Online decision support for an evaporation network

ONLINE DECISION SUPPORT FOR AN EVAPORATION NETWORK José Luis Pi a ch, Ca los Gómez Palacín, Césa de P ada Sys ems Enginee ing and Au oma ic Con ol Depa men , EII, Uni e sidad de Valladolid. C/ Real de Bu gos s/n, 47011, Valladolid.{jose.pi a ch | ca los.gomez | p ada}@au om.u a.es Ma c Kalliski Depa men o Biochemical and Chemical Enginee ing, Technische Uni e si ä Do mund. Emil-Figge-S . 70, 44227, Do mund Ge many. Ma c.Kalliski@bci. u-do mund.de Ch is ian Jasch Lenzing Ak iengesellscha , We ks aße 2, 4860 Lenzing, Aus ia. c.j[email p o ec ed]om Abs ac This wo k p esen s a decision-suppo ool o add ess he model-based op imiza ion app oach o online load alloca ion and scheduling o cleaning ope a- ions in an e apo a ion ne wo k. The aim is imp o - ing he esou ce e iciency by supplying he op imal solu ion o a gi en p oduc ion goal. The app oach includes he semi-au oma ic upda e o e apo a o models, which is based on his o ical da a o minimal modelling e o . The s uc u e o he p oblem is o mula ed ia mixed-in ege p og amming and in e- g a ed in o he plan supe ision sys ems. P oduc ion cons ain s, conce ns abou he p ac ical implemen- a ion and isualiza ion p e e ences a e also aken in o accoun in he design o he p o o ypical ool. Key Wo ds: in eg a ion, decision suppo , isualiza- ion, su oga e model, e apo a ion ne wo k, esou ce e iciency. 1 INTRODUCTION The as changes in global ma ke condi ions and inc easing en i onmen al cons ain s o ce he p o- cess indus y o con inuously adap hei ope a ion o keep compe i i eness. In his way, an agile plan - wise op imiza ion o con inuous and disc e e deci- sions is equi ed o ope a e as e icien as possible while i ing he new condi ions [1]. To ace hese eme ging challenges, plan manage s and ope a o s need o be p o ided wi h compu e - based ools which guide hem o balance p oduc ion and esou ce consump ion [2]. Special a en ion o he e icien de elopmen o plan models needs o be paid, as i is he basis o ad anced con ol and coo - dina ion asks. Mo eo e , he e exis many coding languages and al e na i es o implemen op imiza ion algo i hms in o so wa e modules bu , in he end, hese ools mus be in eg a ed in o he in o ma ion echnology (IT) in as uc u e o he plan s, e.g., ia a neu al deploymen pla o m ha connec s o di e - en IT sys ems [3]. This pape deals wi h hese issues in he e apo a ion ne wo k o Lenzing AG, a iscose ibe p oduc ion ac o y loca ed in Aus ia. A desc ip ion o he ap- p oach and a p o o ypical ool o he op imiza ion o he alloca ion o e apo a o s o p oduc s is p esen ed, wi h he goal o minimizing he o e all speci ic s eam consump ion (SSC). In addi ion, e apo a ion plan s su e om pe o - mance deg ada ion due o ouling inside he hea exchange s. Hence, main enance asks become nec- essa y o eco e e iciency, bu hey in ol e a cos . The e o e, a sui able scheduling o such cleaning asks o e ime o each e apo a o a ises as an addi- ional p oblem o he load alloca ion op imiza ion. This equi es he esolu ion o an economic op imiza- ion in ol ing disc e e and con inuous alues. Ou app oach conside s da a d i en plan models, able o be upda ed in a semi-au oma ic way, and e icien mixed-in ege nonlinea p og amming (MINLP) so wa e o sol e he p oposed op imiza- ions online. The di e en op imiza ions we e p o- g ammed as modules using MATLAB® and MS Excel®, and hen linked o he PI Sys em in he plan . The