Ekwa o-Osi e, Hen y; Bode, Dennis; Ohlendo , Jan-Hend ik; Thoben, Klaus-Die e
A icle — Published Ve sion
Manu ac u ing p ocess ene gy consump ion modeling: a
me hodology o iden i y he mos app op ia e model
Jou nal o In elligen Manu ac u ing
P o ided in Coope a ion wi h:
Sp inge Na u e
Sugges ed Ci a ion: Ekwa o-Osi e, Hen y; Bode, Dennis; Ohlendo , Jan-Hend ik; Thoben, Klaus-
Die e (2024) : Manu ac u ing p ocess ene gy consump ion modeling: a me hodology o iden i y he
mos app op ia e model, Jou nal o In elligen Manu ac u ing, ISSN 1572-8145, Sp inge US, New
Yo k, NY, Vol. 36, Iss. 8, pp. 5673-5693,
h ps://doi.o g/10.1007/s10845-024-02514-z
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/330888
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by/4.0/
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
h ps://doi.o g/10.1007/s10845-024-02514-z
Manu ac u ing p ocess ene gy consump ion modeling: a me hodology
o iden i y he mos app op ia e model
Hen y Ekwa o-Osi e1,2 ·Dennis Bode1,2 ·Jan-Hend ik Ohlendo 2·Klaus-Die e Thoben1,2
Recei ed: 6 Augus 2023 / Accep ed: 12 Oc obe 2024 / Published online: 20 No embe 2024
© The Au ho (s) 2024
Abs ac
This pape in es iga es he app op ia eness o machine lea ning (ML) and o he echniques o modeling manu ac u ing
p ocesses ene gy consump ion by de eloping a compa ison me hodology. Th ee esea ch ques ions a e posed: Fi s ly, how do
p edic ion e o s compa e using di e en echniques wi h a ying complexi y and ML use? Secondly, how does pe o mance
a y wi h di e en amoun s o da a? Thi dly, how do di e en echniques compa e in e ms o equi ed expe ise, e o o build
and in e p e abili y o esul s? To answe hese ques ions, he au ho s de elop a s uc u ed app oach, which is also en isioned
o be useable by p ac icing enginee s and manu ac u e s. Fou modeling ca ego ies a e de ined, anging om simple non-ML
me hods, such as linea eg ession, o complex ML me hods, such as deep neu al ne wo ks. The app oach is e alua ed using
da a om a compound eed manu ac u ing p ocess. The esul s con i m he no ion ha non-ML models a e be e sui ed o
unde s and and model manu ac u ing p ocesses when ew pa ame e s a e p esen , due o hei high in e p e abili y, while ML
models a e ecommended o analyzing p ocesses wi h many po en ially ele an and in e ela ed pa ame e s. In e es ingly
he app oach inds ha he complex ML ca ego y model does no ou pe o m he simple ML ca ego y model in e ms o
p edic ion accu acy, and only has he d awback o equi ing mo e expe ise o build and ha ing lowe in e p e abili y. The
s udy concludes ha he decision o use complex ML o modeling manu ac u ing p ocess ene gy consump ion should be
c i ically ques ioned and ha a simple app oach may be be e sui ed, sugges ing ha he de eloped me hodology would be
o alue o p ac icing enginee s.
Keywo ds Manu ac u ing ·Modeling ·Da a-d i en sus ainabili y ·Ene gy e iciency ·Resou ce e iciency ·Machine
lea ning ·Op imiza ion
In oduc ion
Mo i a ion and objec i e
Machine lea ning (ML) is a powe ul ool o iden i ying
hidden ene gy sa ings po en ials in manu ac u ing p ocesses
(Jamwal e al., 2021). Since he manu ac u ing sec o is
esponsible o o e a hi d o global ene gy consump ion,
BHen y Ekwa o-Osi e
[email p o ec ed]g
1BIBA – B emen Ins i u e o P oduc ion and Logis ics
GmbH, Hochschul ing 20, 28359 B emen, Ge many
2Facul y o P oduc ion Enginee ing, Ins i u e o In eg a ed
P oduc De elopmen (BIK), Uni e si y o B emen,
Badgas eine S aße 1, 28359 B emen, Ge many
app oaches o imp o ing ene gy e iciency o manu ac u -
ing p ocesses a e g ea ly needed (IEA, 2020). The need o
indus y o educe i s ca bon oo p in is omnip esen in bo h
academia and business, and is only g owing in impo ance
e e y yea . The pa allel g owing adop ion and capabili y o
ML, logically aises he oppo uni y o b ing hese wo opics
oge he , and he need o do his sys ema ically. The s eng h
o ML lies in analyzing p ocesses wi h many po en ially
ele an and in e ela ed pa ame e s, and in p o iding ec-
ommenda ions o p edic ions on he ou come (e.g. ene gy
consump ion, esou ce consump ion, p oduc quali y ea-
u es, e c.) o he p ocess. Thus, ML is especially well sui ed
o model manu ac u ing p ocesses and sys ems, o hen op i-
mize he ope a ing pa ame e s o he inpu ma e ials wi h
heobjec i eo educingene gyconsump ion(Ekwa o-Osi e
e al., 2022; Samadiani e al., 2024; Su ind a e al., 2024;
Wal e smann e al., 2021). While ML models ha e p o en o
123
5674 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
be aluable in ecen yea s, hey equi e mo e a e expe ise
han some non-ML modeling app oaches. Simila ly among
he many di e en ML models, he amoun o complexi y
and equi ed expe ise a ies signi ican ly (Engbe s & F e-
i ag, 2024). Addi ionally, ML models a e ypically “black
box”, meaning i canno be explained how he model esul ed
in speci ic p edic ion. Non-ML modeling me hods, such as
simula ions o simple linea eg essions can coun e some
o hese disad an ages o ML me hods, by ypically ha ing
g ea e in e p e abili y and equi ing less expe ise o build.
Non-ML app oaches di e om ML app oaches in ha hey
can be be e sui ed when he main objec i e o he analy-
sis is o ma hema ically ep esen he unde lying physical
phenomena o a p ocess, o when p ocesses wi h limi ed
pa ame e s a e o be analyzed (whe eas an ML model “only”
p o ides a model o he ela ionship be ween p ocess inpu s
and ou pu s, wi hou p o iding a ma hema ical ep esen a-
ion o inne wo kings o his ela ionship). The las decade
has seen a signi ican inc ease in he applica ion o ML in
manu ac u ing indus y, as demons a ed in a ecen su ey
by Google Cloud whe e wo hi ds o manu ac u e s epo ed
using AI in hei egula ope a ions and he mean sha e o IT
spendonAIo allsu eyedcompanieswaso e 30%(Google
Cloud, 2021). The up ake o in e es in AI and ML can also
be seen in academia, whe e AI opics we e p esen in 10% o
as much as 25% o all manu ac u ing publica ions in ecen
yea s, as ound in a s udy by he Eu opean Commission, Join
Resea ch Cen e (2022). The ad an ages and powe ul capa-
bili ies o ML jus i y i s popula i y; howe e , his does no
mean he applicabili y o ML should no be c i ically ques-
ioned, since he e is o en a ade-o o in e p e abili y and
dependence on ML expe s when employing his echnol-
ogy. The s ep o in es iga ing whe he ML eally is he mos
app op ia e me hod o a gi en ask is a ely ca ied ou in
academic esea ch. Al hough he sui abili y o ML s. non-
ML me hods ha e been compa ed o gene ic da a scena ios
as well as some manu ac u ing scena ios, his di e en ia-
ion in he con ex o imp o ing manu ac u ing ene gy and
esou ce e iciency is poo ly unde s ood (Bzdok e al., 2018;
Lago e al., 2018; Mak idakis e al., 2018).
Thus, o summa ize, he co e p oblem his pape in es-
iga es is he di icul y o de e mine when o use ML, o
di e en ypes o ML, o manu ac u ing ene gy consump-
ion modeling. The objec i e o his pape is o de elop
a me hodology o answe he ollowing esea ch ques ions
esul ing om he p oblem s a emen :
(1) How do he p edic ion e o s o di e en manu ac u -
ing ene gy consump ion modeling echniques compa e
when a ying he model complexi y and use o ML?
(2) How do he pe o mances o di e en manu ac u -
ing ene gy consump ion modeling echniques compa e
when di e en amoun s o da a a e used o ain he
models?
(3) How do he di e en manu ac u ing ene gy consump-
ion modeling echniques compa e in e ms o equi ed
expe ise and e o o build he model, as well as in
e ms o in e p e abili y o he esul s?
