ContentslistsavailableatScienceDirect
Journal of Process Control
jo u rn al h om ep age :w w w . e l s e v i e r . c o m / l o c a t e / j p r o c o n t
Distributed control of interconnected Chemical Reaction Networks with delay
L ˝orinc Márton
a,∗, Gábor Szederkényi
b,c, Katalin M. Hangos
b,daDepartmentofElectricalEngineering,SapientiaHungarianUniversityofTransylvania,TirguMures,Romania
bSystemsandControlLaboratory,InstituteforComputerScienceandControl,HungarianAcademyofSciences,Budapest,Hungary
cFacultyofInformationTechnologyandBionics,PázmanyPéterCatholicUniversity,Budapest,Hungary
dDepartmentofElectricalEngineeringandInformationSystems,UniversityofPannonia,Veszprém,Hungary
a r t i c l e i n f o
Articlehistory:
Received12March2018
Receivedinrevisedform25July2018 Accepted5September2018
Keywords:
Processsystems
ChemicalReactionNetworks Multi-agentsystems Passivity
Delaysystems Distributedcontrol Nonlinearsystems
a b s t r a c t
ThispaperintroducesacontrolapproachforaclassofChemicalReactionNetworks(CRNs)thatare interconnectedthroughadelayedconvectionnetwork.First,acontrol-orientedmodelisproposedfor interconnectedCRNs.Second,basedonthismodel,adistributedcontrolmethodisintroducedwhich assuresthateachCRNcanbedrivenintoadesiredfixedpoint(setpoint)independentlyofthedelayin theconvectionnetwork.Theproposedalgorithmisalsoaugmentedwithadisturbanceattenuationterm tocompensatetheeffectofunknowninputdisturbancesonsetpointtrackingperformance.Thecontrol designappliesthetheoryofpassivesystemsandmethodsdevelopedformulti-agentsystems.Simulation resultsareprovidedtoshowtheapplicabilityoftheproposedcontrolmethod.
©2018TheAuthor(s).PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense (http://creativecommons.org/licenses/by/4.0/).
1. Introduction
Thecontrolofplant-wideindustrialprocessesisinthefocus of the researchers for decades [1,2]. The decentralized or dis- tributedcontrolapproachesareadvantageousinprocessnetwork applicationstoreducethecommunicationcostsandpossiblecom- municationhazards that couldarise in the case of centralized control[3].Becauseoftheirpracticalimportanceandchallenging nature,thedifferentnonlinearcontrolapproachesappliedinpro- cesscontrolhavedevelopedtheirownversionofdecentralized, distributedorhierarchicalcontrolarchitectures.
In thepaper [4] the authors proposeda distributed control approachforsuchinterconnectedprocessesthatcanbemodeledas lineartime-invariantsystemsbasedonpassivitytheory.Thiscon- trolmethodalsotakesintoconsiderationthetransportdelayinthe interconnectionsamongtheprocesses.
Apowerfulcontrolapproachofinterconnectedprocesssystems isbasedonthethermodynamiccharacterizationofsuchsystems.
Amodelingframeworkhasbeendevelopedin[5]fornetworksof chemicalprocessesconsideringalsothethermodynamiceffects.
∗Correspondingauthor.
E-mailaddress:martonl@ms.sapientia.ro(L.Márton).
Usingthetheoryofcascade-connectednonlinearsystemsandthe propertiesofMetzlerandHurwitzmatrices,astabilizingdecentral- izedcontrolapproachwasdesignedin[6]utilizingthehierarchical structureofconservationbasedprocessmodels.
Thepopularandpowerfulmodelpredictivecontrolapproach is alsoused in distributed and in hierarchicalframeworks (see thepaper[7]fora review).Thisapproachcanhandlenonlinear interconnectedprocess systems, aswell.A more recent review highlightingfutureresearchdirectionsinthisapproachisavailable in[8].Constrainedcontrolmethodscanbeappliedwheninputor stateboundsshouldalsobetakenintoconsideration,seee.g.[9]or [10].
ChemicalReactionNetworktheoryprovidesefficientmodels andtechniquestodescribeandanalyzenotonlythedynamicsof chemicalreactions[11],butamorewideclassofnonlinearprocess systems.Thestudy[12]dealswiththemodelingofinterconnected reactorsandshowsthatthetransportmechanismcanbedescribed byalinearCRNmodel.TheCRNmodelsdescribepositivesystems andcanefficientlycapturecomplexnonlineardynamicalphenom- ena. The paper[13] offersa modeling approach for CRNs with inflowsandoutflowsanddiscussestherelationofthesesystems withtheconsensusdynamics.Thestabilityofthesesystemsismost oftenanalyzedusingentropy-basedLyapunovfunctions[14].
Unfortunately,however, onlya fewof thenonlinearprocess controlapproachesthatweredevelopedforcontrollingintercon- https://doi.org/10.1016/j.jprocont.2018.09.004
0959-1524/©2018TheAuthor(s).PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/).
nectedsubsystemsareabletoexplicitlyhandledelaysindynamic analysisandcontrollerdesign.Someexceptionsincludethepaper [15],where thesynchronization problemofaclass ofintercon- nectednonlinearbiochemicalprocesseswasconsideredbytaking aninput-outputmodelingapproach.Importantmodelingandsta- bilityandcontrolrelatedresultsondelayedkineticsystemscanbe foundin[16],[17],[18],[19]and[20].
