ContentslistsavailableatScienceDirect
Journal
of
Process
Control
jo u r n al h om ep ag e :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
Indefinite
linear
MPC
and
approximated
economic
MPC
for
nonlinear
systems
夽
Mario
Zanon
a,∗,
Sébastien
Gros
b,
Moritz
Diehl
a,caDepartmentESAT-STADIUS/OPTECKULeuvenUniversity,KasteelparkArenberg10,3001Leuven,Belgium
bDepartmentofSignalsandSystems,ChalmersUniversityofTechnology,Horsalsvagen11,SE-41296Goteborg,Sweden cDepartmentofMicrosystemsEngineeringIMTEK,UniversityofFreiburg,Georges-Koehler-Allee102,79110Freiburg,Germany
a
r
t
i
c
l
e
i
n
f
o
Articlehistory:
Received16January2014
Receivedinrevisedform30April2014 Accepted30April2014
Availableonline28May2014 Keywords:
EconomicMPC LQR
StabilitytheoryforMPC
a
b
s
t
r
a
c
t
Thestabilityproofforeconomicmodelpredictivecontrol(MPC)reliesonstrictdissipativity,whichisin generalhardtocheck.Inthiscontribution,wewillfirstanalyzethelinearquadraticcasewithpossibly indefinitecost.Theconditionsforstabilitywillberecalledanditwillbeshownthateverystabilizing LQR/MPChasanequivalentpositivedefiniteLQR/MPCwhichyieldsthesameopenloopandclosedloop behavior.ThisanalysiswillthenbeusedtoformulateanapproximatednonlineareconomicMPCscheme, whichhasstabilityguarantees.Anexamplewillbeusedtoillustratetheproposedtechniqueandshow itspotentialintermsofperformance.
©2014ElsevierLtd.Allrightsreserved.
1. Introduction
Economicmodelpredictivecontrol(MPC)hasrecentlygained popularity,asitdirectlyoptimizesagivenperformanceindex,as opposedtotrackingMPC,whichminimizesthedeviationfroma givenreference.ThemainadvantageofeconomicMPCover track-ingMPCbecomesobviousintransients,whenthesystemissteered tosteadystatesoastominimizethegivenperformanceindex.
ThestabilitytheoryofeconomicMPCinitiallyconsidered lin-earsystemsand convexobjectives[1,2].For nonlinearsystems, ananalysisofaverageperformanceboundswasproposedin[3]
andaverageconstraintswereconsideredin[4].Lyapunovstability ofeconomicMPCwasfirstprovenin[5]underastrongduality assumptionand generalized in [6,7] under a strictdissipativity assumption.Thenecessityofstrictdissipativityforoptimal steady-stateoperationhasbeenanalyzedin[11,12].Astabilityproofin theabsenceofterminalconstraintsisprovidedin[8].The exten-sionofthestabilityresultstoperiodicsystemshasbeenconsidered
夽 ThisresearchwassupportedbyResearchCouncilKUL:PFV/10/002Optimization inEngineeringCenterOPTEC,GOA/10/09MaNetandGOA/10/11Globalreal-time optimalcontrolofautonomousrobotsandmechatronicsystems.Flemish Govern-ment:FWO:PhD/postdocgrants;IWT:PhDGrants,projects:EurostarsSMART; BelgianFederalSciencePolicyOffice:IUAPP7(DYSCO,Dynamicalsystems,control andoptimization,2012–2017);EU:FP7-SADCO(MCITN-264735),FP7-TEMPO(MC ITN-607957),ERCHIGHWIND(259166).
∗ Correspondingauthor.Tel.:+4915756296036.
E-mailaddress:mario.zanon@imtek.uni-freiburg.de(M.Zanon).
in[9,10].EconomicMPCschemesbasedonLyapunovtechniques havebeenproposedin[13–15].
Inthenonlinearcase,thestrictdissipativityconditioncanbe extremelyhardtocheck,thusmakingitdifficulttoensurestability. Thiscontributionfocusesonthelinearquadraticcasewithpossibly indefinitestagecosts.Theanalysisbecomessimplerinthiscaseand theconditionsthatensurestabilityfortheLQRcontrollerhavebeen studiedin[16,17].
Themaincontributionsofthispaperare(a)theanalysisof lin-earquadraticeconomicMPCand(b)theconstructionofatracking nonlinearMPC(NMPC)schemewhichcloselymimicstheeconomic NMPC behavior while providing stability guarantees. The main resultsforthelinearquadraticcasearegiveninTheorems1and 2,whichstatethatanypossiblyindefinitestabilizingLQRorlinear quadraticMPCcanbereformulatedasapositivedefiniteonethat deliversthesametrajectory.Thisresultisthenusedforeconomic NMPC,toconstructanapproximatedeconomicNMPCschemeof trackingtypewhichprovidesstabilityguarantees.
Thepaperisstructuredasfollows.Section2proposesa motiva-tionalexamplewhichshowshowtrackingNMPCcanyieldnearly thesameperformanceaseconomicNMPC.InSection3,the con-ditionsforhavingastabilizingLQRarerecalled.Section4shows that,givenanyindefinitestabilizingLQR,thereexistsapositive definite LQRwhichyieldsthesame closedloop trajectory. Sec-tion5istheanalogueofSection4forthecaseofafinitehorizon (MPC).InSection6alinearexampleisproposedtoillustratethe theory.AnonlinearexamplewellknownintheeconomicNMPC literatureisproposedinSection7toillustratehowtheproposed http://dx.doi.org/10.1016/j.jprocont.2014.04.023
0 5 10 15 20 25 30 35 40 45 50 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 c A t 0 5 10 15 20 25 30 35 40 45 50 1 2 3 4 5 6 7 q t
Fig.1. StateandinputprofilesfortheeconomicNMPCscheme(blackline),trackingscheme0(blueline)andtrackingscheme1(redline).(Forinterpretationofthereferences tocolorinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.)
frameworkcanbeusedforapproximatedeconomicNMPCwith stabilityguaranteesinthenonlinearcase.
2. Motivationalexample
Asamotivationtothisarticle,wewouldliketostartwithan examplethatshowshowatrackingNMPCformulationcanyielda goodperformancecomparedtoaneconomicNMPCformulation. We propose the following CSTRexample from [5,7]. It consid-ersasinglefirst-order,irreversiblechemicalreactionA→Bwith reactionratekrcA,wherekr=0.4l/(molmin)istherateconstant. QuantitiescAandcBdenotethemolarconcentrationsofAandB respectively.Theprocessdynamicsaregivenby
˙ cA= q
VR(cAf −cA)−krcA (1)
˙
cB=VRq(cBf−cB)+krcA (2)
wherecAf=1mol/l,cBf=0mol/ldenotethefeedconcentrationsof Aand Brespectively and VR=10lis the volumeof thereactor. Here,theflowthroughthereactorqisacontrolvariable.Inorder toensurelocalcontrollabilityatsteadystate,letusassumethat cA+cB=1.Thestatesarethusdefinedasx=cAandthecontrolsas u=q.Moreover,theflowrateisupperboundedbyq≤20l/min.
