Proceedirgs oftir€ l3th SymposilmofMaldysiAnChefricalEnglnee.s DATA VALIDATION FOR REAL TIME OPTIMISATION CYCLE H. w. L€e' and A. Ahmad Laboratoryof PrccessContol, Fac lty of ChemicalandNatural ResourcesEngineering, Udveffiry TeknologiMalaysia,81310Sludai,Johor,Malaysia. 'Corespondingaurhor.Phone:+607'5535877. Fax:+607'5581463 Email:leehw_78@hotrnail.com ABSTRACT for RTO implementation, daiacollectedfrom a processplant mustbe As a prerequisite lrcated by three importanldata validationstages,i.e.. steadystate detection,data andgrosserrordetection to eliminaleerro$ andinconsistencres. rcconciliation This paper discusses thed€velopment of tbedatavalidationsiagefor realtine optimizationin a fatty acid fraciionation process.To evaluatethe proposeddala validation sysl€m,single and multiple gross errols in the fom1 of leakagesin processstream were imposed to the prccess.The resultsproved the capabilityof the proposedlool in identifyingand reciirying all the grossenors introducedto the process. Ke)ryords: Data Validation,SteadyStateDetection,Data Reconciliation, GrossElTor Detection. 1 INRODUCTION Processindustneshave beell undergoingsubstantialchangesin order to cop€ with new challenges fesultiq from high en€rgy and manpower costs, slrict safety and enviromrenral regulaiions, siringenr product specification and scarcity of reduced as well as stiffcompetitionfrom new players. Due to thesereasons, variationfeedstock real time optimisation (RTO) has been bfought forward for chemical industdes as potentialsolutionsto the increasinglyinlenseproduction challenge. Genichi et al. (1994) proposesthat the second stage jn the real iime optimisation (RTO)cycleis datavalidation.Datavalidationis consisted of threemajorcomponents, which are steadystated€tection,datareconciliationand grosserror detection. RTO moves processesfrom one steady siate operaling condilion to anothersetting that are more profiiable. Therefore,st$dy siate detectionis required to detenninethe conditionof the proc€ssand the tlme requiredfof an optimisationcycle. RTO is only implemeniedif the plant is al sleadystateandthe plant must thereforebe allowedto settle so thatthe desiredconditionjs obtained.DarbyandWlite (1988)proposed thatrhetime period betweentwo real time optimisationexecutionsmust be much longerthan the processsettling time to ensurethat the processr€tums to steady state operationbefore optimisation is conducted agaln. Measurcmentsof process variables such as flow lates, pressurcs,levels, concentations andtemperatures in a chemicalplocessare subjectto enor, bothrandom and gross. These enors are ofien umvoidable and they may foul signals during processingand lransmissionstages. Randomand grossenors can leadto measurements, deteriomtionin the performanceofihe control syst€ms. At the Limil larger eross error canevennullify gainsachievable throughprocessopiimisation.Theseenoneous datacan evendrlvethe processto unecoromicor unsafeoperatingrcgime. As a result,m€asures to eliminate lhese effects must therefore be in place to ensure the successof any optimisation tasks. Data Leconciliation and gross error detection were requir€d to accomplishlheseimportanttasks. Data reconciliation is developed to improve the accuracy of measurementsby reducing the efect of elTorin the daia. Meanwhile, thesefteasurementsafe adjustedto satisfylhe conservation laws and rhe balanceconst.ainis(e.g.energy,mass,etc.).This 3- 168 Engineers Chemical of Malavsi.n of theI8thSymposiun Proceedings lo .enove any glosseror lrom by a grosselaordetectionmechanism was accompanied 2 DATA VALIDATION 2,1STEADY STATEDETECTIONMETHOD A number of methodshave been proposedto certify lhal the required steady state conditionis alfeadyachievedpnor to the executionof the oplimisationcvcle ln this re.tprooosed re.rdri'ricdr ,nd\.r,'or.ehoasr.Fcorsde,Fd lhefif'r sas 'hecornDo' qa' rca' r'e m" hern' | hc 1/ r/ lo8or -heorlol etidence ui \ar".rrnla, 'econc "1o (1q8-) a/ Nalas]mhan bl {MTE)slrsaested "/ c de'jned I