D2.1 – REPORT ON DYNAMIC DATA
RECONCILIATION OF LARGE-SCALE
PROCESSES
José Luis Pi a cha
Césa de P adab
a Resea ch associa e (UVA) – Spain
b P o esso (UVA) – Spain
Oc obe 2018
www.spi e2030.eu/cop o
2
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
P ojec De ails
P
ROJECT TITLE
Imp o ed ene gy and esou ce e iciency by be e coo dina-
ion o p oduc ion in he p ocess indus ies
P
ROJECT ACRONYM
C
O
P
RO
G
RANT
A
GREEMENT
N
O
723575
I
NSTRUMENT
R
ESEARCH AND
I
NNOVATION
A
CTION
C
ALL
H2020-SPIRE-02-2016
S
TARTING DATE OF PROJECT
N
OVEMBER
,
1
ST
2016
P
ROJECT DURATION
42
MONTHS
P
ROJECT COORDINATOR
(
ORGANIZA-
TION
) P
ROF
.
S
EBASTIAN
E
NGELL
(TUDO)
T
HE
C
O
P
RO
P
ROJECT
The goal o CoP o is o de elop and o demons a e me hods and ools o p ocess moni o ing and
op imal dynamic planning, scheduling and con ol o plan s, indus ial si es and clus e s unde dy-
namic ma ke condi ions. CoP o pays special a en ion o he ole o ope a o s and manage s in
plan -wide con ol solu ions and o he deploymen o ad anced solu ions in indus ial si es wi h a
he e ogeneous IT en i onmen . As he e o equi ed o he de elopmen and main enance o ac-
cu a e plan models is he bo leneck o he de elopmen and long- e m ope a ion o ad anced
con ol and scheduling solu ions, CoP o will de elop me hods o e icien modelling and o model
quali y moni o ing and model adap ion.
The CoP o Conso ium
Pa icipan No Pa icipan o ganisa ion name Coun y O ganisa ion
1 (Coo dina o ) Technische Uni e si ä Do mund (TUDO) DE HES
2 INEOS Köln GmbH (INEOS) DE IND
3 Co es o Deu schland AG (COV) DE IND
4 P oc e & Gamble Se ices Company NV (P&G) BE IND
5 Lenzing Ak iengesellscha (LENZING) AU IND
6 F insa del No oes e S.A. (F insa) ES IND
7 Uni e sidad de Valladolid (UVA) ES HES
8 École Poly echnique Féde ale de Lausanne (EPFL) CH HES
9 E hniko Ken o E e nas Kai Technologikis Anap yxis
(CERTH) GR RES
10 IIM-CSIC (CSIC) ES RES
11 LeiKon GmbH (LEIKON) DE SME
12 P ocess Sys ems En e p ise LTD (PSE) UK SME
13 Di is In elligen Solu ions GmbH (di is) DE SME
14 A gen & Waugh L d. (Sabisu) UK SME
15 ASM So S.L (ASM) ES SME
16 ORSOFT GmbH (ORS) DE SME
17 Inno TSD (inno) FR SME
3
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
Documen de ails
D
ELIVERABLE
T
YPE
R
EPORT
D
ELIVERABLE
N
O
2.1
D
ELIVERABLE
T
ITLE
Repo on dynamic da a econcilia ion o la ge
scale p ocesses
N
AME OF
L
EAD
P
ARTNER FOR THIS
D
ELIVERABLE
U
NIVERSIDAD DE
V
ALLADOLID
V
ERSION
2
C
ONTRACTUAL
D
ELIVERY
D
ATE
31
O
CTOBER
2018
A
CTUAL
D
ELIVERY
D
ATE
31
O
CTOBER
2018
Dissemina ion le el
PU Public X
CO Con iden ial, only o membe s o he conso ium (including he Commission)
Abs ac
A ailabili y o eliable p ocess in o ma ion in eal ime is key in any decision-making p ocedu e. Thus,
good indus ial decision-suppo implemen a ions equi e dealing wi h g oss e o s and conside a-
ion o p ocess ansien s in o de o ge a se o measu emen s which will be cohe en wi h he basic
unde lying p ocess dynamics. This epo p esen s dynamic da a econcilia ion me hods and ools
adap ed o he equi emen s o indus ial en i onmen s (la ge-scale sys ems and noisy/ aul y da a).
Mo eo e , basic concep s in li e a u e a e ex ended o a i icially inc ease sys em edundancy as well
as o cope wi h ime- a ying pa ame e es ima ion. The p ocedu e summa ized in his epo has
been es ed in he Lenzing case s udy.
R
EVISION
H
ISTORY
The ollowing able desc ibes he main changes done in he documen since i was c ea ed.
Re ision Da e Desc ip ion Au ho (O ganisa ion)
V0 14/09/2018 Documen c ea ion J.L. Pi a ch (UVA)
V1 24/09/2018 In e nal e iew C. de P ada (UVA)
V2 04/10/2018 Ex e nal e iew A. San ecchia (EPFL)
V2 31/10/2018 Final app o al S. Engell (TUDO)
Disclaime
THIS DOCUMENT IS PROVIDED "AS IS" WITH NO WARRANTIES WHATSOEVER, INCLUDING ANY WAR-
RANTY OF MERCHANTABILITY, NONINFRINGEMENT, FITNESS FOR ANY PARTICULAR PURPOSE, OR
ANY WARRANTY OTHERWISE ARISING OUT OF ANY PROPOSAL, SPECIFICATION OR SAMPLE. Any lia-
bili y, including liabili y o in ingemen o any p op ie a y igh s, ela ing o use o in o ma ion in
his documen is disclaimed. No license, exp ess o implied, by es oppels o o he wise, o any in el-
lec ual p ope y igh s a e g an ed he ein. The membe s o he p ojec CoP o do no accep any lia-
bili y o ac ions o omissions o CoP o membe s o hi d pa ies and disclaims any obliga ion o en-
o ce he use o his documen . This documen is subjec o change wi hou no ice.
4
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
Table o con en s
1
Execu i e summa y .................................................................................................... 5
2
In oduc ion ............................................................................................................... 6
3
De ec ion o ansien and s eady-s a e da a .............................................................. 7