Theeconomicstagecostisgivenby
(x,u)=−(2q(1−cA)−0.5q)+0.1(q−4)2, (3) whichgivesthesteadystatexs=cs
A=0.5mol/landqs=4l/min. Consideranalternativetrackingformulationwherethestage costispositivedefiniteanddefinedas
tr(x,u)=
cA−cs A q−qs T H cA−cs A q−qs , ∈{0,1} (4) H0= 14.227 0.825 0.825 0.066 , H1= 1 0 0 1 (5)Asimulationhasbeenruninordertocomparetheeconomic performanceofthetrackingNMPCschemewith=0andthe eco-nomicNMPC scheme.Forcompleteness,alsotheperformancea morestandardtrackingschemewith=1hasbeencompared.All simulationsareruninPythonusingCasadi[18]todiscretizethe optimalcontrolproblem(OCP)andIpopt[19]tosolvethe nonlin-earprogrammingproblem(NLP).Thecontinuoustimedynamics havebeendiscretizedwithasamplingtime Ts=0.5minusinga fixedstepsizeRunge-Kuttaintegratoroforder4.Thechosen pre-dictionhorizonisN=100andtheendpointconstraintxN=xshas beenintroduced.
The proposed scenario lasts 50s and considers a system at steadystatethatisperturbedevery12.5swiththefollowing per-turbationd=[0.4,−0.4,0.2,−0.2]whichaltersthestateasfollows: cA(k12.5)=cA(k12.5)+d(k),k=1,2,3,4.ThemetricGhasbeen pro-posedin[20]tomeasurethegainobtainedbyusingeconomicMPC overtracking:
G=100Peco−P tr
Ps , =0,1, (6)
wherePeco,Ptr,Psdenotetheaverageprofitobtainedbyeconomic NMPC,trackingNMPCandsteadystateoperationrespectively.
ThesimulationresultsyieldG0=0.0009andG1=1.139,which suggestthatscheme0hasanegligibleperformancelosscompared totheeconomicNMPCscheme,whilescheme1hasasmall per-formancelosswhichishoweversensiblyhigherthantheoneof scheme0.ThestateandinputtrajectoriesaredisplayedinFig.1, whereitcanbeseenthatscheme0deliversaninputtrajectory whichisveryclosetotheonedeliveredbyeconomicNMPC,while thestatetrajectoryisevencloser.Scheme1,ontheotherhand, deliverssignificantlydifferenttrajectories.
Proving stability for economic NMPC schemes can be very involved.Moreover,trackingNMPCproblemstypicallyhavealeast squaresstructurewhichcanbeexploitedinefficientalgorithms.For economicNMPC,onthecontrary,suchstructureisnotpresentand thecomplexityoftheproblemtendstobesignificantlyhigher.This motivatesourinterestintrackingNMPCformulationsthatmimic economicNMPCbehavior.
3. ExistenceofastabilizingLQR
Inthissectionwewillfocusonlinearquadraticdiscretetime infinite horizon OCPs. The linear quadratic regulator (LQR) is definedbythefollowingOCP
V(x0)=inf x,u ∞
k=0 l(xk,uk) (7a) s.t. x0=x0, (7b) xk+1=Axk+Buk, k≥0, (7c) lim k→∞xk=0, (7d)withstatex=[x0,x1,...]andcontrolu=[u0,u1,...].Thestagecost isdefinedasl(xk,uk)=
xk uk T H xk uk ,withmatrixH= Q ST S R symmetricandpossiblyindefinite.Thisoptimizationproblemhas beenwidelystudiedintheliteratureandexcellentanalysescanbe foundin[17,21]fordiscretetimesystemsandin[16]forcontinuous timesystems.Thecontributionsabovedefinea setofnecessaryconditions fortheboundednessofproblem(3)andestablishthelinktothe existenceof solutionstothealgebraic Riccatiequation.In [16], moreover,theconnectionismadebetweenthealgebraicRiccati equation,alinearmatrixinequality,afrequencyconditionanda dissipativityconditionandthepropertiesofasetofoptimalcontrol problems.Mostoftheresultsforcontinuoustimecanbedirectly translatedtodiscrete time.Inthissection,wewillbrieflyrecall theconditionsfortheexistenceofastabilizingLQRwithapossibly indefinitestagecost.
ItiswellknownthatthesolutiontotheLQRproblem(3)can beobtainedbycomputingthestabilizingsolutionofthediscrete algebraicRiccatiequation(DARE),definedas
D(A,B,Q,R,S):={(P,K)| Q+ATPA−P−(ST+ATPB)K=0,
K=(R+BTPB)−1(S+BTPA), (A−BK)<1},
(8)
wheredenotesthespectralradius,i.e.(A)=max{||| iseigen-valueofA}.Ingeneral,theDAREhasmanysolutionsbutonlyone, ifit exists,isstabilizingthesystem(A,B),yieldsthesolutionto theLQRproblemandthecost-to-goisthengivenbyV(x0)=xT
0Px0. Notethat,inthecaseofanindefinitecost,i.e.H / 0,alsothe cost-to-goPcanbeindefinite.
Onenecessarybutnotsufficientconditionfortheexistenceof asolutionto(8)istheinvertibilityofmatrix(R+BTPB).Inthe fol-lowing,wewillrecalltheconditionsforwhichtheLQRproblem(3) hasastabilizingsolution,whichalsoguaranteethat(R+BTPB)0 isinvertibleandtheDARE(8)doeshaveanecessarilyunique sta-bilizingsolution.
Ifstabilizabilityisassumed andanyofthefollowing equiva-lentconditionsissatisfied,thentheLQRhasauniquestabilizing solution.
• Thelinearmatrixinequality(LMI):
∃
Ps.t.Q+ATPA−P ST+ATPB S+BTPA R+BTPB
0; (9)
• Thequadraticmatrixinequality(QMI):
∃
Ps.t.Q+ATPA−P−(ST+ATPB)(R+BTPB)−1(S+BTPA)0, (10a)
R+BTPB0; (10b)
• Thefrequencydomaininequality(FDI):
∀
zs.t.|z|=1 (z):=R+S(Iz−A)−1B+BT(Iz−1−AT)−1ST+BT(Iz−1−AT)−1Q(Iz−A)−1B0; (11) • Thedissipationinequality(DIE):
∃
V:Rnx→R s.t.∀
xk,ukl(xk,uk)+V(Axk+Buk)≥V(xk), (12) withnxthestatedimension.
TheconnectionbetweentheFDIanddetectabilityof(A,Q1/2) hasbeenestablishedin[17].
