P edic i e Main enance o Wind Tu bines:
Le e aging Senso Da a wi h T adi ional and
No el Machine Lea ning Techniques⋆
Juan Vane io, Ped o Casas, Axel Weissen eld
AIT Aus ian Ins i u e o Technology, Vienna, Aus ia
[email p o ec ed]
Abs ac . P edic i e main enance has eme ged as a c i ical s a egy o
imp o e he e iciency and eliabili y o wind u bines. Using ad anced
echnologies such as machine lea ning and da a analy ics, wind a m op-
e a o s can de ec po en ial p oblems ea ly, op imize main enance sched-
ules, and educe down ime. Th ough moni o ing and analysis o indus-
y s anda d Supe iso y Con ol and Da a Acquisi ion (SCADA) senso
da a, p edic i e main enance enables p oac i e in e en ions ha ex end
he use ul li e o u bines and imp o e ope a ional pe o mance.
This pape p esen s a no el anomaly de ec ion amewo k o p edic -
ing he no mal ope a ing ange o gene a o bea ing empe a u e using
co a ia es de i ed om cu en and lagged SCADA signals. Syn he ic
anomalies a e injec ed ac oss he ull a iable ange du ing aining, en-
abling a sel -supe ised XGBoos model o dis inguish be ween no mal
and anomalous beha io . The app oach is alida ed using a one-yea
SCADA da ase om a wind a m comp ising 10 wind u bine gene a-
o s. Resul s highligh he amewo k’s e ec i eness in iden i ying ea ly
signs o gene a o ailu e, suppo ing p edic i e main enance and im-
p o ing o e all u bine pe o mance and eliabili y.
Keywo ds: P edic i e Main enance ·Wind Tu bines ·Machine Lea ning ·Da a
Analy ics ·SCADA Sys ems
1 In oduc ion
Wind u bines a e c i ical asse s in enewable ene gy p oduc ion, bu hei op-
e a ion is o en challenged by mechanical ailu es. P edic i e main enance o e s
a p oac i e app oach o iden i y po en ial issues be o e hey escala e, educing
down ime and main enance cos s while imp o ing ope a ional e iciency.
The wind ene gy sec o aces unique challenges, including he emo e lo-
ca ions o wind a ms (WFs), ha sh ope a ional condi ions, and he la ge-scale
⋆This wo k has been unded by he Aus ian Resea ch P omo ion Agency (FFG)
unde g an No. FO999913202 UNDERPIN and by he Eu opean Commission unde
con ac No. 101123179 UNDERPIN.
deploymen o u bines. These ac o s make main enance a cos ly and logis ically
complex ask, accoun ing o 20-30% o he o al powe gene a ion expenses [1].
T adi ional main enance p ac ices, such as eac i e and pe iodic main e-
nance, ha e signi ican limi a ions ha p edic i e main enance aims o add ess.
Fo ins ance, eac i e main enance in ol es pe o ming main enance only a e
a ailu e occu s, leading o inc eased down ime and addi ional cos s due o un-
planned ou ages. In he case o pe iodically scheduled main enance, he s a egy
may lead o unnecessa y in e en ions and cos s.
On he o he hand, p edic i e main enance is a p oac i e app oach le e ag-
ing senso da a analysis and machine lea ning o iden i y po en ial issues and
p edic componen ailu es in ad ance, allowing ope a o s o ake co ec i e ac-
ions be o e p oblems escala e. P edic i e main enance o e s signi ican bene i s,
including educed down ime, ex ended asse li espan, and enhanced ope a ional
e iciency [2]. Timely ailu e p edic ion o c i ical componen s such as gene a-
o s is i al o consis en elec ici y p oduc ion. Fo example, p e ious s udies
ound ha 68% o down ime in wo Chinese wind a ms s emmed om gene a o ,
con e e , and blade issues [3]. The solu ion p oposed in his pape empowe s
ope a o s o moni o abno mali ies p oac i ely, ensu ing ea ly in e en ion and
imp o ed u bine eliabili y.