3.1 Concep ......................................................................................................................... 7
3.1 P oposed p ocedu e ...................................................................................................... 9
4
Dynamic da a econcilia ion ..................................................................................... 10
4.1 P oblem o mula ion ................................................................................................... 10
4.1.1 Time disc e iza ion ........................................................................................................ 11
4.1.2 Inpu -de ia ion penal y ................................................................................................. 12
4.1.3 Mo ing-ho izon window ............................................................................................... 12
4.1.4 Adap i e noise model .................................................................................................... 12
4.2 Enhanced o mula ion ................................................................................................. 13
4.3 T ea men o g oss e o s ............................................................................................ 14
5
Case s udy: Mul iple-e ec e apo a ion plan .......................................................... 16
5.1 Sys em desc ip ion & modelling ................................................................................... 16
5.2 Reconcilia ion esul s ................................................................................................... 17
5.2.1 Example o g oss-e o de ec ion .................................................................................. 18
5.2.2 Pe o mance du ing ansien s...................................................................................... 19
5.3 Time- a ying pa ame e es ima ion ............................................................................. 21
6
Concluding ema ks ................................................................................................. 23
7
Re e ences ............................................................................................................... 24
5
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
1 Execu i e summa y
The CoP o pa ne UVA is esponsible o Task 2.1 – Da a econcilia ion echniques – wi h he pa ici-
pa ion o pa ne s LENZING, EPFL, CSIC, TUDO, INEOS and P&G. The objec i es o his ask can be
summa ized in wo:
 De elopmen o indica o s o elucida e measu emen e o s and o sugges co ec i e ac-
ions o hose ha a e sys ema ic.
 De elopmen o obus da a econcilia ion algo i hms which a e sui able o la ge-scale con-
inuous plan s, wi h pa icula demons a ion in he p ojec case s udies.
In pa icula , his deli e able ocuses on da a econcilia ion in dynamic si ua ions, and we ha e con-
side ed an e apo a ion plan in LENZING as p oo o concep . Al hough hese plan s a emp o ope -
a e no mally a ound some desi ed poin s, he a iabili y p o ided by he ex e nal ac o s such as he
p oduc income ( a iable empe a u es, lows and concen a ions) o he cooling sys em pe o -
mance (a ec ed by he wea he ), makes he con ol sys em o change se poin s in o de o adap he
plan o each si ua ion while ul illing he desi ed e apo a ion demands. This ansla es in a non-
negligible ime pe cen age whe e he plan is no in s eady s a e, a common ac in many la ge-scale
sys ems in he p ocess indus y.
Task 2.1 ex ends om mon h 7 o 24 and, du ing his pe iod, he pa ne s UVA, EPFL and LENZING
mainly we e he ones conduc ing he wo k on dynamic da a econcilia ion. The ch onog am o he
ask execu ion is as ollows:
1. Li e a u e e iew on dynamic da a econcilia ion me hods and p oposal o adap a-
ions/ex ensions o make hem sui able o he applicabili y o la ge-scale p ocesses.
2. In pa allel wi h he p e ious, selec ion o he mo e sui able case s udy o se e as p oo o
concep . De elopmen o a i s -p inciples model o be he backbone o u he econcilia-
ion algo i hms.
3. Scheduling and execu ion o expe imen al es s onsi e in o de o collec da a om he plan
in di e en ansien s a es.
4. Tes ing he mo e p omising da a econcilia ion me hods wi h he p o ided plan da ase .
The main c i e ia used o selec and o ex end he dynamic da a econcilia ion me hods we e: a) he
compu a ional e iciency, i.e., he abili y o p ocess hund eds o a iables in accep able ime o
online implemen a ions; b) he obus ness agains senso noise and g oss e o s and; c) he abili y o
pe o m consis en es ima ion o ime- a ying a iables and pa ame e s, especially o slow a ying
dynamics ha a e ei he no conside ed o di icul o model. Indeed, his builds he b idge be ween
Task 2.1 wi h asks 2.3 – Model-based so sensing – and 2.4 – Moni o ing o equipmen deg ada-
ion–.