rn';per' Iendrfer rar'(dl'heoqo e\idFnLe'MTI-''r'ed I\evlf in rhefollowinge$ationl ' (1 ) (5'+s ) where ri, is the randon variableswhich obevingthe Hotteling's/ distributionwith degrcesof freedom2'V-2 A level of numeratordegreesof freedom1, anddenominator {x is chosen,and 1i:"-. (o) is calculaiedto be lhe uppervalue of the C significance, to be at sleadvslalewith respectto the distribution.lf tir:7ir,\ r, tireplant is assumed variablei. If /ir>Ii,r ,, theplantis takenasnot beingal a steadyslatecondiiion' 2.2DATA RECONCILIATION Data reconciliation is developedto improve the accuracyof measu€menlsby leduclng areadjustedio satisfythe the effectof elror in the dala Meanwhile,thesemeasurenents (eg rnass) Nonr1al1v,dala energv, conse ation laws and the balanceconstraints that reconciliaiion tvpe ofdata problen The minimisalion is a conslrained reconclliation process when only process units involved is requir€d dependson the problem and flowates are ieconciled, a linear data reconciliation probiem is applied On the other alongwith arereconciled measur€ments or pressure temp€rature hand,whencomposilion, nonlinear-data this case, requir€d In is scheme data r€conciliation nowiates,a noniinear reconciliationis requiredbecauseit is dealing wilh processnonlinearitvsuch as thermodynamicequilibrium relationshipand complex correlation for the thermodvnamlc problemcan be andotherphysicaiproperties.The generalnonlineardatareconclliation 2000): rbnnulatei asa weightealeastsquareminimisaiion (Narasimlan, Mi" v(r, i)r t-r(.v - x) (2J Subjectto f (x) =t c6)<o .ec'oro equalrl cor'LrJiflsg i( q \ \ecro ofireaualrycon\rmrrht $fere/isn\ n:,rix,.t i' n r | \elror ol mea'uea !afiable'J' i( n \ I ;. n 16r'ff6..qsrdriance ^'n measurementsof variables r and t' is n x 1 vector oi vailles of vector of m€asured of variablest of measurements weighting factor 2.3 GROSSERRORDETECTION The most comnonly usedmelhod for detectinggrosserror is statisticalhpothesis tesling a sul.(li. lor'he te'Lwith 'r kno$'roi'r ibtron jnd pe{orndrce Lh.rreoui.e,e.ec.ine q .e5lJaIr'r;Le\ceed'J cri' cal il he comprrreo ch,j'acterinic.. gioqse-or 's oec,dred the statisticalhypoihesistests others' Among value selectedfton a table of disldbuiion l- r69 Pro.e€dings oflhe l8tbSymposlun ofMalarsidChenlcllEnsimers likelihoodratio test test (NT), generalized includeglobaltest (CT), nodalor consrrain! testO{I), (GLR),boundedCenetalized likelihoodmrio (BGLR)merhod,measurements (MIMT) test iterativetneasuremenltesr(IMT) and modified ilerative measuremeni ii is suitablefor In this work, the Measwemen! Test(MT) methodwasusedbecause (3) below: RTO application.Thetestslatisticis asgivenin Equalion Equation(3) alsocanbe *ritten as: erors, tris ihe enors,er is the measurement measurenenl Here,:.! is the standafdised data and irr is the measuremenls. is the reconciled ofmeasurementsJ deviatiotr standard I' 3 PI,ANT SIMULATION C6,c,i, Cj L,C Figure 1. SchemaricdiBgralnof a fany acid fractionationplant The fatty acid fractionation (FAF) plant considercdin this paper is consistedof tl(ee of the columnsis illustratedin packeddistillationcolun s. The sequence connecting throughlhree columns, into its components fatty acids arc separated Figure L Here, clude i.e..Pre-cutColumn(PC).Light-culColumn(LC) andMiddle-cutColumn(MC). ln t\e pi€-cul column, light products(C6,Cs and Cro falry acids) are recoveredjn the overheadThe botlom strealnis then led to the lightcut columr whererhe Cr, falty acid is separated from ihe rest ofthe fatq?acids. The botlom producl ofthis colurnnthen entersthe middlecut colnn wherethe Crr fatty acid is recovercdleaving$e Cr6and Cls fatty acidsas by a dynamicsimulationmodel In this study, the FAF plant was represented purpose the FoI the ofRTO inplementarion, simulaior. developed usingHYSYS.PlantrM The ibrmer iook using the same simulalor. model llas also developed requled steadyslale conditionsif the plant the role ofthe actuaiplantwhilst the latlerprovidedthe expected was actllally at a sieadystare.At the sane tiine, the steadystatemodel neededlo provide sysrenatic search for processoptimisation especially in data reconciliation. Since the qualiiy ofboth nodels playspivotal role in producingthe jtrtended€nd-results,the models mnst thereforebe accurateandrobust. To ensurethe consistencyof the model output,rhe steadystatemodel must therefore be validated against the dynamic model. The results reveal€d lhat close agreement betweenthe modelswith an overall diifercnce of lessthan3% lvas obtainedandiherefore consideredadequarefor this study. 