The li e a u e al eady p esen s se e al e ec i e app oaches using di e en
model ypes, such as LSTM and XGBoos , o p edic ion asks based on SCADA
da a. Howe e , he pe o mance di e ences be ween models a e gene ally small
and o en depend on he speci ic wind u bine [4]. Ou s udy ocuses on a yea -
long da ase o SCADA senso measu emen s, eco ded a 10-minu e in e als,
om a single wind u bine gene a o (WTG) wi hin a wind a m. Du ing he
obse a ion pe iod, i e majo aul e en s we e iden i ied h ough sys em logs
and ope a o epo s. As a esul , he u bine expe ienced educed o hal ed
ope a ion o a o al o 38 days. These indings unde sco e he impo ance o ea ly
ailu e p edic ion o enable imely main enance, he eby minimizing down ime
and associa ed cos s.
2 Me hodology
This sec ion ou lines he me hods used o p ocess he da a o he p oposed
p edic i e main enance app oach, which aims o iden i y po en ial ailu es in
ad ance.
2.1 Da a P e-p ocessing
Fi s , inco ec imes amps and inconsis en o ou -o - ange alues in he o iginal
da a we e iden i ied and co ec ed as pa o basic consis ency cleanup. Da a ows
con aining e oneous da a ha could no be co ec ed a e emo ed.
Da a poin s ou side he ope a ional wind speed ange (5 m/s o 25 m/s),
as well as hose wi h a blade pi ch angle exceeding 40°, o ope a ing unde
es ic ed powe limi s o g id demands, a e excluded. Addi ionally, alues ex-
ceeding i e s anda d de ia ions a e emo ed. To u he e ine he da ase , an
Isola ion Fo es model is employed o iden i y and emo e ou lie s. These exclu-
sions ensu e ha he aining da ase only includes pe iods when he WTG was
likely ope a ing unde no mal condi ions.
Fu he mo e, da a augmen a ion is applied by ma king all da a poin s wi hin
a se en-day window be o e and a e a majo ailu e as away_ om_anomalies
(a a) wi h a alue o 0, and all o he poin s wi h a alue o 1. This new column
e lec s he p oximi y o ailu e, allowing o he exclusion o da a poin s nea
ailu es om he aining se , as hei ea u e alues may be al e ed du ing such
e en s.
2.2 Fea u e Selec ion
A ca e ul explo a ion o he da ase , accompanied by a li e a u e e iew, indi-
ca es ha he senso eadings ha a e ep esen a i e o he WTG heal h can
be p edic ed om a ela i ely small se o ambien eadings and his o ical al-
ues o he p edic ed a iables. Fo ins ance, gene a o bea ing empe a u e is
equen ly used o p edic gene a o heal h [4].
The ambien a iables used as inpu ea u es o p edic ion a e lis ed in we
conside he ambien a iables o Table 1. The emaining inpu ea u es a e
he p e ious alue o he a ge a iables (which we conside new augmen a ion
a iables and append hei names wi h ‘_lag’). The a ge a iables o be lea ned
a e desc ibed in Table 2.
Va iable Desc ip ion
Amb_Temp_A g A e age ambien empe a u e
Amb_WindDi _Abs_A g Absolu e wind di ec ion a e age
Amb_WindDi _Rela i e_A g Rela i e wind di ec ion a e age
Amb_WindSpeed_A g A e age wind speed
Amb_WindSpeed_Max Maximum wind speed
Amb_WindSpeed_Min Minimum wind speed
Amb_WindSpeed_S d S anda d de ia ion o wind speed
Table 1. Ambien a iables used as inpu ea u es o p edic ion.
By aining he model o p edic hese a ge a iables, each da a poin can
be in e ed om i s p eceding s a e and he la es ambien senso eadings. This
app oach enhances obus ness by educing dependence on any single po en ially
noisy o co up ed inpu . Mo eo e , he model implici ly lea ns he unde lying
dynamics o he sys em, knowledge ha can be e ec i ely le e aged o anomaly
de ec ion and p edic i e main enance.
Va iable Desc ip ion
G d_P od_PsblePw _A g Possible powe p oduc ion om he g id
G d_P od_Pw _A g A e age powe p oduc ion om he g id
Blds_Pi chAngle_A g A e age pi ch angle o he blades
Gea _Bea _Temp_A g A e age empe a u e o he gea box bea ing
Gen_RPM_A g A e age o a ional speed o he gene a o
Gen_RPM_S d S anda d de ia ion o gene a o RPM
Hyd_Oil_Temp_A g A e age empe a u e o hyd aulic oil
Gea _Oil_Temp_A g A e age empe a u e o gea box oil
R _RPM_A g A e age o a ional speed o he o o
HVT a o_Phase1_Temp_A g A e age empe a u e o HV ans o me phase
Table 2. Ta ge a iables o p edic ion.