6
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
2 In oduc ion
Decision-suppo sys ems equi e in o ma ion abou p ocess pe o mance, p e e able in eal ime,
no mally in he o m o some e iciency indica o s [1] o be compu ed om p ocess measu emen s.
Howe e , all measu emen s a e subjec o e o s (senso calib a ion, noise, ou o ange si ua ions,
e c.). The e o e, in la ge-scale sys ems, edundan senso ing a e no mally implemen ed ei he ia
ha dwa e (duplica ed measu emen s) o so wa e (so senso s [1]). Based on his las concep o
edundancy, da a- econcilia ion algo i hms aim o p o ide a se o p ocess- a iables es ima es close
o he senso alues, bu cohe en wi h he p ocess dynamics : ul illing basic i s -p inciple laws such
as mass and ene gy balances [1].
This epo ocuses on dynamic da a econcilia ion (DDR), which is sol ing an op imiza ion p oblem
whe e he p ocess (dynamic) equa ions ac as cons ain s o be sa is ied wi hin a ce ain ime in e -
al, and all p ocess a iables (inpu , ou pu o pa ame e ) a e ac ually decision a iables o he op-
imiza ion algo i hm. As a esul , he op imiza ion se up is gene ally la ge, wi h many nonlinea con-
s ain s om he p ocess model. The e o e, he inclusion o dynamic models equi es a ca e ul bal-
ance be ween he added complexi y and he equi ed compu a ion imes o sol ing he associa ed
(dynamic) op imiza ion p oblems. A common ade-o solu ion is making use o models ha com-
bine a de ailed s a iona y p ocess cons ain s wi h addi ional simpli ied dynamics.
In case ha he measu emen e o s a e no mally dis ibu ed a ound hei ue alues, he DDR ap-
p oach is able o p o ide he bes se o es ima ions cohe en wi h he model. Ne e heless, due o
se e al easons such as se ious de ec s in ins umen s o in he communica ion ne wo k, he solu ion
p o ided by he da a econcilia ion is dis o ed. As a consequence, he e o is sp ead h oughou he
es o he a iables, c ea ing a smea ing e ec . These p oblems a e called g oss e o s and hei
de ec ion and ea men is c ucial o ob aining good es ima ions. The p opaga ion o g oss e o s in
measu emen s o he e iciency indica o s mus be a oided because, o he wise, he decision-suppo
sys ems will ecommend w ong ac ions. Hence, p e ious da a ea men in oducing g oss-e o
de ec ion plus he use o obus es ima o s in he da a econcilia ion a e also manda o y. In his way,
his s ep a oids he inclusion o co up ed da a (ou lie s) in u he decision suppo phases and
se es as a de ec o o sys ema ic e o s in senso s/p ocess.
This epo p esen s an enhanced DDR me hodology which consis s on a i s s ep o da a ea men
o exclude ou lie s and de ec ion o ansien measu emen s. A e his s ep, eliable da a is assumed
o be a ailable o pe o m DDR i sel . So, he second s ep is using econcilia ion o expe imen al
iden i ica ion o he model’s g ey pa : ime- a ying pa ame e s and expe imen al pa e ns. Once
he model has been iden i ied, he hi d s ep is alida ion wi h new ansien da a. Las , he p o-
posed me hodology is es ed in a pa icula e apo a ion plan in Lenzing AG, whe e a g ey-box non-
linea model is alida ed h ough eigh mon hs o ope a ion.
7
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
3 De ec ion o ansien and s eady-s a e da a
Iden i ica ion o bo h s eady s a e and ansien s a e in noisy p ocess signals is impo an ei he in
model iden i ica ion and execu ion o eal- ime DDR ou ines. On he one hand, dynamic models
ha e coe icien s ep esen ing ime-cons an s, which should only be adjus ed o i da a om ansi-
en condi ions. The e o e, de ec ion o ansien s igge s he collec ion o da a o dynamic model-
ling. On he o he hand, s a ic cons ain s do no ep esen ansien s, so hey should only in ol e
a iables whose dynamics is negligible wi h espec o he dominan one. No e ha , al hough a pu e
s a iona y model could cope wi h he da a econcilia ion ask, chemical p ocesses a e inhe en ly
nons a iona y, so some model pa ame e s would need o be adjus ed pe iodically o keep he mod-
els ue o he p ocess and unc ionally use ul.
Since p ocess a iables a e usually noisy, DDR needs o "see" h ough he noise and announce p ob-
able s eady s a es o p obable ansien si ua ions. Hence, he employed me hod needs o conside
an app op ia e ime ho izon, longe han he mos ecen pai o samples, in o de o obse e a local
end o con iden ly make any s a emen . So, some s aigh o wa d implemen a ions o s eady-
s a e/ ansien de ec ion would be s a is ical es s o he slope o a linea end (compu ed by linea
eg ession) in he ime se ies o a mo ing ho izon da a window: i he p ocess is in s eady s a e, he
slope will luc ua e nea ze o alues.
3.1 Concep
The me hod ecommended he e is based on he ac ha he a iance o a signal measu es he de i-
a ion om he mean alue and should be cons an in a s a iona y p ocess. In a ansien , he mo ing
a e age o he signal will lag behind he change in he signal and he a iance will inc ease. The
me hod hen uses he R-s a is ic, a a io be ween wo a iances, measu ed on he same se o da a
by wo app oaches [4]. The idea is o ake a ansien condi ion which is ba ely de ec able o decid-
edly inconsequen ial (pe human judgmen ) and se he p obably s eady-s a e h eshold o he R-
s a is ic as an imp obable low alue, bu no so low as o be imp obably encoun e ed when he p o-
cess is uly a s eady s a e.
The concep will be illus a ed h ough Figu e 1 aken om [5] whe e a se o da a is ep esen ed
o e ime by g een do s. The me hod i s calcula es a il e ed alue o he p ocess measu emen s, a
mo ing a e age, indica ed by he ed cu ed line ha lags behind he da a. Then he a iance in he
da a is measu ed by wo me hods. Fi s , a mo ing a iance is compu ed o e he mo ing a e age and
hen he a iance o he s a iona y p ocess (ob ained by di e en ia ion o he da a) is compu ed o
no malize he i s one.
8
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
Figu e 1: Noisy measu emen s (g een diamonds), fil e ed da a (solid ed line) and de iaons (pu ple a ows).
I he p ocess is a s eady s a e, hen he il e ed alue o he measu emen 𝑋

