due to the coupling of both units VI and V2'

at the lower control level. Such a constraint
can be expressed by an equality :
(10) [:'
gK]P] (I 2-h

1 2- h 12
I'l} Er]'"
x'xc'l
Er + K Xr 2
2 2
( 15)
 About a Hi e rarchical Control for Complex Dynamic Systems 1563

with h =
]- 1>
..:.......s>. ~ (~).
2-g
(]6) becomes; with
While we get the equations

( ]8)
(]6)

This system is stable if the characteristic

with matrix A]-B]K]G] and A2-B 2K2 correspond to
stable systems . A2-B K c orresponds to the
T -KI-gK]I] + g(]-g)K]P](I 2-h r;)P t ZZ
system u2 with its or1g1nal tracking control
supposed stable. A]-B]K]G] corresponds to the
W -K]P][I 2-h r; ]K 2 tracking control of the system (A],GI]B])
,L- with the ponderation factors G]Q]G] and R] in
[I 2-h r; ]P t the criteria. This property is true if A] is
z diagonal.
- [I 2-h r; ]K 2
For such a system the Riccati equation is
(]7) shows that if : written :
-] -] -]
a) g o v] is independent of system u2 A] HI+H]A]+G]Q]G]-H]G] B]R] B]tG] H] = 0
t
b) g V2 is independent of system u]
If H is the Lyapunov matrix related to the
c) 0 < g < ], ul and u2 are mutually coupled. tracking controlled system u2, H] = ~Gt is a
solution, because
These properties can eventually be checked -] -] -]
easily by inspection of the scheme of the glo- A] t G] HG] +G ] HG]Al-G ] Q1G] -G ] HG] GIB] R] B] t G] G HG] =
bal system. Cases a) and b) are most valuable, -]
because they give the possibility of getting G](A]tH+HA+Q-HBIR] B]tH)G] = 0
conclusions about the overall stability.
This tracking controlled system is stable be-
cause the pair (A 1G1 ,Q{12) is observable due
STABILITY to the form of G] which is a full diagonal ma-
trix.
g = 0 (~). This property will be used for synthesizing
diagonal systems, with inputs and outputs com-
pensation, which will be used as isolating e-
1ements.
Isolating Element (Fig. 5).
Uz
l
We consider now three systems u] u3 serial-
( ]7) ly coupled, in which the compensat10n for (u]
u2) is applied on the input of and the uz,
K]Xr] compensation for (u2 u3) is appl1ed on the
K Xr output of u2 (fig.s).
2 2
We can now write the input vector of u2 as :
+ +
This system is stable if the characteristic E2 = E r2- K2 (X 2-X r2) (]9)
matrix A]-B]K] and A2-B 2G2K2 correspond to
The reference Er2 is modified according to
stable systems. A]-B]K] corresponds to the
E+r2 for taking account of the constraint bet-
system u] with its original tracking control
supposed stable. A2-B 2G2K2 corresponds to the ween u 1 and u2 and the reference Xr2 is modi-
fied according to X+r2 for taking account of
tracking control of tfie system U;(A2,B2G2)
the constraint between u2 and u . We can now
with the ponderation factors Q2 and G2 R G2 3
Z
in the criteria. For such a system the R1cca-
establish the following relation deduced from
ti equation is written : (11) and (12).
-] -] -] G2E2= PltXI-K2GxX2 - K2P 2K3 X3
A2tH2+H2A2+Q 2-H2B2G2G2 R2 G2 G2 B2t H2=O
-]
+ K2[Xr2+P2Er3+K3xr3]
or A2tH2+H2A2+Q 2-H2B2R2 B2t H2 = 0
with G2 12 + 12
which gives the same Lyapunov matrix H2 , as
that obtained for the tracking control of ~he G
x Ix + Ix (Ix is a unit matrix of
dimension x2' x2)
system (A 2 ,B2,Q2,R 2 ). For the new system u2,
P2 N2t M3
the matrix AZ' Q2 are similar to the corres-
ponding matr1x of u2, so u~ is stable if u2
From the value of E2 it is possible to esta-
is stable and the global system is stable.
blish the new characteristic matrix of the
overall system:
1564 z. Binder e t al.