Inthisstudyacontrol-orientedmodelingapproachisproposed forinterconnectedCRNs,relyingbothoncompartmentalsystems [21], and CRNtheory. Themodel takesinto consideration both thenonlinearnatureofthechemicalprocessesandtheunknown transportdelaysintheconvectionnetwork.Adistributedsetpoint controlalgorithmisintroducedwhichassuresthattheconcentra- tionlevelsofthechemicalsineachreactorreachprescribedfixed points.Theproposedcontrolmethodcanalsoattenuatetheeffect ofunknowninputdisturbancesonthecontrolperformance.The resultingcontrolalgorithmhasaneasilyimplementableform,itis independentoftheCRNskineticsandthedelayintheconvection network.Thestabilityandtrackingperformanceoftheintercon- nectedCRNwiththeproposedcontrolisanalyzedusingtechniques borrowedfromthetheoryofmulti-agentsystems.
2. Interconnectedpassivesystems
2.1. Passivesubsystems
ConsideraninterconnectedsystemconsistingofCsubsystems inwhichtheinputofeachsubsystemmaydependontheoutputs oftheothersubsystems.
EachsubsystemismodeledusingODEs(OrdinaryDifferential Equations)intheform
˙
c(j)=f(j)(c(j))+G(j)(c(j))u(j), c(j)(0)=c(j)◦ , (1) y(j)=h(j)(c(j))
wherec(j) ∈Rn,y(j),u(j) ∈Rmarethestate-,output-andinputvec- tors,f(j)(·),h(j)(·),G(j)(·)aresmoothmappingswithappropriate dimensionssuchthatf(j)(0)=0,h(j)(0)=0,j=1...C.
Definition1. System(1)iscalledpassive,ifthereexistsacontin- uouslydifferentiablefunctionS(j):Rn→Rsuchthat
S(j)(c(j))≥0, ∀c(j), (2)
S(j)(0)=0, (3)
S˙(j)≤y(j)Tu(j), ∀u(j),c(j). (4) Siscalledthestoragefunctionof(1)(see,e.g.[22]).
Theinput-affinesystem(1)ispassiveiffthefollowingconditions hold
∂S(j)
∂c(j)f(j)(c(j))≤0, (5)
∂S(j)
∂c(j)G(j)(c(j))=
h(j)(c(j))
T, (6)
seee.g.[23].
Passivitytheoryplaysakeyrole inanalyzingthestabilityof nonlinearsystemsasitisshownthatpassivityof(1)involvesthe stabilityoftheautonomoussystem ˙c(j)=f(j)(c(j))undermildcon- ditions[22].
2.2. Interconnections
Theunderlyinggraphoftheinterconnectedsystemisadirected graphwithCverticesinwhicheachvertexcorrespondstoasub- system.Thereisadirectededgefromthevertexktothevertexjif
theinputofthejthsubsystemdependsexplicitlyontheoutputof thekthsubsystem.
Neighborsetofthejthvertex(Nj):thekthvertexbelongstoNj
ifthereisadirectededgefromthevertexktothevertexj.
Alikethemodelingconceptsdevelopedforlarge-scalesystems [3],considertheinputofeachsubsystem(u(j))asthesumofalocal controlinput(u(j)L )andantheinterconnectioninputtermwhich hastheform:
i(j)(t)=i(j)
y(k1)(t−Tk1j),...,y(kJ)(t−TkJj)
, (7)
wherey(kl)aretheoutputsofthesubsystemsoftheneighborsetNj
(dim(Nj)=J),and0≤Tkij<∞isaconstanttransportdelayfrom theagentkitotheagentj.
Theoutputsofthesubsystemsintheinterconnectedsystemare synchronizediflim
t→∞|y(j)(t)−y(k)(t−Tkj)|→0,∀j,k.
Assumetheinterconnectioninputsintheform(synchronization protocol)
i(j)(t)=
k∈Nj
wkj(y(k)(t−Tkj)−y(j)(t)), wkj>0, (8)
andu(j)L =0element-wise.Fortheanalysisoftheinterconnected systemswithsuchsubsysteminputsthefollowingfunctionalcan beapplied:
S=
N i=1S(j)+
N j=1k∈Nj
t t−Tkjy(j)T()y(j)()d. (9)
Itisshownin[24,25]that,undercertainassumptionsontheunder- lying graph and the subsystems, the interconnection input (8) assuresthesynchronizationofthesubsystems.
3. Basicmodelingnotions
3.1. ChemicalReactionNetworksandtheirstability
ChemicalReactionNetworks(abbreviatedasCRNs) arecom- posedofelementaryirreversiblereactionsRk:Ci→Cj,k=1...,R, whereCj, j=1,...,marethesocalledcomplexes.AcomplexCj isformallyalinearcombinationofspeciesXi, i=1,...,K,such thatCj=
Ki=1ˇijXi,forj=1,...,m,whereˇijisthenonnegative stoichiometriccoefficientcorrespondingtospeciesXiincomplex Cj.
Theconcentrations ofthespecies arecollectedintoavector c ∈RK sothatci=[Xi] for i=1,...,K.Thedynamics ofa CRN describingthetimeevolutionoftheconcentrationsofthespecies induced by thereactionscan bewritten in thefollowing form assumingconstantvolumeandtemperature(see,e.g.[26]):
˙
c=Mϕ(c)=YAϕ(c), c(0)=c◦ (10)
where c◦ is strictlypositiveelement-wise,and Y ∈RK×m is the complexcompositionmatrixthejthcolumnofwhichcontainsthe stoichiometriccoefficientsofcomplexCj,i.e.Yij=ˇij, ∀i,j.More- over,ϕi(c)= Ki=1cYiik isthemassactionvectorandA ∈Rm×mis theKirchhoffmatrix:
A(i,j)=
⎧ ⎪
⎨
⎪ ⎩
ji, for j=/i
−
=/j
j, if j=i. (11)
wherejiistherateconstantofthereactionRk:Cj→Ci.
The reaction vector of Rk is formed by the corresponding stoichiometricvectors, suchthatek=Y·i−Y·j. Thespanof the reactionvectorsdefinesthestoichiometricsubspaceoftheCRN:
Sc=span
ek
.Thepositivestoichiometriccompatibilityclasses ofaCRNarerepresentedbySc◦=(c◦+Sc)∩RnS+.
