ContentslistsavailableatScienceDirect
European Journal of Control
journalhomepage:www.elsevier.com/locate/ejcon
Approximate local output regulation for a class of nonlinear fluid flows R
Konsta Huhtala
∗, Lassi Paunonen
Faculty of Information Technology and Communication Sciences, Mathematics and Statistics, Tampere University, PO. Box 692, Tampere 33101, Finland
a rt i c l e i nf o
Article history:
Received 16 April 2021 Revised 19 June 2021 Accepted 25 June 2021 Available online 2 July 2021 Recommended by Prof. T Parisini Keywords:
Distributed parameter systems Nonlinear systems
Output regulation Fluid flow systems
a b s t r a c t
We consider output tracking foraclassof viscous nonlinearfluidflows includingthe incompressible 2DNavier–Stokesequations.Thefluidissubjecttoin-domaininputsanddisturbances.Weconstructan errorfeedbackcontrollerwhichguaranteesapproximatelocalvelocityoutputtrackingforaclassofrefer- enceoutputs.Thecontrolsolutioncoverspointvelocityobservationsandassumesasmoothenoughstate space.Efficacyofthecontrolsolutionisillustratedthroughanumericalexample.
© 2021TheAuthor(s).PublishedbyElsevierLtdonbehalfofEuropeanControlAssociation.
ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/)
1. Introduction
In thiswork, weconsider an output trackingproblemforvis- cousnonlinearfluidflowsintheneighborhoodofa(locally)expo- nentiallystablesteadystatesolution.Weformulateourresultsfor theincompressibleNavier–Stokesequationsonasufficientlyregu- lardomain⊂R2 withboundary.Moreprecisely,weconsider controllinganoutputyoftheequations
∂
w∂
t =ν
w−(
w·∇ )
w−∇
q+fw+fu+fd, (1a)0=
∇
·w, w|
=0, (1b)where w(
ξ
,t)is thefluid velocity, q(ξ
,t)is thefluid pressure,ν
is the kinematic viscosity, fw(
ξ
) is a body force, fu(ξ
,t) is the control action and fd(ξ
,t) is the disturbance action. Our goal is to design a controller such that a chosen velocity output of (1) converges to a desired reference output approximately for initial states which are, in a certain sense, “close enough” to a steadystatesolutionof(1).R The research was supported by the Academy of Finland Grant number 310489 held by L. Paunonen.
∗Corresponding author.
E-mail addresses: konsta.huhtala@tuni.fi(K. Huhtala), lassi.paunonen@tuni.fi(L.
Paunonen).
Theoryof output regulationfor nonlinear systems is still un- der development, especially forinfinite-dimensionalsystems.The finite-dimensionalresultsof[11]havebeenextendedtoaclassof infinite-dimensionalsystems in[5]and forco-locatedinputsand outputsin[6],basedonwhichseveralexamplecasesarepresented in[1].Inthiswork,we focusonoutputregulationinanapproxi- matesenseutilizingtheresultsof[13].In[13],theauthorsusean errorfeedback controller designed forrobust output regulation of exponentiallystableregularlinear systemsandshowthat thesame controllerachievesapproximatelocaloutputregulationforaclassof nonlinearsystemswhichthey callregularnonlinearsystems.Simi- larapproachofusinglinearcontrolsolutionsfornonlinearsystems hasbeenutilizedforlocalstabilizationofnonlinearfluid flowsin differentsetups,seee.g. [3]forin-domaininputs,[14]forbound- aryinputsand[10]forobserverdesign.
As the main contribution of this paper, we show that the Eq. (1) can be formulated as a regular nonlinear system(in the sense of[13]) forawide rangeofvelocity observationsincluding thepointobservation.Toachieve this,we considertheEq.(1)on a “lifted” statespace, i.e.wedemand moresmoothness fromthe velocityandthepressurethan wouldtypically berequiredtoe.g.
