VICARED: A Neu al Ne wo k Based Sys em
o he De ec ion o Elec ical Dis u bances
in Real Time
Iñigo Monede o1, Ca los León1, Jo ge Rope o1,
José Manuel Elena1, and Juan C. Mon año2
1 Depa amen o de Tecnología Elec ónica,
Uni e si y o Se ille, Spain
{imonede o, cleon, jmelena}@us.es,
[email p o ec ed]
2 Consejo Supe io de In es igaciones Cien í icas,
Se ille, Spain
[email p o ec ed]
Abs ac . The s udy o he quali y o elec ic powe lines is usually known as
Powe Quali y. Powe quali y p oblems a e inc easingly due o a p oli e a ion
o equipmen ha is sensi i e and pollu ing a he same ime. The de ec ion and
classi ica ion o he di e en dis u bances which cause powe quali y p oblems
is a di icul ask which equi es a high le el o enginee ing knowledge. Thus,
neu al ne wo ks a e usually a good choice o he de ec ion and classi ica ion o
hese dis u bances. This pape desc ibes a powe ul sys em o de ec ion o
elec ical dis u bances by means o neu al ne wo ks.
1 In oduc ion
Powe Quali y (PQ) has been a esea ch a ea o exponen ial inc easing in e es
pa icula ly in he las wo decades [1]. I is de ined as he s udy o he quali y o
elec ic powe lines and has ecen ly sha pened because o he inc eased numbe o
loads sensi i e o powe quali y and become oughe as he loads hemsel es become
impo an causes o he deg ada ion o quali y [2].Thus nowadays, cus ome s demand
highe le els o PQ o ensu e he p ope and con inued ope a ion o such sensi i e
equipmen .
The poo quali y o elec ical powe is usually a ibu ed o powe line dis u bances
such as wa eshape aul s, o e ol ages, capaci o swi ching ansien s, ha monic
dis o ion and impulse ansien s. O en he g ea es damage om hese dis u bances
lies in he loss o c edibili y o he powe u ili ies on he side o hei cus ome s. The
classi ica ion and iden i ica ion o each one o he dis u bances is usually ca ied ou
om s anda ds and ecommenda ions depending on whe e he u ili ies ope a e (IEEE
in he Uni ed S a es, UNE in Spain, e c). Ou own classi ica ion, based on hese
s anda ds and ecommenda ions, is gi en in Table 1.
Table 1. Types o dis u bances
Range Type o
dis u bance
Sub ype o dis u bance Time
Min. alue Max. alue
Sligh de ia ion 49.5 Hz. 50.5 Hz
F equency Se e e de ia ion
10 s
47 Hz. 52 Hz.
A e age ol age 10 min 0.85 Un 1.1 Un
Flicke - - 7 %
Sho 10 ms-1s
Long 1s-1min
Sag Long- ime
dis u bance
> 1min 0.01 U 0.9 U
Sho < 3 min Unde -
ol age Long > 3 min 0.01 U
Tempo a y Sho 10 ms – 1s
Tempo a y Long 1s - 1min
Tempo a y Long-
ime dis u bance > 1 min
1.5 KV
Vol age
Swell
O e - ol age < 10 ms
1.1 U
6 KV
Ha monics - THD > 8 % Ha monics and
o he in o ma ion
signals In o ma ion signals - Included in he o he
dis u bances
2 A i icial In elligence on Powe Quali y
New and powe ul ools o he analysis and ope a ion o powe sys ems, as well as
o PQ diagnosis a e cu en ly a ailable. The new ools o in e es a e hose o
a i icial in elligence (AI) [1], including expe sys ems, uzzy logic and a i icial
neu al ne wo ks (ANNs) [3].
Fo he case o elec ical dis u bances, all he ac o s ha make ANNs a powe ul
ool a e p esen . We ge in o ma ion which is massi e – elec ical signals a e
cons an ly ecei ed – and dis o ioned – he e is an impo an noise componen .