The generalCRN model (10) may have multiple (even infi- nitenumberof)steadystatesinthewholestatespace.Therefore, thestructureandnumberofequilibriaaremostoftenstudiedby restrictingthedynamicstothestoichiometriccompatibilityclasses correspondingtodifferentinitialconditions.
Example1(AsimpleCRN). LetusconsideraChemicalReaction Networkconsistingofthefollowingreactions:
R1:2X3k→122X1k→232X2, R2:2X3[13]
312X2. (12)
Themodelcontainsthreespecies:X1,X2,X3,andthreecomplexes:
C1=2X3,C2=2X1,C3=2X2.Fromthese,thecomplexcomposition matrixcanbewrittenas
Y=
⎡
⎣
00 20 02 2 0 0⎤
⎦
(13)Wecanseefrom(12)thatthenetworkcontainsfourelementary reactions.Theratecoefficientsofthesereactionsarethenon-zero off-diagonalelementsoftheKirchhoffmatrixwhichisgivenby
A=
⎡
⎢ ⎣
−(12+13) 0 31
12 −23 0
13 23 −31
⎤
⎥ ⎦
(14)Thepair(Y,A)iscalledarealizationofakineticsystemwitha givencoefficientmatrixMandreaction-monomialvectorϕ,where thecomplexcompositionmatrixYisdeterminedbyϕ.Itisimpor- tanttonotethatarealizationofakineticsystemmaynotbeunique, i.e.theremayexistmorethanoneKirchhoffmatrixAforakinetic dynamicsgivenbyMandϕ[27].
ThereareimportantstructuralpropertiesofaCRNrealization thatcanbeusedfor determiningthestability propertiesofthe dynamics,thatarethedeficiencyandthereversibilityproperties.
Thedeficiencyof aCRN realizationisdefined ası=dim(KerY∩ ImA).A CRN isweakly reversible iftheexistence of a directed path(i.e.reactionsequence)fromthecomplexCitothecomplexCj impliestheexistenceofadirectedpathfromCjtoCi.
Althoughweakreversibilityandzerodeficiencyisarealization propertyofaCRN,theyhaveimportantimplicationsonthestabil- ityoftheCRNsystem.IfaCRNisweaklyreversibleandhaszero deficiencythenthesystem(10)hasexactlyoneequilibriumpoint (c∗)ineachpositivestoichiometriccompatibilityclass[11]thatis atleastlocallystablewiththefollowingLyapunovfunction:
S(c)˜ =
K i=1ci
lnci c∗i −1
+ci∗
(15)
Anequilibriumpointc∗oftheCRN(10)iscalledcomplexbal- ancedif
Aϕ(c∗)=0 (16)
Itisalsoknownthatifthereexistsa complexbalancedequilib- riuminaCRN,thenallotherequilibriaarecomplexbalanced,too [26].Therefore, complexbalanceis a systempropertyoncethe CRNstructureandparametersarefixed,andthusaCRNitselfcan becalledcomplexbalancedif(16)isfulfilled.However,akinetic differentialequationmayhaveseveralcomplexbalancedandnon- complex-balancedrealizations[27].Itisimportantthatcomplex balanceimpliesweakreversibility. Complexbalanceis strongly relatedtothestabilityofkineticsystems.Accordingtothewell- knownGlobalAttractorConjecture(GAC),complexbalancedCRNs aregloballystablewiththeLyapunovfunction(15).TheGACwas
provedforseveralspecialcases,mostremarkablyforCRNswith onereactiongraphcomponent[28],andaproofforthegeneral problemhasbeenreportedin[29].Asignificantresultinthethe- oryofCRNsisthatanydeficiencyzeroweaklyreversiblenetwork iscomplexbalancedindependentlyofthevaluesoftheratecon- stants[30].Thisensuresarobuststabilitypropertywhichcanbe importantinthegeneraltheoryofnonnegativesystems[31].
Inordertohavepassiveagentsfortheanalysis,weconsiderthe followingsub-classofCRNs.
Assumption1. TheCRN(10)iscomplexbalanced.
Assumption2. TheCRN(10)ispersistent.
Persistencemeansthatthetrajectoriesofthesystem(10)do notapproachtheboundary∂RK+arbitrarilyclose,i.e.∀i=1...Kit stands:ci(t)>0ifci◦>0andt≥0.Itisimportanttonotethat Assumptions1and2arestronglyrelated.InthecaseofCRNswith onegraphcomponent,complexbalanceimpliespersistence[28].
Werecalltheimportantspecialcasethatdeficiencyzeroweakly reversiblenetworksare complexbalanced for anypositiverate coefficients[30].Moreover,accordingto[29](whichisnotpub- lishedofficiallyatthetimeofwriting)thedynamicsofanycomplex balancedCRNispersistent.
3.2. NetworkofCRNsconnectedbyconvectionwithdelay
ItisconsideredthatthemassactionCRNsarelocatedincon- tinuouslystirredtankreactors(CSTRs)thatareconnectedthrough staticconnections.InordertohaveausualCRNmodelconsidered inSection3.1,weassumeconstantvolume,constanttemperatureand constantphysico-chemicalpropertiesineachCSTR.
Aswe shall seelater in Section4.1,the above assumptions togetherwithAssumptions1and2ensurethatthelocalCSTRsare passivewithacertaininput-outputpair.Inpracticetheconstant temperatureassumption–thatimpliesconstantphysico-chemical propertieswithconstantpressure–isapproximatelyvalidformost ofthebiochemicalapplications,wherealsothecomplexbalanced andpersistentnatureofthereactionnetworkisalsovalid.Thecon- stantvolumeassumption,however,putssevererestrictionsonthe convectionnetworkasitwillbedescribedlaterinSection3.2.2.