pape o ganizes as ollows. Nex , a desc ip ion o he applica ion, sys em limi a ions and assump- ions a e gi en. The modelling ou ine is summa ized in Sec ion 3. Then, he op imiza ion o he load allo- ca ion and he cleaning schedule is o mally s a ed in Sec ion 4. The in e aces design oge he wi h he sys em in eg a ion is p esen ed in Sec ion 5. Finally, a summa y o he wo k oge he wi h indica ions o he nex s eps is gi en in he las sec ion. 2 APPLICATION CASE The p oduc ion o iscose ibe s is based on he e- newable esou ce wood. The cellulose con ained in he wood is chemically ea ed and con e ed o a iscose solu ion. The key p oduc ion s ep is he e- gene a ion o his solu ion in o ibe s, in oduced in an acid ba h. Apa om he solid ibe s, his chemi- XXXVIII Jo nadas de Au omá ica 575 cal ea men p oduces sodium sul a e (Na2SO4) and wa e as side p oduc s. This deg ades he acid ba h and, in consequence, he p oduc quali y. The e o e, i is necessa y o cons an ly emo e wa e and sodi- um sul a e om he ba h. A ne wo k o mul iple- e ec e apo a ion plan s linked o a c ys alliza ion sec ion is used o such a ask. Figu e 1 depic s a simpli ied schema o an e apo a- ion plan , whe e he main manipula ed a iables a e he eci cula ion low  and he p oduc empe a u e  a e he hea exchange s. Fo a mo e de ailed iew o he plan he eade is e e ed o [4]. Figu e 1. Simpli ied schema o an e apo a ion plan wi h he impo an con ol a iables F and T. 2.1 NETWORK DESCRIPTION The e apo a ion ne wo k comp ises a o al o 23 plan s o di e en nominal capaci ies. This ne wo k needs o p ocess 5 di e en p oduc s, so some plan s can se e in mo e han one p oduc , bu only a single p oduc a a ime. The changeo e om one p oduc o ano he ge s a cos and equi es ime. The e a e se e al ac o s which a ec he e iciency: x Plan ype: [a] Compac wi h small capaci y, [b] 3-s age e apo a o s and [c] la ge e apo a o s. x Ex e nal in luences: Ambien empe a u e and ai humidi y a ec he cooling owe s. x Fouling: Ba h impu i ies se le wi hin he hea exchange s, educing he hea ans e . x Ope a ing poin : Fo a desi ed e apo a ion se poin , he con ol alues a e no uniquely de ined The esou ce e iciency indica o (REI) [2] chosen o his p ocess is he SSC, de ined o each plan as he a io o esh s eam consumed pe amoun o wa e emo ed om he p oduc . The ask o he plan pe sonnel is o ind an op imal alloca ion o plan s o p oduc s ha ensu es he equi ed e apo a- ion a e pe p oduc wi h he lowes SSC. This op i- mal ope a ion can only be achie ed by conside ing all hese signi ican in luences on he esou ce e i- ciency in he op imiza ion o he e apo a ion ne - wo k. Howe e , he size o he combina o ial p ob- lem and he amoun o in luence ac o s make he p oblem e y challenging. Indeed, model-based op- imiza ion app oaches ha e al eady imp o ed he e iciency in he ope a ion o an e apo a ion plan [4], and we s ill o esee mo e po en ial sa ings in a be e coo dina ion o he whole ne wo k. The e o e, compu e -aided decision suppo (DS) ools need o be p o ided o help ope a o s in his ask, so sui able models need o be de eloped. Fu - he mo e, a sui able compu a ional ime is equi ed o p o ide esul s in accep able ime, o a oid p oduc- ion delays, and o ensu e ope a o accep ance. 2.2 PLANT MODELS The e apo a o se -up is simila o all plan s, bu a ies in he numbe o s ages and p oduc ion capaci- ies. The amoun o e apo a ed wa e depends on he ci cula ing low , he p oduc empe a u e , and cooling wa e empe a u e  which, in u n, is limi ed by he ou doo empe a u e. A mapping o he e apo a ion low achie ed o di e en alues in he manipula ed a iables can be eco ded (Figu e 2). Figu e 2. Speci ic s eam consump ion VS e apo a- ion se poin , ob ained o di e en con ol alues. Ex ensi e expe imen al es s shown ha he e ec s o  and  on he SSC, as well as on he e apo a ion low, can be desc ibed by linea ela ionships. Addi- ionally, i was obse ed ha he mapping in Figu e 2 is shi ed in a linea ashion wi h  and he ouling s a e. Hence, wo linea models we e p oposed o desc ibe he plan beha io , one o he e apo a ion low (EF) and ano he o he SSC, as linea unc- ions o he inpu s ,, and he ouling s a e  ( o be es ima ed): 󰇟 󰇠⋅󰇯   1󰇰 (1) 󰇟 󰇠⋅󰇯   1󰇰 (2) Whe e 󰇝, ,,󰇞 a e cons an pa ame e s o o line eg ession and  a e ime dependen ones o be iden i ied online. The absolu e s eam con- sump ion (ASC) o a plan is compu ed by mul ipli- ca ion o (1) and (2): XXXVIII Jo nadas de Au omá ica 576 ⋅ (3) In his way, gi en a ouling s a e  and a cooling wa e empe a u e , he maximum and minimum e apo a ion capaci ies, deno ed by  and  e- spec i ely, o each plan can be compu ed by (1) wi h he accep able ope a ing anges o  and : ,,, ,,, (4) The goal o he selec ion o he ope a ing poin is minimal speci ic s eam consump ion ul illing he e apo a ion demand ( ed on in Figu e 2). A sel - op imizing con olle (SOC) was implemen ed o ensu e ha ope a ion always lies in his egion. The con olle maximizes he p oduc empe a u e  o i s uppe limi and adjus s he ci cula ing low  o achie e he equi ed e apo a ion low [4]. Thanks o his op imal ope a ion pa e n, we a e able o com- pu e he con ol alues co esponding o he ed bounda y in Figu e 2 gi en a desi ed  and an es- ima ed s a e o ouling : indeed no e ha he  is se o i s uppe bound and  is se o he lowe one achie able by he cooling owe , so  can be com- pu ed di ec ly om (1) and, hus, he SSC om (2). 3 MODELLING ROUTINE The model iden i ica ion ask is implemen ed in MATLAB and comp ises a da a ea men o emo e inconsis en measu emen s, iden i ies s ep changes and pe o ms an i e a i e i ing o he model pa- ame e s. The equi ed da a o m he e apo a o s a e ob ained om he PI sys em ia an OPC-connec ion and addi ional in o ma ion such as he ime window o iden i ica ion, ag labels o he measu emen s in he his o ian, minimal numbe o changes in he EF, accep able noise band in s a iona y ope a ion, la ges ansi ion pe iod du ing s ep change, o he ime window o alida ion is p o ided by he ope a o wi h s anda dized Excel shee s. Finally, he quali y o he model is assessed by a compa ison o he model p edic ions wi h he measu ed EF and SSC. 3.1 DETECTION OF STEP CHANGES Fo he modeling o he s a iona y pa , he con ibu- ion o ouling mus be emo ed om he aining se . This is achie ed du ing he model i ing p ocess bu equi es da a om ope a ional poin s ha a e subjec o he same deg ee o ouling. Thus, he ool iden i ies changes in he EF, because ope a ion poin a ies enough o iden i y he pa ame e s and we can assume ha he ouling s a e does no a y signi i- can ly in one day. In ha way, he da a is scanned o s ep changes as shown in Figu e 3, p o iding in e als a and b. S ep changes a e iden i ied in he case ha : da a in a is a s eady s a e (wi hin a h eshold); s ep change is la ge han he h eshold c and; he ansi ion be ween he wo s eady s a es is comple ed wi hin in e al b. Each o he iden i ied s ep is eco ded and ansla ed in o a da a pai by a e aging he measu ed alues be o e and a e he change. O cou se, enough changes p o oked by , and  a e equi ed o a eliable iden i ica ion o he pa ame e s in (1) and (2). Figu e 3. S a iona y ope a ion be o e and a e he s ep (blue bounds), minimal s ep heigh ( ed a ow), maximal ansi ion in e al (do ed lines). 