Inordertoshowthattheseconditionsareindeednecessaryfor theexistenceofastabilizingLQR,letusadapttwotheoremsfrom
[16]tothecaseofdiscrete-timesystems.Thosetheoremsanalyze theboundednessofsimilarOCPsandrelateittotheconceptof dissipativity.ThisallowsonetoestablishboundednessoftheLQR OCP,thusprovingtheexistenceofastabilizingsolutiontotheDARE
(8).LetusdefinethefeasiblesetZnonthehorizonnas Zn(x0)={(x0,...,xn,u0,...,un−1)|x0=x0,
xk+1=Axk+Buk, k=0,...,n}. (13) Letusintroducethefollowingoptimalcontrolproblems(OCP) Vf+(x0)= inf (x,u)∈Z∞(x0) ∞
k=0 l(xk,uk), (14) V+(x0)= inf (x,u)∈Z∞(x0) ∞ k=0 l(xk,uk), s.t. lim k→∞xk=0, (15) V−(x0)=− inf (x,u)∈Z∞(x0) 0 k=−∞ l(xk,uk), s.t. lim k→−∞xk=0, (16) V+ N(x0)= inf (x,u)∈ZN(x0),N≥0 N k=0 l(xk,uk). (17)OptimalityimpliesthatV+
N ≤0andVN+≤Vf+≤V+.If stabilizabil-ity isassumed, thenV+
N, Vf+, V+<∞and ifcontrollabilityis assumedV−>−∞.
Theorems1and2from[16]establishtheexistenceofa stabi-lizingLQR,byprovingtheboundednessofV+.Itisinterestingto notehowboundednessofV+canbeestablishedbylookingatthe boundednessofproblemsthatcanbequitedifferentfromit.Both theoremscanbedirectlytranslatedtodiscretetimeasfollows.
Proposition1. Assumethatthesystemxk+1=Axk+Bukis control-lable,thenthefollowingconditionsareequivalent:
1
Nk=0l(xk,uk)≥0foreveryN≥0andevery(x,u)∈ZN(0); 2V−(x)≤0;3Vf+(x)≤0; 4V+
N(x)≤0;
5ThereexistsafunctionV(x)≤0whichsatisfiestheDIE(12). Moreover,wheneveranyoftheseconditionsissatisfied,then−∞≤ V−(x)≤V+
N(x)≤Vf+(x)≤V+(x)≤∞.Finally,anyfunctionV(x) satis-fyingtheDIE(12)satisfiesV−(x)≤V(x)≤V+(x)andV−(x)≤V(x)≤ V+
N(x)ifV(x)≤0.
Proposition2. Assumethatthesystemxk+1=Axk+Bukis control-lable,thenthefollowingconditionsareequivalent:
1
Nk=0l(xk,uk)≥0 for every N≥0 and every (x,u)∈ZN(0) with xN=0;2V+(x)≥−∞; 3V−(x)≤∞;
4ThereexistsafunctionV(x)whichsatisfiestheDIE(12).
Moreover,wheneveranyoftheseconditionsissatisfied,then−∞≤ V−(x)≤V(x)≤V+(x)≤∞.
Proof. Theprooffollowsthesameargumentsasin[16].
Remark1. ThefirstconditionofPropositions1and2istrivially satisfiedwhenH0.Boundednessoftheinfimaishencealways guaranteedforpositivesemidefinitecostfunctions,provided con-trollabilityisassumed.
Remark 2. Note that Condition 1 in both Propositions 1 and 2impliesthat thesystemisoptimallyoperated atsteadystate, thoughourdefinitionofoptimaloperationatsteadystatediffers slightlyfromtheonegivenin[22,11].
TheimportanceoftheDIE(12)fortheexistenceofastabilizing LQRisestablishedinProposition2.ForquadraticfunctionsV(x)= xTPx,theDIE(12)isequivalenttotheLMI(9).Thisiseasilyseenby replacingthedynamicequationsxk+1=Axk+BukintheDIE,which directlygivestheLMI.TheconditionV+>−∞isthensatisfiedifand onlyifthereexistsarealsymmetricsolutionP=PTtotheLMI(see
[16],Theorem3).TheQMIistheSchurcomplementoftheLMI,thus itisequivalenttotheLMIifR+BTPB0holds.Notealsothatthe FDIimpliestheLMIanditallowstoestablishtheboundednessof theinfimaapriori,giventhesystemmatricesandthestagecost only.
Providedthatthepair(A,B)isstabilizable,theexistenceofa necessarilyuniquesymmetricsolutionP=PTtoD(A,B,Q,R,S)such that|(A−BK)|<0isguaranteedif,forall|z|=1,theFDIissatisfied, i.e.(z)>0,see[17,21].
Remark3. Notethat,inordertochecktheLMIcondition,the matrixPthatsatisfiesitdoesnotneedtobeasolutiontotheDARE. ThisconditioncanhencebecheckedbysolvingtheLMIforP:if asolutionexists,thentheconditioncanbesatisfied,thesystem isdissipativeandastabilizingLQRexists.Onthecontrary,ifthe solutiondoesnotexist,theconditioncannotbesatisfiedandthe systemisnotdissipative,thusthesystemcannotbestabilizedusing theproposedcostfunction.Notethatthischeckcanbeformulated asaconvexproblemthatisefficientlysolvable.
Remark4. IftheLMIconditionisnotsatisfied,onemightbe inter-estedindefininganewcostforwhichtheconditionholds.Inorder tomodifytheinitialcostaslittleaspossible,onecansolvethe followingconvexproblem
min P,TT 2 (18) s.t.
Q+ATPA−P ST+ATPB S+BTPA R+BTPB +T0, (19) whereT= TQ TT S TS TRand·canbeanymatrixnorm.Thenew costisthendefinedby ˜Q=Q+TQ, ˜R=R+TR, ˜S=S+TS.
Inthefollowing section,wewillassume thatthe aforemen-tionedconditionsaresatisfied,thusthatastabilizingLQRexists.
4. EquivalentLQRformulations
InSection3,theconditionsfortheexistenceofastabilizingLQR havebeengiven.In general,thestagecostdoesnot needtobe positivedefinite,i.e.H
Ɑ
0andthesolutionPtotheDAREcanalso beindefinite.InthissectionwewillshowthatanystabilizingLQR canbereformulatedasapositivedefiniteone,i.e.P0andH0, whichyieldsthesametrajectory.Let us introduce the following notation. The DARE
D(A,B,Q,R,S) delivers a solution (P, K), where P=PT is a real symmetric matrix and K is a feedback gain. Let us define the LQR trajectory as (x0):=P(A,B,Q,R,S), i.e. the LQR prob-lem P(A,B,Q,R,S) has associated DARE (P,K)=D(A,B,Q,R,S) and delivers the closed loop trajectory (x0)=(xk)k≥0 with xk+1=(A−BK)xk. In the following, we will consider two problems equivalent if they deliver the same trajectory, i.e.
P(A1,B1,Q1,R1,S1)=P(A2,B2,Q2,R2,S2).