The ques ions will be in es iga ed using a compa ison
me hod de eloped by he au ho s, applied o an e alua ion
scena io o da a om a compound eed manu ac u ing p o-
cess. The o iginali y o his wo k lies in he me hodology he
au ho sde elop,whichallows o sys ema iciden i ica iono
an app op ia e modeling echnique and da a amoun o selec
o a gi en manu ac u ing p ocess. The me hodology poses
a aluable con ibu ion in he ield o indus ial enginee ing,
in ha i p o ides guidance on he p ope applica ion o ML,
o p ac i ione s wan ing o model he ene gy consump ion
o a p ocess; whe e signi ican da a science and modeling
expe ise is cu en ly needed o disce n he bes app oach,
he p oposed me hodology can be applied by hose wi h less
expe ience, in o de o ob ain an indica ion on he da a and
model ype bes sui ed o hei case. An addi ional unique
aspec o he me hod is ha unlike exis ing app oaches, his
one is designed o be applicable ac oss di e en manu ac u -
ing domains and p ocesses, as will be shown in he ollowing
chap e s.
Theo e ical backg ound and exis ing li e a u e
In his sec ion he au ho s summa ize he undamen al con-
cep s o modeling o manu ac u ing ene gy e iciency. Nex
a classi ica ion o modeling app oaches ideally sui ed o his
compa a i e s udy is p oposed. The app oach is de i ed om
exis ing classi ica ions, a selec ion o which a e men ioned in
his publica ion. Finally selec ed s udies ha ha e compa ed
ML and non-ML modeling app oaches a e p esen ed. An
ex ensi e sys ema ic li e a u e e iew was conduc ed a ound
his opic, esul ing in 60+ pape s, howe e , since his a icle
is no a e iew pape , only selec ed e e enced and indings
om he ull li e a u e e iew a e p esen ed.
Modeling o manu ac u ing ene gy e iciency
The pu pose o employing modeling o imp o e manu ac-
u ing ene gy e iciency is o p edic how pa ame e s o he
sys em a ec he ene gy consump ion o he sys em. In his
con ex , he sys em can be an en i e manu ac u ing plan , a
p ocess o a single machine. This is done by c ea ing a model
o he ela ion among sys em pa ame e s (such as machine
se ings, p ocess se ings and inpu ma e ial cha ac e is ics)
and he ene gy consump ion. Fu he o he ou pu cha ac-
e is ics which a e cons ained, such as p oduc quali y and
p oduc ion ime,a e ypicallyalsoincluded.Va iousme hods
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5675
can be used o c ea e such a model. Ma hema ical eg ession
models can be calcula ed, including: s a is ical eg essions
such as linea o polynomial, classical ML eg essions such
as decision ee, suppo ec o machine (SVM), XGBoos
o deep ML models such as con olu ional neu al ne wo ks
(CNN) o long sho - e m memo y ne wo ks (LSTM). Fo
examples o some o hese app oaches see Al ela Nie o e al.
o Zhang and Ji (Al ela Nie o e al., 2019; Zhang & Ji, 2020).
Fuzzy logic can be used o build a ep esen a ion o he man-
u ac u ing sys em, as demons a ed by Lau e al. ( 2008).
O , a simula ion o he sys em can be c ea ed, as e alua ed
by Thiede and Die mai and Ve l (Die mai & Ve l, 2009;
Thiede, 2012). In his pape he au ho s p opose a ca ego iza-
ion o he a ious modeling me hods a ailable, p esen ed in
henex sec ion.Onceamodelo hesys emhasbeenc ea ed,
his model is used o p edic how changes in sys em pa am-
e e s will in luence ene gy consump ion. This can p o ide
insigh s on oppo uni ies o imp o e ene gy e iciency.
A inal s ep ha is o en done is o use he model in an op i-
miza ion algo i hm o ind a p ocess pa ame e combina ion,
gi en ce ain cons ain s, which esul s in he lowes possi-
ble ene gy consump ion. Weiche e al. lis se e al examples
o such app oaches in hei li e a u e e iew o ML o he
op imiza ion o p oduc ion p ocesses (Weiche e al., 2019).
Fo suchmul i-objec i eop imiza ionp oblemsgene icalgo-
i hms a e widely used (Zolpaka e al., 2021). Fo sake o
comple ion, such an op imiza ion will be demons a ed a he
end o his pape , hough no elabo a ed upon in de ail.
Modeling ca ego ies
As seen in "Modeling o manu ac u ing ene gy e iciency"
sec ion, he e a e a a ie y o modeling app oaches a ail-
able.Fo he compa a i ein es iga ionplanned o hiss udy,
he au ho s equi e a ca ego iza ion wi h he ollowing cha -
ac e is ics: should di e en ia e be ween ML and non-ML
app oaches, should di e en ia e based on model complex-
i y as well as e o o implemen and should be applicable
omanu ac u ingene gyconsump ionmodeling.Theau ho s
oundse e alexis ingca ego iza iono modelingapp oaches
ha ul ill some o hei equi emen s, bu did no ind a clas-
si ica ion ha was ully app op ia e o he planned analysis.
Many pape s on manu ac u ing ene gy consump ion model-
ing men ion ha modeling app oaches can be di ided in o
wo b oad ca ego ies: he ca ego y called model-based o
“whi e-box”, in which models a e c ea ed using o mulas
o simula ions o he physical p ocesses unde lying he sys-
em, and he ca ego y called empi ical da a-based, in which
models a e c ea ed using s a is ics o ML on da a collec ed
om he manu ac u ing sys em (Abdoune e al., 2023; Ljung
& Glad, 2016). Some pape s which ocus only on machine
ool ene gy consump ion modeling also men ion a hi d ca -
ego y called s a e-based modeling which uses he di e en
ope a ional s a es (e.g. amp-up, p ocessing, idle, e c.) o a
machine ool o p edic he ene gy consump ion (Abdoune
e al., 2023; Li e al., 2022).
Occasionally sub-ca ego ies a e p o ided, hough hese
a e ypically ocusedonlyonMLo lackdi e en ia ionbased
on model complexi y. Fo example Máša e al. ca ego ize
di e en le els o ma hema ical models used o manu ac-
u ing ene gy sa ing measu es, wi h ecommenda ions o
applica ions o each (Máša e al., 2018). The classi ica ions
a e based pa ially on complexi y, howe e ocus on mod-
eling analyses and use cases when he en i e sys em is o
be modeled. The 5-le el py amid hey de ine con ains basic
balance models a le el 1, eg ession and balance models o
sys em componen s a le el 2 (using ML algo i hms such
as a i icial neu al ne wo ks, as well as eg ession models
using o dina y leas squa es), s a ic simula ion models o
he en i e manu ac u ing sys em a le el 3, dynamic (lin-
ea ) ma hema ical models o sys em componen s a le el 4
and model op imiza ion a le el 5. Le els 1, 2 and 3 we e
inco po a ed by he au ho s o his s udy, howe e , mo e di -
e en ia ion was needed wi hin le el wo since complexi y
can a y signi ican ly among di e en ML models. S udies
ound by he au ho s, which di e en ia e ML models do so a
a e y high g anula i y, and ypically do no explici ly di e -
en ia e based on model complexi y. Fo example Lago e al.
compa e a a ie y o s a is ical models including ones wi h
and wi hou eg esso s, and a la ge selec ion o ML mod-
els including, neu al ne wo ks, suppo ec o eg esso s and
ensemble models, o aling 27 di e en models. Though he
g oupings o models used by Lago e al. p o ide a s a ing
poin o de ining di e en ca ego ies o models, he high
numbe o ca ego ies and lack o di e en ia ion based on
complexi y and implemen a ion e o make i inapp op ia e
o he compa ison planned o his s udy (Lago e al., 2018).
In hei li e a u e e iew, Renna and Ma e i p o ided a ho -
ough o e iew o he ma hema ical me hods used o imp o e
ene gy e iciency in manu ac u ing sys ems, g ouping anal-
ysis in o he ca ego ies o exac nume ical, app oxima ion
s a egies, heu is ics/me a heu is ics, expe imen al and eal
ime da a analysis and simula ion (Renna & Ma e i, 2021).
This classi ica ion, hough ho ough, does no p o ide a basis
o compa ingcomplexi yande o equi ed o hedi e en
ca ego ies.
Thus, he au ho s o his pape c ea ed a ca ego iza ion
based on adap a ions om published app oaches, as well as
hei own expe ience and hinking, as will be p esen ed in
"Me hod" sec ion.
Exis ing compa a i e s udies
Whilemul iplemanu ac u ings udiesexis inwhichdi e en
ML app oaches a e compa ed, he au ho s ound no manu-
ac u ing s udies compa ing ML and non-ML app oaches.
123
5676 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
Howe e , some pape s om o he disciplines we e ound
which do such a compa ison. A selec ion is p esen ed in his
sec ion.