3.2 PARAMETER ESTIMATION Based on he assump ion ha he ouling is di e en om s ep o s ep bu emains cons an du ing he s ep i sel , pa ame e s , can be es ima ed by com- pa ison o he model p edic ions ( , 󰇜 wi h he ac ual alues (SSC, EF). The esul ing ouling ac o is subsequen ly used o bo h ope a ing poin s ha a e conside ed (be o e and a e s ep changes). Hence, an i e a i e LS op imiza ion o e he o e all da a se a ises, which is yields an in e media e se o s eady-s a e pa ame e s , used a e wa ds o cal- cula e new alues o , be o e sol ing an upda ed LS op imiza ion o ind he nex model gene a ion. These i e a ions con inue un il he esidue  (objec- i e unc ion) does no imp o e any mo e.  󰇛,󰇜           (5) He e  is he se o alues o he manipula ed a ia- bles ,  and ,  is he numbe o iden i ied s ep changes and ,  a e no malizing ac o s. The iden- i ica ion p ocedu e is summa ized in Algo i hm 1. Algo i hm 1. Pa ame e es ima ion o e apo a ion plan s. 1. P o ide an ini ial guess o  and se 0. 2. A e age he measu ed s eady-s a e alues o he SSC and EF be o e and a e he s ep change. 3. Adjus ouling ac o s  by compa ing he SSC and EF om S ep 2 wi h he model p edic ion. 4. Minimize (5) wi h  as decision a iables o ind he bes i o all s ep changes. 5. I  se 1 and go o S ep 3, else he algo i hm s ops. XXXVIII Jo nadas de Au omá ica 577 In p ac ice, o a easonable ini ial guess, he model pa ame e s  con e ge a e a ew i e a ion s eps. 3.3 VALIDATION The alida ion s ep in he modelling ou ine is pe - o med o assess he quali y o he model on he basis o an independen se o s ep changes ha has also been ob ained acco ding o Sec ion 3.1. Fo each iden i ied s ep change, he ouling ac o is also ad- jus ed in he models o ma ch he a e age alues o eco ded da a be o e he load change. The model wi h he upda ed ouling ac o s is hen used o simu- la e he plan o he same inpu s applied du ing change. The esul ing absolu e e o is hen no mal- ized wi h he heigh o he s ep change in he ASC, o yield a ela i e measu e o he model e o . The obse ed ela i e e o s a e ypically below 10%. These alues we e accep able, since he ne wo k op imiza ion is pe o med pe iodically and mis- modelling is educed om un o un by an online es ima ion o he ouling s a e. In some cases ela i e e o s o up o 30% ha e been obse ed due o a poo choice o load changes (e.g. non-s a iona y ope a ion alsely iden i ied as s eady s a e). A manual selec ion o s ep changes, choosing an al e na i e modeling ho izon, o an adjus men o he modeling se ings was su icien o imp o e he model i . 4 NETWORK OPTIMIZATION The objec i e is he minimiza ion o he ASC o he en i e ne wo k, gi en a desi ed e apo a ion demand. The o e all ASC is calcula ed as he sum o (3) o all e apo a o s. Two main ac o s which a ec he ASC a e objec o op imiza ion: he load alloca ion and he cleaning policy. 