Definition 1. Problem P(A2,B2,Q2,R2,S2) is a positive def-inite reformulation of the possibly indefinite LQR problem
P(A1,B1,Q1,R1,S1) if
Q2 ST 2 S2 R2 0 and P(A2,B2,Q2,R2,S2)= P(A1,B1,Q1,R1,S1).Theorem1. AnystabilizingLQRproblemcanbereformulatedasa positivedefiniteLQRproblem.
Beforeprovingthetheorem,twolemmasareintroducedthatwill helpinestablishingthedesiredresult.Thefollowinglemma con-sidersthecaseofasystem(A,B)whichispre-stabilizedusinga linearfeedbackK.Itprovesthat,bysuitablymodifyingthecost,an LQRcanbedefinedforthepre-stabilizedsystem(AK,B),whichis equivalenttotheoriginalLQRdefinedfor(A,B).
Lemma1. LetusassumethatastabilizingLQRexistsforsystem(A, B),withweightingmatricesQ,R,S.TheequivalenceP(A,B,Q,R,S)=
P(AK,B,QK,R,SK)holdsforanyarbitraryfeedbackmatrixK,when choosing AK=A−BK, QK=Q−STK−KTS+KTRK, S
K=S−RK. Moreover, the real symmetric matrix PK, computed as (PK,K)=
D(AK,B,QK,R,SK),satisfiesPK=Pwith(P,K)=D(A,B,Q,R,S).The feedbackgainKKisgivenbyKK=K−K.
Proof. TheprooffollowsbysubstitutionofAK,QKandSKinto(8)
andsimplification[17].ThedetailsareprovidedinA.Notethatthe reformulationproposedinLemma1isequivalenttochoosingfor thecontrolinputuk=−Kxk+
v
k,i.e.xk+1=(A−BK)xk+Bvk=AKxk+Bvk. (20)
Remark5. Notethat,fromuk=−Kxk+
v
kandv
k=−KKxk,one obtains−Kxk=uk=−Kxk+
v
k=−Kxk−KKxk, whichalsoentailsKK=K−K.Corollary 1. By choosing K=K, the DARE reduces to QK+ AT
KPAK−P=0. Thisentails that, forany arbitrarily chosen matrix RK such that RK+BTPB0, the solution remains unchanged, i.e.
P(AK,B,QK,R,SK)=P(AK,B,QK,RK,SK).Oneobtainsthusa prob-lemwhosesolutionisKK=0and
v
k=0.Proof:TheDAREcorrespondingtoP(AK,B,QK,RK,SK)isgivenby QK+ATKPAK−P−(SKT+AKPB)(RK+BTPB)−1(SK+BTPAK)=0.
Thelasttermisalwayszero,ascanbeseenbyreplacingthe expres-sionforK: SK+BTPAK= S−RK+BTPA−BTPBK = S+BTPA−(R+BTPB)K = S+BTPA−(R+BTPB)(R+BTPB)−1(S+BTPA) = 0.
RegardlessofthechoiceofmatrixRK,theDAREsimplifiesandPis definedby
QK+AT
KPAK−P=0.
Remark6. TheconditionRK+BTPB0isrequiredfortheLQRto exist,seee.g.(3).Thisconditioncanalternativelybederivedfrom dynamicprogrammingarguments.
ThefollowinglemmastatesthattheLQRcostcanbemodified suchthatthetrajectory(x0)remainsunchangedwhilethe cost-to-gomatrixPcanbeanyarbitrarysymmetricmatrix.
Lemma2. LetusassumethatastabilizingLQRexistsforsystem(A,B), withweightingmatricesQ,R,S.GivenanyrealsymmetricmatrixP,the equivalenceP(A,B,Q,R,S)=P(A,B,QP,RP,SP)holdswhenchoosing QP=Q+ATPA−P,S
P=S+ATPBandRP=R+BTPB.Thesolution PPtoD(A,B,QP,RP,SP)isgivenbyPP=P−Pandtheresulting feed-backgainisKP=K.
Proof. BywritingtheDARED(A,B,QP,RP,SP),andreplacingthe expressionsforQP,RP,SP,oneimmediatelyobtainsD(A,B,Q,R,S), whichissolvedbymatrixPandfeedbackgainK.
BydefinitionP=P+PPandonecanwrite
0=Q+ATPA−P−(ST+ATPB)(R+BTPB)−1(S+BTPA) =Q+ATPA−P+ATP PA−PP−(ST+ATPB+ATPPB) ×(R+BTPB+BTP PB) −1 (S+BTPA+BTP PA) =QP+ATP PA−PP−(STP+ATPPB)(RP+BTPPB) −1 (SP+BTP PA) NotethatchoosingP=P,oneobtainsP
P=0.
Corollary 2. By successively applying Lemma 1 using K=K andLemma2,oneobtainsthatP(AK,B,QK,RK,SK)=P(AK,B,QK,P, RK,P,SK,P), where QK,P=QK+AT
KPAK−P, SK,P=SK+ATKPB and RK,P=RK+BTPB.Notethat,asRKcanbechosenarbitrarilylarge,also RK,Pcanbechosenarbitrarilylarge.
Usingtheselemmasandcorollaries,itiseasiertoproveTheorem 1.
Proof. (Theorem1)UsingtheequivalencegiveninCorollary2,it followsthatP(A,B,Q,R,S)=P(AK,B,QK,P,RK,P,SK,P)andKK,P=0. As we are interested in possibly indefinite but stabilizing LQR formulations, by hypothesis, the feedback gain K is stabilizing, i.e.(AK)=(A−BK)<1,wheredenotesthespectralradius,i.e. (A)=max{|||iseigenvalueofA}.ByLyapunovstabilitytheory, for anysymmetric matrix QL0, there exists a matrix PL such thatAT
KPLAK−PL+QL=0.ChoosingP=P−PL,andrecallingthe resultofCorollary1,oneobtainsAT
KPAK−P=QL−QK.ThenQK,P= QK+AT
KPAK−P=QL0.Onelaststepisneededtomakematrix HK,P=
QK,P ST K,P SK,P RK,Ppositivedefinite.ByCorollary2,onecan arbi-trarilychoosematrix RK,P arbitrarilylargewithoutchangingthe solution(x0), northecost-to-goP. Theresult HK,P0is then
obtainedbyselectingRK,PSK,PQ−1
K,PSTK,P.Whilethisreformulated problemalreadyyieldsanLQRwithpositivedefinitecost,itapplies tothe system(AK, B).By applying Lemma1 using K=−K one obtainsanLQRwithpositivedefinitecost,whichappliestothe sys-tem(A,B),i.e.P(A,B, ˜Q, ˜R, ˜S)=P(A,B,Q,R,S),where ˜Q =QK,P+ ST
K,PK+KTSK,P+KTRK,PK, ˜S =SK,P+RK,PKand ˜R =RK,P.Thiscanbe shownbyapplyingtheSchurcomplementtocheckpositive defi-niteness.Matrix ˜R ispositivedefinite,soitremainstoprovethat
˜
Q ˜STR˜−1˜S.Thisfollowsimmediatelyfromthedefinition: ˜
Q=QK,P+ST
K,PK+KTSK,P+KTRK,PK,
˜STR˜−1˜S =ST
K,PR−1K,PSK,P+SK,PT K+KTSK,P+KTRK,PK. Theresult ˜Q ˜STR˜−1˜SdirectlyfollowsfromQ
K,PSTK,PR−1K,PSK,P.