Bzdok e al. come om he ield o biology and compa ed
a s a is ical app oach (di e en ial exp ession analysis) and a
ML app oach ( andom o es ) by a emp ing o iden i y ce -
ain genes in a gi en RNA sequence (Bzdok e al., 2018).
They an he analysis wi h bo h me hods and hen compa ed
which iden i ied mo e genes co ec ly o e he cou se o
1000 simula ions. They ound ha he s a is ical app oach
pe o med almos as well as he ML app oach. They con-
cluded ha his was pa due o he p e-exis ing knowledge
hey had which helped hem design a good s a is ical model,
andino he pa due o he ela i elysmall numbe o ea u es
in he da a se . They p edic ed ha he di e ence in pe o -
mance would be g ea e i he da a se we e mo e complex
and hey did no ha e he p eexis ing knowledge.
Lago e al. did a simila compa ison in he ield o ene gy
and economics, by o ecas ing spo elec ici y p ices (Lago
e al., 2018). They compa ed a a ie y o s a is ical models
including ones wi h and wi hou eg esso s, and a la ge selec-
iono MLmodelsincluding, neu al ne wo ks,suppo ec o
eg esso s and ensemble models. As men ioned in "Model-
ing ca ego ies" sec ion, in o al he de ailed s udy compa ed
27 models and ound ha he ML models mos ly pe o med
be e . The s udy p esen ed in his pape will ocus less on
compa ing such a wide ange o speci ic models, bu a he
will aim o compa e models o di e en classes o implemen-
a ion complexi y, o quan i y he ade-o in pe o mance
and e o . The g oupings o models used by Lago e al. how-
e e p esen a s a ing poin o de ining he a o emen ioned
classes o complexi y.
Mak idakis e al. also did a compa ison o o ecas ing
abili y, bu applied o he M3 o ecas ing compe i ion da a
se , which con ains ime se ies om i e domains o demo-
g aphics, mic o- and mac oeconomics, indus y and inance
(Mak idakis e al., 2018). They es ed eigh popula amilies
o ML models and eigh adi ional s a is ical ones. In e -
es ingly hey ound ha he s a is ical models pe o med
signi ican ly be e han he ML models, while also being
compu a ionally less complex. I should be no ed ha one
eason hey gi e is ha he ML models migh ha e been
o e - i ed, which is a guably an issue ha could be educed
wi h u he ea u e enginee ing. They also aise he con-
side a ion ha he da a se has ela i ely ew ea u es (only
6), and hus may be “ oo simple” o ML. This is some-
hing he au ho s o he s udy p esen ed in his pape plan o
in es iga e. Mak idakis e al. end wi h a call o esea che s
o do mo e empi ical es ing o ML o ecas ing me hods,
a he han only publishing s udies in which ML o ecas s
a e claimed o be sa is ac o y wi hou compa ing o simple
s a is ical me hods o benchma ks. Though no con ined o
o ecas ing, his pape aims o answe p ecisely his call o
ac ion.
I should be no ed ha he s udies lis ed in his sec ion
only ca ego ize ML models as supe ised o unsupe ised.
Ma e ials and me hods
Me hod
Theau ho sha ede elopedas uc u edapp oach ocompa e
he e ec i eness o manu ac u ing p ocess ene gy consump-
ion modeling echniques o di e en complexi y le els, wi h
and wi hou ML. The app oach consis s o h ee dimensions,
ou lined as ollowing:
Me hod dimension 1: model dimension
Di e en ypes o ML and non-ML modeling echniques o
a ying deg ees o complexi y a e compa ed (see blue ex in
Fig. 1). Since as discussed in "Modeling ca ego ies" sec ion
exis ing modeling classi ica ions a e no ideal o he planned
compa a i e analysis in his pape , he au ho s de ine ou
modeling ca ego ies, adap ed om di e en classi ica ions
in li e a u e.
The ca ego iza ion con ains ou gene al ca ego ies o
models, wi h he pu pose o c ea ing a collec i ely exhaus-
i e ca ego iza ion ha will allow o sys ema ic compa ison
o di e en app oaches. Such a ca ego iza ion is impo an
as i will allow o a sys ema ic compa ison o modeling
app oaches, wi hin he applica ion o manu ac u ing ene gy
e iciency,basedonmodelcomplexi yanduseo ML.Table1
summa izes heca ego iza ion.Theau ho sadop ed hecom-
mon wo high le el classi ica ions ypically seen in li e a u e
in "Modeling ca ego ies" sec ion, “model-based” and “em-
pi ical”, in o hei classi ica ion. “Model-based” was used
la gely used o de ine he ca ego y called “Complex Non-
ML” and a ious app oaches wi hin he “empi ical” amily
we e di ided among he o he h ee ca ego ies o he au ho s’
classi ica ion.
Each ca ego y is u he de ined in he ollowing ex :
Model ca ego y: simple non-ML Li e a u econ ains a ious,
pa ially con lic ing de ini ions o when a eg ession algo-
i hm is conside ed o be a ML algo i hm and when i is no .
This ambigui y p omp s he need o es ablish a wo king de -
ini ion o a oid any po en ial con usion. Thus, in he con ex
o his pape , linea eg ession, also known as o dina y leas
squa es, is ca ego ized as he p ima y non-ML eg ession
algo i hm. This is because i in ol es i ing da a o a s aigh
line and does no inhe en ly in ol e “lea ning”. O he mod-
eling app oaches ha a e conside ed non-ML in his hesis
a e no algo i hms ha i models o da a, unlike all he o he
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5677
Fig. 1 Concep diag am o he model compa ison app oach o be used o answe he esea ch ques ions (Colo igu e online)
Table 1 Ca ego iza ion o modeling analyses, anging in complexi y and me hodology, o imp o ing ene gy e iciency o manu ac u ing sys ems
Simple Non-ML Complex Non-ML Simple ML Complex ML
De ini ion Linea eg ession T adi ional manual
modeling app oaches ha
do no equi e ML
Classical ML models wi hou
hype pa ame e uning
Ad anced/mode n models
wi h ex ensi e
hype pa ame e uning
Example
analyses
• O dina y leas squa es
linea eg ession
• Mechanis ic p ocess o
sys em simula ion
• Fuzzy logic modeling
• Physics-based model o
he sys em
• Decision ee
• Random o es
• XGBoos
•SVM
• Deep lea ning models
(mul ilaye pe cep on,
con olu ional NN,
ecu en NN, LSTM)
• Rein o cemen lea ning
models
Selec ed
p ope ies
High in e p e abili y and
explainabili y; i.e.
“whi e-box”
High in e p e abili y and
explainabili y; i.e.
“whi e-box”
Some in e p e abili y and
explainabili y; i.e.
“g ey-box”
Almos no in e p e abili y
and explainabili y; i.e.
“black-box”
Can be ca ied ou in no/low
code en i onmen s
Can some imes be ca ied
ou in no/low code
en i onmen s
Can be ca ied ou in no/low
code en i onmen s
Canno be e ec i ely
ca ied ou in no/low
code en i onmen s
Requi es less ime o build
han complex non-ML and
ML
Typically equi e mos ime
o build ou o all
ca ego ies
Requi es less ime o build
han complex non-ML and
ML
Requi es mo e ime o
build han simple
non-ML and ML
Less da a science expe ise
needed
Li le o no da a science
expe ise needed, bu high
expe ise in physics o
simula ion ools needed
Some da a science expe ise
needed
Mos da a science
expe ise needed
123
5678 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
h ee ca ego ies in Table 1. Ra he , hey a e manual model-
ingapp oaches ha a ec ea ed wi hou heuse o algo i hmic
me hods. See model ca ego y “Complex Non-ML”. The de -
ini ion o linea eg ession as u ilized in his hesis can be
ound in he ex book o Hope, who no ably also does no
clea ly s a e whe he linea eg ession is ML o no . Lin-
ea eg ession is one o he simples and mos widely used
me hod o assessing he ela ionships be ween con inuous
p edic o and esponse a iables (Hope, 2019). I s popula i y
can be a ibu ed o i s high in e p e abili y and s aigh o -
wa d implemen a ion.
Model ca ego y: complex non-ML Models in his ca ego y
do no employ ML, bu s ill demand a high le el o expe -
ise in he manu ac u ing sys em being modeled, as well as a
deep unde s anding o complex modeling app oaches. When
i comes o modelingmanu ac u ingsys ems and hei ene gy
e iciency wi hou ML o linea eg ession, h ee modeling
echniques a e equen ly employed. These echniques a e
mechanis ic p ocess o sys em simula ion, uzzy logic mod-
eling,andphysics-basedmodeling.I is impo an o no e ha
all h ee o hese modeling me hods equi e cus omiza ion
and adap a ion o sui he speci ic cha ac e is ics and equi e-
men s o he sys em being modeled. This cus omiza ion is
necessa y o ensu e ha he esul ing models accu a ely ep-
esen he in icacies o he gi en sys em.
