Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
PLEASE SCROLL DOWN FOR ARTICLE
This article was downloaded by: [MIT Massachusetts Institute of Technology]On: 1 July 2010Access details: Access Details: [subscription number 917268020]Publisher Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of Computational Fluid DynamicsPublication details, including instructions for authors and subscription information:http://www.informaworld.com/smpp/title~content=t713455064
Feedback Control of a Thermal Fluid Using State EstimationJ. A. Burnsa; B. B. Kingb; D. RUBlOaa Center for Optimal Design and Control. Interdisciplinary Center for Applied Mathematics, VirginiaPolytechnic Institute and Stale University, Bliicksburg, VA, USA b Department of Mathematics, OregonState University, Corvallis, OR, USA
To cite this Article Burns, J. A. , King, B. B. and RUBlO, D.(1998) 'Feedback Control of a Thermal Fluid Using StateEstimation', International Journal of Computational Fluid Dynamics, 11: 1, 93 — 112To link to this Article: DOI: 10.1080/10618569808940867URL: http://dx.doi.org/10.1080/10618569808940867
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial orsystematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply ordistribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contentswill be complete or accurate or up to date. The accuracy of any instructions, formulae and drug dosesshould be independently verified with primary sources. The publisher shall not be liable for any loss,actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directlyor indirectly in connection with or arising out of the use of this material.
http://www.informaworld.com/smpp/title~content=t713455064http://dx.doi.org/10.1080/10618569808940867http://www.informaworld.com/terms-and-conditions-of-access.pdf
nero. 1\)98, VoL I L pp. 93- 112Reprints available directly from the publisherPhotocopying permitted by license only
It') 1998 OPA (Overseas Publishers Association) N.V.Published by license under
the Gordon and Breach SciencePublishers imprint.
Printed in Muluysiu.
Feedback Control of a Thermal Fluid UsingState Estimation
J. A. BURNS a .• , B. B. KINGh .! and D. RUBIO"'··
a Center for Optimal Design ami Control. Interdisciplinary Center for Applied Mathematics. Virginia Polytechnic Instituteand Slale University. Blacksburg. VA 2406/-053/. USA:
b Department of Mathematics, Oregon State University, Corvallis. OR 97331-4605. USA
In this paper we consider the problem of designing a feedback controller for a thermalfluid. Any practical feedback controller for a fluid flow system must incorporate sometype of state estimator. Moreover, regardless of the approach, one must introduceapproximations at some point in the analysis. The method presented here uses distri-buted parameter control theory to guide the design and approximation of practical stateestimators. We usc finite clement techniques to approximate optimal infinite dimensionalcontrollers based on linear quadratic Gaussian (LQG) and MinMax theory for theBoussinesq equations. These designs arc then compared to full state feedback. We pre-sent several numerical experiments and we describe how these techniques can also beapplied to sensor placement problems.
Keywords: Fluid flow control. state estimation. finite clement approximation
I. INTRODUCTION
In recent years considerable attention has been
focused on thc problem of active feedback control
for fluid tlow problems. In addition, there have
bcen tremendous advances in computational tools
for simulation and design of such systems, How-
ever, it is seldom when these tools are used by
control engineers to help desing feedback con-trollers.
If one employs only finite dimensional control
theory, then the size of the numerical (finite
element) model prohibits the practical application
of existing control algorithms. Thus, one approach
*Corresponding author. This research was supported in part by the Air Force GOlec of Scientific Research under grant F49620-96-1-0329 and the National Science Foundation under grant DMS-95G8??3.
t This research was supported in part by the Air Force Office of Scientific Research under grant F49620-96-1-0329 while the authorwas a visiting scientist at the Center for Optimal Design and Control. Virginia Polytechnic lnstitute and Stare University, Blacksburg.VA 24061-0531. by the Alexander von HumboldtStiftungwhile the author held a fellowship at Universitiit Trier. Germany.and bythe National Science Foundation under grant DMS-9622842.
**This research was supported in part by the Air Force Office or-Scientific Research under grant F49620-96-I·0329.
93
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
94 J. A. BURNS £'1 al.
to controller design starts with the developments
of a "reduced-order" model (i.e., a low orderordinary differential equation) that mayor maynot have much to do with the physics of fluids.
Although this approach leads to a finite dimen-sional (not necessarily low order) controller, it
often requires considerable robustness in order to
oil-set the unrnodclcd dynamics.At the other extreme, the literature is full of
papers that usc "large" finite clement models and
work to build special control algorithms for such
models. This approach may capture more physics
and when numerical methods for large scale control
problems arc available, it is a useful design method.