Remark7. Inordertocomputetheequivalentpositivedefinite LQRmatrices ˜Q, ˜R, ˜S onecansolvethefollowingSDP
min ˜ P, ˜Q,˜R,˜S, ˜H ˜P −I 2 + ˜H −I2 (21a) s.t. ˜H=
˜ Q ˜ST ˜S R˜ (21b) ˜ H0 (21c) ˜ P0 (21d) ˜ Q +ATPA˜ − ˜P−(˜ST+ATPB)K˜ =0, (21e) ( ˜R +BTPB)K˜ −(˜S+BTPA)˜ =0. (21f) Due toour choice of objective, theproblem is convexand, byTheorem1itisalsofeasible,thusthesolutiontothisSDPalways existsandisunique.OneadvantageofsolvingtheproposedSDPto computematrices ˜Hand ˜Pisthatthenumericalpropertiesofthose matricesareoptimalinthesenseofthechosenmatrixnormand reference.
5. EquivalentMPCformulations
In this section, we will analyze the MPC problem. We will restrictourattentiontothecaseofnoactiveinequalityconstraint atsteadystate.Notethatthisdoesnotexcludethepossibilityof havingstateandinputconstraintsintheMPCformulation,aslong astheyarenotactiveatsteadystate.WhenconsideringanMPC problem,thetimehorizonbecomesfiniteandtheproblemtobe solvedis V(x0)=inf x,u N
k=0 l(xk,uk)+xTNPNxN (22a) s.t. x0=x0, (22b) xk+1=Axk+Buk, k=0,...,N−1, (22c) with l(xk,uk)= xk uk T H xk uk , matrices H= Q ST S R and PN symmetricandpossiblyindefinite.Whennoinequalityconstraints are imposed, the solutionto this problem can be obtainedbypropagatingbackwardsthediscreteRiccatiequation(DRE)
RN(A,B,Q,R,S,PN):={(P0,P1,...,PN,K0,K1,...,KN−1) |Pk−1 =Q+ATPkA−(ST+ATPkB)Kk−1,
Kk−1 =(R+BTPkB)−1(S+BTPkA),
k=N,...,1.}. (23a)
TheclosedloopMPCtrajectorycanthenbecomputedby simulat-ingforwardintimethesystem
xk+1=AK0xk=(A−BK0)xk,
whiletheMPCpredictedtrajectorycanbecomputedbysimulating forwardintime
xk+1=AKkxk=(A−BKk)xk.
Remark8. Note that,iftheterminal costmatrix is chosenas PN=P,wherePisobtainedbysolvingtheLQR,thenthetheoryfrom Section4applies.ThissectionfocusesonthecasePN /= P.
Let us define the MPC predicted trajectory as (x0):=PN(A,B,Q,R,S,PN), i.e. the associated DRE is (P0,P1,...,PN,K0,K1,...,KN−1)=RN(A,B,Q,R,S,PN) and the closed loop trajectory is given by (x0)=(xk)0≤k≤N with xk+1=(A−BKk)xk.Inthefollowing,we willdefinetwo problems equivalent if they deliver the same predicted trajectory, i.e.
PN( ˜A, ˜B, ˜Q, ˜R, ˜S, ˜PN)=PN(A,B,Q,R,S,PN).
Definition 2. Problem PN( ˜A, ˜B, ˜Q, ˜R, ˜S, ˜PN) is a posi-tive definite reformulation of the possibly indefinite MPC problem PN(A,B,Q,R,S,PN) if
˜ Q ˜ST ˜S R˜ 0, PN˜ 0 and PN( ˜A, ˜B, ˜Q, ˜R, ˜S, ˜PN)=PN(A,B,Q,R,S,PN).Theorem1 considered thecase of an infinite horizon linear quadraticOCP(LQRproblem).Thefollowingtheoremconsidersthe caseofafinitehorizonOCPwithaterminalcostwhichisdifferent fromtheLQRcosttogo(MPCproblem).
Theorem2. EverystabilizingpossiblyindefiniteMPCschemeofthe form(5)withtimeinvariantmatricesA,B,Hcanbeformulatedasan equivalentpositivedefiniteMPCscheme.
Beforeprovingthetheorem,letusfirstintroducethefollowing lemma.
Lemma3. TwoDREsRN(A,B,Q,R,S,PN)andRN(A,B, ˜Q, ˜R, ˜S, ˜PN) yieldthesamefeedbacksequenceK0,...,KN−1if
˜
PN− ˜P=PN−P, ˜R+BTPB˜ =R+BTPB, ˜S +BTPA =˜ S+BTPA, (24) withPand ˜PcomputedbysolvingtheassociatedDAREsD(A,B,Q,R,S) andD(A,B, ˜Q, ˜R, ˜S).
Proof. Let us first consider the two equivalent LQR prob-lems P(A,B,Q,R,S)=P(A,B, ˜Q, ˜R, ˜S) with associated DAREs (P,K)=D(A,B,Q,R,S) and ( ˜P,K)=D(A,B, ˜Q, ˜R, ˜S). Defining k=Pk−P and ˜k= ˜Pk− ˜P, the two DREs can berewritten as
RN(A,B,Q,R,S,P+N)andRN(A,B, ˜Q, ˜R, ˜S, ˜P+ ˜N),i.e. P+k−1=Q+ATPA+ATkA−(ST+ATPB+ATkB)Kk−1, (25a) Kk−1=(R+BTPB+BTkB)−1(S+BTPA+BTkA), (25b) and ˜ P+ ˜k−1= ˜Q +ATPA +˜ AT˜kA−(˜ST+ATPB +˜ AT˜kB)Kk−1, (26a) Kk−1=( ˜R+BTPB˜ +BT˜kB)−1(˜S +BTPA +˜ BT˜kA). (26b)
Theidentityk= ˜kisobtainedifcondition(24)issatisfied.This canbeseenbyreplacingthevaluesofPand ˜PfromtheDAREinto Eqs.(Proof)and(Proof),whichgives
k−1=ATkA−(ST+ATPB)K−(ST+ATPB+ATkB)Kk−1, (27a) Kk−1=(R+BTPB+BTkB)−1(S+BTPA+BTkA), (27b) and ˜ k−1=ATkA˜ −(˜ST+ATPB)K˜ −(˜ST+ATPB +˜ ATkB)Kk−1,˜ (28a) Kk−1=( ˜R +BTPB +˜ BTkB)˜ −1(˜S +BTPA +˜ BTkA).˜ (28b) Byreplacingcondition(24)into(Proof)oneimmediatelyobtains
(Proof).