A his poin he au ho s p o ide jus i ica ion o why he
modeling ca ego y “Complex Non-ML” is omi ed om he
compa ison me hodology. Rega ding simula ions, as Thiede
highligh s, he e a e ew o no comme cial simula ion ools
ha a e designed o simula e ene gy consump ions h ough-
ou a gi en manu ac u ing p ocess. This would mean he
au ho s would ha e o ind a e y speci ic so wa e o
de elop one hemsel es, as Thiede did (2012). On op o
his, a ho ough simula ion can be expec ed o cos signi -
ican ime and esou ces o c ea e (Banks, 2010; Chung,
2004). As Mou zis e al. ound in hei li e a u e e iew
o simula ions in manu ac u ing, mos comme cial so wa e
a e designed o speci ic manu ac u ing p ocesses, meaning
di e en so wa e would po en ially need o be used when
modeling di e en manu ac u ing p ocesses, as he au ho s
in end o do (Mou zis e al., 2014). Physics-based mod-
els a e sui ed o low-unce ain y low-complexi y sys ems,
a he han high-unce ain y high-complexi y sys ems such
as manu ac u ing p ocesses wi h mul iple s eps (Wang e al.,
2022). A physics-based model which does no accoun o
all physical dynamics in a sys em will likely be inaccu-
a e (E ge & an Oo , 2022). Mos examples o s udies
in which physics-based models a e c ea ed o imp o ing
manu ac u ing ene gy e iciency, a model is made only o a
single p ocess s ep, e.g. a cu ing, o a ing, o he mal p o-
cess s ep, and is used o ine- une he p ocess s ep (A am &
Xi ouchakis, 2011; Cubillo e al., 2016; Imani As ai e al.,
2018;Osa a,2019). Expanding he scope o such a model,
would equi e de i ing and adding equa ions o mul iple
physical phenomena, as well as in e ac ions among hese.
The esul ing inc easeincomplexi yo hesys em mos o en
hen canno be modeled wi h physics equa ions, o equi es
ex ensi ecompu a ional esou ces, no o men ion much ime
and expe ise o de i e. Mo eo e , in eg a ing he s ochas-
ic na u e o a machining p ocess, such as ool wea , in o
physics-based models poses a signi ican challenge (Sealy
e al., 2016). Ra he physics and da a-d i en hyb id models
a e some imes used (Wang e al., 2022). Re e ing o he clas-
si ica ion p oposed in his pape , such an app oach would be
a combina ion o “Complex Non-ML” wi h “Simple ML” o
“Complex ML”, al hough he au ho s will no conside ca e-
go y combina ions. Fuzzy logic modeling quickly becomes
e y ime consuming, when mo e pa ame e s a e added o
he model. Lau e al. demons a ed a uzzy logic app oach
o o ecas o e all ene gy consump ion o a clo hing p oduc-
ion ac o y, and included only h ee pa ame e s (daily o al
mass o inished p oduc s, o al labo hou s o ope a o s in
he plan in one day and he o al unning ime o equip-
men ) (Lau e al., 2008). When applied in he con ex o
manu ac u ing, uzzy logic can only p ac ically be applied
o conside he o e all manu ac u ing sys em, a he han
indi idual ene gy consuming p ocess s eps, as Thiede s a es
(2012). The abo e-men ioned limi a ions led he au ho s o
conclude ha he “Complex Non-ML” modeling app oaches
(simula ion, physics-based and uzzy logic) a e no well
sui ed o his compa a i e s udy and will hus be excluded.
Model ca ego y: simple ML ML models in his ca ego y
consis o all hose which do no ely on neu al a chi ec u es
o laye s. They encompass a ange o classical ML me h-
ods,including Suppo Vec o Machines,S ochas icG adien
Descen , Nea es Neighbo s, Nai e Bayes, Decision T ees
and ensemble me hods like G adien Boos ing and Random
Fo es s. These algo i hms a e widely accessible h ough pop-
ula ML lib a iessuch assciki -lea n(Ped egosae al., 2011).
One no able ad an age is hei applicabili y e en when one
has limi ed p io knowledge in ML. They can be e ec i ely
u ilized, especially when employing no/low code solu ions
like RapidMine and KNIME, making hem accessible o a
b oade audience o a ious applica ions (Be hold e al.,
2007;Mie swa&Ral ,2023).
Model ca ego y: complex ML Deep lea ning (DL), which
haswi nesseda su geinpopula i y, s andsou asa undamen-
al depa u e om classical lea ning me hods and se es as
he di e en ia o in his classi ica ion sys em. I is cha ac e -
ized as complex due o he black-box na u e o i s algo i hms,
as opposed o classical ML models, which o e some
in e p e abili y when using me hods such as ea u e impo -
ance analysis, ecu si e ea u e elimina ion and decision
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5679
ee isualiza ion o name some in e p e abili y echniques.
E ec i elyapplyingDLalgo i hmsnecessi a esa leas some
g asp o he algo i hm’s hype pa ame e s and a chi ec u e.
Mo eo e , p ope de elopmen o a DL algo i hm ypically
equi es signi ican ly mo e ime han de elopmen o a clas-
sical ML algo i hm. Building a obus model o en equi es
mul iple i e a ions. While Au oML has s eamlined some
aspec s, i has no ye made DL as quick and s aigh o wa d
o apply as classical ML algo i hms.
A his poin a no e should be made on DL and i s applica-
ion on non- abula da a, such as image, ex and sound. DL
has ound widesp ead acclaim in hese applica ions, whe e i
consis en ly ou pe o ms classical ML me hods. Howe e ,
i is impo an o cla i y ha his s udy does no di ec ly
conside image, ex o sound ecogni ion applica ions o
manu ac u ing p ocess modeling. Ins ead, hese applica ions
may be employed o con ibu e addi ional abula da a poin s.
Fo ins ance, an image ecogni ion algo i hm may assess
p oduc quali y and p o ide esul s such as “good quali y”
o “poo quali y” o a gi en p oduc ba ch o imes amp,
which is can hen be in eg a ed in o a da a able con aining
o he p ocess measu emen s. This able is subsequen ly used
o ain a model o he manu ac u ing sys em. Thus, in his
s udy, DL is applied only o abula da a, which is an appli-
ca ion a ea whe e he supe io i y o DL o e classical ML
is no sel -e iden and has no been ex ensi ely explo ed in
p e ious esea ch.
Fo he “Complex ML” ca ego y wo common ypes o
neu al ne wo ks a e p ima ily conside ed: mul ilaye pe cep-
on (MLP) and con olu ional neu al ne wo k (CNN), bo h
aken om he Ke as Py hon lib a y (Cholle , 2015). See he-
o e icalbackg oundchap e o de ailson healgo i hms.The
hype pa ame e s o he CNN models will be uned manually
by ial and e o o each da ase , using Tenso Boa d, while
he hype pa ame e s o he MLP models will be uned using
a la ge g id sea ch including he ollowing hype pa ame e s:
ba ch size, epochs, d opou a e, numbe o laye s, numbe
o nodes be ween laye s and op imize .
Me hod dimension 2: use case dimension
The speci ic objec i e o each analysis will depend on he
use case, i.e. applica ion scena io, unde conside a ion. The
me hodology p esen ed in his s udy is in ended o be used
on use case in which he goal is o model a manu ac u ing
sys em in o de o unde s and he ela ion be ween he sys-
em pa ame e s and he ene gy consump ion. The mo i a ion
o modeling in hese use cases is ypically o use he model
o iden i y pa ame e changes ha could imp o e he ene gy
e iciency (see ed ex in Fig. 1). The au ho s chose o limi
he analyses o modeling echniques in o de o allow o
a di ec quan i a i e compa ison o he di e en analyses,
namely he p edic ion e o . This p o ides a basis on which
analysis can be compa ed. The long- e m plan o he au ho s
is o do hese compa isons o a ious manu ac u ing ene gy
use cases, in o de o gain insigh s on which analyses a e
sui ed o which use cases. The me hodology can be applied
oanymanu ac u ingdomaininwhichmanu ac u ingsys em
pa ame e s can be used o p edic ene gy consump ion and
o he ele an sys em ou pu s. Examples o ene gy in ensi e
use cases he au ho s a e wo king on, in addi ion o he one
which will be p esen ed in his pape , include: wa e ex ac-
iondu ing gela inp oduc ion, ca bodyd ying inau omobile
manu ac u ingandwas ema e ialballingin heplas ics ecy-
clingindus y.Thispape willp esen he esul o heme hod
applied o one use case. Fu u e wo k will summa ize he ind-
ings om mul iple use cases.