Both methods can be described as approximate-
then-design approaches to feedback control.In this paper we describe the beginnings of an
approach that uses distributed parameter control
theory to help guide the design and computation
of practical feedback controllers. In particular, weconcentrate on the design of linite dimensionalstate estimators (observers) by combining finite
clement schemes with distributed parameter
theory. The resulting controller docs not require
full state feedback and hence is practical for flowcontrol problems. The basic idea is based on the
observation that the real goal is to construct finite
dimensional controllers (hopefully of low order)
that will be applied to the physical flow. Rather
than do model reduction followed by control
design, we suggest that it sometimes better to do
control design and then controller reduction. Thisapproach makes maximum usc of powerful CFD
tools now available (with minor modifications)and, at the same time, offers new insight into issues
such as optimal placement of sensors.In this paper we illustrate these basic ideas by
applying the method to a convection loop control
problem. We usc linite element approximations
to build linite dimensional state estimators that
arc the kcy to feedback control when direct state
measurements arc not available. We do not ad-
dress the secondary issue of controller reduction.
However, we close with a brief discussion of this
matter and provide references where this approach
has been implemented on vibration control and
thermal control problems.
1.1. The PDE Model
A thermal convection loop consists of a viscous
fluid contained in a circular pipe standing in a
vertical plane (see Fig. I). When the loop is heated
from below and cooled from above it creates a
temperature gradient opposite gravity and the
fluid tends to flow due to the buoyancy force. On
the other hand, viscosity and thermal diffusivity
resist this motion. The fluid motion is createdwhen the buoyancy force overcomes these dis-
sipative terms. Experiments show that when the
difference in temperature between the top and
the bottom of the loop is large enough, the fluid
exhibits unstable motion which may also be
chaotic (sec [7, 19,21 D.We consider the problem of boundary control
and state estimation for this two-dimensional
thermal fluid flow problem. In particular, weapply a temperature control at the outer boundary
of the loop and measure the heat flux at the inner
wall. The interior radius of the pipe is denoted
by R 1 and the exterior radius by R2. The radialposition of a fluid particle is denoted by r E [RioR2 ] and the angular position, denoted by
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 95
the pipe is small as compared with the interiorradius, i.e., Rz -R 1 « R I . In this case, the fluidmay be considered as flowing in a straight pipe ofwidth Rz - Rio so that the velocity depends onlyon the radial coordinate. Moreover, we assumethat the fluid flow has circular streamlines so thateach fluid particle flows at a fixed distance fromthe center of the pipe. Therefore, if v denotes thevelocity vector, then
In the convection loop problem under consid-eration, the body force, Fg , is due to gravitational
acceleration, g, and the buoyancy force, F", Thebuoyancy force is the vertical force produced bychanges in density and for a unit mass, is givcn
by F" = (Po - p)(-g). The relation (3) yields Fh =Po(3(T- To)(-g). Hence, the external force perunit mass is given by
v(t, r, 'P) = v([,r)el' ( I )(4)
where e., is a unit vector in the direction ofincreasing 'P.
The equations of motion for a Newtonianviscous flow are given by
OV I Ic--I-(v·\7)v =-FK--\7p-l-v6vat p p
-I- ('1.-1- 11 ) \7(divv)OP -I-div(pv) =0,
p a[(2)
Since variations in density are considered only inthe buoyancy force (all other changes in densityare neglected), the continuity Eq. (2) is reduced todiv v = O. This, we have the incompressible Na-vier - Stokes eq uations
i)v I I'" -I- (v· \7)v = - FK - - \7p -I- 116v, (5)ut p f'
div v ee fl (6)
where p is the density of the fluid, F, denotes thebody force per unit mass, v = liff' is the kinematic
viscosity, I' and II are viscosity coefficients andp is the pressure. Although the Navier-Stokesequations are valid for the general problem, thecomplexity of these equations can be reduced forcertain flows by making additional assumptionsand introducing approximations. We shall con-sider the Boussinesq approximation which as-sumes that the system parameters are constantexcept for density in the buoyancy term. This is areasonable assumption since changes in tempera-ture produce changes in density and these changescause fluid motion. In particular, it is assumed thatthe relation between the density p and the tem-perature Tis linear and has the form
A complete description of the fluid behavior isobtained when heat transfer is included. The heatequation
describes this transfer. In particular, heat is trans-ferred by convection and diffusion, represented by
the terms v· \71' and X6T, respectively. Theconstant X introduced in (7) is the coefficient ofthermal diffusivity. Finally, combining Eqs.(4)- (7),we have the Boussinesq equations
p = Po(1 - (3(1' - To», (3)
aT---I-v·\7T=x6Ti)[
iJv -iJ[ -I- (v· \7)v = g (I -I- (3(1'0 - 1'»
I--\7p-I-1/6,',
P
(7)
(8)
where To is a reference temperature, Po IS thereference fluid density and (3 is the thermal ex-pansion coefficient. Usually, the reference tem-
perature To is taken to be the average temperatureof the fluid.
div v = 0,
01'---I-v·\7T= v6T.(it A
(9)
(10)
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
96 J. A. BURNS et ot.
1.2. Abstract Formulation
We shall assume a measured output of the form
where 11(1, rp) is the Dirichlet boundary control atthe outer wall. The initial data is defined by
Dy(I,
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 97
In order to define the state operator A, we firstdefine two operators Au and A I. Lei
the Dirichlet problem
6.y = 0 in ll,1'=0 on I', ,
y = II on I'j.