EachCSTRhasaninletandanoutletportwithvolumetricflow ratesvIiandvisuchthatvIi=vi, i=1...C,wherethenumberof CSTRsisdenotedbyC.
We also introduce a pseudo-CSTR (CSTR0) for describing the environment.Becauseoftheconstantvolumeassumptionofeach internalCSTR,thisassumptionalsoholdsfortheenvironment,such thatvI0=v0.
Assumption3. InCSTR0theconcentrationvector(c(0))isconstant andstrictlypositiveelement-wise.
Intheusualpracticallyimportantcasesthevariationoftheinlet feed–usuallyitscompositionbutsometimesevenitsflowrate ischanging–isthemajordisturbancetoaplant,therefore,the constantconcentrationsassumptioninCSTR0thatrepresentsthe environmentdoesnotalwaysholdinpractice.However,onecan relaxthis assumptionifadisturbancerejectiveextensionofthe controlschemeisdeveloped:thistechniqueisusedforourpro- posedcontrolmethodinSection4.3.
3.2.1. OpenCRNmodel
Anumber ofC mass-actionChemical ReactionNetworksare consideredinthesystemunderinvestigation.LetY(j)andA(j) be thestoichiometricandKirchhoffmatricesoftheODEmodelofthe CRNthattakesplaceinthejthCSTR:
˙
c(j)=Y(j)A(j)ϕ(j)(c(j)). (17)
Thenthecomponentmassbalancecontainingthein-andout- flowconvectivetermsofthejthCSTRreadsas
dc(j) dt = 1
Vj
C=0
˛jvc()−vjc(j)
+Y(j)A(j) ϕ(j)(c(j)) (18)
wherej=1...C.
3.2.2. Theconvectiveconnections
Connectionsaresetupbetweenthereactorssuchthattheoutlet oftheithreactorisdividedintofractionswiththefractioncoeffi- cients˛ijthatarefedintothejthreactor.Thismeansthat
C=0
˛i=1, i=0,...,C, (19)
vIj=vj=
C =0˛jv, j=0,...,C. (20)
Becauseconstantvolumeisassumedineveryregion,thesum ofconvectiveinflows
v0=
C=0
˛0v (21)
isequaltothesumoftheconvectiveoutflowsoftheprocesssys- tem.Itisimportanttonotethatthecompositesystemconsistingof theoriginalprocesssystemanditsenvironmentisclosedwithC+1 regionseachofconstantvolume.
With thenotations above we can formulate theconvection matrixasfollows:
CC=
⎡
⎢ ⎢
⎢ ⎣
−(1−˛00)v0 ˛10v1 ˛20v2 ... ˛C0vC
˛01v0 −(1−˛11)v1 ˛21v2 ... ˛C1vC
···
˛0Cv0 ˛1Cv1 ˛2Cv2 ... −(1−˛CC)vC
⎤
⎥ ⎥
⎥ ⎦
(22) ThismatrixwillbetermedKirchhoffconvectionmatrix.
TheconstantvolumeassumptionimpliesthatCC1=0which is also a consequence of Eq. (19).Here 1=(11...1)T and 0= (00...0)T.Moreover,Eq.(20)impliesthat1TCC=0T.
TheabovetwoequationsandthesignpatternofCCshowsthat Kirchhoffconvectionmatricesarebothrowandcolumnconserva- tionmatrices.WhenthereisnoflowfromtheithCSTRtothejth CSTR,theparameter˛ij=0inCC.TheKirchhoffmatrixdescribes thestructureoftheunderlyingdirectedgraphGCoftheconvection network:theweightedLaplacianofGCis−CC.OnGCthefollowing assumptionismade:
Assumption4. GC contains adirectedspanningtreewithroot CSTR0.
Thisassumptionmeansthatthesupplyfromtheenvironment reacheseachindividualreactor,therearenoreactorswhichare unreachablefromtheCSTR0intheprocessnetwork.
Thisassumptioniseasytoverifyinpracticebasedontheflow sheetoftheplant,butitmaynotholdinallpracticalcasesforthe overallplant.Atthesametime,ifthesupplyisreallynecessaryfor theproduction,thentheplantcanbenaturallydecomposedinto sub-plantsthatobeyAssumption4individually,andthecontroller designcanbedoneforthemseparately.
3.2.3. Connectedmodelwithconvectiondelays
Considerthattransportdelaysarepresentintheinterconnec- tionsamongtheCSTRs.Denotethedelayvaluebetweenthethand
jthCSTRasTj.Thenthestateequationof(18)obtainsthemodified form:
Vjdc(j)i dt =
C=0
˛jvc()i (t−Tj)−vjc(j)i +VjY(j)A(j)ϕ(c(j)), (23)
c()i ()=()i (), −Tj≤≤0.
Herej=1...Candi()()∈C+isaninitialconditionfunction.
4. Distributedcontrollerdesign
Thedesignofthedistributedcontrolschemeisbasedonthe passivityanalysisoftheinterconnectedCSTRsthatisgiveninthe nextsection. Thereafter the distributed setpoint control design willbeintroducedandfinally,itsdisturbancerejectiveversionis described.
4.1. StoragefunctionandpassivityanalysisofopenCRNs
LetusconsiderthattheinputofthejthopenCRN(23)isthe vectoru(j) ∈RKsothat:
dc(j)
dt =Y(j)A(j)ϕ(j)(c(j))+ 1
Vju(j). (24)
NotethattheabovegeneralformisderivedfromEq.(18)bycon- sideringthedifferencebetweentheconvectivecomponentmass in-andoutflowtermasaninput.However,becauseoftheconstant volumeassumptionineachCSTRonecanarbitrarilymanipulate onlythelocalinletconcentrationsfromtheenvironment,butnot theflowrates.