In addi ion, he signal mus be p e-p ocessed o ge a ea u e ex ac ion by means
o wa ele ans o m and o he ma hema ical echniques which p o ide a unique
cha ac e is ic which can ep esen e e y single PQ dis u bance. I is ca ied ou by
means o a di e en esolu ions analysis using he echnique called mul i- esolu ion
signal decomposi ion o mul i- esolu ion analysis. In mul i- esolu ion analysis he
signal is decomposed in a se o app oxima ion wa ele coe icien s and ano he se o
de ail wa ele coe icien s.
The de ail coe icien s o he lowes le els s o e he in o ma ion om he as es
changes o he signal while he highes ones s o e he low- equency in o ma ion.
Thus, wi h he help o hese new ma hema ic ools he de ec ion o he elec ical
dis u bances has ended o be easy bu hei classi ica ion is s ill a di icul ask in
which ANNs play an impo an ole [4-11].
3 Neu al Ne wo k Real-Time Classi ie
We ha e de eloped a p o o ype o a eal- ime sys em o he de ec ion and
classi ica ion o elec ical dis u bances. The sys em is a de ec o o powe line
dis u bances whose de ec ion ke nel is based on a i icial in elligence echniques (in
pa icula , a i s e sion based on ANNs). The sys em consis s o a PC applica ion
which includes he AI ke nel and an acquisi ion ca d.
A. En i onmen
The en i onmen o he applica ion shows he in o ma ion which is acqui ed and
egis e ed by he sys em. I consis s o se e al windows whe e he acqui ed signal is
ep esen ed by means o he VRMS o he h ee signal phases, and a neu al. O he
windows show he las de ec ed dis u bance, a ba diag am ha epo s he numbe
and he ype o de ec ed dis u bances and a window wi h a his o ic which egis e s he
da e and ime o he di e en e en s.
We also ha e mo e op ions like a ba diag am epo ing a empo al g aphic iew o
he dis u bances, a mo e de ailed ep esen a ion o he las de ec ed dis u bance o a
iphasic diag am and ep esen a ion o he signal.
The acquisi ion ca d ob ains 640 samples e e y 100 milliseconds. These samples
a e shown on he cha and p ocessed by he AI ke nel. When one o mo e
dis u bances a e de ec ed in he 100 milliseconds, he co esponding egis e s a e
upda ed, changing he co esponding windows o he las dis u bance, he ba
diag ams and he his o ic.
B. Ke nel
In o de o ain he ANN, we ha e o gene a e he maximum possible numbe o
signals ep esen ing pa e ns o elec ical signals which include all he abo e-
men ioned dis u bances, so we ha e designed a signal gene a o wi h his aim. In ac ,
we ha e gene a ed o e 27,000 signals including one-dis u bance signals and wo-
dis u bance signals. The de ec ion sys em uses Wa ele ans o m o he acqui ed
signal o he gene a ion o signal ea u es [4-10]. The aim o ea u e ex ac ion by
Wa ele ans o ms is o p o ide a unique cha ac e is ic which can ep esen e e y
single PQ dis u bance.
The inpu ec o s o he ANN a e gene a ed ca ying ou a numbe o ope a ions on
he Wa ele ans o m. I is known ha he Wa ele ans o m de ec s be e he low-
equency componen s in he las de ail le els and as a ia ions in i s le els. Thus,
ou solu ion is based on he concep ha he ampli ude dis u bances would be be e
de ec ed in he i s le els o Wa ele ans o m while he equency dis u bances
would be be e de ec ed in he las le els. The e o e, we decided o use pa allel
neu al ne wo ks as i is shown on Figu e 1.
The signal is p e-p ocessed using he wa ele ans o m as i has been said abo e.
The esul o his a e he inpu s o all he ANNs. Fi s o all, hese inpu s a e gi en o
he dis u bance de ec o ANN, which ou pu is ei he 0 - no dis u bance - o 1 -
dis u bance -. I he e is a dis u bance, he ANN inpu s a e gi en o ano he h ee
Dis u bance ANN
F equency ANN
Vol age ANN
Ha monics ANN
P e-p ocessing
Signal
ANN
Inpu s
Dis u bance
No dis u bance
Vol age
dis u bance
F equency
dis u bance
Ha monics
dis u bance
Fig. 1. Block diag am
ANNs, each one specialized in he de ec ion o a di e en ype o dis u bance. In he
same way, he ou pu s o hese ANNs a e 0 o 1, depending on he ac ha he e is o
no ha kind o dis u bance. These ypes a e, acco ding o Figu e 1 abo e, ol age
dis u bances – sags, swells, unde ol ages and o e ol ages -, equency dis u bances –
sligh and se e e de ia ions – and ha monics dis u bances.