(20)
and for
[~] ED(Au),
define Ao by
A [I'] = [v 6., 0"1 [v] = [v 6., v]o T 0 X6." T X 6.T .
Let A I : H -+ H be the linear operator given by
where L : L 2(1l) -+ LZ(lld be the integral operatordefined by
g{3 rh[[II](r) = 2rr./o cos 0 such 111([/
IIS(I)II :::; /1'11'-1'.
Let A' denote the Hilbert adjoint of A =
Ao + A I. Since A generates an analytic semigroupon H, A' is a densely defined closed operator.
Now we turn to the task of defining the input
operator, B. To do this, we first need to consider
where 2 = [~] E W. The dual space of W is denotedby W'- Also let VI = HA(lld and V2 = HA(Il) de-
note the energy spaces with inner products
(fJv fJV)(v, v)VI = (v, v)u(n,) + -;:;-:, ",vi en 1.1(1/,)
and
- - (fJT fJT)(T, T)" = (T, T) L'(!l) + !'>:"'!'>:"vi vi /)(11)
/ I fJT I fJT\
+ \ -; fJ
J. A. BURNS et al,
respectively. Finally, set V = VI X V2. If H isidentified with H', then there exist continuousdense inject ions
W VH = H ' V ' W'.
This Gclfund quintuple plays a fundamental rolein thc development or the finite c1emcnt approxi-mations discussed in Section 3.
Now, we lift the operator ~ with domain D(~) =H 2(n) n H,I,(n) to /}(n). Let'& : /}(n) -> [D(~')]'be defined by
'&Z = W, E [D(~')]'ir and only if
Observe that ( is well defined since (;.) E V =HI>(n l ) x HI>(n) implies that 1'(.) is continuous andfJ/fJepn,,) belongs to L2(n). Thus, v(·)/rfJ/fJepT(,·)belongs to L2(0.).
The output operator is defined by D(C) =D(Ao) and
C,=C[;]=C2%,T(r,ep)1 . (24)I r~R,
Observe that C: D( C) S;; IJ -> H I / 2 (f l ) S;; L2(f .)is linear and bounded from H\n) into L2(f I) == Y.
The controlled Boussinesq equations defined by
the system (11)-(17) can now be formulated as astate space system in Wi given by
where A, B, ((z), and C are defined in (19), (21),(23) and (24), respectively.
(Z, ~'IF) = Ij',(ll') = (w" 1I')!"D(u')]'x"D(u'j'
1'01' all II' E D(~'). In addition, we lift A toA: H-> w'. As above, A is defined by
where
(z,A'II') = 'I/J,(II') = (~)" 1\') Wi X W,lor all I\' E W. The input operator 8: /}(L'2) -> W'is defined by
i(l) = AZ(I) +((Z(I)) + BU(I), I> 0,
with initial data
Z(O) = Zo E He W',
and measured output
Y(I) = C(I),
(25)
(26)
(27)
((z) = [ -(I'(-)/r) (2T/iJ
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 99
where x"E H,uE V = L\r2) and the operators Aand 8 are defined as in (19) and (21), respectively.
If LQR control is applied to this model, then theresulting (optimal) controller has the form
In particular, given Q = Q* 2: 0 and R = R* > 0we seek a controluop,(t) E LZ(O, 00; V) to minimizeJ(u) subject to (28). As noted above in (29), if
UOpl(t) exists, then UOpl(t) has the form
(29)
where K: H -. V is a bounded linear operator.This operator is called the feedback gain operator
and plays a fundamental role in the development
of practical low order controllers. However, since
XOpl(t) is a distributed state, it is not possible to
measure all of X"pl(t) so that complete state infor-mation is not available in practical flow control
designs. Thus, one must build a state estimator to
approximate XOpl(t). It is extremely important to
note that only part of the state XOpl(r) may be
needed in (29). For example, any part of XOpl(r)
that belongs to the null space of K does notcontribute to Uopl(I). With this in mind, we look
for an estimate of '~OPI(l), say Y,(I), such that
The estimate, xc(l), and the corresponding control
law
u(t) = -K.~c(t)
are constructed using LQG and MinMax designs.
Finally, a nonlinear state estima tor, ze(t), and corre-sponding nonlinear feed back controller
are constructed by extending the linear design.
2.1. The LQR Design
Both LQG and MinMax controllers are based on
LQR type designs. Thus, we start our discussion
with the LQR problem defined by the quadratic
cost
.I(u) = ("" [(QAt), x(t)) II + (Ru(t), u(t))uJdt../0
where K: H -. V is the feed back gain operator.For the convection loop problem we set,
where the operators h,(ll,j, IL,(lll denote the iden-tity operators in L 2(n l) and L2(n), respectivelyand q., qT are positive constants. Similarly, the
control weighting operator is given by R = r,,!u,where I u denotes the identity operator onV = e(fz) and q; is a positive constant. Observethat Q = Q* > 0 and R = R' > 0 are boundedlinear operators.