Remark9. Lemma3providessufficientconditionsforthe equiv-alenceoftwo DREsbut doesnot provideanyguarantee onthe existenceoftwoDREswhich satisfythoseconditions.It willbe clearintheproofofTheorem2thatthisisindeedpossible.
Proof. (Theorem2)Theprooffollowsfromcombiningtheresults ofTheorem1andLemma3.Inparticular,conditions(24)canbe easilyenforcedinTheorem1.FromCorollary1wehavethatST
K+ BTPAK=ST
K,P+BTPPAK=0,foranychoiceofmatrixP.Asmatrix RK,P canbearbitrarilychosen,onecanchooseRK,P=R+BTPB− BTP
PB.
AsinTheorem1,wecanchooseP=P−PL,toobtainPP=PL.In ordertoguaranteepositivedefiniteness,letuschooseQLarbitrarily butpositivedefiniteandscaleitbyafactor >0.TheLyapunov equationbecomes AT
KPLAK− PL+ QL=0and,subsequently, ST
K,P =− BTPLAK, RK,P =R+BTPB− BTPLB.
TheconditionforensuringpositivedefinitenessRK,PST
K,PQK,P−1SK,P thusbecomes
R+BTPB BTPLB+ (BTAT
KPLB)QL−1(BTPLAKB), (29) whichholdsfor sufficientlysmall.Thetransformationbackfrom system(AK,B)tosystem(A,B)canthenbedoneasinTheorem1.
Remark10. InthesamewayasforTheorem1,alsoforTheorem 2,theMPCreformulationcanbecomputedbysolvingthefollowing SDP min ˜ P, ˜Q,˜R,˜S, ˜H˜P−I 2 + ˜H −I2 (30a) s.t. ˜H=
˜ Q ˜ST ˜S R˜ (30b) ˜ H 0 (30c) ˜ P0 (30d) ˜ Q +ATPA˜ − ˜P−(˜ST+ATPB)K˜ =0, (30e) ( ˜R +BTPB)K˜ −(˜S +BTPA)˜ =0, (30f) ˜ R+BTPB˜ =R+BTPB, (30g) ˜S+BTPA˜ =S+BTPA. (30h)Thenewterminalcostisthengivenby ˜PN=PN−P+ ˜P.
Remark11. ThereformulationobtainedinTheorem2yieldsa convexifiedstagecostwhichisindependentofthechosenterminal cost.Itcanbeverifiedthattheresultisalsovalidforthecaseofan endpointconstraintintheMPCscheme.
6. Numericalexample
Inordertoillustratetheresultspresentedintheprevious sec-tion,letusconsiderthelinearsystem
xk+1 =Axk+Buk, A =
⎡
⎣
−0.3319−0.3393 00.7595.1250 10.5399.4245 −0.5090 0.9388 0.8864⎤
⎦
, B=⎡
⎣
0−1.3835.1060 −0.1496⎤
⎦
. ThestagecostischosenindefiniteanddefinedasH=
Q ST S R =⎡
⎢
⎣
−1.0029 −0.0896 1.1050 −0.0420 −0.0896 1.6790 −0.5762 0.2112 1.1050 −0.5762 −0.4381 −0.2832 −0.0420 0.2112 −0.2832 0.6192⎤
⎥
⎦
. (31)ToobtainapositivedefiniteequivalentLQRonecaneitherapply theproceduredevelopedintheproofofTheorem1orsolveSDP(7), whichresultsinmatrices ˜Hand ˜Pwiththefollowingproperties
( ˜HTh1) =
⎡
⎢
⎢
⎣
615.9651 1.0000 0.9516 0.0017⎤
⎥
⎥
⎦
, ( ˜HSDP) =⎡
⎢
⎢
⎣
1.3444 0.6252 0.0010 0.0010⎤
⎥
⎥
⎦
, (˜PTh1) =⎡
⎣
3310.0529.4083 1.0004⎤
⎦
, (˜PSDP) =⎡
⎣
10.0140.2345 0.0383⎤
⎦
.Inparticular,itcanbeseenthatsolvingSDP(7)yieldsbetter con-ditionedmatrices,asaneffectofthechosencostfunction,which aimsatfindingmatrices ˜Hand ˜Pasclosetotheidentityas possi-ble.Inthisexample,theFrobeniusnormwasusedintheobjective function.
Tovisualizetheresults,asimulationhasbeenrunstartingfrom x0=[1,0,0].TheresultingtrajectoryandthevalueofthethreeLQR formulationsaredisplayedinFig.2.Itcanbeseenthatall formu-lationsstabilizethesystemandtheobtainedtrajectoriescoincide. However,onlythetworeformulatedLQRyieldapositivedefinite valuefunction,whichisalsoaLyapunovfunctionforthesystem.
7. ApproximatedeconomicNMPC
EconomicNMPCconsistsinoptimizingacostwhichisnot neces-sarilypositivedefinite,thusstandardstabilityproofsdonothold. AstabilityproofforeconomicNMPChasbeenproposedin[7].It reliesontheconceptofstrictdissipativity,whichisunfortunately extremelyhardtocheckinthegeneralcase.Thismotivatesasearch forNMPCschemesthatapproximateeconomicbehavior,butfor whichitisrathereasytoprovestability.
Localinformationaboutthesystemcanbeobtainedbylooking atthesecondorderexpansionatsteady-state.Theoptimalsteady stateiscomputedbysolvingthefollowingNLP
min
x,u (x,u) s.t. x=f(x,u), (32) wherexandudenotethestateandcontrolvectorsandfunctionf(x, u)istheconsidereddiscretetimesystem.UsingwT=[xTuT],the LagrangianofProblem(32)isgivenbyL(w,)=(x,u)−T(x− f(x,u)),whereisthevectorofLagrangemultipliersassociated withtheequalityconstraints.Thesecondorderexpansionatsteady stateisthenobtainedbyusingthelinearsystem(A,B)with
A =
∂
f(x,u)∂
x ws , B=∂
f(x,u)∂
u ws ,wherexsandusdenotethestateandcontrolvectorsatsteadystate. The cost function is given by 2(x,u)=1/2(w−ws)TH(w−ws), wherematrixHisgivenby
H=
Q S ST R =∂
2 L(w,)∂
w2 ws .NotethatmatrixHisingeneralindefinite,henceanNMPCscheme using2(·,·)asstagecostisingeneralnotoftrackingtypeand stabilityisguaranteedifthesystemisdissipative.Checkingifthe linearsystemisdissipativeisstraightforward:iftheLMI,QMIorFDI conditionissatisfied,thenthesystemisdissipative.Alternatively onecancheckifthecost,rotatedasin[5]isconvex.