The exis ence o a dis u bance de ec o ANN p e ious o he o he ANNs is due o
he g ea e impo ance o he de ec ion o dis u bances compa ed wi h he classi ica ion
o hem. Besides, he dis u bance ANN ac s as a il e o he nex ANNs. Some o he
possible mis akes commi ed by he h ee pa allel speci ic neu al ne wo ks a e
elimina ed by he dis u bance de ec o ANN. The eason o using se e al ANNs and
no only one is ha a unique ANN wi h se en ou pu s – one o e e y ype o
dis u bance – needs oo many neu ons o wo k p ope ly and, consequen ly, mo e
memo y esou ces.
C. Neu al Ne wo k building
To be able o ain each neu al ne wo k, i s ly we mus decide he con enien lea ning
me hod. The e a e basically wo lea ning me hods mainly used, supe ised lea ning
and sel -o ganised lea ning. The i s one ge s inpu da a and associa es hem wi h a
de e mined ou pu while he second one makes i s own inpu da a classi ica ion. Fo
ou case, we wan a di e en ou pu acco ding o he exis ence o no o a dis u bance
o i s ype, depending on he chosen neu al ne wo k. Tha is he eason why a
supe ised lea ning wo ks be e . Besides, we ha e o choose a pa icula ype o
supe ised lea ning. The bes op ion is he backp opaga ion (BP) lea ning me hod
using he mul ilaye pe cep on due o i s be e a e be ween simplici y and e iciency.
Impo an ea u es o he neu al ne wo ks a e he s udy o he necessa y inpu
alues, he neu al ne wo k s uc u es, ans e unc ions and lea ning algo i hms.
In pa icula , we ha e used he ollowing alues as inpu ec o o he ampli ude
neu al ne wo k: he V
RMS
o he signal, he in eg al, he maximum and he V
RMS
o he
de ail wa ele coe icien s o 1, 2 and 3 le el. In o de o ge a as e con e gence and
be e esul s hese da a we e scaled so ha minimum is -1 and maximum is 1.
The chosen kind o ANN is a mul ilaye pe cep on wi h 3 hidden laye s wi h
di e en numbe o neu ons, depending on he ANN and i s numbe o ou pu s. The
ou pu unc ions o he laye s ha e been chosen wi h a loga i hmic sigmoid ans e
unc ion o all he laye s.
All he inpu s, s uc u es, unc ions and aining algo i hms ha e been eached a e
es ing wi h di e en ones. The bes esul s un il now ha e been ob ained o neu al
ne wo ks shown in Table 2.
Table 2. Neu al Ne wo k s uc u e
Neu al
Ne wo k ype
Numbe o hidden
neu ons
Numbe
o ou pu s
T ans e unc ions T aining algo i hm
Dis u bance 20, 14 & 8 1 loga i hmic sigmoid Le enbe g-Ma qua d
Vol age 12, 9 & 6 4 loga i hmic sigmoid Le enbe g-Ma qua d
F equency 16, 12 & 7 2 loga i hmic sigmoid Le enbe g-Ma qua d
Ha monics 20, 14 & 8 1 loga i hmic sigmoid Le enbe g-Ma qua d
D. P og amming Tasks
Fo p og amming asks we ha e used he MATLAB ool o es he di e en
possibili ies in he p e-p ocessing o he signal and in he s uc u e o he ke nel. We
used his ool due o he powe ul oolboxes wi h specialized unc ions con ained in i
u ilizing he signal and he wa ele oolboxes o he p e-p ocessing ask and he
neu al ne wo ks oolbox o he design o he ke nel [11].
Once we ca ied ou he es and ound a good code o he p e-p ocessing and he
AI ke nel, we p og ammed hem in C++ language in o de o op imize he execu ion
ime. The es s ca ied ou in execu ion ime abou he p e-p ocessing ime a e a ound
he 0.1 milliseconds o he wa ele ans o m.