4.1 OPTIMAL ALLOCATION Fi s , gi en a se o ∈ p oduc s o be p ocessed in ∈ e apo a ion plan s, he p oblem is o allo- ca e plan s o p oduc s and hen dis ibu e he e- qui ed o al demand pe p oduc  in a way ha he o e all ASC in he ne wo k is minimized. Two se s o decision a iables a e de ined o his aim: x : Bina y a iables which link he p oduc  o he plan . x : Real a iables de ining he e apo a ion low o be achie ed in a plan  p ocessing he p oduc . Now, ecalling (4), assuming ha he ouling s a e  o each plan will be es ima ed,  is measu ed and con olled, a se o maximum and minimum capaci- ies o each plan ∈ is p o ided. Mo eo e , ol- lowing he op imal con ol pa e n explained in Sec- ion 2.2 o se ing  o each plan o i s uppe limi , om (1)-(3) we ge :  (6)         (7) Thus, eeding his in o ma ion, he op imal alloca ion o p oduc s o plan s is ound by sol ing he mixed in ege quad a ic p og amming p oblem below: min ,  ≔ ∈∈ s. .: (8) 1 ∈ ∀∈ (9)  ∈ ∀∈ (10) ⋅ ∀∈,∀∈ (11) ⋅ ∀∈,∀∈ (12) 0 󰇛,󰇜∉  (13) Whe e  is cons ained in (13) o ind easible solu ions wi hin he se  o allowed connec ions be ween plan s and p oduc s. 4.2 CLEANING SCHEDULE A complemen a y op imiza ion is p oposed o deal wi h he issue o ouling, which akes ad an age o he al eady de eloped decision suppo : once op imal e apo a ion se poin s a e compu ed o each plan , he idea is o sugges he nex cleaning cycle by bal- ancing he cos s o ope a ion o e ime wi h he cleaning cos s in an op imal ashion. This ask equi es models o he e olu ion o he ouling o e ime. Ex ensi e expe imen al es s ha e been pe o med measu ing he SSC in he e apo a- o s unning a e e ence ope a ion poin s be ween consecu i e cleaning cycles. This allows isola ing he e ec o ouling on he SSC inc ease, hence meas- u emen s a e compa able. In his way, app oxima e linea e olu ions o he ouling beha io could be iden i ied by eg ession, see Figu e 4. Thus, he ouling con ibu ion  in (1)-(2) becomes: 󰇛󰇜α⋅ (14) Whe e  s ands o he ime (in days) ha a plan is in ope a ion,  is he ini ial o cu en es ima ion o XXXVIII Jo nadas de Au omá ica 578 he ouling s a e, and  is he slope o he linea model. In his way, p edic ions o he u u e SSC (hence cos s) can be compu ed gi en a desi ed . Figu e 4. Measu ed e olu ion o he SSC ( ed) and ou pu o he eg ession model (blue). In o de o lump esou ces o di e en na u e (s eam, manpowe , cleaning p oduc s, e c.) in a single e i- ciency indica o , an agg ega ion based on cu ency is used. Hence, using p ices and cos s o u ili ies, he No malized A e age Cos pe Time (NAC) is de ined as an REI, and indica es he uni a y cos (€/d) in- cu ed o ope a e a plan be ween wo consecu i e cleaning asks (ope a ion cycle): ≔󰇡∑󰇛󰇜⋅   Δ⋅ Ws Wa ⋅Δ⋅⋅1.1/ (15) He e  is he sugges ed u u e day o pe o m he cleaning ope a ion, Δ is he ime equi ed o com- ple e a cleaning ope a ion, and , and  a e he cos s o he esh s eam, manpowe and was e wa e . No e ha , once he EF o each plan is se om (8)-(13), he 󰇛󰇜 is compu ed ia (7) and (14). No e also ha when an e apo a o is s opped o cleaning, i s load mus be assumed by o he s, so an app oxima e cos ac o o a 10% inc ease o e he nominal ope a ion cos is added in (15). The NAC is o be minimized wi h espec o  o each e apo a o o compu e a pe iodic “indi idually op imal” cleaning policy, which a emp s o be a “nea ly op imal” one o he whole ne wo k. Howe - e an issue appea s in using (15) as objec i e unc- ion: he cos o ope a ion is a disc e e sum which ge s  e ms, being  unknown a p io i, as i is deci- sion a iable. To exp ess his cos in a sui able way, we make use o he o mula ound by Gauss in he la e 1700’s o his ype o a i hme ic se ies [6]: ⋅123⋯⋅1⋅ 2(16) Mo