solve similar control problems for linear systems. To formulate (1)as a regularnonlinear system,we rely onthe fluid being vis- cous and assume that the domain together with the bound- ary conditions are such that the system can be formulated on the “lifted” state space.These properties,together withthe type (x·
∇
)xofthenonlinearitytypicalforfluidflows,characterizethe fluidflowsforwhichtheresultscanbeapplied.https://doi.org/10.1016/j.ejcon.2021.06.025
0947-3580/© 2021 The Author(s). Published by Elsevier Ltd on behalf of European Control Association. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ )
Usingtheresultsof[13],weshowthatintheneighborhoodof asteadystatesolution,velocity observationson(1)approximately converge toany desired“small enough” periodic referencesignal ofthetype
yr
(
t)
=a0+qs
i=1
aicos
( ω
it)
+bisin( ω
it)
inthesense thatforsmallenoughinitialdata,afinitenumberof chosen harmonicsof the systemoutput andthe referenceoutput are thesame.Here thecoefficient vectorsai,bi∈Rpy maybe un- known butwe expect toknowthe frequencies
ω
i. Thecontroller introducedin[13]alsorejectsperiodicin-domaindisturbancesig- nalsofthetypeud
(
t)
=c0+qs
i=1
cicos
( ω
it)
+disin( ω
it)
withsmallenough amplitude, whereagainci,di∈Rd are allowed to be unknown.Note thatseveralcontrollers havebeen designed for robust outputtracking ofsimilar signal classesin the caseof linearsystems,seee.g.[15,16].
Rest ofthepaperis organizedasfollows.InSection 2,we re- calltheconceptsofregularnonlinearsystemsandapproximatelo- caloutputregulation.InSection3,weshowthattheNavier–Stokes equationswithin-domaincontrolandpointobservationfitintothe framework ofregularnonlinearsystemsona suitablestate space.
InSection 4,weconstruct,basedon[13],a controllerforapprox- imate localoutputregulationfortheNavier–Stokesequationsand then illustrate the controller’s performance through a simulation exampleinSection5.Finally,thepaperisconcludedinSection6.
ThroughoutthepaperwedenotebyL(X,Y)thesetofbounded linear operators froma Hilbertspace X to a HilbertspaceY. For a linear operator A:D(A)→X, D(A) is the domain ofA,
ρ
(A) is the resolvent set of A and TA is the strongly continuous semi- group generated by A on X. For a fixed s∈ρ
(A), we denote by X−1 thecompletionofX withrespecttothenormxX−1=(sI− A)−1xX and define X1=D(A), equipped with the norm xX1= (sI−A)xX.Finally,theL2-innerproductoveradomainisde- notedby(·),(·)L2().2. Regularnonlinearsystemsandoutputregulation
Output regulationforfluid flowsystems coveredby thiswork isbasedontheconceptsofregularnonlinearsystemsandapproxi- matelocaloutputregulation,whichwereintroducedin[13].These conceptsarepresentednext,withthedefinitionofregularnonlin- earsystemsformulatedinaslightlyrestrictedsettingbyexcluding partsthatarenotrelevanttothiswork.
Definition 1. Let X, U, Y and V be Hilbert spaces, and let C defined by Cx=lims→+∞Cs(sI−A)−1x with D(C)=
{
x∈ X|
theabovelimitexists}
be the -extension of the observation operatorC,see[21].Thesystemx˙
(
t)
=Ax(
t)
+Bu(
t)
+Bdud(
t)
+QF(
x(
t))
, x(
0)
=x0∈X,y
(
t)
=Cx(
t)
,whichwedenotebyF,iscalledaregularnonlinearsystemifthe followinghold.
(i) The operator A generates an exponentially stable strongly continuoussemigroupTA onX.
(ii) It holds that B∈L(U,X−1), Bd∈L(Ud,X−1), Q∈L(V,X−1) and C∈L(X1,Y), and the triples (A,B,C), (A,Bd,C) and (A,Q,C)areregularlinearsystemsinthesenseof[21].
(iii) The nonlinear map F:X→V satisfies F(0)=0 andis lo- cally Lipschitz. Thatis, forevery bounded set O⊂X, there existsaconstantLOsuchthatforallx1,x2∈O
F(
x1)
−F(
x2)
V ≤LOx1−x2X.Furthermore,foreach
γ
>0thereexistsaζ
>0suchthatif sup xXx∈O<
ζ
,thenLO<γ
.Togeneratetheplantinput,weuseanerrorfeedbackcontroller oftheform
˙
z
(
t)
=G1z(
t)
+G2e(
t)
, z(
0)
=z0∈Z, (3a)u
(
t)
=Kz(
t)
, (3b)wheree(t)=y(t)−yr(t)istheregulationerrorandZisaHilbert space. Coupling the controller with a regular nonlinear system yieldstheclosed-loopsystemEdefinedby
˙
xe
(
t)
=Aexe+Bewext(
t)
+QeF(
x(
t))
, xe0∈Xe, e(
t)
=Cexe(
t)
+Deyr(
t)
on the Hilbert space Xe=X×Z, where xe=[x, z]T, wext= [ud, yr]T,
Ae=
A BK G2C G1, Be=
Bd 0 0 −G2, Qe=
Q 0, Ce=
C 0 , De=
0 −I .