Me hod dimension 3: da a dimension
A deciding ac o o da a-d i en sus ainable manu ac u ing
is ha ing a good da a basis. Howe e , he amoun o da a,
bo h in e ms o numbe o ea u es and amoun o da a sam-
ples needed is o en unknown and answe ed wi h “ he mo e
he be e ”. The au ho s do no plan o answe his ques ion
wi h a de ini i e numbe , bu aim o gain insigh s o di e en-
ia e he ela i e equi ed da a amoun s ac oss he modeling
ca ego ies. To in es iga e how well di e en analyses wo k
wi h di e en amoun s o da a, he da a amoun pe use case
will be a ied in wo ways: a ying he numbe o ea u es
and a ying he amoun o da a samples.
Fi s ly, he numbe o ea u es, i.e. columns o da a, will
be sys ema ically a ied by once using a selec ion o 3, 10
and inally all ea u es. The o al numbe o ea u es a ies
depending on he use case, bu is ypically app oxima ely
30 o 60, based on pas modeling he au ho s ha e done
and examples om li e a u e. The subse s a e selec ed using
ecu si e ea u e elimina ion (RFE). RFE is a ML ea u e
selec ion/ educ ionalgo i hmwhicha emp s o ind hemos
in luen ial ea u es by s a ing wi h all ea u es in he aining
da ase and sys ema ically elimina ing hem un il he speci-
ied numbe is eached (Kuhn & Johnson, 2016). I is known
as a “w appe ” selec ion me hod because i in ol es aining
aMLmodel o e e y possible combina iono ea u es. Todo
his, he RFE algo i hm akes a speci ied ML algo i hm and
i e a i ely emo es and eplaces each ea u e in he model,
checking he change in p edic ion e o each ime. Fea u es
a e hen anked by g ea es educ ion in p edic ion e o . The
leas impo an ea u e is disca ded and he model is e i ed
wi h he emaining ea u es. This cycle is epea ed un il he
speci ied numbe o ea u es emain. The RFE class om he
sciki -lea n package is used in his s udy (Ped egosa e al.,
2011). The au ho s choose RFE o selec he ea u e subse s
since i is an objec i e and obus selec ion me hod. Supe -
ised ea u e selec ion can be done wi h ei he w appe , il e
123
5680 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
o embedded me hods. Fil e me hods use s a is ical ech-
niques o sco e he impo ance o each ea u e o he a ge .
Embedded me hods a e buil -in ea u e impo ance me hods
o ce ain ML algo i hms, and can be e ie ed a e a model
is calcula ed. Ce ain il e me hods such as Spea man’s ank
coe icien , could echnically ha e been used, bu would be
less sui ed o his s udy since mos il e me hods a e no
in ended o be used o selec a speci ic numbe o ea u es
o a e complex o implemen o mul i a ia e (mul iple a -
ge s) eg ession models. In insic/embedded me hods we e
no used since hey a e only a ailable o ce ain ML algo-
i hms. An unsupe ised ea u e selec ion me hod, which
dis ega ds he a ge a iable, was no chosen because he
a ge a iables a e speci ied o all modeling scena ios con-
side ed in his s udy. Fo mo e de ails on ea u e selec ion
see Kuhn and Johnson (2016). No only will his p o ide
insigh on he ela ion be ween numbe o ea u es and model
pe o mance, bu i may also e eal whe he ce ain model
ca ego ies pe o m be e o wo se wi h mo e o less ea u es.
In he con ex o modeling manu ac u ing p ocess ene gy
consump ion, ea u e selec ion can also e eal which ea-
u es ha e he mos impac on he ene gy consump ion o an
analyzed p ocess.
Secondly, lea ning cu es will be used o compa e he
changes in model accu acy o di e en amoun s o da a
samples, i.e. ows o da a, om 10 o 100% o all a ailable
da a. Lea ning cu es consis o lines showing he aining
and alida ion e o s o a model, o di e en amoun s o
aining da a. Analyzing hese cu es can p o ide a ious
insigh s such as whe he a model is o e i ing o unde i -
ing,whe he mo epa ame e smay imp o e he pe o mance
o whe he mo e aining da a could imp o e he pe o -
mance. The au ho s will aim o ha e a leas 6 mon hs’ wo h
o eco ded da a o each use case. The sampling a e will
a y depending on he manu ac u ing sys em unde in es i-
ga ion, so he absolu e amoun o da a samples will also a y.
The au ho s will decide on how o add ess his once mul i-
ple use cases ha e been in es iga ed. Va ying he amoun o
da a samples will p o ide insigh s o he gi en use case on
how much aining da a is needed o he di e en modeling
app oaches o ob ain a gi en model (see g een ex in Fig. 1).
Calcula ion o p edic ion e o compa ison
The me ic used o compa ing he p edic ion e o is he
mean absolu e e o (MAE). No e ha he e m “accu acy” is
pu poselya oidedsincein hecon ex o p edic ion,accu acy
is he a e o co ec p edic ions, which is ele an o clas-
si ica ion p oblems, a he han eg ession p oblems, whe e
he amoun by which a p edic ion e s om he co ec alue
is o in e es . MAE is chosen because, a e no malizing all
da a om ze o o one, i allows o in e p e a ion o he MAE
as a pe cen age o he p edic ed da a, which is mo e in ui i e
o in e p e ; e.g. a manu ac u e can be e es ima e wha an
accep able pe cen age de ia ion in ene gy consump ion o
hei p ocess is, han an accep able oo squa ed e o . The
coe icien o de e mina ion (R2)and oo meansqua ede o
(RMSE) howe e could also easily be calcula ed du ing exe-
cu ion o he me hodology. Ra he han simply s a ing ha
he e o o one model is g ea e han ha o he o he , he di -
e ence in pe o mance will be compa ed wi hin he con ex
o he labels. This is an impo an dis inc ion o make when
conside ing mo e han only he e o o p edic ion models. In
his s udy, he complexi y o he models is also conside ed,
so i o example, he e o o a “Complex ML” model is
no no ewo hily lowe han ha o a “Simple ML” model,
he conclusion could be made ha he added complexi y p o-
ides no p edic ion ad an age and is only less in e p e able
and mo e cumbe some o c ea e. Model p edic ion compa -
isons in all p e ious li e a u e ha he au ho s encoun e ed
ei he only check which model has a lowe p edic ion e o ,
o only es whe he he di e ence in model pe o mance
is s a is ically signi ican by using he modi ied pai ed S u-
den ’s es (aka - es ) combined wi h 2 ×5 c oss alida ion,
as p oposed by (Die e ich, 1998). The Akaike and Bayesian
In o ma ionC i e ions ha e his samelimi a ion, hough hey
ake he numbe o model pa ame e s (i.e. ea u es) in o con-
side a ion. These app oaches do no p o ide in o ma ion on
whe he he p edic ion e o di e ence in wo models is la ge
enough o wa an using one model o e he o he , a he only
whe he he di e ence is s a is ically signi ican . The au ho s
o his pape use he logical assump ion ha he mo e he al-
ues o be p edic ed by he models a y, he less no ewo hy
a gi en di e ence in p edic ion e o be ween o models is.
Acco dingly, in he so-called no ewo hiness es p esen ed
in his pape he di e ence in MAE be ween models will
be compa ed o he a e age s anda d de ia ion o he alues
being p edic ed (i.e. he labels). No e ha he au ho s con-
sciously a oid using he e m “signi ican ” since his has a
speci ic meaning in he con ex o s a is ics; hus hey use
he e m “no ewo hy”. The a e age o he s anda d de ia-
ion o all he labels is aken, since he labels a y among
use cases; e.g. o one use case, elec ici y consump ion and
a quali y measu emen could be used as labels, while o
ano he use case gas consump ion and empe a u e could be
used as labels. A guably, he analysis could be made mo e
accu a e by gi ing di e en weigh s o he s anda d de ia-
ions o di e en classes o labels, e.g. ene gy- ela ed labels
a e weighed mo e han quali y- ela ed labels, howe e , his
would equi e indi idual e alua ion o each use case, and
wouldbedi icul oimplemen inan objec i emanne .Thus,
he a e aging app oach was chosen by he au ho s, o ensu e
he me hodology can be applied consis en ly ac oss di e en
manu ac u ing ene gy scena ios. When he da a is ini ially
no malized o a ange o 0 o 1, he a e age s anda d de i-
a ion and MAE a e uni -less and can be di ided by each
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5687
O e u iliza ion(>100%)occu encesdec ease o he239
ba ches 21.0%.
Figu e 8shows ha he ene gy consump ion is educed
o almos e e y ba ch.
A e age ene gy dec ease pe ba ch 6.0%.
O e all, he op imized pa ame e s achie ed an expec ed
a e age sa ing o 5.3% in ene gy cos s compa ed wi h
he o iginal pa ame e s, and a 6.0% a e age dec ease in
o alene gyconsump ion.Addi ionally,o e u iliza iono he
p ess machine is educed by 21.0%.