If the LQR problem has a solution, then the
optimal feedback gain operator K has the form
where II solves the Algebraic Riccati Equation(ARE)
(Il,. AJ II + (A" I1w ) II - (R- 18*Ilz , 8' Ilw ) u+ (Q"W)II =0.
(30)
for each z, wE D(A).The following theorem establishes the existence
of UOpl, and may be found along with its proofin [16].
THEOREM 2 Let A and 8 b« the operators defined
in (19) and (21), respectively. There exist a sel]»adjoint, non-negative definite operator Il E £(H)that satisfies the algebraic Riccati equal ion (30).Moreover,
I. (A*)(I-Eln E [(H), VE: > 0,2. R- 1 8'11 E [(H, V),
3. J(x", u"",) = (nxo, xo)x .
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
100 J. A. BURNS et al.
Further, lite LQ R problem has a solution of theform
where .1'''1'' is lite correspondingsolution 10 lite linearsvstetn (2R) with II = 11,,1'"
Note that the nonlinear term I(z) is not used inthe LQR design. It may be viewed as a model
uncertainty. In order to sec how well LQR con-
troller performs, we apply this linear controller
(31) to the full nonlinear system (25) and obtain
the closed-loop nonlinear system
£(1) = (A - I1K)z(l) +/(Z(I)), I> 0, z(O) = Zoo
Numerical simulations of this system will be
discussed in Section 4.
2.2. LQG and MinMax Designs
The LQG and MinMax designs not only provide a
methodology for constructing state estimators, but
also they allow for the existence of some unrno-
delled dynamics and measurement errors. Thesefeatures arc included in the system by system and
measurement error terms,11 and ~, respectively.
As noted above, even the nonlinear term may be
viewed as a model error.As before, linearizing (25) about z; = 0 yields
a linear distributed parameter control system
defined on W' by
.i·(I) = AX(I) + BII(I) + C'I(I), .1'(0) = .1'0 (32)
with sensed output
.1'(/) = CX(I) + E~(I). (33)
As mentioned above, rather than using full state
feedback we design a state estimator, Xe(I), saris-lying a linear system
.~,.(I) = AeXe(l) + J·~.y(I), xe(O) = Xeo
where An Fe and K are operators to be deter-mined. If (34) is inserted into the linear system
(32), then one has the linear 'closed-loop system
defined by
d[ .1'(1) ] [A -11K] [X(I) ]dl Xe(l) = FcC Ae. xe(t)
[c 0] [11(1)]
+ 0 FeE ~(I)'
The infinite dimensional controller defined
above is completely determined by the three opera-
tors An Fe and K. To construct these operators,we will use both LQG and MinMax designs.
It is well known (see [12,14, 15,20]) that thisapproach is equivalent to finding the optimal
solution to a quadratic differential game. How-
ever, we do not wish to devote time here to the
discussion of this aspect of the problem. Our
interest lies only in the fact that the MinMax
controller stabilizes the system and attenuates
the disturbance to controlled output map. Thus,
we shall present only the items essential to the
construction of the MinMax controller and refer
the interested reader to those references for a dis-cussion of the theory. The theory for finite dimen-
sional systems may be found in ([I, 18]). Here, we
take E = 11.'(1',) and
so that N = EE' = JU(!',) and M = CC' > 0 andbounded. The constants, g ; gT are weights. For
each e;::: 0 and .v, II' E V(A), consider the Rieeatiequations
(Il.v, Aw) + (Ax, rill') - (R- 1B'l1x, B'l1w)
-e2(C ' rlx , C ' rlw) + (Qx , w) = 0 (35)
and
and usc the linear feedback law
11(1) = -KXe(I), (34 )
(Px,A',I') + (A'x, PlI') - (N-1CPx,CPw)
- e2(QPx , PlI') + (Mx, w) = O. (36)
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 101
Note that Q, M, Rand N are bounded. Therefore.the theory in [14, 15] can be applied to show that
for e sufficiently small the Riccati equations (35)and (36) have minimal solutions no 2: 0 and Po 2: 0,respectively. If the operator [1- e2 Po IT iI ] is posi-tive definite, then one defines
Ke = R- 1 B'ITo, (37)
A(} = A - BKo - FoC + e2M ITo. (39)The corresponding controller has the form
UO(I) = -W'B'IToxc(l) = -KoXc(I). (40)
For f} > 0 this controller is 'called the MinMaxcontroller. The special case with e = 0 is theclassical LQG controller, most commonly knownas the Kalman filter. The larger the value of e, themore robust the MinMax controller should be to
certain unstructured perturbations. However, if f}
is too large, Eqs. (35), (36) are not solvable.
2.3. The Infinite Dimensional
Nonlinear Controller
The nonlinear controller defined by (41) and (42) is
infinite dimensional and, before one can make use
of this structure, approximations must be intro-duced. However, considerable practical informa-
tion can be extracted from this infinite dimensional
controller that greatly enhances the developmentof "good" low order approximations (see [4]).Finally, we note that for f} = 0, the nonlinearcontroller in (42) is called an extended Kalman
filter.