Before introducing theoutput tracking problem, we recall the concept ofharmonics ofa function. Consider a function f= fp+ fe,wherefp∈L2loc([0,∞);Cn)isT-periodicand fe∈L2([0,∞);Cn). Foranon-negativeintegerl,thelthharmonicof f isthefunction
fl
(
t)
=α
lsin2π
ltT
+β
lcos2π
ltT
, t≥0,where
α
l= limk∈N,k→∞
2 kT
kT 0
f
(
t)
sin2π
ltT
dt,
β
l= limk∈N,k→∞
2 kT
kT 0
f
(
t)
cos2π
ltT
dt,
thusforfrequencies ofthe harmonics,we have
ω
i=2π
li/T. Now the problem of achieving approximate local output regulation is statedasfollows.Problem 2. Let T>0 be a constant. Assume that yr and ud are T-periodic functions and let V=
{
l0,l1,...,lnv}
be a finite set ofnon-negativeintegers.Designanerrorfeedbackcontroller(3)such that:
1. Theclosed-loopsystemEisaregularnonlinearsystem.
2. There exist positive constants cy, cd and ce such that if
yrL2[0,T]≤cy,udL∞≤cdandxe0Xe≤ce,thenxeconverges asymptoticallytoaT-periodicfunctionandthelthharmonicof y−yris0foreachl∈V.Theoutputsatisfiesy=yp+ye,where ye∈L2([0,∞);Y) andyp∈L2loc([0,∞);Y)is a T-periodicfunc- tion.Accuracy of output tracking by solving the above problem clearlydependsonhowdominanttheharmonicsincludedinVare.
In manycases,ensuring that the first few harmonics ofy match thoseofyr resultsinasmalltrackingerror,sincehigherharmon- icsoftheoutputaretypicallysmall.
3. TheincompressibleNavier–Stokesequationsasanabstract controlsystem
The goal of this section is to formulate the Navier–Stokes Eq.(1) asaregular nonlinear system. Westart by findinga suit- able state space forthe formulation andthen verify that there- quirementsofDefinition1arefulfilled.
3.1. Choosingthestatespace
Translating the Eq.(1) to the vicinity of a steady state solu- tion(
v
e,pe)usingthechangeofvariablesv
(ξ
,t)=w(ξ
,t)−v
e(ξ
), p(ξ
,t)=q(ξ
,t)−pe(ξ
)yields∂ v
∂
t =ν v
−( v
e·∇ ) v
−( v
·∇ ) v
e (4a)−
( v
·∇ ) v
−∇
p+fu+fd, (4b)0=
∇
·v
,v |
=0 (4c)withthe initialstate
v
(ξ
,0)=v
0(ξ
).We assumethat thecontrol andthedisturbancearedefinedbyfu
( ξ
,t)
=g1
( ξ )
g2( ξ )
· · · gm( ξ )
u(
t)
, (5a)fd
( ξ
,t)
=g1
( ξ )
g2( ξ )
· · · gd( ξ )
ud(
t)
, (5b)where each g1,...,gm and g1,...,gd is an R2-valued function on , u(t)∈U:=Cm is the finite-dimensional control input and ud(t)∈Ud:=Cd is the finite-dimensional disturbance input. Ad- ditionally, we assume that the output space Y is also finite- dimensionalandY=Cpy withpy≤m.
Forsimplernotation,wedefinethespaces X˜=
v
∈(
L2())
2∇
·v
=0,( v
·n) |
=0, H˜ =
v
∈(
H1())
2∇
·v
=0,v |
=0andthebilinearandtrilinearforms
a
( v
,ψ )
=2ν ( v )
,( ψ )
L2()∀ v
,ψ
∈H˜,b
( v
,φ
,ψ )
=( v
·∇ ) φ
,ψ
L2()∀ v
,φ
,ψ
∈H˜,where
(
v
)=1/2(∇ v
+(∇ v
)T).Assumption3. Weassumethefollowing:
(i) TheboundaryisofclassC3and fw∈(H1())2. (ii) Thelinearizationof(4)isexponentiallystable.