ChoosethestoragefunctionforthejthCRNastheweightedform oftheLyapunovfunction(15):
S(j)= Vj
Lnc(j)−Lnc(j)∗
Tc(j)−1T
c(j)−c(j)∗
. (25)
Here ∈R+isapositivefiniteconstant,Lnisthenaturallogarithm appliedelement-wisetoavector,andc(j)∗ isa(generallyinitial conditiondependent)equilibriumpointofthejthCRN.
Forthepassivitytheoutputmappingofthesubsystemhasto satisfytherelation(6).Fromthemodel(24)itresultsthattheoutput ofthejthsubsystemhastohavetheform:
y(j)= 1 Vj
∂S(j)∂c(j)
T. (26)
ByEqs.(25)and(26)yieldsthepassiveoutputvectorofthejth CRN:
y(j)=
Lnc(j)−Lnc(j)∗
. (27)
Lemma1. TheopenCRNsystem(24)withAssumption1ispassive fromtheinputu(j)totheoutputy(j)inEq.(27).
Proof. Considerthestoragefunction(25).Thetimederivativeof S(j)readsas:
S˙(j)(c(j))= Vj
Lnc(j)−Lnc(j)∗+1
T˙
c(j)−1Tc˙(j)
, (28)
S˙(j)(c(j))= Vj
Lnc(j)−Lnc(j)∗
T1
Vju(j)+Y(j)A(j)ϕ(c(j))
. (29) ByAssumption1,ifu(j)=0, ˙S(j) isnon-increasing,seee.g.the study[11].Hence,
Lnc(j)−Lnc(j)∗
TY(j)A(j)ϕ(c(j))≤0.Ityields:
S˙(j)(c(j))≤y(j)Tu(j), (30)
i.e.thesystem(24)ispassive.䊐
Remark1. ForpassivityanalysisofopenCRNsasimilarapproach was taken in [32]: the input was chosen proportional to the positive input flow rate and the passive output was taken as (Lnc−Lnc∗)T(c−cIN),wherecINistheinletconcentration.How- ever,forcontrolpurposesitismorebeneficialtochoosetheinput u(j)asin(24)sincethecontrolcanbeimplementedbymodifying theinletconcentration.
4.2. Distributedsetpointcontrol 4.2.1. Controlproblemstatement
ConsideraprocessnetworkconsistingofCsubsystems(CRNs) thatareinterconnectedthroughadelayedconvectionnetwork.The reactionsthattakeplaceinthereactorsaredescribedbytheopen CRNmodel(24).
Controlobjective:LetthesetpointofthejthCSTRbec(j)SP that belongstotheequilibriumpointsetof(17).Designadistributed controllerforeachCRNsuchtoassurethatc(j)→c(j)SPast→∞,∀j= 1...C.
4.2.2. Localcontrolinput
Toimplementthecontrol,augmenttheinputofeachCRNwith alocalcontrolinflow(vLjc(j)L )inwhichthespecies’concentrations (c(j)L )arespecifiableandtheyareusedaslocalcontrolinputs,and vLj ∈R+istheconstantlocalinputflowrate,j=1...C;vL0=0.
As it has already been noted before, the constant volume assumptionineachCSTRimpliesthattheflowratescannotbefreely manipulated,buttheyarerelated.Therefore,thelocalinletspecies concentrations,thataredifferentfromtheoverallfeedconcentra- tionassumedtobeconstantinAssumption3,arethemanipulable localinlet concentrations. In practical schemes,the volume(or overallmass)ofeachCSTRisheldconstantonalowerlevelofa hierarchicalcontrolschemebymanipulatingtheoutputflowrates [6,33]whiletheconcentrationsarecontrolledonahigherlevel:
thatisconsideredinthispaper.
Inthecontrolledprocessnetworktheinputvectorofthejth openCRNhastheform:
u(j)=
C =0aj˜vc()(t−Tj)+aLj˜vjc(j)L −v˜jc(j) (31)
wherev˜j=vj+vLj,aLj=vLj/˜vj.
Inthecontrolledprocessnetworkthebalanceequationsinthe convectionnetwork(similartoEq.(19))readas:
C =0aj=1, j=0,...,C, (32)
˜vj=
C=0
aj˜v+aLj˜vj, j=0,...,C. (33)
Letusdistinguishintheinflowvectoru(j)theinterconnection term(i(j))andthelocalcontrolterm(u(j)L )asfollows:
u(j)=i(j)+u(j)L , (34)
i(j)=
C =0ajv˜c()(t−Tj)−(1−aLj)˜vjc(j), (35)
u(j)L =aLj˜vj
c(j)L −c(j)
. (36)
Theinterconnectiontermcontainstheflowsfrom-andtotheother openCRNs.Thevalueoflocalcontrolflowcanbespecifiedthrough c(j)L .
Fig.1.Interconnectionsbetweentheenvironmentandprocessnetwork.
Theflowsbetweenthecontrolledprocessnetworkandtheenvi- ronmentareshowninFig.1.
4.2.3. Theconceptofsynchronization-basedcontrol
Thepropertiesofthebalanceequations(32)and (33)canbe exploitedtoachieve thedistributedcontrolobjectivebytracing backtheformulatedcontrolproblemtothesynchronizationprob- lem,discussedinSection2.
Ifthesynchronizationcanbereachedintheprocessnetwork,the steady-stateoutputsoftheCRNstakethesamevalue,i.e.y(j)=y(k) ast→∞∀j,k=1...C.
Asthesupplyhasconstantconcentrationvector,i.e.c(0)=c(0)∗ itcanbeconsideredthaty(0)(t)=0,see(27).