will coincide wi h he
a e age o he da a. Then, a p ocess a iance 𝜎



es ima ed wi h he mo ing a e age 𝑋

will be
ideally equal o 𝜎



es ima ed o he s a iona y p ocess. Thus, he a io o he a iances 𝑟




≅
1 . Al e na i ely, i he p ocess is in a ansien , he il e ed alue 𝑋

lags behind he p ocess da a and
he a iance as measu ed by 𝜎



will be much la ge han he one es ima ed by d
1
, so 𝑟




≫1.
The il e ed alue which p o ides an es ima e o he da a mean is compu ed by
𝑋

󰇛𝑘󰇜𝜆

𝑋󰇛𝑘󰇜󰇛1𝜆

󰇜⋅𝑋

󰇛𝑘1󰇜
󰇛1󰇜
whe e 𝑋󰇛𝑘󰇜 is he p ocess a iable a ime sample 𝑘 and 𝜆

is a i s -o de il e ac o . Simila ly, he
me hod o measu e he a iance 𝜎



is compu ed by 𝜈 as:
𝜈

󰇛𝑘󰇜𝜆

󰇡𝑋󰇛𝑘󰇜𝑋

󰇛𝑘1󰇜󰇢

󰇛1𝜆

󰇜⋅𝜈

󰇛𝑘1󰇜
󰇛2󰇜
The p e ious alue o he il e ed measu emen is used ins ead o he mos ecen ly upda ed alue
o p e en au oco ela ion om biasing he a iance es ima e, keeping he equa ion o he a io
simple. In con as , he a iance 𝜎



is es ima ed by 𝛿 using ano he il e based on sequen ial da a
di e ences as a way o con e a possible non-s a iona y p ocess in a s a iona y one:
𝛿

󰇛𝑘󰇜𝜆

𝑋󰇛𝑘󰇜𝑋󰇛𝑘1󰇜

󰇛1𝜆

󰇜⋅𝛿

󰇛𝑘1󰇜
󰇛3󰇜
The a io o a iances, he R-s a is ic o be compa ed o i s c i ical alues, may now be compu ed by
he ollowing simple equa ion
𝑅󰇛2𝜆

󰇜⋅𝜈

󰇛𝑘󰇜
𝛿

󰇛𝑘󰇜
󰇛4󰇜
as i ollows a 𝐹 dis ibu ion. No e ha he coe icien 󰇛2𝜆

󰇜 in 󰇛4󰇜 is equi ed o scale he a io o
ep esen he ac ual a iance a io [6]. Recommended alues o he weigh ing ac o s a e 𝜆

0.2,
𝜆

 𝜆

0.1 [7], which e ec i ely mean ha he mos ecen 45 da a poin s a e used o calcula e
he R-s a is ic.
C i ical alues o 𝑅 a e selec ed by he le el o signi icance 𝛼, al e na ely he con idence le el 1𝛼,
ha he end use desi es o achie e. I he compu ed R-s a is ic is g ea e han R-c i ical, hen he e
9
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
is a 100󰇛1𝛼󰇜 pe cen con idence ha he p ocess is no a s eady s a e. Consequen ly, a alue o
𝑅 less han o equal o R-c i ical means he p ocess may be a s eady s a e.
3.1 P oposed p ocedu e
The abo e concep can be implemen ed by an algo i hm ha combines s eady-s a e and an-
sien iden i ica ion o p e en immedia e sequencing o compu a ions i he sys em passes
h ough a ime whe e i is p obably no a s eady s a e. A e he dynamic beha io is de ec -
ed, he algo i hm will allow he nex se o condi ions o begin a e he p ocess e u ns o
s eady s a e. In addi ion, a ime limi o any one un in he expe imen al p ocedu e can be
explici ly included in o de o iden i y any occu ence o ei he 1) a change was made and no
de ec ed o 2) he change made he p ocess so uns able ha s eady s a e could no be ob-
ained. Figu e 2 illus a es he logic used o he au oma ic algo i hm [7], which wo ks as ol-
lows.
Fi s , p ocess da a is ob ained and he R-s a is ic alue is compu ed. I his alue is la ge han
i s uppe c i ical alue, he algo i hm de e mines ha he p ocess is p obably no in s eady
s a e (pa h Y1), so he ansien a iable TS is se o 1. This is ollowed by a check o de e mine
whe he o no he ime limi has been exceeded. I no (pa h N7), he algo i hm wai o he
nex sample o ge new da a. I he ime limi has been exceeded (pa h Y8), hen he nex un
is implemen ed, he poin o change (POC) ime is
eco ded o analyse he eco ded se o da a, and
he nex sampling is obse ed.
I no in a ansien , he algo i hm checks whe he he
p ocess is de ini i ely a s eady s a e (pa h Y4) o no
(which means in an inde e mina e s a e) by compa -
ing wi h he lowe c i ical alue o he R-s a is ic. I
inde e mina e (pa h N3), and he ime limi has no
been exceeded (pa h N7), he algo i hm wai o he
nex sample o ge new da a. I he ime limi has
been exceeded, (pa h Y8), he nex un is implemen -
ed, he poin o change (POC) ime is eco ded and
he nex sampling is obse ed.
I s eady s a e has been de ec ed (pa h Y4) and TS=1
( o me ly he p ocess was in a ansien s a e, pa h Y5),
TS is se o ze o, and he nex un is implemen ed. How-
e e , i TS was 0 (pa h N6) he p ocess has no been in a
TS, which means ha he ecen implemen a ion o new
condi ions has no aken e ec ye . In his case, i he
ime limi has no been exceeded (pa h N7), he nex
sampling is obse ed and new da a is analyzed. I he
ime limi has been exceeded, (pa h Y8), he nex un is
implemen ed, he poin o change ime is eco ded o
analyze he eco ded se o da a, and he nex sampling is
obse ed.
Figu e 2: Algo i hm o s eady-s a e and
ansien idenficaon.
16
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
5 Case s udy: Mul iple-e ec e apo a ion plan
One o he bigges e apo a ion plan s a Lenzing AG is chosen as p oo o concep o he DDR
me hods summa ized in his epo . This plan gi es se ice a ached o he ibe -p oduc ion
p ocess and i s goal is o egene a e an acid low coming om he spinning p ocess, whe e
ibe s ake o m om he p e iously p ocessed cellulose pulp.
5.1 Sys em desc ip ion & modelling
The Lenzing e apo a ion plan conside ed o his epo is basically o med by se e al e apo-
a ion chambe s and hea exchange s a anged in se ial connec ion, a s eam condense and a
cooling sys em. Figu e 7 depic s a simpli ied scheme o he line, in which some single equip-
men (e apo a ion chambe s and exchange s) ha e been lumped due o lack o measu emen s
in be ween hem.
Figu e 7: Simplified diag am o he e apo aon plan wi h locaon o exis en ins umen aon:
ansduce s (blue) and con olle s (g een).
The sys em wo ks as a mul iple-e ec e apo a ion, achie ed on he one hand hanks o he p essu e
d op in he chambe s 𝑉