The first part ofthe assumption guarantees sufficient regular- ityofthesolutionsoftheNavier–StokesEq.(1),whilethesecond partisrequiredforDefinition1.(i)toholdandissatisfiedforlarge enough
ν
,c.f.[3].As the first steptowards choosing thestate spaceX, we con- sidersemigroupgenerationforthelinearizedversionof(4)onX˜.A weak formulationforthestationary,linearizedversion of(4)sub- jecttozerocontrolanddisturbanceinputsisgivenby
0=−a
( v
,ψ )
−b( v
,v
e,ψ )
−b( v
e,v
,ψ ) ∀ ψ
∈H˜.Lemma4. TheoperatorA˜definedby A˜=A˜2+A˜1,
A˜2x,ψ
L2()=−a(
x,ψ )
, A˜1x,ψ
L2()=−b(
x,v
e,ψ )
−b( v
e,x,ψ )
, D(
A˜)
=D(
A˜2)
=
x∈H˜
∀ ψ
∈H˜,ψ
→a(
x,ψ )
is X˜-continuousgeneratesanexponentiallystableanalyticsemigrouponX˜.
Proof. We start by showing that a(·,·) is H˜-bounded and H˜- coercive, i.e.H˜ can be continuously and densely embedded inX˜ andthereexistc1,c2,
λ
>0suchthatforeveryφ
,ψ
∈H˜|
a( φ
,ψ ) |
≤c1φ
H˜ψ
H˜, (6a)a
( φ
,φ )
≥c2φ
2H˜−λ φ
2X˜. (6b) Sincethenorm(·)X˜ isequivalenttothe·H˜ normthrough Korn’sandPoincare’sinequalities,weimmediatelyhaveforacon- stantc1>0andforany
v
∈H˜a
( v
,v )
=2ν ( v )
2X˜≥c1v
2H˜.Similarly,foraconstantc2>0andany
v
,φ
∈H˜,|
a( v
,φ ) |
≤2ν ( v )
X˜( φ )
X˜ ≤c2v
H˜φ
H˜.Assuch,a(·,·)isH˜-boundedandH˜-coercive,whichimpliesthatA˜2
generatesananalyticsemigroupTA˜2 onX˜ [2,Section2].
Regarding the trilinearform b(·,·,·), Assumption 3.(i) guaran- tees that
v
e∈H˜, c.f. [12, Ch. 5]. We have forconstants c3,c4>0 usingintegrationbypartsandSobolevembeddings|
b( v
1,v
2,ψ ) |
≤| v
1,( v
2·∇ ) ψ
|
+| v
1·n,v
2·ψ
|
≤c3
v
1L4()v
2L4()ψ
H˜≤c4
v
1H˜v
2H˜ψ
H˜∀ v
1,v
2,ψ
∈H˜.NowA˜1∈L(H˜,X˜),thusperturbationtheoryofsemigroups,seee.g.
[7,Ch.III],impliesthatA˜generatesananalyticsemigroupTA˜ onX˜. ByAssumption3.(ii),TA˜ isexponentiallystable.
The fact that we may choose
λ
=0 in (6b) implies that the semigroupTA˜2 isexponentiallystableforanyν
>0.Furthermore, A˜2 is self-adjoint andthe fractional powers (−A˜2)δ are well de- fined. Domains of the fractional powers are defined by, c.f. [18, Ch.2],[14],D
((
−A˜2)
δ)
=v
∈(
H2δ())
2∇
·v
=0,( v
·n) |
=0, 0≤
δ
<14,D
((
−A˜2)
δ)
=v
∈(
H2δ())
2∇
·v
=0,v |
=0, 1
4<
δ
≤1.Thenorms correspondingto domainsofthefractional powersfor thefullrange
δ
∈Raregivenby xD((−A˜2)δ)=(
−A˜2)
δxX˜.Wenextutilizedomains ofthefractionalpowerstofinda“lifted”
statespaceX suchthatinparticularDefinition1.(iii)issatisfiedby (4).