It is important to note that a finite dimensional
nonlinear controller can be constructed by ap-
proximating the nonlinear state estimation equa-
tion (41) and the feedback gain operator in (42).In principle, it is not even required that the state
equation (18) be approximated. However, any simu-lation of the closed loop system will require that
the coupled system (43) be approximated.
In the next section, we concentrate on a directapproach to approximating the control law de-
fined by (41) - (42). In particular, we develop a
finite element scheme to produce those approxi-
mating operators needed in the construction ofthe discretized versions of the Riccati equations
(35)-(36). The resulting finite element versionshave the form
If we now substitute the Ko, Fo and Ao as definedabove in (37)-(39) into the nonlinear system, thenthe resulting nonlinear state estimator becomes
The resulting closed loop linear system is given by
The corresponding nonlinear feedback law is given
by
d[Z(I)] [A -BKO ] [Z(/)]cll Z,(I) = FoC Ao Zc(l)
[F(Z(I» ] [G 0'1['7(1)]
+ F(z,(I» + 0 FoE, ((I) .(43)
Z,(I) = AoZc(l) + /(Z,(I» + FoY(I),z,(O) = z'n'
U"/(I) = -Kozc(t).
(41 )
(42)
(fI"x, A"II') + (A"x, re'lI')
-- ([R-11"!B']"fI"x, [B'j"n"",)
-- e2([G 'j"n"x, [G 'j"n"lI')+ (Q"x, 11') = 0 (44)
and
ir'», !A']"",) + ([A']"x, P"",)- ([N-'j"C" p"x, c' P"",)_ e2(Q "p "x , 1'''",)+ (M"x, 11') = 0 (45)
respectively. Therefore, a quick glance at (35)-(40) reveals that using this approach one needs
finite element approximations of the operatorsA.B,C,G,R-'.Q,N-'.M,/and E. In tho next
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
102 J. A. BURNS et al.
section, we construct these approximations:A", B". c". G ". [R-'] ". Q': [N- '] ". Mil, f",ande". These operators arc then used to constructapproximations n3 and P~I and using the defini-tions (37)-(40) one obtains a finite dimensionalcontroller.
tions wE V(A o) = [HZ(O,) n Hb(O,)] x [HZ(O)nHb(O)).
Thus, for wE W, we obtain
(Z(/),w) = ((-Ao)'/Zz(i), (-Ao)'/Zw)
+ (A,z(/),w) + (/(Z(f)),W)+ (Gry(f),W) + (U(f), B'w). (46)
is well defined. Thus, if UEH 1/2(rZ) then BUE V'and
The last term in this equation is the only difficult
term. In order to relax the smoothness on w onemust define (11(1), B' w) for wE Hb(ll,) x HMll).Recall (22) that
T· Iholds for w = (WI. wz) with WI E Ho(lll), Wz EHZ(ll) n IIb(ll). Moreover, if lI(f) E H'IZ(f z) andWz E H'(ll), then the trace theorem implies that.aWz!all E H-I/Z(fz) and
1 awzU(f, cp) ~(Rz, cp)dcp1', UII(
awz )=0 U(f),-a'll I/'/'(r,} x 1/-1/'(1',)
(47)
(Bu,w) =(u,B'w)
g(3Rz l= - -4- u(cp) cos iptt . 1'2
lR'
w, (r)dr - dcp - xRzR,
l aw,u(cp) -iJ- (Rz, cp)dcp.1', II
,g/3Rzl l R'(lI(f), B w) = --4- lI(f,cp)COSCP WI (r)drdcp1f r 2 RI
1 awz .- xRz u(/,cp)~(Rz,cp)dcp,1', on
for wE 110 (ll , ) x IIMll). Consequently (47) is welldefined for any u E H'/2(1'2) and wE 11M!!,)xHb(ll). Thus, in this case, piecewise linear func-tions in two dimensions that vanish on theboundary may be considered in the finite elementscheme. Also, we shall use continuous piecewise
3. APPROXIMATIONS OF THECONTROL LAW
Therefore, for w E W = V(A o), it follows that
Note that if a finite clement scheme is based on theabove weak problem, then one needs tcst Iunc-
(Z(/),w) = (Az(/),w) + (/(z(/)),w)+ (GII(/),W) + (Bu(/),w)
=(Aoz(/),w) + (A,z(/),w) + (/(z(/)),w)+ (GII(/)W) + (Bu(/), w).
Z(f) = AZ(f) +f(z(/)) + Grl(f) + Bu(/), / > O.
During the past ten years convergent numericalalgorithms have been developed to compute feed-
back operators (sec [2, 3,5,8,10,12,16]). Many ofthese schemes arc based on finite clement methodswhich usc splines to compute numerical approxi-mations of the feedback gains. Such schemes canbe used 10 approximate an infinite dimensional
controller. Also, as illustrated in [4]. these schemescan be used to construct practical finite dimen-sional controllers of low order.