c ea ed by he condense , and in he o he hand o acuum pumps con-
nec ed o he e apo a ion chambe s labelled as 𝑉

. The e apo a ion plan , when connec ed o he
main p ocess, ecei es an inpu liquid mix u e o wa e wi h acid and o he chemical componen s
plus esidual o o ganic ma e ial. The goal is o concen a e he solu ion by emo ing ce ain amoun
o wa e . To achie e his, he acid ba h goes h ough he line o hea exchange s 𝑊

,𝑊

in coun e
cu en wi h sa u a ed-s eam lows (some coming om he e apo a o s 𝑉

and o he om a esh
s eam gene a ed in a boile ) o inc ease i s empe a u e. Then, he ho mix u e en e s sequen ially o
he low-p essu e chambe s 𝑉

, which o ces a pa ial e apo a ion o wa e . A e wa ds, an addi ional

17
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
e apo a ion phase is pe o med in he las se o chambe s 𝑉

hanks o he condense , which sucks
ou s eam by condensing i wi h cold wa e om a cooling owe . Finally, pa o he concen a ed
liquid lea es he p ocess and he es mixes wi h he inpu , being eci cula ed h ough he p ocess.
A nonlinea s eady-s a e model o his sys em (whose co e pa is based in i s p inciples) was p e i-
ously de eloped o eal- ime op imiza ion pu poses [21]. The model equa ions a e omi ed o b e -
i y (see he abo e e e ence o de ails) bu can be summa ized as ollows :
 Equa ions o ene gy and mass balances aken in he e apo a o s 𝑉

,𝑉

, hea exchange s
𝑊

,𝑊

, s eam condense s, s eam sa u a o and in he cooling owe .
 Densi y ela ionships be ween mass and olume ic lows o he liquid mix u e, wa e and
s eam as a unc ion o empe a u e and/o p essu e.
 Hea ansmission be ween luids in he exchange s: 𝑄𝑈𝐴 𝐿𝑀𝑇𝐷, whe e 𝑄 is he
ansmi ed hea , 𝑈𝐴 is he hea - ansmission coe icien ( o be es ima ed), and he loga-
i hmic mean empe a u e di e ence (𝐿𝑀𝑇𝐷) has been compu ed using he Chen's app ox-
ima ion [22].
 Phase equilib iums in he e apo a ion chambe s as a unc ion o empe a u es, p essu e and
concen a ions.
 Psychome ic condi ions in he cooling owe and expe imen ally ob ained cooling pe o -
mance depending on he empe a u e di e ence o he cool wa e wi h he ambien .
 Rela ionship be ween he ai low h ough he cooling owe wi h he an speed, including he
expe imen ally iden i ied con ec ion e ec due o in-ou empe a u e di e ence.
In o de o pe o m DDR, app oxima e i s -o de dynamics a e added o he ene gy balances in he
e apo a ion chambe s, s eam condense s and he cooling owe , ying o ep esen he ene gy ac-
cumula ion due o he luids esidence ime in such equipmen :
𝑚⋅𝑐