Fo he design and p og amming o he ool en i onmen he selec ed ool has been
Bo land C++ Builde 5 which is a powe ul ool o he de elopmen o isual
applica ions as well as a obus C++ compile .
4 Resul s
Be o e embedding he ke nel in he classi ie ool we selec ed he bes aining
me hod o he con igu a ion o ANNs. Thus, o he aining o he ne wo ks we used
80% o he gene a ed signals as aining pa e ns and 20% as es pa e ns. On he
o he hand, h esholds we e de ined in he ANN ou pu s in o de o dis inguish i a
pa icula ou pu alue may be conside ed as a dis u bance o no . The de ined
h esholds we e 0.3 and 0.7 and hus, ou pu alues abo e 0.7 we e conside ed as
dis u bances and below 0.3 ideal signals. Values ound be ween 0.3 and 0.7 we e
aken as e o s in he de ec ion o he inpu pa e n. The dis ance be ween he ou pu
ne wo k and he desi ed alue was de ined as a sa e y coe icien in he de ec ion.

We a e going o conside wo di e en kinds o esul s: he gene al ones, i is o
say, he pe cen age o success in e e y ANN – see Table 3 - , and he pa icula ones,
which a e mo e in ui i e and conside some pa icula cases o ailu e in one o he
neu al ne wo ks. In addi ion, we ha e he esul s o only one dis u bance signals and
he esul s o wo-dis u bance signals which include he one-dis u bance signals oo.
Table 3. One-dis u bance signal esul s
Type o ANN Numbe o
ou pu s
Tes
signals
Numbe o
e o s
Co ec ly
de ec ed %
Dis u bance 1 334 1 99.70
Vol age 4 334 20 94.01
Ha monics 1 334 1 99.70
F equency 2 334 5 98.50
The i s conclusion we ob ain is ha he highe numbe o ANN ou pu s we use,
he highe numbe o e o s we ge . This is due o he highe complexi y in oduced
by he necessi y o ixing all he ou pu s a he same ime. The second conclusion is
ela ed o he in luence o hese e o s. As said abo e, he mos impo an ANN is he
one which de ec s dis u bances – he exis ence o no o a dis u bance is much mo e
impo an han i s ype – so we ha e ocused ou e o s on i s co ec wo king. On he
one hand, we mus say ha hese pe cen ages a e e e ed o all he es ing signals bu
we also ha e o bea in mind ha some o he signals ha ail a e il e ed by he
dis u bance ANN. On he o he hand, we ha e o analyze wha kind o signals end o
ail. To illus a e his poin we ha e conside ed some o he signals ha ail in he
ol age ANN – able 4-.
Table 4. E o s in analysis
Real signal De ec ed e en
Ideal signal wi h a small 9% sag Sag
10.8% and 11 ms o e ol age Swell
99.7% and 35 ms unde ol age Sag
98% and 10 ms sag Unde ol age
Ideal signal wi h a small 8% o e ol age O e ol age
80% Swell and 9 ms O e ol age
97% and 10 ms sag Ideal signal
Analogously, esul s a e simila o o he signals and o he ANNs. We obse e ha
he signals which ail a e nea he limi o a dis u bance, so i is no a big mis ake o
conside hem as he neu al ne wo k ells us.
In able 5, we ha e he esul s ob ained o wo-dis u bance signals, wi h
app oxima ely 27700 signals, abou 5500 o he es and he es o aining.
T aining pe o mance o he dis u bance ANN is shown in Figu e 5.
Table 5. Two-dis u bance signal esul s
Type o ANN Numbe o
ou pu s Tes signals Numbe o
e o s
Co ec ly
de ec ed %
Dis u bance 1 5523 70 98.73
Vol age 4 5523 575 89.58
Ha monics 1 5523 57 98.97
F equency 2 5523 278 94.97
Conclusions a e simila o he ones we ha e achie ed o one-dis u bance signals.
Resul s a e sligh ly wo se due o he g ea e complexi y o he signals.