eo e , he e a e wo ypes o cleaning asks, deno ed by  (big) and  (small), eaching di e en eco e ies , booking di e en imes Δ and using mo e o less was e wa e . Thus, each ask will ge di e en ixed cos s in he NAC (15) so he op i- mize mus choose which op ion minimizes he cos s. Thus, he p oposed economic op imiza ion o p edic he op imal cleaning policy o one plan eads: min ,  ≔⋅|󰇛1󰇜⋅| (17) s. .: 01; 0 (18) He e no a ion | s ands o (15) e alua ed wi h alues Δ,  and Ws Wa co esponding o a big cleaning ope a ion (| is analogous o a small cleaning). No e ha his op imiza ion o choose be- ween disc e e al e na i es can be handled ia NLP because (17) is mono onous w. . . , so i s minimum is loca ed in an ex eme, ei he 0 o 1. In his way, he bes cleaning (big o small) is chosen. 5 SYSTEM INTEGRATION The modelling as well as he load alloca ion modules we e implemen ed using MATLAB®. The cleaning schedule op imiza ion was coded in di ec ly in MS Excel. These choices a e jus i ied since he equi ed licenses and expe ience o he enginee ing depa - men a Lenzing AG a e a ailable o he sus ainable main enance o he decision-suppo solu ion. Figu e 5 depic s a schema o he eal- ime op imiza- ion (RTO) implemen ed o cope wi h he load allo- ca ion ask, which is execu ed each 30 min. The use dashboa d is included as a P ocess Book in he PI sys em. I shows he esul s and allows o manually igge he op imiza ion in case o signi ican changes in he e apo a ion . A e he ac i a ion, he s a ic in o ma ion (ne wo k in o ma ion, model pa ame e s, e c) is ead om an Excel in e ace. P oduc ion con- s ain s change dynamically, so hey a e ei he di ec - ly supplied by he da a his o ian o in e ed om measu emen s. Then, an upda e o he ouling pa am- e e s  and  is pe o med and sa ed o a ile. The ouling pa ame e s o he inac i e equipmen a e no upda ed in he ile. Figu e 5. RTO concep o he e apo a ion ne wo k. Execu ion o he op imiza ion esul s in he alloca- ions o e apo a o s o p oduc s, he load dis ibu ion, XXXVIII Jo nadas de Au omá ica 579 he SSC and ASC o each plan . This in o ma ion is w i en back o he sys em using special PI-Tags which a e displayed o he ope a o s ia dashboa d applica ion in he PI P ocess Book. Hence, he ope a- o s should adjus he e apo a o loads and alloca- ions acco dingly. Howe e , he op imiza ion o he load dis ibu ion may some imes esul in in easibili y, since app oxi- ma e models (1) migh sligh ly unde es ima e he e apo a ion capaci y o some plan s. This si ua ion migh lead o op imiza ion p oblems ha a e o e all in easible based on he model p edic ion, e en hough he eal plan is capable o ul il he desi ed e apo a ion low. 5.1 HANDLING INFEASIBILITIES I is impossible o p o ide eliable decision suppo o he ope a o s wi hou a easible solu ion om he op imize , because ha d cons ain s migh be iola - ed. To a oid hese si ua ions, a easibili y check is pe o med i s , ha e alua es whe he he cu en ly measu ed e apo a ion low o each e apo a o can be achie ed wi h he models unde he same ex e nal cons ain s (wea he , cooling wa e empe a u es and ne wo k a ailabili y). Then, o he iden i ied in ea- sible plan s, he MIQP cons ain s (11) a e so en wi h slack a iables ∈ as ollows: ⋅ ∀∈,∀∈ (19) Whe e  is he se o plan s which a e iden i ied in easible a e he easibili y check. Then, he sum o e all slack a iables is included as a penal y e m in o he objec i e unc ion (8) as: min ,,  ≔ ∈∈ ⋅ ∈ (20) The