Ifthenonlinearsystemislocallycontrollableatsteadystatebut notdissipative,steady-stateoperationisnotoptimal[12].If nev-erthelessoneisinterestedinstabilizingthesystematsteadystate, theobjectivecanberegularizedsoastomakethesystem dissi-pative,seee.g.[5,7].Addingregularizationcanmakethesystem dissipative,butthestagecostcanneverthelessstillbeindefinite.As thesystemisdissipative,theindefiniteLQRorlinearMPCscheme willbestabilizingforthelinearsystem.AccordingtoTheorem2, onecanconvexifytheindefinitestagecosttocomputeapositive definitestagecostandtheconvexifiedLQRandMPCproblemsare thenequivalenttotheoriginalones.Thisnewpositivedefinitestage costcanthenbeusedinanNMPCcontexttostabilizethenonlinear plant.Asthecostisquadraticpositivedefinite,theNMPCschemeis oftrackingtype.TheclassicalNMPCstabilityproofsarethen appli-cable[23].Moreover,thestagecostisasecondorderapproximation ofthenonlineareconomicstagecost.
Notethattheconvexificationprocedureproposedinthispaper reliesonasecond orderapproximationoftheeconomicNMPC, whichfailstoexistifthesystemisnotlocallystabilizableatsteady state,i.e.if(A,B)isnotstabilizable.Techniquestoovercomethis limitationarethesubjectofongoingresearch.
7.1. Numericalexample
Inordertoillustratetheseconcepts,letuslookattheCSTR exam-plefrom[5,7].AnalogouslyasinSection2,wewillassumethat cA+cB=1inordertohavelocalstabilizabilityatsteadystate.The realeconomicstagecostis
(x,u)=−(2q(1−cA)−0.5q), (33) andthequadraticterm (q−4)2with =0.1isaddedinorderto makethesystemdissipative.WhileinSection2weconsidered eco-nomicNMPCwiththeregularizedstagecost,letusnowconsider theoriginaleconomicstagecost(33).
TheoptimalsteadystatecomputedbysolvingProblem(32)is cA=0.5mol/Landq=4L/min.Thesecondorderexpansionatsteady stategives A =0.6757, B=0.0203, H =
Q S ST R = 0 1.1832 1.1832 −0.0229 .As anticipated, this system is not dissipative, hence, the LMIcondition (9) cannot besatisfied. In [5], theregularization 0.505(cA+cB+q) is added to the cost. In [7], the regularization (q−4)2isaddedtothecostwith =0.1.Intheframework pro-posedinthispaper,theregularizationcanbechosenastheminimal one which is needed in order to satisfy the LMI(9) using the approachproposedinRemark4.Choosingtorestrictthe regular-izationtotheform (q−4)2asproposedin[7]andsolvingSDP(7), oneobtainsthattheneededregularizationforthelinearsystemis =0.0275.Notethatthisvalueisonlyvalidforthelinearsystem anddoesnotguaranteestabilityoftheeconomicNMPCschemea
0 10 20 30 40 50 60 −1 0 1 x1 0 10 20 30 40 50 60 −0.5 0 0.5 x2 0 10 20 30 40 50 60 −1 0 1 x3 t 0 10 20 30 40 50 60 −2 −1 0 VIndef 0 10 20 30 40 50 60 0 0.2 0.4 VSDP 0 10 20 30 40 50 60 0 5 10 VTh3 t
Fig.2.TrajectoryandvaluefunctionforthethreeLQRformulations.
priori.ItisinsteadguaranteedthatthetrackingNMPCusingthe reg-ularizedandconvexifiedcostisstabilizingforthenonlinearplant. ByconvexifyingthecostintheframeworkofTheorem2usingthe minimalregularization =0.0275,oneobtains
H2 =
25.0142 0.5528 0.5528 0.0132 .Inthefollowing,wewillcallscheme2theNMPCschemewhich usesatrackingstagecostdefinedbymatrixH2andregularized eco-nomicNMPCtheeconomicNMPCschemewhichusesthestagecost proposedin[7],i.e.cost(33)regularizedwith0.1(q−4)2.Using =0.1,onewouldobtainmatrixH0usedinSection2.Thisexplains whyscheme0yieldsstateandinputprofilesthatcloselymimicthe onesobtainedwiththeregularizedeconomicNMPC(cf.Section2
andFig.1).
ThetrackingNMPCscheme2hasbeencomparedtothe reg-ularized economic NMPC formulation in the same scenario as in Section2. In this second example, though, gain G is com-putedusing theaverageprofit ofeach schemecomputedusing the actual stage cost (33). For the tracking scheme that uses
H2,oneobtainsG2=−8.0599,whichindicatesthattheeconomic NMPC formulation, regularized with the term 0.1(q−4)2 per-forms worse than the tracking NMPC formulation which uses H2.ForthetrackingimplementationwhichusesH1 instead,one obtainsG1=2.1907,whichindicatesthattheregularizedeconomic NMPCformulationperformsbetterthanthissecondtrackingNMPC formulation.
ThestateandinputprofilesaredisplayedinFig.3,whereitcan beseenthatthetrackingscheme2outputstrajectoriesthatare sig-nificantlydifferentfromtheregularizedeconomicNMPCscheme. Inparticular,trackingscheme2hasamoreaggressivebehaviorand itactivatestheinputboundsinseveraloccasions.
Providedthateithertheterminalcostandconstraintare cho-senadequately[23] orthehorizon ischosen longenough[24], thetrackingNMPCschemehastheadvantageofimmediately pro-vidingstabilityguarantees.Moreover,theunderlyingOCPcanbe efficientlysolvedbynumericalalgorithmsforfastNMPC,e.g.the realtimeiteration(RTI)scheme[25]withtheGauss-Newton Hes-sianapproximation.ForeconomicNMPCinstead,theunderlying OCPisingeneralmoredifficulttosolve.
0 5 10 15 20 25 30 35 40 45 50 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 cA t 0 5 10 15 20 25 30 35 40 45 50 0 2 4 6 8 10 12 14 16 18 20 q t
Fig.3.StateandinputtrajectoriesobtainedwiththeeconomicNMPCscheme(blackline),withthetrackingscheme2(blueline)andwiththetrackingscheme1(redline). (Forinterpretationofthereferencestocolorinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.)
8. Conclusions
In this paper we have analyzed economic MPC for the lin-earquadraticcase.Boththecasesoffiniteandinfiniteprediction horizonshavebeenconsideredandithasbeenshownthatevery stabilizingLQRorMPCschemecanbereformulatedsuchthatthe costis positive definite, whilepreserving thesameclosed loop behavior.
Approximatedeconomic NMPCwithstability guaranteeshas beenproposed,basedonthepreviousresults.Theapplicationto a simple well knownexample has shown thepotential of this approachalsointermsofeconomicperformance.
Futureworkwillfocusonextendingtheproposedframework tomoregeneralsettings, includingperiodicsystemsand active constraintsatsteadystate.