As a inal s ep be o e accep ing he esul s, i is necessa y
o implemen he p oposed p ocess pa ame e s in he eal-li e
p ocessand measu e he ac ual educ ionin ene gy consump-
ion. This is a necessa y s ep since bo h he op imiza ion
and he e alua ion a e based on he ML model. Thus, he
calcula ed imp o emen s may in pa be due o inaccu a-
cies in he model p edic ion and may be less o g ea e in
p ac ice. This scena io is a possible applica ion o ene gy
consump ion modeling o inc ease ene gy e iciency o a
manu ac u ing p ocess, and illus a es he e iciency gains
ha can be achie ed using an op imiza ion app oach.
Conclusion
Summa y
This pape demons a es a no el me hod o in es iga ing
how complex o a p edic ion model is needed o ob ain
eliable p edic ions o he ene gy consump ion o a man-
u ac u ing p ocess. The au ho s de ine ou ca ego ies o
analysis anging in complexi y and usage o ML. A compa -
ison me hod is cons uc ed a ound hese ca ego ies so ha a
sys ema ic e alua ion on he e o , model pe o mance and
equi edda a o di e en modelingapp oachescan bemade.
This me hod is es ed on a use case om a compound eed
manu ac u e and esul s in he conclusion ha a “Simple
ML” model using he 10 mos in luen ial ea u es and ~ 700
da a samples, is su icien o achie e a useable p edic ion
Fig. 6 Compa ison o p edic ed ene gy cos s be o e and a e op imiza ion. P edic ion based on he ac ual pa ame e combina ions in blue, and he
p edic ion based on he op imized pa ame e s in o ange. Y-axis alues omi ed due o con iden iali y (Colo igu e online)
Fig. 7 Compa ison o p edic ed machine u iliza ion be o e and a e op imiza ion. P edic ion based on he ac ual pa ame e combina ions in blue,
and he p edic ion based on he op imized pa ame e s in o ange (Colo igu e online)
123
5688 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
while equi inglesse o andp o idingmo ein e p e abili y
han a “Complex ML” model. To demons a e he p ac ical
alue o modeling he ene gy consump ion a manu ac u -
ing p ocess, a no el op imiza ion app oach is applied using
he iden i ied model. Ene gy consump ion sa ings o 6% a e
p edic ed o be achie able i he op imal p ocess pa ame e
se ings a e implemen ed.
Upon e isi ing,i isclea ha hede eloped me hodology
has he capabili y o answe he esea ch ques ions posed a
he s a o he s udy, o a gi en manu ac u ing p ocess.
Fo he manu ac u ing p ocess used in he e alua ion, and
answe s o he ques ions can be summa ized as ollows:
(1) How do he p edic ion e o s o di e en manu ac u -
ing ene gy consump ion modeling echniques compa e
when a ying he model complexi y and use o ML?
By de ining di e en ca ego ies o modeling ech-
niques based on complexi y and use o ML, he au ho s
we e able o make a s uc u ed compa ison. Fo he
conduc ed expe imen , inc easing model complexi y
esul ed in lowe p edic ion e o s, howe e wi h sig-
ni ican ly diminishing e u ns.
(2) How do he pe o mances o di e en manu ac u -
ing ene gy consump ion modeling echniques compa e
when di e en amoun s o da a a e used o ain he
models?
The au ho s we e able o in es iga e he e ec s o da a
amoun on he di e en models, by sys ema ically a y-
ing he numbe o ea u es, as well as he numbe o
da a samples, used o build he models. Fo he in es-
iga ed use case, adding mo e ea u es imp o ed he
pe o manceo allmodels,bu wi hdiminishing e u ns;
only he inc ease om 3 o 10 ea u es educed p e-
dic ion e o by a no ewo hy amoun . Models wi h
mo e ea u es equi ed mo e da a samples. As expec ed,
he “Simple non-ML” model equi ed he leas da a
samples.Unexpec edly, o he10and67 ea u escena -
ios, he “Simple-” and “Complex ML” models equi ed
app oxima ely he same numbe o da a samples o min-
imize p edic ion e o .
(3) How do he di e en manu ac u ing ene gy consump-
ion modeling echniques compa e in e ms o equi ed
expe ise and e o o build he model, as well as in
e ms o in e p e abili y o he esul s?
The wo “Simple” models equi ed a simila e o o
build, and signi ican ly less han he “Complex ML”
model. The “Simple non-ML” model (linea eg es-
sion) was he mos in e p e able. The wo ML models
we e less in e p e able, hough he “Simple ML” model
(XGBoos ) had a some in e p e abili y, as opposed o
he “Complex ML” model (MLP).
Limi a ions
One limi a ion o he wo k p esen ed he e is ha he “Com-
plex non-ML” ca ego y was omi ed in he analysis, due o
applicabili y. Including his ca ego y would p o ide a mo e
comp ehensi e compa ison; howe e he compa ison would
hen only be applicable o a smalle selec ion o use cases.
The me hodology is pu posely cons uc ed o be as simple
as possible, while s ill p o iding use ul insigh s. Wi h ha
being said, a second limi a ion is ha he me hodology does
no di e en ia e modeling app oaches beyond he ou ca -
ego ies. The e a e dozens o di e en models wi hin each
o he ou ca ego ies, which may di e sligh ly om each
o he in e ms o complexi y, in e p e abili y and equi ed
Fig. 8 Compa ison o p edic ed o al ene gy consump ion be o e and a e op imiza ion. P edic ion based on he ac ual pa ame e combina ions in
blue, and he p edic ion based on he op imized pa ame e s in o ange (Colo igu e online)
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5689
e o o build. An imp o emen could be made by de in-
ing sub-ca ego ies in o de o di e en ia e e en mo e among
modeling app oaches.
Con ibu ion and u u e esea ch
The p ima y con ibu ion o his pape is a s uc u ed epli-
cable model compa ison me hodology, which enables he
sys ema iciden i ica ion o anapp op ia emodel andamoun
o da a o a gi en manu ac u ing ene gy consump ion mod-
eling scena io. The o iginali y o he me hodology lies in i s
manu ac u e -o ien ed design, applicabili y o a ious man-
u ac u ing scena ios and sys ema ic s uc u e. The no el ee
based op imiza ion app oach is a u he con ibu ion o his
pape . The compa ison me hodology can p o ide aluable
insigh s in he ields o indus ial enginee ing and da a ana-
ly ics because such di e en ia ion can help manu ac u e s
and p ac i ione s a he s a o a modeling ini ia i e decide
whichapp oach o akeand oes ima e he esou ces oin es .
The eby acili a ing manu ac u e s in imp o ing he ene gy
e iciency o hei p ocesses. This con ibu ion ela es o
Sus ainabili y De elopmen Goal #9, “Build esilien in as-
uc u e, p omo e inclusi e and sus ainable indus ializa ion
and os e inno a ion” (UN Gene al Assembly, 2015). The
compa isonme hodology isdesigned obegene ic andappli-
cable o a a ie y o manu ac u ing p ocesses and indus ies,
a he han being de eloped o only a speci ic p ocess. Once
he compa ison me hodology has been applied o u he
use cases, he au ho s expec o be able o gene alize which
model ca ego y and da a amoun s a e mos app op ia e o
di e en ypes o manu ac u ing ene gy modeling use cases.
Then i would no be necessa y o execu e he ull compa -
ison me hodology o a gi en use case each ime, in o de
o ob ain an indica ion which model class and da a amoun
wouldlikelybe su icien .These insigh swouldbee en mo e
widely accessible among p ac i ione s, han i hey would
i s need o be calcula ed each ime.
Appendix
Appendix 1: Model p edic ion pe o mance,
wi h u he me ics
See Table 4.
Table 4 Model p edic ion
pe o mance, wi h u he me ics
123
5690 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
Appendix 2: Lea ning cu es
See Figs. 9,10, and 11.
Fig. 9 Lea ning cu es o h ee ea u es scena io. ~ 200 Da a samples
su icien o linea eg ession. > 900 o XGBoos . ~ 500 o CNN.