ForthetranslatedNavier–StokesEq.(4),nonlinearityintheab- stractframeworkF isdescribedby
F
( v )
=−P( v
·∇ ) v
, Q=I, (7)
wherePistheLerayprojector,seee.g.[8,14].Thedomainsofdef- initionforFandQ aredictatedbythefollowingLemma.
Lemma5. For a (small)
δ
>0,choose X=D((−A˜2)1/2+δ) andV= D((−A˜2)δ).ThenDefinition1.(iii)holdsforF.Proof. The proof is based on the “propertiesof multipliers”, see [17,Ch.4.6.1,Thm.1],[4,Lemma5.4],whichstatethatif
s2>s1, s2> d
2 , (8)
wheredisthespatialdimension,then Hs1·Hs2→Hs1
isacontinuousembedding,where Hs1·Hs2:=
f g
f ∈Hs1, g∈Hs2.
Since d=2, we choose s1=2
δ
and s2=1+2δ
, and apply theaboveresult.Nowforaconstantc1>0and
φ
,ψ
∈Xφ
i∂
jψ
kHs1()≤c1φ
iHs2()∂
jψ
kHs1() (9a) fori,j,k∈{ ξ
1,ξ
2}
,thusforsomeconstantsc2,c3>0also( φ
·∇ ) ψ
V≤c2φ
X∇ψ
V≤c3
φ
X(
−A˜2)
1/2ψ
V=c3
φ
Xψ
X. (9b) Utilizing (9), forv
1,v
2∈X and some constants c4,c5>0 we have F( v
1)
−F( v
2)
V=
−P( v
1·∇ ) v
1−( v
2·∇ ) v
2V
=
P(( v
1−v
2)
·∇ ) v
1+( v
2·∇ )( v
1−v
2)
V≤c4
( v
1−v
2)
X∇ v
1V+v
2X∇ ( v
1−v
2)
V≤c5
( v
1X+v
2X) v
1−v
2X,thusF islocallyLipschitz.ClearlyF(0)=0,andif
v
1X,v
2X<γ
c for a large enough constant c>0, then (
v
1X+v
2X)<γ
,whichcompletestheproof.
DuetoLemma5,wechooseforafixed(small)
δ
>0 X=D((
−A˜2)
1/2+δ)
as the state space for our abstract system presentation and de- note
Xs=D
((
−A˜2)
1/2+δ+s) ∀
s∈R,withthecorrespondingnormsdefinedaccordinglyby
xXs=(
−A˜2)
sxX =(
−A˜2)
1/2+δ+sxX˜. NowV=X−1/2,F:X→V andQ=IV ∈L(V). 3.2. TheabstractsystemformulationWedefinetheoperators
A=A2+A1:D
(
A)
→X, (10a)A2=
ν
P, A1
v
=−P( v
e·∇ ) v
+( v
·∇ ) v
e , (10b)D
(
A)
=D(
A2)
=D((
−A˜2)
3/2+δ)
. (10c) Now A2v
=A˜2v
and A1v
=A˜1v
forv
∈D(A). To verify that Definition 1.(i) holds on the state space X, we note that A gen- eratesastronglycontinuoussemigroupTAonX,c.f.[7,Ch.5].The semigroupTAisexponentiallystable,sinceforx∈X TAxX=(
−A˜2)
1/2+δTAxX˜=
TA˜(
−A˜2)
1/2+δxX˜≤
TA˜L(X˜)xX.We still need to verifyDefinition 1.(ii).We do so forcontrols anddisturbancesoftheform(5)andobservationsuptothe“level of unboundedness” of a point observation. Using integration by parts,wehavefortheX-adjointofA1
A∗1
φ
=P( v
e·∇ ) φ
−( ∇ v
e)
Tφ
.
Properties ofmultipliers withthechoicess1=1+2
δ
,s2=2+2δ
tosatisfy(8)imply,similarlyto(9),forany
φ
,ψ
∈X1/2andacon- stantc>0 P( φ
·∇ ) ψ
X≤cφ
2X1/2ψ
2X1/2, P( ∇φ )
Tψ
X≤cφ
2X1/2ψ
2X1/2.SinceAssumption3.(i)implies
v
e∈X1/2⊂(H2+2δ())2 [12,Ch. 5], wehaveA1,A∗1∈L(X1/2,X).Assuch,theoryofadmissiblecontrolandobservationoperators, see[20,Ch.4–5],nowstatesthat
• LetY˜ beaHilbertspace.IfC˜∈L(X1/2,Y˜),thenC˜isanadmissi- bleobservationoperatorforTA2 anditsadjointC˜∗∈L(Y˜,X−1/2) isanadmissiblecontroloperatorforTA2.