5 Conclusions
Wha we ha e de eloped is a eal- ime sys em o he de ec ion o elec ical
dis u bances based on a i icial neu al ne wo ks. Wi h his sys em we a e capable o
de ec ing he exis ence o no o dis u bances and hei ype wi h a e y high
possibili y o success and bea ing in mind ha mos o he mis akes a e commi ed
wi h no e y common signals in eal li e – hose in he edge o a dis u bance.
The use o C++ language makes i possible o achie e he objec i e o making ou
sys em a eal- ime one. This may allow elec ical companies o de ec dis u bances
wi h ime enough o ind possible oubles and ake s eps o a oid u he p oblems.
Ou cu en wo k is ocused on ca ying ou es s wi h he sys em wo king in eal
ime in he powe line in o de o imp o e ou esul s wi h eal signals. Ano he line
o ou in es iga ion is he s udy o he u iliza ion o di e en s uc u ed pa allel
ne wo ks o he same ype using a o ing sys em which will allow us o achie e be e
esul s.
Re e ences
[1] W.R. Anis Ib ahim and M.M. Mo cos, “A i icial In elligence and Ad anced
Ma hema ical Tools o Powe Quali y Applica ions: A Su ey”, 668 IEEE T ansac ions
on Powe Deli e y, Vol. 17, Ap il 2002.
[2] W.E. Kazibwe and H.M. Sendaula, “Expe Sys ems Ta ge s. Powe Quali y Issues”,
IEEE Compu . Appl. Powe , ol. 5, pp 29-33, 1992.
[3] G. Zheng, M.X. Shi, D. Liu, J. Yao, Z.M. Mao, “Powe Quali y Dis u bance
Classi ica ion Based on Rule-Based and Wa ele -Mul i-Resolu ion Decomposi ion”,
P oceedings o he Fi s In e na ional Con e ence on Machine Lea ning and Cybe ne ics,
Beijing, 4-5 No embe 2002.
[4] A.Elmi wally, S. Fa ghal, M.Kandil, S. Abdelkade and M.Elka eb, “P oposed wa ele -
neu o uzzy combined sys em o powe quali y iola ions de ec ion and diagnosis”, IEEE
P oc-Gene . T ans. Dis ib. Vol. 148 No 1, pp. 15-20, Janua y 2001.
[5] P. K. Dash, S. K. Panda, A. C. Liew, B. Mish a, and R. K. Jena, “New app oach o
moni o ing elec ic powe quali y,” Elec . Powe Sys . Res. ol. 46, no. 1, pp. 11–20,
1998.
[6] A.K. Ghosh and D. L. Lubkeman, “The classi ica ion o powe sys em dis u bance
wa e o ms using a neu al ne wo k app oach”, IEEE T ans. Powe Deli e y. Vol. 10. pp.
671-683, July 1990.
[7] J.V. Wijayakulasoo iya, G.A. Pu us and P.D. Minns, “Elec ic powe quali y dis u bance
classi ica ion using sel -adap ing a i icial neu al ne wo k”, IEEE P oc-Gene . T ans.
Dis ib. Vol. 149 No. 1, pp. 98-101, Janua y 2002.
[8] R. Daniels, “Powe quali y moni o ing using neu al ne wo ks,” in P oc.1s In . Fo um
Applica ions Neu al Ne wo ks Powe Sys ., 1991, pp. 195–197.
[9] S. San oso, J. P. Edwa d, W. M. G ady, and A. C. Pa sons, “Powe quali y dis u bance
wa e o m ecogni ion using wa ele -based neu al classi ie - Pa 1: Theo e ical
ounda ion,” IEEE T ans. Powe Deli e y, ol. 15, pp. 222–228, Feb. 2000.
[10] S. San oso, J. P. Edwa d, W. M. G ady, and A. C. Pa sons, “Powe quali y dis u bance
wa e o m ecogni ion using wa ele -based neu al classi ie —Pa 2: Applica ion,” IEEE
T ans. Powe Deli e y, ol. 15, pp. 229–235, Feb. 2000.
[11] M. Mallini and B. Pe unicic, “Neu al ne wo k based powe quali y analysis using
MATLAB,” in P oc. La ge Eng. Sys . Con . Powe Eng., Hali ax, NS, Canada, 1998, pp.
177–183.