𝑑𝑇
𝑑𝑡𝐹

⋅𝐻󰇛𝑇

,𝐶

,𝑃

󰇜𝐹

⋅𝐻󰇛𝑇

,𝐶

,𝑃

󰇜

󰇛20󰇜
Whe e, simpli ying a lo o b e i y, 𝐹

and 𝐹

ep esen he inle and ou le mass lows wi h hei
espec i e s eam ea u es ( empe a u e, concen a ion and p essu e), 𝐻󰇛⋅󰇜 s a es o he speci ic
en halpy unc ion, 𝑚 is he o al mass inside he equipmen , 𝑐

ep esen s he speci ic hea and 𝑇
can be he a e age o ou le empe a u e o he medium whe e he ene gy is accumula ed. No mal-
ly, he dominan dynamics in hese equipmen is he one o he liquid phase, so we choose he mass
𝑚, speci ic hea 𝑐

and empe a u e 𝑇 acco dingly. O cou se he concen a ions and absolu e mass
o he liquid inside an equipmen can a y wi h ime oo, bu we neglec ed adding such dynamics in
he mass balances because: a) i is as e han he one on he empe a u e and/o b) concen a ions
a e no measu ed, so such dynamics canno be checked.
Indeed, 𝑚⋅𝑐

can be lumped in a ime cons an 𝜏 and ea ed as a ime- a ying pa ame e o be
es ima ed ia enhanced DDR (Sec ion 4.2).
5.2 Reconcilia ion esul s
A da a his o ian o eigh ecen mon hs o ope a ion wi h he plan has been p o ided by Lenzing AG
o es DDR p ocedu es. The da ase p o ides alues om all senso s depic ed in Figu e 7, eco ded
wi h a sampling ime o i e minu es. This sampling equency is conside ed as enough o co e he
dominan plan dynamics, which akes a ound 30 minu es o s abilize a e a se poin change.
18
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
1000 1200 1400 1600 1800 2000 2200 2400
Sample nº
105
110
115
120
125
130
135
ºC
W2 Sa u a ed s eam empe a u e
Reconciled
Measu ed
The sys em dynamics has been disc e ized by o hogonal colloca ion [10] using 2-deg ee in e pola -
ing polynomials. A mo ing ime-window o 𝐻7 (co esponding o 35 min.) and he inpu -de ia ion
app oach (Sec ion 4.1.2) a e employed o de ine an enhanced DDR schema (Sec ion 4.2). Some ob-
ained esul s a e p esen ed and discussed below.
5.2.1 Example o g oss-e o de ec ion
A selec ed econcilia ion window o a week o ope a ion is shown in Figu e 8, whe e he measu e-
men s o he ci cula ing low o acid ba h and he empe a u e o he sa u a ed s eam be o e en e -
ing 𝑊

hea exchange s a e depic ed, oge he wi h i s econciled alues ob ained om DDR.
A he beginning, i can be obse ed ha some biases be ween he econciled and measu ed alues
appea om ime o ime, especially in wha seems he plan is nea o s eady s a e. This could be
because he plan model in s eady s a e is no pe ec o some co up ed da a a ec ed he pa ame-
e es ima ion. Ne e heless, his is no conside ed a big issue, as hese biases e en ually educe o
accep able alues.
None heless, no e ha , since he six h day onwa ds, a sys ema ic e o is de ec ed in he s eam em-
pe a u e: indeed his e o is no cons an bu seems o inc ease wi h he ime. Howe e , no sensible
de ia ion is obse ed in he ci cula ing low. This is a clea indica o o ha some hing has happened
a ound he las s age o hea exchange s: i could be a g oss e o in he measu emen due o a aul
in he empe a u e senso , o a physical change in he p ocess (equipmen aul o ope a ion mode).
I would be desi able o c oss da a wi h he main enance his o ian o elucida e he o igin o his e o
and, in case o online DDR, i would be ecommendable o send a main enance o de o check he
s a e o such senso i he p oblem pe sis s in ime.
Figu e 8: DDR execuon du ing 8 days o ope aon.
1000 1200 1400 1600 1800 2000 2200 2400
Sample nº
100
120
140
160
180
200
m3/h
Ci cula ing low
Reconciled
Measu ed
Possible de ec ion
o a g oss e o
19
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
5.2.2 Pe o mance du ing ansien s
In o de o e alua e he pe o mance du ing ansien beha io , a ain o se poin changes in he
ci cula ing low and con ol empe a u e was scheduled in he plan , see Figu e 9 and Figu e 10.
Some de ia ions be ween he measu emen s and he econciled inpu s can be obse ed jus a e
some se poin changes, especially in he ci cula ing low. This can be an indica o ha we conside ed
some slow dynamics in he model bu hey a e as e in he eal plan , so he es ima ed inpu s by
DDR a e modi ied o i he es o he plan measu emen s. I could be also possible due o neglec -
ing he senso s dynamics which, some imes, migh be impo an .
1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400
Sample nº
130
140
150
160
170
180
190
200
210
220
m
3
h
Ci cula ing low
Reconciled
Measu ed
Figu e 9: Induced changes in he ci culang flow (con ol inpu ).
1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400
Sample nº
90
92
94
96
98
100
102
104
106
108
110
ºC
Ba h con ol empe a u e
Reconciled
Measu ed
Figu e 10: Induced changes in he empe a u e o he acid ba h lea ing 𝑾
𝟐
(con ol inpu ).
20
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
Figu es 11 o 13 below show he econciled alues o some in e media e empe a u es in he plan ,
o which we ha e measu emen s o compa e. In gene al, he ob ained es ima es a e spikie han
he co esponding measu emen s, which could be again a p oblem o plan -model misma ch in ou
simple app oxima ions o he sys em dynamics, o jus a smoo hing e ec by he senso s due o hei
own dynamics (no mally empe a u e senso s ac as a low-pass il e ).
Special men ion needs o be done o Figu e 11. The e, a ecu en gap o abou 2 deg ee C is ob-
se ed be ween he econciled alues and he measu emen s. This may indica e a p oblem in ou
model wi h he s eady-s a e equa ions in equipmen a ound he acid-ba h inle , o an indica o o
ha his senso needs ecalib a ion.
1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400
Sample nº
20
22
24
26
28
30
32
34
36
38
40
ºC
W1 spinba h inle empe a u e
Reconciled
Measu ed
Figu e 11: Va iabili y in he empe a u e inle o he ba h eci culaon.
1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400
Sample nº
76
78
80
82
84
86
88
90
92
ºC
W1 spinba h ou le empe a u e
Reconciled
Measu ed
Figu e 12: Tempe a u e o he acid ba h be ween exchange s ages 𝑾
𝟏
and 𝑾
𝟐
.
21
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400
Sample nº
95
100
105
110
115
120
125
130
ºC
W2 Sa u a ed esh s eam empe a u e
Reconciled
Measu ed
Figu e 13: Tempe a u e o he sa u a ed s eam a he 𝑾
𝟐
inle .
5.3 Time- a ying pa ame e es ima ion
Apa om ob aining es ima ions o he model a iables ha a e unmeasu ed, pa ame e es ima-
ions a e p o ided by DDR as a byp oduc . Among hese, especially in e es ing a e he es ima ions
o slow- a ying pa ame e s, because hey ep esen indeed he long- e m dynamics which has no
been conside ed ini ially in he model, such as ouling, deg ada ion, ca alys deac i a ion, e c. This
kind o dynamics is no mally e y di icul o model by i s p inciples o a pa icula sys em: he e
a e se e al in luencing ac o s and he unde lying physics is complex. Hence, es ima ions p o ided by
DDR can se e as “so senso s” om which da a-d i en equa ions can be ob ained by eg ession o
pa e n iden i ica ion among o he known a iables.
This is he case in he Lenzing e apo a ion plan , whe e he plan e iciency dec eases wi h ime due
o p og essi e ouling in he hea exchange se s 𝑊