• The sets of admissible control (observation) operators for TA
andTA2 arethesame.
NotethataboveweassumedforY˜tobeself-dual.
Wefirstsearchforadmissibleobservationsfor(4)onthestate spaceX byconsidering observationssuchthatC∈L(X1/2,Y).Typ- icallythe“mostunbounded” observationofinterestwouldbe the pointobservation
Cpx
( ξ
,t)
=x( ξ
p,t)
(11)forsome
ξ
p∈.BySobolevembeddings,when ⊂R2,Hs()⊂ C()¯ for s>1,thus Cp∈L(X,C). Thatis, all the observations of interestfor(4)areboundedoperatorsfromX toY.Assuch,C=C andifU,Ud andQ are admissiblecontrol operatorsforTA2,then Definition1.(ii)holds.Consider next admissible control operators for TA2, thus also for TA. We start with the operator Q=IV =IX
−1/2. Note that in this case the “input space” V is not self-dual, but instead the correct dual is the X-dual of X−1/2, i.e. V=X1/2. Thus we have Q∗∈L(X1/2,V) and Q∈L(V,X−1/2), i.e. the triple (A,Q,C) is a regularlinearsystem.
Forasinglecontrolinputofthetype(5a),wehaveB=g(
ξ
).If g∈Xs,thenB∈L(C,Xs).Thatis,ifg∈X−1/2=D
((
−A˜2)
δ)
,thenBisanadmissiblecontroloperatorforTA2.
We concludethe section by gatheringourfindings inthe fol- lowingresult.
Theorem6. GivenAssumption3,assumethatthecontrolshapefunc- tionsgiandthedisturbanceshapefunctionsgjsatisfygi,gj∈X−1/2= D((−A˜2)δ)foreach i=1,2,...,m, j=1,2,...,d andasmall
δ
>0. ThenthetranslatedNavier–StokesEq.(4)withthedynamicsoperator (10), thenonlinearity(7),the control(5a),thedisturbance (5b)and py point observations (11) form a regular nonlinear system on the statespaceX=D((−A˜2)1/2+δ).4. Thecontroller
Weusealow-gain-typecontrollerdesignintroducedin[13]to solveProblem2.Theonlysysteminformationrequiredtoconstruct thecontrolleristhetransferfunctiongains
G
(
±iω
k)
=C(
±iω
kI−A)
−1Bofthe linearizedsystem(A,B,C) forthefrequencies
ω
k=2π
lk/T foreach lk∈V.A goodestimate for thesegainsof thelinearized systemcanbeobtainedexperimentallyfromthegainsofthenon- linearsystemN,see[13],androbustnessofthecontrollermeans that the approximate gains can be used to achieve the output trackinggoal.The controller consistsoftwo finite-dimensional systems. The firstsystemF isdescribedbythetransferfunction
GF
(
s)
=IY−nv
k=0
s2+
ω
2ks2+
ε
s+ω
2kIY,
Fig. 1. The closed-loop system.
where
ε
>0 is the control tuning parameter. The second system RisdescribedbythetransferfunctionGR
(
s)
=nv
k=0
s2+
ω
2ks2+2s+
ω
2k×
nv
k=0
ρ
kRk s−iω
k+
ρ
−kR−k s+iω
k ,where
Rk=G∗
(
iω
k)(
G(
iω
k)
G∗(
iω
k))
−1, R−k=G∗(
−iω
k)(
G(
−iω
k)
G∗(
−iω
k))
−1,ρ
k=nv
j=k,j=0
ω
2j−ω
2k+2iω
kω
2j−ω
k2,
ρ
−k=nv
j=k,j=0
ω
2j−ω
2k−2iω
kω
2j−ω
k2 .WedenoteastatespacerealizationofGF onXF=CnF by
˙
xF
(
t)
=AFxF(
t)
+BFuF(
t)
, xF(
0)
=xF0∈XF, yF(
t)
=CFxF(
t)
,andastatespacerealizationofGRonXR=CnR by
˙
xR
(
t)
=ARxR(
t)
+BRuR(
t)
, xR(
0)
=xR0∈XR, yR(
t)
=CRxR(
t)
+DRuR(
t)
.After coupling the two subsystems of the controller as depicted in Fig.1,i.e.bysettinguF =y−yr+yF anduR=yF,we havethe structure ofan errorfeedback controller (3) withz=[xF, xR]T∈ XF×XR,
G1=
AF+BFCF 0 BRCF AR, G2=
−BF 0, (12a)
K=
DRCF CR . (12b)
The followingresultisobtainedin[13]fortheclassofregular nonlinear systemsandwe formulateit fortheincompressible2D Navier–Stokesequations.