Finally, o pe o m p edic i e main enance, we aim o p edic in ad ance
he occu ence o anomalous e en s. Fo his, we p opose a no el me hod ha
di ec ly add esses he high imbalance be ween no mal ope a ion samples and
anomalous samples, he la e o which may e en no be p esen in he da a.
These no mal samples a e iden i ied by he a iable a a aking a alue o 1,
and a e expec ed o concen a e in speci ic egions, i.e., o li e in some mani-
old(s) o he da a space. The e o e, i we whe e o uni o mly injec syn he ically
c ea ed anomalous samples ( agged wi h a a equal o 0) spanning he whole da a
ange o he no mal samples, he a io o no mal s anomalous samples ac oss
egions o he da a space would ei he see a mix u e o no mal and anomalous
samples (in he mani old egion) o a la ge majo i y o anomalous samples (ou -
side o i ).
Figu e 1 shows an illus a i e examples o he desc ibed se up. I can be
seen he e ha he anomalous samples can be loca ed anywhe e, e en inside
he no mal da a egions. Howe e , di e en ly o m no mal da a poin s hey a e
expec ed no o concen a e anywhe e, and so di e en densi y egions eme ge
and can be lea ned. We p opose o use a abula machine lea ning model o lea n
o disc imina e be ween hese pa e ns, which maps o disc imina e be ween
no mal and anomalous da a egions.
2.3 Model T aining
The da ase p uned acco ding o he ules om sec ion 2.1 con ains da a con-
side ed as ‘no mal’, and is he e o e spli in o wo consecu i e agmen s o 60%
and 40% o he o al leng h o aining and es pu poses, espec i ely, o ensu e
he models gene alize well o unseen da a.
Fi s , a His og am-Based G adien Boos ing Reg esso (HGBR) model is
buil o p edic ing each one o he a ge s om Table 2. These models a e
implemen a ions o he Ligh GBM ee-based model [5], and a e ained on he
Fig. 1. Anomaly de ec ion me hod. The igu e ep esen s a wo-dimensional da a
space, whe e no mal samples obse ed om he eal da a a e ep esen ed as g een
ci cles and a i icially gene a ed anomalous samples a e ep esen ed as ed iangles.
The g een egion is he eal egions whe e no mal samples may be ound, and he solid
black lines show he obse ed da a ange o no mal samples. No mal samples na u ally
concen a e in a ew egions while syn he ic anomalous samples sp ead ac oss he da a
space, which help in iden i ying he no mali y egion.
p e-p ocessed and il e ed aining da ase . C oss- alida ion was used o une
hype pa ame e s and selec he bes -pe o ming models.
Second, a no mali y de ec o is deployed o p edic he alues om a a as a
p oxy a iable o he absence o ailu es. A common challenge in anomaly de-
ec ion is he sca ci y—o comple e absence—o g ound u h labels o anoma-
lous e en s. In his s udy, while a p oxy a iable p o ides some indica ion o
anomalies, hese samples a e limi ed in di e si y and exhibi high empo al co -
ela ion, as hey span con iguous ime in e als. Consequen ly, hey ep esen
only a subse o possible anomaly ypes. In con as , g ound u h da a o no -
mal ope a ing condi ions is ypically abundan . To add ess his imbalance, he
p oposed me hod in oduces syn he ic anomalies ha span he en i e ea u e
space and le e ages he highe densi y and consis ency o no mal samples on
hei unde lying mani old o dis inguish anomalous om no mal beha io .
To his end, he inpu ea u es a e i s s anda dized using a quan ile scale
o no malize hei alue anges. Nex , we in oduce a no el i e a i e echnique
designed o compu e a sco e ha e lec s he likelihood o imminen ailu es:
1. Conside each o iginal da a alue, ep esen ing a no mal ope a ion sample,
as a ‘posi i e’ ins ance and pe u b i wi h Gaussian noise a 5% o he
ea u e de ia ion,
2. Syn hesize ‘nega i e’ samples ( ep esen ing abno mal ope a ion poin s) by
sampling alues uni o mly in he da a ange o each ea u e in he aining
da ase and assigning a ze o o hei a a pa ame e (i.e., ma king hem as
anomalous),
3. Use he newly c ea ed da ase o (i e a i ely) ain an XGBoos Classi ie [6]
model o p edic he nega i e class,
4. Repea o a ixed numbe o s eps o i he F1-sco e on aining does no
inc ease.