o In the general case, a hierarchy may include

an arbitrary number of levels. When we go up
the hierarchy, more and more agregated models
must be used. These models musl be such as
outputs of level (i+l) must furnish all refe-
rences of level (i). The method is well adap-
The overall system is stable if the characte· ted to a modular treatment, every system being
ristic matrix A1-B I KI , A3-B3K3' ~2-B2G2IK2Gx submitted to a similar treatment. So, it is
correspond to stable system. It 1S true for possible to coordinate progressively accor-
the first and the second one which correspond ding to higher and higher levels a complex
to the initial systems with a tracking control system. We can mention that a level of the
supposed stable by hypothesis. The third sys- hierarchy represents a coherent simulation of
tem corresponds to the tracking control of the complex system. Constraints exist between
t~e system u~+(A2,G~IB2G2) with the pondera- the systems of a same level. One level tracks
t10n factors GxQ2Gx and G2R2G2. For such a references values issued by the next higher
system the Riccati equation is written : level.
A H + HA2 + GxQ2Gx - Let us notice that the respect of interactions
2t is effective without the necessity of using
-HG-I -I -I -I -I
x B2G2G2 R2 G2 G2B2t Gx H o (21 ) an iteration who would cover two successive
levels of the hierarchy. This property inclu-
des the independance of hierarchy levels,
If A2 is diagonal a solution of this equation
from the stability point of view.
is

because (21) can be written now: APPLICATION TO A DISTILLATION

-I
Gx[A2tH2 + H2A2 + Q2-H 2B2R2 B2tH2lGx = 0 PILO'!' PROCESS

H2 is a solution for the system u 2 with the In our laboratory we have the possibility to
tracking control. use a distillation pilot plant, made of two
columns, coupled serially. Three products,
The system u~+ is stable because the pair
distillated water, ethanol and methanol are
(A).,G xQ I/2) is observable due to form of Gx ,
Z
which 1S a full diagonal matrix. This system
separated. The first column separates water
from the two alcools, the second separates
:.solates uI and u3 from any mutual influence.
the two alcools. This pilot has a semi-indus-
trial character. It has the possibility to
use a maximum flowrate of 170 liters per hour.
EXTENSION TO COMPLEX SYSTEMS
Heating powers of both columns are respective-
ly of 72 Kw and 56 Kw. The whole set of mea-
We can now conclude that constraints in a cas-
surements issued by sensors is centralized on
cade system (fig. 6) may be compensated by u-
a minicomputer. This minicomputer can also
sing corrected inputs (case g=O). For this ca-
elaborate and send to actuators the different
se the second system is coupled to the first
one, the third to the second and so on ... set points.
From the stability point of view this fact is The hierarchical control of this process will
a~ing if the cascade is closed according to include two stages, one supervision stage whe-
a loop. For this case the use of a diagonal re global objectives could be taken in charge,
isolating system, called N in figure 7, gives one control stage which elaborates working re-
the possibility of suppressing every mutual ferences of the three process units. As an
influences between elements situated on both example the process used is made of the two
sides of N. This isolating system can corres- distillation columns and of one m1x1ng unit
pond to one of the physical units concerned, which gives the possibility to recycle dis-
or may artificially be realised by a diagonal tillated products.
of short time constants, when compared to the
The global model uo which has been chosen in-
main time constant of the system. Let us noti-
cludes 4 state variables 5 inputs and 4 out-
ce finally that a set of inputs may be subjec-
puts. The inputs are : (Fig. 9).
ted to constraints issued from several origi-
nated systems (fig. 8). The same case may ap- flowrate of the rough input pro-
pear for subset of outputs (fig.B). For both duct (composition is in ethanol
cases the correction is applied through sub- percentage).
set of inputs to the receiving system.
heating power and reflux flowrate
Finally, we can propose the following rules of the first column.
for the compensation of system networks :
heating power and reflux flowrate
a) Each connection between two systems must of the second column.
include one correction.
are the effective active va-
b) Each system is generally input-corrected. A riables which are elaborated
diagonal system may be either input or output by uO.
or input-output corrected.
The four output variables are :
c) Each loop of systems, must include at least
Ldr2 X1r2: flowrate and purity of the ethanol.
one isolating element.
 About a Hierarchical Control for Complex Dynamic Systems 1565

flowrate and composition (in per- model used are different.