As noted above, many of these methods are
based on constructing those finite dimensionalapproximating operators required to produceapproximations of the control law (41)-(42).The lirst step in this process usually begins byapproximating the operators defining the evolu-tion equation (18) (i.e., f;A,G and B) and theoutput (33) (i.e., e and E). For the problemconsidered here, the most difficult task is theapproximation of the unbounded operators Band C.
The abstract form of the PDE model in W' is
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 103
linear elements to approximate the control spaceLz(rz) so that d' E H'(rz) c HI/z(rz).
velocity V"(I, r) can be expressed as a linearcombination of {Z(r)} ,
3.1. Finite Element Approximationor the System Operators
N,
v"(t,r) = L:>Z(I)
104 J. A. BURNS et al.
approximated by a linear combination,
A" ]I,
A'f. and
TI . A" A" Ah I h c: c: S,,· dte operators \', 1'1 I ~,' 1" \,' Tl \' ans'}. are represented by the matrices
respectively.
I'; = [('li1, 'Ii)) O(!!)] ..1.;=1 ... ,NT
where I~'r denotes the finite element approximationof the identity operator on X.
In order to compute we need the matrix repre-
sentations or these operators. To be clear, we usethe boldface notation F to denote the matrixrepresentation or a finite dimensional operator F.Although it is straightforward to construct the
system matrices from the above discussion, weinclude the exact form below so that an interestedreader may easily reproduce our numerical results.
Let Ih= [~ I~] denote the mass matrix where
(50)e" = [0 e';J, E"
1, " . I"\ = ru 1.'(1',)
Q" = [1J,IjO')(I!I) 0]qT/~)(I!)
[ g ~o/ ~: 0]/II" = g ~./~/'N = I~.'(r,)
with' l/~'(/) E IR, I::; k::; NT and UZ(I, cp) satisfies" d' . h( 0) h( ') )the periodic con iuon 11 I, = u I, _rr .
Substituting these approximations to the state
(4X) and control (49) into the variational form inEq. (46) and letting the test functions w range overthe basis vectors, we obtain the approximating
operators (and the matrix representations listedbclow) A/',/',e h , E h where
N,.
I/h(/,cp) = I>Z(/)IZ(cp), (49)k=[
1'01' e and E, respectively.Finally. this process produces the remaining
operators required to form the approximatingRiecati equations (44)-(45). For the case consid-ered here, one has
Likewise, let {7~, I :s k :s Nr} be a basis for thespace Y("I~ H~cr(l'I) = {IV E HI (Tj}: 11'(0) = 1I'(2rr)}of piecewise linear functions satisfying the periodicboundary condition 11'(0) = 1I·(2rr). This yieldsapproximating operators
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 105
defined by (51)-(52) is finite dimensional, and ifthis control is applied to the nonlinear evolution
Eq. (18) with output (33), then the infinitedimensional closed loop system is given by
Equation (53) represents the closed loop systcmfor the full PDE, but with a finitc dimensional
controller. As h --; 0, this system approaches (43).In particular, one can show that for sufficiently
small mesh the system (53) approximates thesystem (43) for given initial conditions Zo and z
106 J. A. BURNS et at.
several numerical runs for parametric values that
allowcd larger control values. As expected, per-
forrnuncc can be enhanced by stronger controls.
These results arc not included because they require
physically infeasible control inputs and wouldserve only to show that one can obtain better
"performance" ir one allows more control input.
Wc consider a pipe with the same dimensions as
the one used in thc experimental work by Wang,
Singer and Bau [21]. The fluid is assumed to be
water at room 'j' = 600F. The initial run used thesystem parameters for this case which are given inTablc I bclow.
The weighting coefficients arc given in Table II.
Figure 3 contains the plots or initial conditions.
The init ial data for thc runs presented here was
selected to represent an initial quadratic channel
flow. For initial velocity, we chose "o(r) =144.33(r - RI)(r - R2). The initial temperature is
62°F on the top hall' or the loop and 58°F on the
bottom hair. Again, we selected this particular
data because it is typical or others runs and
because a discontinuous initial temperature profile
excites several modes, This is clear in the second
set or numerical experiments where "thermal
waves" can be observed. For all the runs. the
initial data for the stale estimator is given by
1"'" = 1.1,." and r., =0.85To.Our discussion focuses on the dilTerence between
LQR with tull state reed back control, LQG with
state estimation, and MinMax design with stateestimation, Wc use a grid with 5 nodes in the radial
direction, and 20 nodes in the angular direction.
This levcl or approximation was sufficient to cap-
ture the opcn-Ioop and closed-loop responses.
Initial Condition
1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29radius (tt)
J:~,. 3 4
radius (II) 1.2 0 ' 2angle (rad)
FIGURE 3 Initial condition.
The two sets or simulations we present are based
upon two choices or viscosity. For the first set,viscosity is chosen to be 1/ = 1.22 x 10- 5 rt2/s . Inthe second set, a smaller value or viscosity will be
specified.