and 𝑊

. The ouling e ec can be obse ed indi-
ec ly as an inc ease o he speci ic s eam consump ion (measu emen ) o e a mon h o ope a ion,
o di ec ly by he e olu ion o he hea - ansmission coe icien s 𝑈𝐴 (unmeasu ed). The e o e, we
ha e un enhanced DDR in he es da a conside ing he coe icien s 𝑈𝐴 o 𝑊

and 𝑊

as pa ame-
e s o be es ima ed. The esul s o e a week o ope a ion a e displayed in Figu e 14.
Al hough he es ima ion is no e y smoo h ( ypical beha io when es ima ing his kind o coe i-
cien s om noisy measu emen s) we can clea ly obse e a co ela ion be ween 𝑈𝐴 wi h espec o
he ci cula ing low. This e ec is cohe en wi h he physics, as he hea ans e by con ec ion is
expec ed o inc ease wi h he low (and ice e sa).
In addi ion, we can obse e a slow- a ying end o e he ime, whe e he hea - ansmission coe i-
cien dec eases in a e age. This co esponds o he abo e-men ioned ouling e ec , so ha we could
now decouple his da a om he con ec ion e ec and y o iden i y a ouling model o u he use
in decision suppo sys ems (i.e., o p edic op imal main enance policies [21]).