MAE > 5% should be conside ed high. Small gap indica es low a i-
ance. High aining e o indica es high bias o all models. Thus, mo e
da a samples will no help, bu a he mo e ea u es a e needed, o a
mo e complex model, since cu en models a e unde i ing he da a
Fig. 10 Lea ning cu es o 10 ea u es scena io. Linea eg ession
model has high bias and low a iance, and hus i he i educible e o
has no been eached, he model ei he equi es mo e ea u es o is no
complex enough a model o he p ocess. A e 500 da a samples he
model does no imp o e. Low aining e o o XGBoos indica es low
bias. La ge gap indica es high a iance, and ha he XGBoos model is
o e i ing. Valida ion cu e is pla eaued, so adding mo e da a samples
would likely no imp o e he model. CNN e o lines ha e also con-
e ged, and pla eaued, so ei he he model has eached he i educible
e o , o mo e ea u es o complexi y is needed
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5691
Fig. 11 Lea ning cu es o all 67 ea u es scena io. Linea eg ession
modelhashighbias and mode a e a iance,and hus hemodelisunsui -
able o he da a se (mo e ea u es ha e only inc eased he a iance o
he model). Fo XGBoos , low aining e o indica es low bias. La ge
gap indica es high a iance, and ha he model is o e i ing. Valida-
ion cu e is no qui e pla eaued, hus adding mo e da a samples may
imp o e he model. CNN has con e ged and jus pla eaued, so mo e
da a samples would likely no imp o e he model
Au ho con ibu ions Concep ualiza ion: Hen y Ekwa o-Osi e;
Me hodology: Hen y Ekwa o-Osi e, Dennis Bode; Fo mal analysis
and in es iga ion: Hen y Ekwa o-Osi e, Dennis Bode; W i ing—o-
iginal d a p epa a ion: Hen y Ekwa o-Osi e; W i ing— e iew and
edi ing: Dennis Bode, Jan-Hend ik Ohlendo , Klaus-Die e Thoben,
Hen y Ekwa o-Osi e; Funding acquisi ion: Dennis Bode; Supe ision:
Klaus-Die e Thoben.
Funding Open Access unding enabled and o ganized by P ojek
DEAL. This esea ch has been unded by he Ge man Fede al Min-
is y o Economic A ai s and Clima e Ac ion (BMWK) h ough he
P ojec “ecoKI” [03EN2047A]. The au ho s wish o acknowledge he
unding agency and all p ojec pa ne s o hei con ibu ion.
Da a a ailabili y The da ase sgene a eddu ing and analyzed du ing he
cu en s udy a e no publicly a ailable due o a con iden iali y ag ee-
men wi h he company om whose manu ac u ing acili y he da a
was collec ed. Da ase s a e a ailable om he co esponding au ho on
easonable eques .
Decla a ions
Con lic o in e es The au ho s decla e ha hey ha e no known com-
pe ing inancial in e es s o pe sonal ela ionships ha could ha e
appea ed o in luence he wo k epo ed in his pape .
Open Access This a icle is licensed unde a C ea i e Commons
A ibu ion 4.0 In e na ional License, which pe mi s use, sha ing, adap-
a ion, dis ibu ion and ep oduc ion in any medium o o ma , as
long as you gi e app op ia e c edi o he o iginal au ho (s) and he
sou ce, p o ide a link o he C ea i e Commons licence, and indi-
ca e i changes we e made. The images o o he hi d pa y ma e ial
in his a icle a e included in he a icle’s C ea i e Commons licence,
unless indica ed o he wise in a c edi line o he ma e ial. I ma e ial
is no included in he a icle’s C ea i e Commons licence and you
in ended use is no pe mi ed by s a u o y egula ion o exceeds he
pe mi eduse,youwillneed oob ainpe missiondi ec ly om hecopy-
igh holde . To iew a copy o his licence, isi h p://c ea i ecomm
ons.o g/licenses/by/4.0/.
Re e ences
Abdoune, F., Ragazzini, L., Noui i, M., Neg i, E., & Ca din, O. (2023).
Towa d digi al win o sus ainable manu ac u ing: A da a-d i en
app oach o ene gy consump ion beha io model gene a ion.
Compu e s in Indus y, 150, 103949. h ps://doi.o g/10.1016/j.
compind.2023.103949
Al ela Nie o, M. T., Naba i, E. G., Bode, D., Redecke , M. A.,
Decke , A., & Thoben, K.-D. (2019). Enabling ene gy e iciency
in manu ac u ing en i onmen s h ough deep lea ning app oaches:
Lessons lea ned. In F. Ame i, K. E. S ecke, G. on Cieminski & D.
Ki i sis (Eds.), IFIP ad ances in in o ma ion and communica ion
echnology, 1868-4238: Ad ances in p oduc ion managemen sys-
ems: P oduc ion managemen o he ac o y o he u u e. APMS
con e ence, 2019 (Vol. 567, pp. 567–574). Sp inge . h ps://doi.
o g/10.1007/978-3-030-29996-5_65
A am, O. I., & Xi ouchakis, P. (2011). E alua ing he use phase ene gy
equi emen s o a machine ool sys em. Jou nal o Cleane P o-
duc ion, 19(6–7),699–711.h ps://doi.o g/10.1016/j.jclep o.2010.
10.010
Banks, J. (2010). Disc e e-e en sys em simula ion (5 h ed.). P en ice
Hall.
Be hold, M. R., Ceb on, N., Dill, F., Gab iel, T. R., Kö e , T., Meinl,
T., Ohl, P., Sieb, C., Thiel, K., & Wiswedel, B. (2007). KNIME:
The Kons anz In o ma ion Mine . In S udies in classi ica ion, da a
analysis, and knowledge o ganiza ion (G KL 2007). Sp inge .
Bzdok, D., Al man, N., & K zywinski, M. (2018). S a is ics e sus
machine lea ning. Na u e Me hods, 15(4), 233–234. h ps://doi.
o g/10.1038/nme h.4642
Cholle , F. (2015). Ke as (compu e so wa e). Gi Hub. h ps://gi hub.
com/ cholle /ke as
123
5692 Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693
Chung, C. A. (2004). In C. A. Chung (Ed.), Simula ion modeling
handbook: A p ac ical app oach.Indus ial and manu ac u ing
enginee ing se ies. CRC P ess.
Cubillo, A., Pe inpanayagam, S., & Espe on-Miguez, M. (2016). A
e iew o physics-based models in p ognos ics: Applica ion o
gea s and bea ings o o a ing machine y. Ad ances in Mechani-
cal Enginee ing, 8(8),168781401666466.h ps://doi.o g/10.1177/
1687814016664660
Dehning, P., Thiede, S., Mennenga, M., & He mann, C. (2017). Fac-
o s in luencing he ene gy in ensi y o au omo i e manu ac u ing
plan s. Jou nal o Cleane P oduc ion, 142, 2305–2314. h ps://
doi.o g/10.1016/j.jclep o.2016.11.046
Die mai , A., & Ve l, A. (2009). A gene ic ene gy consump ion model
o decision making and ene gy e iciency op imisa ion in manu-
ac u ing. In e na ional Jou nal o Sus ainable Enginee ing, 2(2),
123–133. h ps://doi.o g/10.1080/19397030902947041
Die e ich, T. G. (1998). App oxima e s a is ical es s o com-
pa ing supe ised classi ica ion lea ning algo i hms. Neu-
al Compu a ion, 10(7), 1895–1923. h ps://doi.o g/10.1162/
089976698300017197
Ekwa o-Osi e, H., Bode, D., Thoben, K.-D., & Ohlendo , J.-H. (2022).
Iden i ica ion o machine lea ning ele an ene gy and esou ce
manu ac u ing e iciency le e s. Sus ainabili y, 14(23), 15618.
h ps://doi.o g/10.3390/su142315618
Engbe s, H., & F ei ag, M. (2024). Au oma ed model selec ion o
mul i a ia e anomaly de ec ion in manu ac u ing sys ems. Jou -
nal o In elligen Manu ac u ing.h ps://doi.o g/10.1007/s10845-
024-02479-z. Ad ance online publica ion.
E ge, O., & an Oo , E. (2022). Combining physics-based and da a-
d i en modeling in well cons uc ion: Hyb id luid dynamics
modeling. Jou nal o Na u al Gas Science and Enginee ing, 97,
104348. h ps://doi.o g/10.1016/j.jngse.2021.104348
Eu opean Commission, Join Resea ch Cen e. (2022). AI wa ch: AI
up ake in manu ac u ing. Eu opean Commission, Join Resea ch
Cen e. h ps://doi.o g/10.2760/267198
GoogleCloud.(2021).A i icial In elligence accele a ion among manu-
ac u e s. Google Cloud. h ps://cloud.google.com/blog/p oduc s/
ai-machine-lea ning/ esea ch-on-ai- ends-in-manu ac u ing
Has ie, T., Tibshi ani, R., & F iedman, J. H. (2009). In T. Has ie, R. Tib-
shi ani & J. F iedman (Eds.), The elemen s o s a is ical lea ning:
Da a mining, in e ence, and p edic ion. Sp inge se ies in s a is ics
(2nd ed.). Sp inge .