4.1. Case I: v = 1.22 x lO- s ft 2/sec
This run corresponds to the problem in [21]. WhenlJ = 1.22 x 10- 5 rt2/sec , the largest feasible Min-Max design parameter is () = 7.5 X 10- 4 As shownin Figure 4, the loop velocity decays to zero in
approximately 1200 seconds. Therefore, we use thistime interval to compare the open loop and closedloop responses. Figure 4 also shows the closed
loop response when full state feedback (LQR)
control is applied. Observe that LQR reed back
docs enhance performance.
TAIILE t System parametersOpen Loop Sy~lom Velocity LQR Design Sy$lem Velocity
TABLE II Weighting parameters
S.O x 10-';01' 1.514 x 10 "ft'ls
III
1.1975in
II,
1.2959in
'I,
Ii X
g, c,
0.3
~ 0.22:' 0.1
~ 0> -0.1
radius (tI)
KinematIC ViscoSlly" 1.22e-05
1.5 x 10-' ISOli 50 30 10 5000FIGURE 4 Velocity, uncontrolled and LQR design system.
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
FEEDBACK CONTROL-OF A THERMAL FLUID USING STATE ESTIMATION t07
Figure 5 shows the closed loop responses forthe LQG and MinMax controllers along with theestimates for these responses based on the stateestimator (51). Observe that very little perfor-mance is lost in both cases. This is due in part tothe excellent estimation of the velocity field.
Figure 6 contains plots of the uncontrolled(open loop) temperature field at the four radialvalues (corresponding to nodes), r = 1.21719,1.23688, 1.25656, '.27625. The bottom right plot
rs the temperature field closest to the outer wallwhere control is applied. Observe that 1200 se-conds, the open loop temperature profile has notyet settled to zero.
rn Figure 7 we show the closed loop responsesfor the LQR control. Here we see that the full statefeedback controller drives the temperature fieldto zero in about 1000 seconds. Thus, LQR controldrives the full system to zero by 1200 seconds.
Figures 8 and 9 contain plots of the closed loopLQG responses and corresponding temperature
LOG Design Sysle:n Velocity MlnM
5'----::-:--angle (rad)
5 '..----.
angle (rad)
~
;:: 2N~ 0
~-2.... 0
Min Max Design VelOCity Estimate
radius ("1
laG Design VelocityEstimate
0.3
~ 0,2z- 0.1B 0.> -0.1
500time(s)
gfnemauc ViscoSIty'" 1,226-05
Temperatures. system from LQR design.FIGURE 7
5"L-~::::--angle (rad)
0.3
~ 0.2~ 0.1g 00;::> -0.1
raous (N)
FIGURE 5 Velocity, system and estimates from LQG andMinMax designs.
xmemeuc ViSCOSity'" 1.22e-05
Open loop System Temperatures lOG Design System Temperatures
5angle (rad)
5angle (rad)
ro::: 2~
~ 0
E-2.... 0
5 '---~.angle (rad)
~
~ 2!i 0~-2
... 0
5angle (rad)
~
;:: 2N~ 0
~-2... 0
~
~ 2!i 0~-2... 0
5 "----saoangle (fad) time (s)
Kinematic Viscosity :: 1.22e-05
5 ,----angle (rad)
500time (s)
Kinematic Viscosity = 1.22e-05
FIGURE 6 Temperatures. uncontrolled system. FIGURE 8 Temperatures. system from LQG design.
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
lOX
LOG Dosign Temperatures Estimate
J. A. BURNS el ul,
MlnMax Design Temperatures Estimate
~
;::: 2 ..
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION 109
LOG Design System Velocity MinMax Design System VelocIty LOR Design System Temperatures
m
E 2N
,,---angle (tad)
LOG Design Velocity Estimate MinMax Design Velocity Estimate
s '" ---~,OOangle (rad) time (5)
Kinemallc ViSCOSity '" 2.44e-06
1000500time (5)radius {ttl
~e- 0.1
~> -0.1
Kinematic Viscosity '" 2.44e-06
radius (til
1.2
0.'g 0,2{; 0.1
~ 0> -0.1
FIGURE 13 Velocity. system and estimates from LQG andMinMax designs.
FIGURE 15 Temperatures, system from LQR design.
Open Loop System Temperatures LOG Design Syslem Temperatures
sangle (rad)
~
:g 2:;ll! 0~-2>- 0
,,---angle (rad)
m;: 2N
~ 0
e-2.>- 0
,~.:...-~~--;
angle (rad)
ec:g 2:;l~ 0
~-2>- 0
5angle (rad)
m
~ 2~ a~-2.>- 0
s ~ -~~~----;angle (rad)
:00time ($)
sroemcuc Viscosity", 2.44e-06
,.JL_---angle (rad)
~
~ 2
~ 0
~-2>- 0-500 1000 5time (5) angle (rad)
xtnemanc Viscosity '" 2,440-06
5angle (rad)
FIGURE 14 Temperatures, uncontrolled system. FIGURE 16 Temperatures. system from LQG design.
important differences in system behavior occurbecause or the lower viscosity.