4 EFFf,CT OF GROSSERROR in this section,$oss elrors were createdby simulating processleak in pipeiines. Ii *as implem€ntedby le3king the flowrate for the distillale stream at 60 minutes after the processin steadystate condition. The goss elror was detectedas the disiillate flowrate was one ofthe variables in d1edala r€conciliation. The amo nts of th€ leakagewere 3-t?0 Elgrneen Chemical ofMalavsian Proceedings of thel3thSynposium highly dependenton the dislillate flowrate During implementation'carewas takennot to letthe disiillateflowrateio dry out. Theamountsofleakagewer€shownintablebelow: strerm Table L.Tfe smorntofrhe Leakage (Ks/h) 41.6519 301.1?80 t24.16-t0 MC L.akase Ihere weretwo waysof implementalionsThe first involvedcreatinga singLegross eror by leaking the distillate stream either in PC, LC oL MC colunln The second approachintroducedmlLitipl€ gross errors by lealdng all the distillate stueamsin the PC, oflhese bothcasesweresimilar,eitherin Lt anrtMC columns.The dyrlamicresponses 2 illusfates the dlnamic rcsponseofthe gross Figurc errors situation single or multiple 1othepla t gloss were introduced erroN proaess whenmuftipie ofleakageanddistillatesteams Figure2. The dlaurnicrcsponse ,1.I STEADYSTATEDETECTION were selected,ie' tbe llowrateof lhe PC distillate ln this case,6 key measurements stread, PC botlom siream, LC distillate stream,LC bottom slr€am' MC distillate stream was collectedover a pedodof30 minutes and MC bottomsiream. Eachmeasuremenl theory of evidencemethodwas mathenatical The (dynamic model) from the plant oodel was used to iest the difference Hotteling t' and the iask Lr" i"tended *"J to Ju..v o"t at 0 05% and the specified d was significance. leveL of The the two m€ans. beiween of r,l, for each 1. The value Equation iom r;.--? (a) is 47.04 as determined is lowerthanthe 7l', , (o). FAf processachievedsleadvstatecondition measurem€nt wilhin th€ period of 30 minutes. Therefore' a l0 mi11ltesor bigger cycle time can thereforeus€tuIfor RTO. 4.2 DATA RECONCILIATION It is important io reitera!€here tliat the purposeof data reconciliation in this stldy is to supportthe RTO implementationwhere measureddatamust be reconciledto satisli mass is at a tue steadystalecoMition This is doneonly whentheprocess andenergybalaaces. be and any;djustmentrequir€dto the process&!a would madeas smallas possibl€to 3- l l l Prcce.dings ofthe1StISyi'tosinmofMalaysian Chenicet Elgineeis guaranieethat lhe reconciledconditionsarc still accuraretyrepresentingrhe rfte operating condilion.Resultsofthis inplemertarionar€shownin Tabte2. Table2. Theresultofleakagestudybeibreandafrerdatareconciliation BeforeDataReconciliarion 402.E6 ]l 64.]0 t 78.92 l 952.ti Alter DataReconciliation 16:1.04 t071.77 816.03 i ri65.t 0 The valueof the objectiveftnction beforedatarcconcitiarion $.ashightydepenitent to lhe amountof the distillareiost. It was provenfrom l|e rable 2. The value of the objectivetunction was large when all d1edisti[ate steamswere ]eakedbefore data reconciliaiion is cafied our Ir is thenfollowedby the teat(age of rheLc_disrinatesteanr. It is dueto rheamountof rhedisrillaretostwerc grearerifconpared to the losrofpc and MC distillatestreans.The d-ata reconciliarion problem\l,assolved sins the Successive OrdorarrcP ogrrmn-nCr\QP, dtgo.irhmo o\ ded o\ H\ Sys ;p imi.e". Ddrd reconciliation was ableto feducerhesegrossenors and the detectionol the qrcsseffor wascaniedout by grosselTorderector. 