Hope, T. M. (2019). Linea eg ession. In A. Mechelli & S. Viei a
(Eds.), Machine lea ning: Me hods and applica ions o b ain dis-
o de s (pp. 67–81). Academic. h ps://doi.o g/10.1016/B978-0-
12-815739-8.00004-3
IEA. (2020). T acking indus y 2020. IEA. h ps://www.iea.o g/ epo s/
acking-indus y-2020
Imani As ai, R., Newman, S. T., & Nassehi, A. (2018). A mechanis ic
model o ene gy consump ion in milling. In e na ional Jou nal o
P oduc ion Resea ch, 56(1–2), 642–659. h ps://doi.o g/10.1080/
00207543.2017.1404160
Jamwal,A.,Ag awal,R.,Sha ma,M.,Kuma ,A.,Kuma ,V.,&
Ga za-Reyes, J. A. A. (2021). Machine lea ning applica ions o
sus ainable manu ac u ing: A bibliome ic-based e iew o u u e
esea ch. Jou nal o En e p ise In o ma ion Managemen .h ps://
doi.o g/10.1108/JEIM-09-2020-0361. ahead-o -p in .
Ka make ,S.K.,Hassan,M.M.,Smi h,M.J.,Xu,L.,Zhai,C.,&Vee a-
machaneni, K. (2022). Au oML o da e and beyond: Challenges
and oppo uni ies. ACM Compu ing Su eys, 54(8), 1–36. h ps://
doi.o g/10.1145/3470918
Ke s en, J., Rohde, H.-R., Ne , E., & Almann, H. (2010). Mis-
ch u e he s ellung: Rohwa e, P ozesse, Technologie (3. Au l.).
Misch u e -Biblio hek.Ag imedia.
Kuhn, M., & Johnson, K. (2016). Applied p edic i e modeling (co -
ec ed a 5 h p in ing). Sp inge . h ps://doi.o g/10.1007/978-1-
4614-6849-3
Lago, J., de Ridde , F., & de Schu e , B. (2018). Fo ecas ing spo elec-
ici y p ices: Deep lea ning app oaches and empi ical compa ison
o adi ional algo i hms. Applied Ene gy, 221, 386–405. h ps://
doi.o g/10.1016/j.apene gy.2018.02.069
Lau, H., Cheng, E., Lee, C., & Ho, G. (2008). A uzzy logic app oach o
o ecas ene gy consump ion change in a manu ac u ing sys em.
Expe Sys ems wi h Applica ions, 34(3), 1813–1824. h ps://doi.
o g/10.1016/j.eswa.2007.02.015
Li, H., Yang, D., Cao, H., Ge, W., Chen, E., Wen, X., & Li, C. (2022).
Da a-d i enhyb id Pe i-Ne basedene gyconsump ion beha iou
modelling o digi al win o ene gy-e icien manu ac u ing sys-
em. Ene gy, 239, 122178. h ps://doi.o g/10.1016/j.ene gy.2021.
122178
Ljung, L., & Glad, T. (2016). Modeling and iden i ica ion o dynamic
sys ems. S uden li e a u AB.
Mak idakis, S., Spilio is, E., & Assimakopoulos, V. (2018). S a is i-
cal and Machine Lea ning o ecas ing me hods: Conce ns and
ways o wa d. PLoS ONE, 13(3), e0194889. h ps://doi.o g/10.
1371/jou nal.pone.0194889
Máša, V., S ehlík, P., Touš, M., & Vond a, M. (2018). Key pilla s o
success ul ene gy sa ing p ojec s in small and medium indus-
ial en e p ises. Ene gy, 158, 293–304. h ps://doi.o g/10.1016/
j.ene gy.2018.06.018
Mie swa, I., & Ral , K. (2023). RapidMine S udio ( e sion 10.3) (com-
pu e so wa e). h ps://www. apidmine .com
Mou zis, D., Doukas, M., & Be nidaki, D. (2014). Simula ion in man-
u ac u ing: Re iew and challenges. P ocedia CIRP, 25, 213–229.
h ps://doi.o g/10.1016/j.p oci .2014.10.032
Osa a, J. A. (2019). The modynamics o manu ac u ing p ocess-
es—Thewo kpieceand hemachine y.In en ions, 4(2),28.h ps://
doi.o g/10.3390/in en ions4020028
Ped egosa, F., Va oquaux, G., G am o , A., Michel, V., Thi ion,
B., G isel, O., Blondel, M., P e enho e , P., Weiss, R.,
Dubou g, V., Vande plas, J., Passos, A., Cou napeau, D.,
B uche , M., Pe o , M., & Duchesnay, É. (2011). Sciki -lea n:
Machine lea ning in Py hon. Jou nal o Machine Lea ning
Resea ch, 12(85), 2825–2830. h ps://jml .csail.mi .edu/pape s/
12/ped egosa11a.h ml
Renna, P., & Ma e i, S. (2021). A li e a u e e iew o ene gy e iciency
and sus ainabili y in manu ac u ing sys ems. Applied Sciences,
11(16), 7366. h ps://doi.o g/10.3390/app11167366
Samadiani, N., Ba na d, A. S., Gunasega am, D., & Fayyazi a , N.
(2024). Bes p ac ices o machine lea ning s a egies aimed a
p ocess pa ame e de elopmen in powde bed usion addi i e
manu ac u ing. Jou nal o In elligen Manu ac u ing.h ps://doi.
o g/10.1007/s10845-024-02490-4. Ad ance online publica ion.
Sealy, M. P., Liu, Z. Y., Zhang, D., Guo, Y. B., & Liu, Z. Q. (2016).
Ene gy consump ion and modelingin p ecision ha d milling. Jou -
nal o Cleane P oduc ion, 135, 1591–1601. h ps://doi.o g/10.
1016/j.jclep o.2015.10.094
Su ind a, M. D., Al a isy, G. A. F., Caesa end a, W., Pe a, M. I., P ase-
yo, T., Tjahjowidodo, T., K ólczyk, G. M., Glowacz, A., & Gup a,
M. K. (2024). Use o machine lea ning models in condi ion mon-
i o ing o ab asi e bel in obo ic a m g inding p ocess. Jou nal
o In elligen Manu ac u ing.h ps://doi.o g/10.1007/s10845-024-
02410-6. Ad ance online publica ion.
Thiede,S.(2012).Ene gye iciencyinmanu ac u ingsys ems.InZugl.:
B aunschweig, Technical Uni e si y, Disse a ion, 2011 (2012nd
ed.) Sus ainable p oduc ion, li e cycle enginee ing and manage-
men . Sp inge . h p://www.loc.go /ca di /enhancemen s/ y1409/
2012935578-d.h ml
123
Jou nal o In elligen Manu ac u ing (2025) 36:5673–5693 5693
UN Gene al Assembly. (2015). T ans o ming ou wo ld: The 2030
Agenda o Sus ainable De elopmen (Resolu ion 25). UN Gen-
e al Assembly. h ps://sus ainablede elopmen .un.o g/pos 2015/
ans o mingou wo ld/publica ion
Wal e smann, L., Kiemel, S., S uhlsa z, J., Saue , A., & Miehe, R.
(2021). A i icial in elligence applica ions o inc easing esou ce
e iciencyinmanu ac u ingcompanies—Acomp ehensi e e iew.
Sus ainabili y, 13(12), 6689. h ps://doi.o g/10.3390/su13126689
Wang, J., Li, Y., Gao, R. X., & Zhang, F. (2022). Hyb id physics-based
and da a-d i en models o sma manu ac u ing: Modelling, sim-
ula ion, and explainabili y. Jou nal o Manu ac u ing Sys ems, 63,
381–391. h ps://doi.o g/10.1016/j.jmsy.2022.04.004
Weiche , D., Link, P., S oll, A., Rüping, S., Ihlen eld , S., & W o-
bel, S. (2019). A e iew o machine lea ning o he op imiza ion
o p oduc ion p ocesses. The In e na ional Jou nal o Ad anced
Manu ac u ing Technology, 104(5–8), 1889–1902. h ps://doi.o g/
10.1007/s00170-019-03988-5
Zhang, C., & Ji, W. (2020). Edge compu ing enabled p oduc ion
anomalies de ec ion and ene gy-e icien p oduc ion decision
app oach o disc e e manu ac u ing wo kshops. IEEE Access, 8,
158197–158207. h ps://doi.o g/10.1109/ACCESS.2020.3020136
Zolpaka ,N.A.,Yasak,M.F.,&Pa hak,S.(2021).A e iew:Useo e o-
lu iona y algo i hm o op imisa ion o machining pa ame e s. The
In e na ional Jou nal o Ad anced Manu ac u ing Technology,
115(1–2), 31–47. h ps://doi.o g/10.1007/s00170-021-07155-7
Publishe ’s No e Sp inge Na u e emains neu al wi h ega d o ju is-
dic ional claims in published maps and ins i u ional a ilia ions.
123