The XGBoos model is selec ed due o i s s ong pe o mance in classi ica ion
asks on abula da a and i s suppo o inc emen al (i e a i e) aining. In his
wo k, he model is ained o ou pu he p obabili y ha a gi en da a poin
co esponds o a pe iod o no mal ope a ion.
This se up, e e ed o as he No mali y De ec o (ND), is ained on a
da ase in which no mal and syn he ically gene a ed anomalous samples pa -
ially o e lap. Consequen ly, he model’s ou pu p obabili ies a e con ined o a
na owe sub ange wi hin [0, 1], a he han co e ing he en i e spec um. De-
spi e his comp ession, lowe p obabili y alues a e s ill in e p e ed as indica i e
o no mal ope a ing condi ions.
To p e en o e i ing o he syn he ic anomalies, a low-complexi y model
is used. This encou ages apid egula iza ion and be e gene aliza ion by e ec-
i ely isola ing egions o no mal ope a ion in he ea u e space, while ea ing
he emaining egions as anomalous.
2.4 Model E alua ion
The e alua ion o he models o p edic ing a ge a iables is pe o med on he
p e iously spli es ing da ase , whe e he da a is expec ed o be ep esen a i e
o no mal ope a ion. The no mali y de ec o is ins ead deployed on all a ailable
da a – ecall ha i was ained only on no mal da a om he ini ial 60% o he
a ailable ime.
3 P elimina y E alua ion
Fo e alua ion, his s udy ocuses on mul i a ia e ime-se ies da a collec ed om
a single wind u bine gene a o (WTG) in an onsho e wind a m ha has been
ope a ional since 2009. The da ase comp ises SCADA measu emen s eco ded
h oughou he yea 2016 a 10-minu e in e als. Whe e applicable, each ime
s ep includes s a is ical agg ega es such as minimum, maximum, a e age, and
s anda d de ia ion alues o key pa ame e s. Du ing he obse a ion pe iod, i e
majo gene a o ailu es we e eco ded, each lagged by high-se e i y ala ms.
To isually ep esen he a ailable da a, Figu e 2 shows he sca e plo o
he empi ical da a o ambien wind speed agains he WTG p oduced powe .
The da a do s in blue co espond o da a ha has been labeled by he ule-
based logic as no mal ope a ion poin s, while he ed do s show da a ou side
o he no mal ange. To con as his in o ma ion, he g een cu e shows he
heo e ically maximum powe ou pu acco ding o he endo speci ica ions. I
can be seen ha he no mal da a ollows he heo e ical cu e closely, while he
da a ou o ange alls a away om he expec ed cu e.
Fig. 2. Powe cu es in he da ase : Empi ical powe cu e in blue, endo -p o ided
heo e ical maximum capaci y powe cu es in g een, ed da a poin s do no co espond
o no mal ope a ion in e als.
3.1 P edic ion o a ge a iables
The s udy i s acknowledges he in o ma ion sha ed be ween di e en a iables
o implemen a nex -s ep p edic ion o mul iple a iables om a subse o all
a ailable a iables in he p e ious da a-poin . The a ionale is ha an accu a e
p edic ion o he nex da a poin (s) would allow o un he anomaly de ec ion
mechanism on he p edic ed da a and he e o e iden i y he anomalies in ad-
ance. A he same ime, i would con i m ha he e exis inhe en ela ionships
in he da a ha can be used no only o p edic ions, bu also o de ec ing
ou -o -dis ibu ion samples.
Figu e 3 shows he pe o mance o he HGBR model in p edic ing he a ge
a iable o he WTG powe ou pu , by plo ing he p edic ed ou pu ho izon-
ally and he ac ual da a e ically. I can be seen ha he sca e poin s all e y
close o he iden i y line (in g een), ep esen ing good quali y in he p edic ions.