AppendixA. ProofofLemma1
By replacing the definitions of AK, QK and SK in
D(AK,B,QK,R,SK),oneobtains 0=QK+AT KPKAK−PK−(STK+ATKPKB)(R+BTPKB) −1 (SK+BTP KAK) =Q−STK−KTS+KTRK+(A−BK)TP K(A−BK)−PK −(ST−KTR+(A−BK)TP KB)(R+BTPKB) −1 ×(S−RK+BTP K(A−BK))
Thelasttermofthesumcanberewrittenas (ST−KTR+(A−BK)TP KB)(R+BTPKB) −1 (S−RK+BTP K(A−BK)) =−KT(R+BTP KB)(R+BTPKB) −1
I (S−RK+BTP K(A−BK)) +(ST+ATP KB)(R+BTPKB) −1 (S−RK+BTP K(A−BK)) =−KTS+KTRK−KTBTP KA+K T BTP KBK +(ST+ATP KB)(R+BTPKB) −1 (S+BTP KA) −(ST+ATP KB)(R+BTPKB) −1 (R+BTP KB) I K =−KTS+KTRK−KTBTPKA+KTBTPKBK −STK−ATP KBK+(ST+ATPKB)(R+BTPKB) −1 (S+BTP KA). ReplacingthisexpressioninD(AK,B,QK,R,SK)andsimplifyingone obtains Q+ATP KA−PK−(ST+ATPKB)(R+BTPKB) −1 (S+BTP KA)=0, whichdirectlyentailsPK=P.TheresultKK=K−Kisobtainedasfollows KK =(R+BTPB)−1(S K+BTPAK) =(R+BTPB)−1(S−RK+BTPA−BTPBK) =(R+BTPB)−1(S+BTPA)−(R+BTPB)−1(R+BTPB)K =K−K. References
[1]J.Rawlings,D.Bonne,J.Jorgensen,A.Venkat,S.Jorgensen,Unreachable set-pointsinmodelpredictive control,IEEETrans.Autom. Control53(2008) 2209–2215.
[2]J.B.Rawlings,R.Amrit,Optimizingprocesseconomicperformanceusingmodel predictivecontrol,in:ProceedingsofNMPC08Pavia,2009.
[3]D. Angeli, R. Amrit, J. Rawlings, Receding horizon costoptimization for overlyconstrainednonlinearplants,in:ProceedingsoftheJoint48thIEEE ConferenceonDecisionandControland28thChineseControlConference, 2009.
[4]D.Angeli,J.Rawlings,Recedinghorizoncostoptimizationandcontrolfor non-linearplants,in:Proceedingsofthe8thIFACSymposiumonNonlinearControl Systems,2010.
[5]M.Diehl,R. Amrit,J. Rawlings, ALyapunovfunction foreconomic opti-mizingmodelpredictivecontrol,IEEETrans.Autom.Control56(3)(2011) 703–707.
[6]D.Angeli,R.Amrit,J.Rawlings,Enforcingconvergenceinnonlineareconomic MPC,in:Proceedingsofthe50thIEEEConferenceonDecisionandControland EuropeanControlConference(CDC-ECC),2011.
[7]R.Amrit,J.Rawlings,D.Angeli,Economicoptimizationusingmodelpredictive controlwithaterminalcost,Annu.Rev.Control35(2011)178–186.
[8]L.Grüne,Economicrecedinghorizoncontrolwithoutterminalconstraints, Automatica49(2013)725–734.
[9]R.Huang,L.T.Biegler,E.Harinath,Robuststabilityofeconomicallyoriented infinitehorizonNMPCthatincludecyclicprocesses,J.ProcessControl22(2012) 29–51.
[10]M.Zanon,S.Gros,M.Diehl,ALyapunovfunctionforperiodiceconomic opti-mizingmodelpredictivecontrol,in:Proceedingsofthe52ndConferenceon DecisionandControl(CDC),2013.
[11]M.Müller,D.Angeli,F.Allgöwer,Onconvergenceofaveragelyconstrained economicMPCandnecessityofdissipativityforoptimalsteady-stateoperation, in:ProceedingsoftheAmericanControlConference,2013.
[12]M.Müller,D.Angeli,F.Allgöwer,Onnecessityandrobustnessof dissipa-tivityineconomicmodelpredictivecontrol,UniversityofStuttgart,2013, preprint.
[13]M.Heidarinejad,J.Liu,P.D.Christofides,Economicmodelpredictivecontrol usingLyapunovtechniques:handlingasynchronous,delayedmeasurements anddistributedimplementation,in:Proceedingsofthe50thIEEE Confer-enceonDecisionandControlandEuropeanControlConference(CDC-ECC), 2011.
[14]M.Heidarinejad,J.Liu,P.D.Christofides,Economicmodelpredictivecontrolof switchednonlinearsystems,Syst.ControlLett.62(2013)77–84.
[15]M.Ellis,P.D. Christofides,Economicmodel predictivecontrol with time-varyingobjectivefunctionfornonlinearprocesssystems,AIChEJ.60(2014) 507–519.
[16]J.C.Willems,LeastsquaresstationaryoptimalcontrolandthealgebraicRiccati equation,IEEETrans.Autom.ControlAC16(6)(1971)621–634.
[17]B.P.Molinari,ThestabilizingsolutionofthediscretealgebraicRiccatiequation, IEEETrans.Autom.Control20(1975)396–399.
[18]J.Andersson,J.Åkesson,M.Diehl,CasADi–asymbolicpackageforautomatic differentiationandoptimalcontrol,in:S.Forth,P.Hovland,E.Phipps,J.Utke,A. Walther(Eds.),RecentAdvancesinAlgorithmicDifferentiation,LectureNotes inComputationalScienceandEngineering,Springer,Berlin,2012.
[19]A.Wächter,AnInteriorPointAlgorithmforLarge-ScaleNonlinearOptimization withApplicationsinProcessEngineering,CarnegieMellonUniversity,2002, Ph.D.thesis.
[20]R. Amrit, J.B. Rawlings, L.T. Biegler, Optimizing process economics online using model predictive control, Comput. Chem. Eng. 58 (2013) 334–343.
[21]A.C.M. Ran, H.L. Trentelman, Linear quadratic problems with indefinite cost for discrete time systems, SIAM J. Matrix Anal. Appl. 14 (1993) 776–797.
[22]D.Angeli, R.Amrit,J. Rawlings,Onaverageperformanceandstabilityof economicmodelpredictivecontrol,IEEETrans.Autom.Control57(2012) 1615–1626.
[23]J.Rawlings,D.Mayne,ModelPredictiveControl:TheoryandDesign,NobHill, 2009.
[24]L.Grüne,NMPCwithoutterminalconstraints,in:ProceedingsoftheIFAC Con-ferenceonNonlinearModelPredictiveControl2012,2012.
[25]M.Diehl,H.Bock,J.Schlöder,R.Findeisen,Z.Nagy,F.Allgöwer,Real-time optimization and Nonlinear Model Predictive Control of Processes gov-ernedbydifferential–algebraicequations,J.ProcessControl12(4)(2002) 577–585.