IftheoutputsofallCRNsaresynchronized(y(j)=y(i) ∀i,j),it yieldsthaty(j)=y(0)=0 ∀j∈N0,i.e.c(j)=c(j)∗ ∀j∈N0.More- over,ifAssumption4holds,y(j)=0,i.e.c(j)=c(j)∗ ∀j=1...C.
Accordingtothesynchronizationprotocol(8)theinputu(j)of thesubsystemhastobedesignedasafunctionofthepassiveoutput (y(j))ofthesubsystemandtheneighboringsubsystems.
However,intheprocessnetworktheinterconnectioninput(i(j)) dependsonthestatevectors(c(j)),seeEq.(35).
Definethedesiredinputvectoru(j)y oftheCRNsasafunctionof thepassiveoutputs:
u(j)y =
C=0
ajv˜y()(t−Tj)−(1−aLj)˜vjy(j). (37)
Herey(j)=
Lnc(j)−Lnc(j)SP
wherec(j)SPistheprescribedequilib- riumpoint(setpoint)∀j=1...C.
Thelocalcontrolinputsu(j)L oftheCRNsin(34)havetobefor- mulatedsuchthattheycompensatethedifferencesbetweenthe desiredinterconnection(necessarytoreachthesynchronization withtheprescribedsetpoint)andtherealinterconnections.
Fromtherelations(35),(36),(37)wecancomputetheexplicit formofthelocalcontrolinputwhichsatisfiesu(j)=i(j)+u(j)L =u(j)y :
c(j)L =y(j)+ 1 aLj˜vj
C=0
aj˜v
y()(t−Tj)−c()(t−Tj)
−v˜j
y(j)−c(j). (38)
It is visible from (38) that the computation of the control law of each subsystem requires the knowledge of the outputs oftheneighboring(connected)subsystems.Thisnecessitatesthe localcommunicationofthecontrollersinastraightforwardimple- mentation.Therefore,theproposedapproachcanbeclassifiedas distributedcontroldesignasitisdefinede.g.in[7,34].
4.2.4. Stabilityandsteady-statesoftheclosedloopsystem
Theorem1. ConsiderasystemofinterconnectedCRNsinwhicheach subsystemmodelisgivenby(24),(34),(35),(36)andtheintercon- nectionparametersaredefinedasin(32)and(33).IfAssumptions 1–4holdandc(j)L in(36)canbechosensuchthatu(j)=u(j)y element- wise∀j=1...C,thenc(j)(t)isboundedfor t≥0andc(j)→c(j)SP as t→∞∀j>0.
Proof. BasedonthestoragefunctionS(j),givenin(25),definethe followingLyapunov-Krasovskiifunctional:
S=2
C j=0S(j)+
C j=0 C =0aj˜v
tt−Tj
y()Ty()d. (39)
Thetimederivativeofitreadsas S˙=2
C j=0S˙(j)+
C j=0 C=0
aj˜v
y()Ty()−y()T(t−Tj)y()(t−Tj)
.
(40)
ByLemma1ityieldsthat ˙S(j)≤y(j)Tu(j) ∀j.Accordingly:
S˙≤2
C j=0y(j)Tu(j)+
C j=0 C=0
aj˜v
y()Ty()
−y()T(t−Tj)y()(t−Tj)
. (41)
By choosing c(j)L such that u(j)=i(j)+aLjv˜j
c(j)L −c(j)
=u(j)y ∀j=1...C,andsincey(0)=0,ityields:
S˙≤2
C j=0y(j)T
C=0
ajv˜y()(t−Tj)−(1−aLj)˜vjy(j)
(42)
+
C j=0 C =0aj˜v
y()Ty()−y()T(t−Tj)y()(t−Tj).
Asv˜j=aLj˜vj+
C=0aj˜v=
C=0ajv˜j (the sum of inflows is equaltothesumoftheoutflows),itresultsthat
S˙≤
C j=0 C=0
aj˜v
−y(j)Ty(j)+2y(j)Ty()(t−Tj)
−y()T(t−Tj)y()(t−Tj)
,
S˙≤−
C j=0 C =0ajv˜
y(j)−y()(t−Tj)
Ty(j)−y()(t−Tj)
≤0.
(43)
Letusintroducethenotation
e(j,)i (t)=yi(j)(t)−y()i (t−Tj). (44) As ˙S≤0 it yieldsthat S(∞)=limt→∞S(t)<∞ for finite S(0).Accordingly,by(43)ityieldsthat:
C j=0 C =0ajv˜
K i=1 ∞t=0
y(j)i ()−yi()(−Tj)
2d≤S(0)
−S(∞)<∞. (45)
Hence,e(j,)i ∈L2 ∀i,j,.
SinceS(t)<∞andcSPi(j)isafinite,strictlypositiveconstant∀i,j, ityieldsthatc(j)i andconsequentlye(j,)i ,y(j)i ∈L∞ ∀i,j,.Hence,all theentriesofc(j)arebounded∀j.
c(j)i ∈L∞andy(j)i ∈L∞impliesthatu(j)i ∈L∞.ByEq.(24)itcan beseenthat ˙ci(j) ∈L∞ ∀i,j.Thetimederivativeofy(j)i readsas ˙y(j)i =
˙
ci(j)/c(j)i .ByAssumption2ityieldsthat ˙yi(j) ∈L∞.Hence ˙e(j,)i ∈ L∞ ∀i,j,.
Ase(j,)i ∈L2,e(j,)i ∈L∞and ˙e(j,)i ∈L∞,byBarbalat’slemma,it yieldsthatlimt→∞e(j,)i =0∀i,j,.
SinceTjarefinite,limt→∞yi(j)(t)=limt→∞y()i (t), ∀i,j,.From Assumption3itresultsthaty(0)i =0.Hence,byAssumption4,it yieldsthatlimt→∞ci(j)(t)=ciSP(j) ∀i,j.䊐
It is importanttonotethat S in (39)and thelocalcontrol input(38)areindependentofthekineticparametersoftheCRN subsystems.