weigh  is oughly chosen o be g ea e han he la ges possible alue o (8), i.e., wi hou he con ibu ion o he slack a iables. Thus, he sol e will only p o ide he absolu ely necessa y amoun o cons ain iola ion. No e ha he ne wo k ope a ion will no esul in cons ain iola ions on he con ol inpu s, since he SOC is in place o each plan , and in easibili y is only a esul o a plan -model mis- ma ch. Mo eo e , a wa ning can be passed o he supe iso (plan enginee ). Thus, depending on he se e i y o he plan -model misma ch, co ec i e ac ions can be aken, e.g., a model upda e acco ding o he p ocedu e in Sec ion 3. The inal op imiza ion p oblem can be coded in MATLAB and sol ed wi h an MILP sol e ia suc- cessi e linea app oxima ions [7], o di ec ly wi h a MINLP sol e like BONMIN [8] ia he open sou ce OPTI-Toolbox, al hough his op ion migh be less compu a ionally e icien . 5.2 DECISION-SUPPORT INTERFACES The isualiza ion in e ace is adap ed o he al eady exis ing concep ha was designed o gi e an o e - iew o he e apo a ion p ocess du ing p oduc ion. On he one hand, he ope a o s a e supplied wi h he dashboa d depic ed in Figu e 6 ha shows he com- pu ed op imal solu ion o he cu en ime. The e - ical columns ep esen he 23 plan s in he ne wo k and he ows ep esen 5 p oduc s. Ligh g ey boxes a e he alloca ion possibili ies o plan s o p oduc s. I a plan is assigned o one o hese possible combina- ions, he box becomes g een. Plan s ha a e cu en - ly assigned o a p oduc bu a e no in ope a ion (un- de main enance o cleaning) a e shown by ed iles. Figu e 6. In e ace o he p o o ypical ool o online op imiza ion o he e apo a ion ne wo k. The alloca ion plan acco ding o he op imiza ion esul s is indica ed wi h yellow (pa ial load) o g een ci cles ( ull load) a he co esponding posi ion in he ma ix ep esen a ion. The op imal load dis ibu ion o plan s is di ec ly gi en nex o he cu en alue a he op o he ma ix. Small pic og ams show he necessa y di ec ion o he change in e apo a ion se poin s. On he igh hand side, he cu en and op i- mal alues o a o al p oduc e apo a ion low and he ACS a e lis ed. Finally, he p edic ed ne wo k- wide sa ings po en ial is shown in € sa ed pe hou , in o de o c ea e an incen i e o he ope a o s o apply he p edic ed e apo a ion se poin s o plan s. On he o he hand, he cleaning p edic ion module o Sec ion 4.2 has been implemen ed by an Excel-based ool, pa ially coded in Visual Basic and using he OpenSol e [9] add-on, whose cu en e sion in- cludes BONMIN as op imiza ion engine. This ool complemen s he one abo e, by ecei ing he load alloca ion o each e apo a o as inpu da a. The in e ace is o med by se e al shee s: one o each plan and a gene al o e iew o he ne wo k. In each plan shee he e is a se o alues o be se : du a ion o cleaning asks, cos s o esou ces, ene gy p ices, model pa ame e s and con ol se poin s , (see Figu e 7). The ool p o ides a bu on in each shee o igge he op imiza ion (17)-(18), dis- playing hen when he e apo a o should be cleaned and which ype o ope a ion is bes , as well as he cos componen s and cu en alue o he NAC o he sugges ed policy. Mo eo e , he ool se es also XXXVIII Jo nadas de Au omá ica 580 as a simula o o wha -i analysis, because he use is allowed o manually se he nex cleaning day and he ype o cleaning. In his way, he ool in o ms he ope a o abou he po en ial losses in €/d incu ed wi h espec o he op imally compu ed NAC, en- cou aging him/he o apply he sugges