Theorem7. AssumethatG(i
ω
k)issurjectiveforeachk=1,2,...nv andtheassumptionsofThm.6hold.Thereexistsε
∗>0suchthatan error feedback controller(3)with theoperatorschosenas (12)with 0<ε
≤ε
∗solvesProblem2forthesystem(10),(11).Proof. The proof follows directly from Theorem 6 and [13, Sec- tion5.2].
5. ANumericalExample
Let be the unit disk and consider the Navier–Stokes Eq.(1) with
ν
=1/25around a steadystate solutioncorrespond- ingtothebodyforcefw
( ξ
1,ξ
2)
=ξ
2(
1−ξ
12−ξ
22)
, −ξ
1(
1−ξ
12−ξ
22)
T∈H˜ and fu=0, fd=0.Our output trackinggoal isto have thepoint observationy
(
t)
=Cv
1( ξ
,t) v
2( ξ
,t)
=
v
2 0.4, −0.4 ,t∈L
(
X,R)
Fig. 2. The steady state velocity field ve(ξ), where color depicts speed of the fluid.
trackthereferencesignal
yr
(
t)
=0.5sin(
2t)
(13)despitethedisturbance fd
( ξ
,t)
=Bdud(
t)
=P0,
χ
d( ξ )
T(
1+cos(
2t))
,where
χ
d is the characteristicfunction and d=[−0.4,−0.1]× [−0.4,−0.1].The outputtrackingistobeachieved,approximately andlocally,byusingthecontrolfu
( ξ
,t)
=Bu(
t)
=Pχ
u( ξ )
, 0 Tu(
t)
where u=[−0.6,−0.3]×[0.1,0.4]. Now U=Ud=Y=R and since
χ
u,χ
d∈Hs()for any s<1/2 [19], also B∈L(U,X−1/2) andBd∈L(Ud,X−1/2).Assuch,ifthesteadystate(v
e,pe)islocally exponentiallystable,thenthetranslatedsystem(4)formsaregular nonlinearsystemonthestatespaceX.WeusetheTaylor–Hoodfiniteelementspatialdiscretizationfor theNavier–Stokes equations.Withthe help offunctionsincluded inthe MatlabPDEtoolbox, theunit diskis approximatedby 694 triangleswiththemaximumedge lengthof≈0.1,whichleadsto approximationorderof1453 foreach ofthevelocity components and380forthe pressure.The steadystate solution (
v
e,pe),with thesteadystate velocitydepictedinFig.2,iscalculatedusingthe Newton’s method, and we assume pe(0)=0 to obtain a unique steadystatepressure.We check numericallythat linearizationof the translatedsys- tem(4)isexponentiallystable. Thenwe designanerrorfeedback controller(12)withV=
{
0,1,2,3}
andchooseasthecontroltun-ingparameter
ε
=0.095toroughlymaximizethestabilitymargin ofthelinearizedclosed-loopsystem. Forthesimulation,werelax the incompressibility condition by using a penalty method with thepenaltyparameterp=10−5,seee.g.[9,Ch.5.2],todecouple the fluid pressurefromthe fluid velocity.As the initial state, we use
xe0=
v
e−v
e1/2, 0 T∈X×Z,where
v
e1/2(ξ
) is the steady state velocity corresponding to the body force fw1/2=0.5fw and fu=0, fd=0. Evolution of the closed-loop system is then solved using Crank–Nicolson method withthetimestept=0.01togetherwithNewtoniteration.Output tracking performance of the controller is depicted in Fig.3 anda snapshotof thefluid velocity at thetime t=120 is showninFig.4.
Thecontrollerachievesoutputtrackingof(13)withsatisfactory performance forthe chosen initial state despite the disturbance.
The effectofthedisturbance isnot clearlyvisibleinFig. 3,since