1.3 GROSSERRORDETECTION Thesame6 measurements fiar usedin thesreadysatederedionwereadoDted in thissrudv .gr n. Tl e pla"rld. ruo f- e.J'rofrnedlrare.oncrtdr:on$ereconpdredba.edor rr; " cnticaivalue C The critical value,C was deteminedbasedon the overa[ siqnificarl! Ievel,o whichwasspecifiedas 0.05(e.g.95%ot confidentialinte al). andthe v;he was 2.8044from the standardnomal disrribulionwirh accumrtated probabilityar 0.9989.If the value test staiistic, ] z, j I exceedsthe citicat value C, then rhis measurementis said to contain gross €1ror. Ofieruise, this measufemen!is consideredftee of gross en'or. Measuremenrs containing srosserrorarenufkedunderline. LC-Le.kaee (Ke/h) PCBottom PCDistillate fC Bo$om LC Distiuate MC Bottom MC Distillate 9400.81 9i3998 r,89.91 9t00.8i 9139.98 61753 446t.71 589.91 448i .59 4904.59 6175l 4461.71 4890.82 29E2.89 l:r94.16 )961.+9 9:100.83 9186.46 622.39 4 4 3 i .i 9 4904.59 2 9 8 2 .8 9 i 4 9 4 .3 6 4890.81 2967.19 r174.77 1414.77 MC-L€akAse (Kdh) PCBoton PCDistdlJte LC Boftom lC D i s ti l l a t€ 9400.81 637.59 448i.49 4 9 0 4 .1 0 MC Bottod 2983.42 MC Distillate .lll!.04 9i86.46 622.39 4175.74 4826.07 2959.09 r47|.62 637.59 4181.49 490410 293102 4475.71 1826.07 2959.09 1.170.04 147t.62 ln thePc,leakage case,one€uor,i.e.,thepC distillateflowrarewasdeteded.On the otherhand.LC distillateflowratelvasdetecred containingelTorin the Lcleakage case. Similarly,fie lvlc distiilateflowratewasidentifiedin Mcjeakage case.On the contrary, all the distiilareslreamswere detededto containgrosserror whenthese *"i" "tthe "u-s subJected to leakage.By compadryto the threesingiegrosserror cases, distillate sireamsafter leakagewere the sameas the caseof muliiple grosserror.Alt the gross enorsin distlllatestreams wercdetected eitherin sineleor in D ltjole situarions. Tle din ll. e 01 e-ch.oll|mn srs ioer.fied.o, d,r;nggjo". enoruher rhe 'uedn leakagevalveofthe colunrnwasopened.As menrioned earti€r,whenihe leakag€vaLve 3,172 ofMalaysian ftoceedings ofde lSthSymposium ChenicaL Engineers was opened,th€ leakage strcam started to flow causing lhe redlrction in the distitlate in the distillat€ flowrat€. The effectof$oss enor causedby the leakagewas detecred 5 CONCLUSIONS Data validation is crucial to the succ€ssof the rca1 time optimisation. Steady state detection hadbe€nappliedto detectthepfocesscondirionbasedon 6 keysneasuements. hadbeenapplied Wtrenlhe processachievedits steadysiatecondition,datareconciliation in orderto tulfil massand energybalances. to adjustand reconcilethe measurcments Gross elTor detection was used to identify ihe ristlng of the goss elTor jn th€ measuf€ments.As expected,gross errof was deiectedin the distillate streamwhen the leakageof the distillationsteam was occrmed.Theseresultsexposedthe abillty of the dalar€conciliationand grossenors in the neasurements. ACKNOWLEDGMENTS This project is tundedby the Ministy of Science,Technologyand the Environment ttrougl IRPA researchgrant. Our headiest appreciaiionsare for everybody who has to thesuccess ofthisproject. directlyor indirectlycontribute REFERENCES Uni!." ofComplexProcess Da$y, M. L., andWhite,D. C. (1988)."On line Optimization Progress, 84(i0), 51 59. ChenicalEngineering Cenichi,E., Yodhinori,O., Hitoshi,M., Morinusa,O., Doughlas,B., Rosalyn,F. P., and John, S. A. (1994). "IntegratedAdvancedControl and Closed-LoopReal-Time Optimizationof an OlefinsPlant."Proceedi4 of IIAC AdvancedCootroiof Chemical Processes, Kyoto,Japan,95 100. Narasimllan,S., Chen,S. K., and Mah, R. S. H. (1987).DetectingChangesof Steady TheoryofEvidence.AICHEJaumal.33(lI): 1930 1932. StateUsingtheMalhematical Narasinrlun,S. and Haril:umar,P. (1993).'A Methodio lncoryorateBoundsin Data Problem." andGrossE1Tor Detection L TheBoundedDataReconciliatioD Reconciliation l7(11), 1115-1120. Compuie$andChemicalEngineering, An & GrossEror Detection: S. andJordache, C. (2000).Datareconciliation Na.rasinlhan, Comp3ny, Texas. Inteiligen!Us€ofprocessData.CulfPublishing 3 1,73 j