ions. Figu e 7. In e ace o he p o o ypical ool o he imp o ed scheduling o cleaning ope a ions. 6 CONCLUSIONS & OUTLOOK The modelling, op imiza ion and isualiza ion con- cep s p esen ed in his pape suppo he ope a o s o ake be e decisions in eal ime o imp o e he ne - wo k ope a ion. The modelling ool execu es an au- oma ized model upda e based on his o ical da a and use inpu s. The esul ing models a e inco po a ed in he RTO scheme ha sol es a MIQP p oblem acco d- ing o he cu en p oduc ion cons ain s and he plan s ouling s a es. The esul s a e isualized in he daily p oduc ion en i onmen , including p edic ions o he po en ial mone a y sa ings, incen i izing hus he ope a o s o apply he ecommenda ions. Models o long- e m ouling e ec s we e iden i ied by ex ensi e expe imen a ion, o be hen used in an economic op imiza ion. The inco po a ion o such unc ionali y allows inding he bes cleaning policy o each plan . This p o ides addi ional bene i s in e ms o ene gy and cos s associa ed o he cleaning. The de eloped DS ools a e cu en ly unde e alua- ion a Lenzing AG: he implemen a ion in o he exis ing sys ems and ope a ional policies is pe - o med s ep by s ep o ge expe ience in li e es ing and o ensu e accep abili y om he plan pe sonnel Abou a yea o no mal ope a ion is equi ed o as- sess he impac , bu p elimina y es s wi h his o ical da a e ealed a ound 10% ASC po en ial sa ings. Fu he imp o emen s in he modelling app oach a e expec ed i di e en plan models a e used o he summe and win e pe iods. I he impac assessmen shows su icien imp o emen , o he heu is ics and decomposi ions o he op imiza ion p oblems will be e alua ed o ake in o accoun unce ain y in model pa ame e s and/o ex e nal ac o s. Acknowledgemen This esea ch is unded by he Eu opean Union’s Ho izon 2020 p og am, unde g an nº 723575, and by he MINECO/FEDER (DPI2015-70975-P). Re e ences [1] S. Engell and I. Ha junkoski, "Op imal ope a ion: Scheduling, ad anced con ol and hei in eg a- ion," Compu e s & Chemical Enginee ing, pp. 121-133, 2012. [2] S. K äme and S. Engell, Resou ce E iciency o P ocessing Plan s: Moni o ing and Imp o emen , (In p ess): Wiley, 2017. [3] LeiKon, "D4.1 Requi emen speci ica ion o he in eg a ed deploymen pla o m," Ou comes o he MORE P ojec , 2014. [4] J.L. Pi a ch, C.G. Palacín, C. de P ada, B. Vogla- ue and G. Sey iedsbe ge , «Op imisa ion o he Resou ce E iciency in an Indus ial E apo a ion Sys em,» Jou nal o P ocess Con ol, ol. 56, pp. 1-12, 2017. [5] M. Kalliski, B. Beisheim, D. K ahè, U. Ens e, S. K äme and S. Engell, "Real- ime esou ce e i- ciency indica o s," a p edi ion - Au om. P axis, ol. 58, pp. 64-71, 2016. [6] D.M. Bu on, Elemen a y Numbe Theo y, Bos- on: MA: Allyn and Bacon, 1989, pp. 80-81. [7] C. Bliek, P. Bonami and A. Lodi, "Sol ing Mixed In ege Quad a ic P og amming p oblems wi h IBM-CPLEX: a p og ess epo ," in P oc. o he 26 h RAMP Symposium, Tokyo, 2014. [8] P. Bonami, L. T. Biegle , A. R. Conn, G. Co nue- jols, I. E. G ossmann, C. D. Lai d, J. Lee, A. Lodi, F. Ma go and A. Waech e , «An Algo- i hmic F amewo k o Con ex Mixed In ege Nonlinea P og ams,» Disc e e Op imiza ion, ol. 5, nº 2, pp. 186-204, 2008. [9] A. Mason, "OpenSol e – An Open Sou ce Add- in o Sol e Linea and In ege P ogammes," in Ope a ions Resea ch P oceedings 2011, D. Kla e, H. La hi and K. Schmedde s, Eds., Sp inge Be lin Heidelbe g, 2012, pp. 401-406, h p://opensol e .o g. XXXVIII Jo nadas de Au omá ica 581