4.3. Distributedsetpointcontrolwithdisturbanceattenuation Thecontrollaw,presentedintheprevioussubsections,requires theknowledgeoftheconvectionnetwork parameters.Theaug- mented controlproposedin this subsectionfollows theideaof high-gaincontrolmethodstoattenuatetheeffectsofunmodelled disturbancesanduncertaintiesintheinterconnectionsonthecon- trolperformances:instablecontrolloops,withsufficientlyhigh feedbackgaintheeffectoftheboundeddisturbancesonsteady- stateperformancescanbemadearbitrarilysmall(seee.g.[35]).
ConsiderthattheCRNsintheinterconnectedsystem,originally describedbythemodel(24),aresubjecttoadditivedisturbances, i.e.theycanbemodelledas
dc(j)
dt =Y(j)A(j)ϕ(j)(c(j))+ 1
Vju(j)+d(j), j=1...C (46) wherethedisturbanceinputd(j)(t)∈RK.
It isimportant tonoticethat theabovegeneralformis also derivedfromEq.(18)byconsideringthedifferencebetweenthe convectivecomponentmassin-andoutflowtermasaninput,and byseparatingtheinflowtermoriginatingfromtheenvironmentas thedisturbance,i.e.d(j)=˛0jv0c(0).Byassumingthedisturbance changingintime,wehaverelaxedtheconstantenvironmentcon- centrationconditionsinAssumption3inordertodesignacontroller thatrejectsitseffectonthecontrolledoutput.
Notethatthepassivityproperty,discussedinLemma1,ispre- servedfromd(j)toy(j)whenu(j)=0.Itisbecausethedisturbance inputvectorhasthesamedimensionandthesamelinearinput structureasthelocalcontrolinputvector.
Assumption5. Thedisturbanceinputd(j)iscontinuousand
d(j)2≤d(j)M (47)
wheredM(j) ∈R+isafiniteconstant.
Definetheinputandoutputvectorsofallthesubsystems
u=(u(1)T...u(C)T)T, (48)
d=(d(1)T...d(C)T)T, (49)
y=(y(1)T...y(C)T)T. (50)
ByAssumption5ityields
d2≤dM, where dM=
C
j=1
d(j)M
2. (51)
Fig.2.ExampleofcontrolledinterconnectedCRNs.
Toattenuatetheeffect ofdisturbances,augmentthedesired input,originallygivenin(37),as
u(j)d =
C =0aj˜vy()(t−Tj)−(1−aLj)˜vjy(j)−
2y(j) (52)
where ∈R+istheconstant,finitecontrolgain.
Theorem2. ConsiderasystemofinterconnectedCRNsinwhicheach subsystemmodelisgivenby(46),(34),(35),(36)andtheinterconnec- tionparametersaredefinedasin(32)and(33).IfAssumptions1–5 holdandc(j)L in(36)canbechosensuchthatu(j)=u(j)d element-wise
∀j=1...Cwith >1+dM
ε , 0<ε<∞, (53)
thenyconvergestowardtheset{y|y2≤ε}.
Proof. ConsidertheLyapunovfunctioncandidateSgivenin(39).
WiththeCRNmodel(46),byfollowingsimilarargumentsasin theproofofTheorem1,thetimederivativeofSreadsas:
S˙≤−
C j=0 C =0ajv˜e(j,)Te(j,)−y22+2yTd (54)
wheree(j,)=y(j)(t)−y()(t−Tj).
By applying that yTd≤|y|T|d|≤(y22+d22)/2, and by Assumption5ityields:
S˙≤−
C j=0 C=0
ajv˜e(j,)Te(j,)
+(dM+(−1)y2)(dM−(−1)y2). (55)
Since >1+dεM, it yields that dM+(+1)y2>0 and, if y2>ε, dM−(−1)y2<0. Accordingly, ˙S<0 i.e S is a decreasingstoragefunctionify2>ε.
By(39)thedecreaseofSinvolvesthedecreaseofS(j) orthe decreaseof
tt−Tjy()Ty()dterms.
AsS(j)=0iff y(j)i =0,i=1...K,j=1...C thedecreaseof S(j) involvesthatsomey(j)i tendto0.
As Tj are finite and positive ∀,j, the decrease of
tt−Tjy()Ty()dalsoinvolvesthatsomey(j)i tendto0.
Theconvergenceofy(j)i towardszeropersistsuntily2≥ε,i.e.
untiltherelation(53)issatisfied.䊐
4.4. Restrictiononconvectionflowrates:positivity
During thecontrol design thephysical meaning of thecon- trolinputs,thatareconcentrations,shouldbetakenintoaccount.
Thisimplies,thatthecontrolshouldalwaysbepositive,i.e.c(j)L ≥0 element-wise∀j=1...C.Thesepositivityconditionsimplyrestric- tionsontheachievablesetpointsthatwillbederivedbelow.
Considertheequationsusedforcontroldesignu(j)=i(j)+u(j)L = u(j)y,andi(j)+u(j)L =u(j)d respectively.Insteady-stateu(j)y =u(j)d =0 sincey(j)=0∀j=1...C.Hence,thesteady-statevalueofthecontrol is
c(j)L ∗= 1 aLj˜vj(−i
(j)
SP+aLjv˜jc(j)SP)= 1 aLjv˜j
−
C =0aj˜vc()SP+˜vjc(j)SP
. (56)
Thepositivenessofthecontrolinputinsteadystateisassured ifthefollowinginequalitiesholdelement-wise:
C =0aj˜vc()SP ≤˜vjc(j)SP, ∀j,=1...C. (57)
Theaboveinequalityexpressesthecomponentmassconservation conditions,similarlytothosefortheoverallmassinEq.(33).Eq.