22
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
0 500 1000 1500 2000 2500
Sample nº
130
140
150
160
170
180
190
200
210
220
m
3
/h
Ci cula ing low
Reconciled
Measu ed
(a) Induced changes in he ci culang flow.
0 500 1000 1500 2000 2500
Sample nº
500
600
700
800
900
1000
1100
Jm/K
Hea ansmission coe icien
(b) E oluon o he hea - ansmission coefficien .
Figu e 14: Esmaon o he hea - ansmission coefficien a 𝑾
𝟏
du ing 9 days o ope aon.
23
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
6 Concluding ema ks
This deli e able has summa ized he main ideas on dynamic da a econcilia ion o be applied in
la ge-scale sys ems, pu ing special emphasis in allowing online implemen a ions. In his way, he
mo e sui able heo e ical app oaches ha e been e iewed and some ha e been es ed in he
mul iple-e ec e apo a ion plan o Lenzing AG.
As a i s a emp , he ob ained esul s we e sa is ac o y enough o p oo he DDR concep s in a la ge
sys em. Ne e heless, he applica ion o DDR o la ge p ocessing plan s has demons a ed o be
challenging bo h om he heo e ical and implemen a ion aspec s. One o he main di icul ies
encoun e ed is ela ed o he modelling : a ep esen a i e dynamic model o he plan is equi ed as
a s a ing poin o eally us DDR esul s, bu la ge plan s in ol e se e al complex p ocesses di icul
o model, e en in s eady s a e.
Ano he limi ing ac o is he quali y o he senso s da a, which needs o be checked ca e ully be o e
econcilia ion. In his epo some ideas ha e been p esen ed o pallia e his issue. Howe e , when
he p ocess is in an unconside ed ope a ion mode o aul y da a, he NLP op imiza ion does no ge a
easible solu ion, so es ima ions a e los du ing such ime ins an s. In ac i has happened o us in
ou case s udy, whe e some ully uncohe en peaks can be obse ed in igu es 11 and 12 om
sample nº 1730 o 1830 app oxima ely. Mo eo e , he ea men o all excep ions which can appea
ela ed o his issue is c ucial (and also ime consuming o he designe ) o ge a eliable se o
alues.
24
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
7 Re e ences
[1]
M. Kujanpää, J. Hakala, T. Pajula, B. Beisheim, S. K äme , D. Acke sho , M. Kalliski, S. Engell, U.
Ens e and J. L. Pi a ch, Success ul Resou ce E iciency Indica o s o p ocess indus ies, ol. 290,
Espoo: VTT Technology, 2017, p. 78.
[2] P. Kadlec, B. Gab ys and S. S and , "Da a-d i en So Senso s in he p ocess indus y,"
Compu e s & Chemical Enginee ing, ol. 33, no. 4, pp. 795-814, 2009.
[3] M. Leibman, T. Edga and L. Lasdon, "E cien da a econcilia ion and es ima ion o dynamic
p ocesses using nonlinea p og amming echniques," Compu e s & Chemical Enginee ing, ol.
16, pp. 963-986, 1992.
[4] N. Sh ow i, K. Vilanka and R. Rhineha , "Type-II C i ical Values o a S eady-S a e Iden i ie ,"
Jou nal o P ocess Con ol, ol. 20, no. 7, pp. 885-890, 2010.
[5] T. Huang, S eady S a e and T ansien S a e Iden i ica ion in an Indus ial P ocess, MS Thesis.
Oklahoma S a e Uni e si y, 2013.
[6] R. Rhineha , M. Su and U. Manimegalai-S idha , "Leap ogging and Synop ic Leap ogging: a
new op imiza ion app oach," Compu e s and Chemical Enginee ing, ol. 40, pp. 67-81, 2012.
[7] R. Rhineha , "Au oma ed S eady and T ansien S a e Iden i ica ion in Noisy P ocesses," in
Ame ican Con ol Con e ence (ACC), Washing on, DC, USA, 2013.
[8] J. Taylo and M. Laylabadi, "A no el adap i e nonlinea dynamic da a econcilia ion and g oss
e o de ec ion me hod," in IEEE Con . on Con ol Applica ions, Munich, 2006.
[9] H. Bock, M. Diehl, D. Leinewebe and J. Schlöde , "A Di ec Mul iple Shoo ing Me hod o Real-
Time Op imiza ion o Nonlinea DAE P ocesses," in Nonlinea Model P edic i e Con ol, F.
Allgöwe and A. Zheng, Eds., Basel, Sp inge , 2000, pp. 245-267.
[10] L. T. Biegle , A. M. Ce an es and A. Wäch e , "Ad ances in simul aneous s a egies o dynamic
p ocess op imiza ion," Chemical Enginee ing Science, ol. 57, no. 4, pp. 575-593, 2002.
[11] J. Taylo and R. Mo eno, "Nonlinea Dynamic Da a Reconcilia ion: In-dep h Case S udy," in IEEE
In e . Con . on Con ol Applica ions (CCA), Hyde abad, 2013.
[12] V. M. Za ala, C. D. Lai d and L. T. Biegle , "A as mo ing ho izon es ima ion algo i hm based on
nonlinea p og amming sensi i i y," Jou nal o P ocess Con ol, ol. 18, no. 4, pp. 876-884, 2008.
[13] S. Bai, J. Thibaul and D. D. McLean, "Dynamic da a econcilia ion: Al e na i e o Kalman il e ,"
Jou nal o P ocess Con ol, ol. 16, pp. 485-498, 2006.
[14] J. Taylo , "S a is ical pe o mance analysis o nonlinea s ochas ic sys ems by he mon e ca lo
me hod," T ans. on Ma hema ics and Compu e s in Simula ion, ol. 23, no. 1, pp. 21-33, 1981.
[15] M. Bendig, In eg a ion o O ganic Rankine Cycles o Was e Hea Reco e y in Indus ial
P ocesses, PhD Thesis: École Poly echnique Fédé ale de Lausanne, 2015.
25
Deli e able 2.1
Repo on dynamic da a econcilia ion o la ge-scale p ocesses
[16] G. Heyen, E. Ma échal and B. Kali en ze , "Sensi i i y calcula ions and a iance analysis in plan
measu emen econcilia ion," Compu e s & Chemical Enginee ing, ol. 20, no. S1, pp. S539-S544,
1996.
[17] P. Hube , "Robus s a is ics," in In e na ional Encyclopedia o S a is ical Science, M. Lo ic, Ed.,
Sp inge Be lin, 2014, pp. 1248-1251.
[18] N. A o a and L. Biegle , "Redescending es ima o s o Da a Reconcilia ion and Pa ame e
Es ima ion," Compu e s & Chemical Enginee ing, ol. 25, no. 11-12, p. 1585–1599, 2001.
[19] G. She lyako , S. Mo gen hale and A. Shu ygin, "Redescending M-es ima o s,"
J
ou nal o
S a is ical Planning and In e ence, ol. 138, pp. 2906-2917, 2008.
[20] M. Fuen e, G. Gu ie ez, E. Gomez, D. Sa abia and C. de P ada, "G oss e o managemen in da a
econcilia ion," in 9 h In e na ional Symposium on Ad anced Con ol o Chemical P ocesses,
Whis le , B i ish Columbia, Canada, 2015.
[21] J. L. Pi a ch, C. G. Palacín, C. de P ada, B. Voglaue 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.
[22] J. J. J. Chen, "Commen s on imp o emen s on a eplacemen o he loga i hmic mean,"
Chemical Enginee ing Science, ol. 42, no. 10, pp. 2488-2489, 1987.