centage of ethanol) of the distil-
late product issued of the top of CONCLUSION
column (I).
The isomomorphism between the real topology
At the global or coordination level, produc- of the computer system and that of algorithms
tion objectives are Ldr2 and Xdr2 related to is a quite interesting point in this concep-
the flowrate and the purity of the ethanol. tion of control systems. Its importance is em-
Reference values which will be transmitted to phasized for large industrial systems where
the lower level of the hierarchy are : the geographical setting of physical units
needs a distributed control system. Moreover
inputs : Wrl Trl Wr2 Tr2 it gives the possibility to implement progres-
outputs: Ldrl Xdr l Ldr 2 Xdr 2 . sively layer by layer the control. The algo-
rithm used limits the information tranfers
Inputs are determined by a state feedback between decision centers. The stability study
which gives the possibility to track Ld 2 and has given way to some elementary rules for
Xd 2 · assembling decision centers, which gives the
possibility to control quite complex proces-
Ponderation matrix Q and R are chosen for gi- ses. The first tests performed on a semi-in-
ving a high weight to the errors connected on dustrial units have been successful.
one side with the wished objectives, (Ldr2 -
Ld2) and (XJr 2 - Xd 2 ) and on the other side
to fluctuations of energetic inputs (Trj-T I ) REFERENCES
and (Tr2-T2). The second level of the h~erar­
chi ca I control is made of two systems uI (co- Barrero, L., Z. Binder, J.P. Eynard, D. Rey
lumn (I», u2 (column (2». uI and u2 will be (1979). A microprocessor controller for
of dimension 4. Input vectors will be made of decentralised hierarchical control. IFAC/
W+T+, that is to say, boiling puwer, and re- IFIP Symposium SOCOCO 79, Prague.
flux flowrat e . Output vectors will be mad e of
Binder, Z., O. Badr, R. Perret (1975). Trac-
Ld, Xd, flowrate and composition of th e dis-
king approach to the control problems. Pre-
tillated product.
prints of 6 th World Congress IFAC, ISA,
par. IC-37.3, pp. 1-10.
Constraints exist between
Binder, Z. , R. Perret (1978). Synthesis of
u I and u2 : Ld l Lf2 complex objects as an integrated system.
Xd l Xf 2 In KLIR (ed) Applied General System Re-
search, Plenum Press, pp. 471-485.
The respect of the established rules given Findeisen, W. and colI. (1978). On line hie-
precedently, for obtaining a stable simulation rarchical control for stady-state systems"
level, has led to corrections on the inputs I.E.E.E. Trans. AC, vol. 23, pp. 189-209.
of u2. The application of tracking control to
uI u2 gives the inputs references Wl+ T I + W2+ Hagras, A.N., Z. Binder (1980). Three level
T2+ and outputs references Ld l + Xd l + Ld 2+ cooperative decentralised control for in-
Xd 2+ for the process itself. The process may terconnected dynamic system. Preprints of
then be controlled also by a tracking method. IFAC Symposium "Large Scale Systems, Theo-
ry and Applications", Pergamon Press, pp.
A lot of experiments have been performed on 149-157.
the pilot process. As a typical example figu-
res 10 and I I show the effect of the succes- Lefkowitz, I. (1975). System control of che-
sive application of two reference steps on the mical and related process systems. Procee-
output flowrate Ld 2 . These reference steps are ding of 6 th World Congress IFAC, pp. 38.2/
applied at the level of the coordinate model. 1-7.
The figure show the evolution of the liquid Sandell, N.R., P. Varaiya, M. Athans (1978).
flowrates and composition for both columns (I) A survey of decentralised control methods
and (2), at the level of the coordinator model, for large scale systems. I.E.E.E. Trans.
of the local model and of the process itself. AC-21, pp. 79-84.
Convergence of the system is good.
Singh, M.G., A. Titli (1977). Hierarchical
feedback control for large dynamical sys-
DISTRIBUTED COMPUTER CONTROL tems. Int. Journ. Syst. Sc., 8, pp. 31-47.
Titli, A., and colI. (1979). Analyse et com-
The method which has been developed has a modu- mande des systemes complexes. CEPADUES.
lar character. The same tracking algorithm is
applied to all systems (u). Moreover informa-
tion exchanges between systems (u) is limited
to interaction information. These properties
make easy the implementation of the algorith-
mic structure according to a microcomputer =:=:=:=:=:=
network structure. By associating to each sys-
tem (u) a micro-computer, decision center have
been created. So the control program used is
similar on every decision center. Only the
eS T 2 _ AA ·
1566 Z. Binder et al .

9
[
~1f--.D --+-{J
I.
'1)
~-----D -~f--
L]
(r)

f i r.""· J.

I ; I

o cb .0
~. "t
,I
}-I ~
, qY'
il ~ :I-
1 [ ") !

Fi g ur e 6.

I"j ,., or .. 1.

1" '" j 9," "~ 'I

--D----~
LJ 11 --~ Ti
L2_r-----
j'" ,~''' '" J t' ",. :~j
, 1 , I/1)- - - - ~ - - - ~"
- -D ,- ':'II-- ---I
L2J 1 --
l2j- J _-v-----'-'~- 1,1
jP
, "\ ..
Or ; ,:..,- - -~-
i , '-' ~- :'- ~:.r"

- _ • ..;:cbcl.l :r.,,:':t.
'I - --- ~ocai ~.;;'::;e ~

- :>r,J C~S 3

wr'~IOJ±:
'
l '[I - -
WT 2
Tr '
u
()
-
- - Ldr,
. Cd ,.- .,
.. '. "7
.

2 - ----. Xcl r ,

Ld I xrd I

,- , .
o , , '0 T'",,,,
Fi. g u!"~ i:

Fi fo il! ~ ~ .
For Discussion see page 1576