(57)putsaconditionontheachievablesetpointinthejthCSTR dependingonthesetpointintheCSTRsconnectedtoitsincoming flows.
Therelation(56)canalsobeappliedtoderiverelationsamong thesetpoints of thecontrolledsubsystem and thesteady state upperboundofthecontrolinputs.
5. Acasestudy
5.1. InterconnectedCRNmodelforsimulations
AninterconnectedCRNnetworkwasconsideredconsistingof threedifferentCRNsandtheenvironment.TheCRNsinthethree subsystemswerechosenas
CRN1: R11:2C→312A→122B, R12:2C32
[23]2B. (58)
CRN2:R2: 2A12
[21]2B. (59)
CRN3:R3: A+C45
[54]
B. (60)
Itis easytoseethat allthreeCRNsarereversibleorweakly reversibledeficiencyzeronetworkswithasinglegraphcomponent (linkageclass).Therefore,basedon[28],Assumptions1and2(com- plexbalanceandthuspersistenceofthedynamics)arevalidfor themindependentlyoftheirreactionratecoefficients.Itisimpor- tanttoremarkthatcomplexbalanceisoftenfulfilledinpracticefor thermodynamicalreasons,especiallyinthecaseofpurechemical reactions[36].
For thesimulations, therate parameters werechosen tobe ij=1mol/m3/s, ∀i,j.Thereactionstakeplaceinreactorshav- ingconstantvolumes(Vj=1m3).ForeachCRNthestatevector, incorporatingtheconcentrationofspeciesA,BandC,isdefinedas c(j)=(cAcBcC)T, ∀j.
TheinterconnectionsamongtheCRNswereimplementedasit ispresentedinFig.2withthefollowingparametersv˜=0.1m3/s, a31=0.1. Thedelays in theinterconnectionswerechosen Tij= 10s, i,j=1,2,3.
Theflowratesintheinterconnectionsbetweentheenvironment andtheprocessnetworkare: fromthesupplytoCRN1 theflow rateis(1−a31)˜v;thecontrolflowratesoftheCRNsarevL1,vL2, vL3;theproductflowratefromtheCRN3 totheenvironmentis
˜vP=(1−a31)˜v+vL1+vL2+vL3.
TheoutflowsoftheopenCRNsare:v˜1=v˜+vL1,v˜2=v˜1+vL2,
˜v3=˜v2+vL3.
Theconstantsupplyconcentrationwaschosen c(0)=(0.02 0.02 0.02)Tmol/m3.
TheinitialstatesoftheCRNswerechosenas c(j)(0)=(0.05 0.1 0.15)Tmol/m3,j=1,2,3.
Thesetpointswerechosenas:
c(1)SP =(√ 0.016 √
0.032 √
0.016)Tmol/m3, c(2)SP =(0.180.18√
0.016)Tmol/m3, c(3)SP =(0.875 0.8752 0.875)Tmol/m3.
5.2. Simulationexperiments 5.2.1. Nocontrolcase
ThesimulationswereperformedinMatlab/Simulinkenviron- ment.First,thebehavioroftheCRNnetworkwasexaminedwithout control(u(j)L =0, j=1,2,3).Fig.3showsthetrajectoriesforthis case.ThebehaviorsoftheinterconnectedCRNsdependboth on thedynamicsoftheindividualchemicalreactionsandontheflows throughtheinterconnectionsamongthereactors.Thetrajectories oftheuncontrolledsubsystemsinterconnectedwiththeconvection networkconvergetoinitialvaluedependentequilibriumstates.
Fig.3.Simulationresults–interconnectedCRNswithoutcontrol.
Fig.4. Simulationresults–interconnectedCRNswithdistributedsetpointcontrol.
5.2.2. Distributedsetpointcontrol
Second,thedistributedcontrolwascomputedandimplemented based on the relation u(j)L =u(j)y −i(j) with vLj=0.01m3/s j= 1,2,3,seethecontrolsignaldefinedin(38).Duringthissimulation experiment,nodisturbancewasconsidered.Thecontrolledtrajec- toriesoftheCRNsarepresentedinFig.4.Fig.4showsthat,withthe proposedcontrol,theprescribedsetpointsarereached.Theoutput calibrationconstantin(27)waschosen =0.8mol/m3.
Theparameter ,introducedintheoutputEq.(27),actsasagain inthecontroller.Itseffectonthecontroltransientperformances ispresentedinFig.5.Thisexamplepresentsthetrajectoryofthe controlledconcentrationc(1)A fordifferentvaluesof .
5.2.3. Comparisonwithamodelpredictivecontroller
ForthesameinterconnectedprocessnetworkaModelPredic- tiveController(MPC) wasdesigned usingthe MPC Controller Simulinkblock of MatlabMPC Toolbox [37]. Thisperforms the linearizationaroundthesetpoint,andthediscretizationofthelin- earizedmodel.Thesamplingperiodfordiscretizationwaschosen 0.1s.TheMPCblocksolvesaconstrainedoptimizationproblemin eachsamplingperiodtoobtainthecontrolinput,see[38].Allthe outputweightsandinputcontrolrateweightsofthecostfunction werechosen0.1.Thelowerboundconstraintsforalltheoutputs andcontrolinputswerechosen0.
Fig.5.Simulationresults–effectof ondistributedsetpointcontrol.
The simulation results withthe MPC is presented in Fig. 6.
By comparing the simulations results of the proposed control withtheMPC,itcanbeaffirmedthatsimilarsettlingtimesand trackingcontrolperformancescanbeobtained.However,forthe implementationoftheproposeddistributedcontrolalgorithmonly the convection network parameters should be known. For the implementationoftheMPC,besideit,theparametersoftheCRN subsystemsshouldalsobeknown.
Fig.6.Simulationresults–interconnectedCRNswithMPC.