3.2 P edic i e Main enance - anomaly de ec ion in ad ance
To illus a e he pe o mance o he no mali y de ec ion mechanism, Figu e 4
p esen s a ime se ies plo dis inguishing no mal and abno mal pe iods. The a i-
able a a, indica ing no mal ope a ion, is shown in g een, while ac ual eco ded
anomalies a e ma ked wi h dashed ed lines. The ou pu o he no mali y de ec-
o – a con inuous anomaly sco e – is smoo hed using an exponen ially weigh ed
Fig. 3. P edic ion and ac ual da a o wind u bine powe ou pu .
mo ing a e age (EWA) wi h a hal -li e o app oxima ely hal a week and is de-
pic ed by he o ange cu e. In his implemen a ion, highe sco e alues indica e
a g ea e likelihood o a ailu e o anomalous beha io in he nea u u e.
The exac h eshold o igge he ala m o upcoming anomalous da a can
be use -de ined acco ding o i s own isk a e sion, as long as i s ays be ween he
minimum and maximum obse ed alues, which in ou expe imen s we e 0.5 and
0.9 espec i ely. These alues a e induced by he smoo hing. A highe h eshold
igge s less alse posi i es bu may igge uncom o ably close o he ac ual
ailu e. Finding an op imal alue o he h eshold could possibly be o mula ed
as a cos op imiza ion p oblem and is le o u u e esea ch.
The in e p e a ion o said ou pu is ha lowe alues indica e pe iods o
no mal ope a ion, while he highe alues ep esen he p oximi y o anomalies.
I can be seen ha he no mali y p edic o is able o an icipa e he occu ence
o ailu es many days in ad ance o he ailu e e en s.
4 Conclusion
As he demand o enewable ene gy ises, p edic i e main enance is poised o
play a c i ical ole in ensu ing he eliabili y and e iciency o wind a m ope -
a ions. This s udy demons a es he e ec i eness o le e aging SCADA senso
da a and machine lea ning o ea ly aul de ec ion in wind u bines.
The p oposed me hodology in eg a es da a p ep ocessing, ea u e enginee -
ing, and ad anced model aining o p edic key ope a ional a iables and iden-
i y anomalies well in ad ance. By enabling ea ly in e en ion, his app oach
Fig. 4. Time se ies da a o he wind u bine. Speci ic ailu e poin s a e e ical dashed
ed lines. The g een cu e ep esen s no mal ime pe iods wi h alue 1 and abno mal
pe iods by p oximi y o ailu es wi h a 0. The o ange cu e is he EWA il e ed ou pu
sco e om he no mali y de ec o .
suppo s imp o ed u bine pe o mance, educed down ime, and mo e sus ain-
able wind ene gy p oduc ion.
Fu u e wo k will ocus on deploying he me hodology in eal-wo ld wind a m
moni o ing sys ems, expanding i o mul i- u bine scena ios, and inco po a ing
eal- ime eedback o u he enhance accu acy and esponsi eness.
Re e ences
1. Ma ía Isabel Blanco. The economics o wind ene gy. Renewable and sus ainable
ene gy e iews, 13(6-7):1372–1382, 2009.
2. Shu Liu, Siwei Ren, and Hongliang Jiang. P edic i e main enance o wind u bines
based on digi al win echnology. Ene gy Repo s, 9:1344–1352, 2023.
3. Yingying Zhao, Dongsheng Li, Ao Dong, Dahai Kang, Qin L , and Li Shang. Faul
p edic ion and diagnosis o wind u bine gene a o s using scada da a. Ene gies,
10(8):1210, 2017.
4. Wisdom Udo and Ya Muhammad. Da a-d i en p edic i e main enance o wind
u bine based on scada da a. IEEE Access, 9:162370–162388, 2021.
5. Guolin Ke, Qi Meng, Thomas Finley, Tai eng Wang, Wei Chen, Weidong Ma, Qiwei
Ye, and Tie-Yan Liu. Ligh gbm: A highly e icien g adien boos ing decision ee. In
I. Guyon, U. Von Luxbu g, S. Bengio, H. Wallach, R. Fe gus, S. Vishwana han, and
R. Ga ne , edi o s, Ad ances in Neu al In o ma ion P ocessing Sys ems, olume 30.
Cu an Associa es, Inc., 2017.
6. Tianqi Chen, Tong He, Michael Benes y, Vadim Kho ilo ich, Yuan Tang, Hyunsu
Cho, Kailong Chen, Ro y Mi chell, Ignacio Cano, Tianyi Zhou, e al. Xgboos :
ex eme g adien boos ing. R package e sion 0.4-2, 1(4):1–4, 2015.