Figures 12 and 13 illustrate the performance orthe feedback controllers and the velocity stateestimators. There is little new in these plots exceptone can clearly see the import or the thermalmodes on the velocity field. Figure 14 shows theopen loop temperature field. Again observe the"thermal waves". Figure 15 shows that full state
LQR reedback dampens out the thermal field inapproximately 1000 seconds.
When LQG and MinMax controllers arc ap-plied to this system one sees a definite loss inperformance. Again, the state estimators "under-estimate" the temperature field. However, unlikein Case I, the MinMax estimator does a slightlybetter job or estimating the temperature fieldand this results in a small improvement in
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
110 J. A. BURNS ct al.
LaG oesign Temperatures Estimate MinMax Design Temperatures Estimate
~
;:: 2
~:!I 0~-2... 0
5angle(fad)
~
;:: 2N! 0~-2.... 0
5 \,~~.:..------;;;:angle (rad)
S ,V'~~-s-O-O--;;1000 5'- '-~'----;angle (rad) time (5) angle (rad)
Kinematic Viscosity .. 2.44e-06
5 '-----angle (rad)
Kinematic VIscosity .. 2.440-06
~ 2"!!! 0~-2... 0
FIGURE 17 State estimate for temperatures. LQG design. FIGURE 19 State estimate for temperatures, MinMax de-sign.
MinMax O!lslgn SystemTemperatures
FIGURE IS Temperatures, system lrom MinMax design.
• In principle, one only has to construct a finitedimensional approximation of the nonlinear con-troller (41) - (42).
• As long as one has the computational tools tosolve the discrete Rieeati equations, then it ispossible to construct a finite dimensional con-troller. However, this controller may be extre-mely large and additional controller reductionmay be required.
• For the thermal convection loop problem con-sidered here, state estimation as a means tofeedback control design is practical and leads toonly a minor loss of performance.
As noted above, before this approach can bedeveloped into a realistic design tool for fluidcontrol, additional controller reduction is re-
5. CONCLUSIONS
In this paper we presented a computationalapproach to feedback design that makes use ofmeasured outputs only. In particular, we use finiteelement approximations to construct state estima-tors rather than require full state feedback. Thereare several important points about this approachthat are worth noting.
5angle (rad)
Kinematic Viscosity .. 2.448-06
~ 2"!~ 0
~-2... 0
5~--'--anglo (rod)
~
;:: 2
~:!I 0~-2... 0
performance. These results arc presented inFigures 10-19.
We close this section by noting that all of thenumerical results presented here assumed no dis-turbances (i.e., 11(1) = ~(l) = 0) so that it is fair tocompare LQR, LQG and MinMax designs. It isnot clear that these results remain the same whendisturbances arc added. This is the subject of on-going work.
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010
FEEDBACK CONTROL OF A THERMAL FLUID USING STATE ESTIMATION It I
FIGURE 20 Approximate Functional Gains. Grid size: 10radial nodes, 50 angular nodes.
[Kez](~) = j' h,.(~, r)v(r)rdr\/,
+ r hT(E"r,
112 J. A. BURNS ct III.
Theory and Biomathematics, Clement, P. and LUl11cr. G.Eds .. Marcel Dckker. New York. NY. PI'. 377-403.
[151 McMillian. C. and Triggiani. R. (1994). "Min-max GameTheory and Algebraic Riccuti Equations for BoundaryControl Problems with Continuous lnput-Solution Map.Part II: The General Case", Applied Mathematic." andOntunizatiou, 29. 1-65.
[16] Rubio. D. (1997). Distributed Parameter Control ofThermal Fluids. Ph.D. Disscmnton. Virginia PolytechnicInstitute and State University.
117] Rubio. D.• "A Computational Study of the Represen-tation Problem for Flow Control", J. Marh. .Systems.1::,\'Ii"l(lI;OI1 IIl1d Control. 10 appear.
118J Rhee. I. and Speyer. J. L.. "A Game Theoretic Controllerand irs Relationship to H(XJ and Lincur-Exponcnuul-Gaussian Synthesis", Proceedings of the 28,h IE/:.-/::Conference 011 Decision and Control. Tampa. FL, Decem-ber 1989.909-915.
[19] Singer. J.• Wang. Y. and Buu, H. H. (1991). "Controllinga Chaotic System". Ph...s, Rei'. Lett .. 66.1123-1125.
[20] van Kculcn, B. (1993). H=-colllrol for Infinite Dimen-sional Systems: a Stale Space Approach. Ph. D. Disserta-1;011, University or Groningen, the Netherlands.
121] Wang. Y.. Singer. J. and Bau, H. (1992). "ControllingChaos in a Thermal Convection Loop", J. Fluid. Mech ..237.479-498.
Downloaded By: [MIT Massachusetts Institute of Technology] At: 14:27 1 July 2010