Vous êtes sur la page 1sur 6

IEEE TRANSACTIONS ON POWER APPARATUS AND SY STEM S, VOL. PAS- 8 9 , NO.

6 , JULY /AUGUST 1 9 7 0
Testing for the Cumulative Flashover Distribution
GORDON W. BROWN, M EM BER, IEEE
Abstrac t- M ethod s for minimiz ing the error in an estimate of the
voltage at a given p robability of flashover ( e.g., 3 sigma w ithstand ) VOLTAGE VS.
are p resented . The varianc e in the estimate of a voltage is minimiz ed POB8 LTER
w ith resp ec t to the ap p lied voltage and the number of ap p lic ations of
that voltage. It is show n that testing at ap p rox imately 5 - p erc ent and
9 5 - p erc ent p robability of flashover y ield s max imum information
about suc h q uantities as 2 sigma and 3 sigma w ithstand voltages, and
W
stand ard d eviation. I . - I .- - V .X 1 - .- I- .-
w 3 - 2 lI0 3
INTRODUCTION
NORM AL
VARIABLE
Y ( s
INTROD' UCTION ,
A is w ell k now n, there ex ists a relationship betw een the /
p robability
that a
given
insulation
sy stem
w ill flashover v
0
( e.g.,
transmission- line insulation
string)
and the
voltage ap p lied
to the insulation. If, as is usually d one, this relation is assumed
to be the c umulative normal d istribution, then, for a given volt-
- ~ ~~~~~~~ ~ ~~ ~ ~~ ~ ~~ ~ ~~~~~ I n ftn an go go
age w aveshap e
( 1 )
1 1 ( v- V5 o) /a
e2 2 d
p
=
2 jc o
e d y
w here
p p robability of flashover at given ( c rest) voltage V
V c rest of ap p lied voltage of given w aveshap e, k V
V5 0 c ritic al flashover voltage, k V
uf stand ard d eviation of flashover voltages, k V
y variable of integration.
Defining the normal freq uenc y func tion
e- y 2 /2
Z ( y )
=
e
and the stand ard normal variate
y
V - V6 0
( 3 )
then ( 1 ) bec omes
P
=
Z ( y d y .
_oD
This d istribution is show n in Fig. 1 .
M ost often, the d istribution is found for a given insulation and
given voltage w aveshap e by ap p ly ing voltage a number of times
at eac h of N voltage levels.
Pap er 7 0 TP8 4 - PWR, rec ommend ed and ap p roved by the Trans-
mission and Distribution Committee of the IEEE Pow er Group for
p resentation at the IEEE Winter Pow er M eeting, New Y ork , N.Y .,
January 2 5 - 3 0 , 1 9 7 0 . M anusc rip t submitted June 1 2 , 1 9 6 9 ; mad e
available for p rinting Dec ember 2 , 1 9 6 9 .
The author is w ith the Westinghouse Elec tric Corp oration, East
Pittsburgh, Pa. 1 5 1 1 2 .
Fig. 1 . Cumulative normal d istribution.
Let
N number of voltage levels
nk number of ap p lic ations of voltage at the k th voltage level
x k number of flashovers at the k thvoltage level.
The values of x k /nk c an be c onsid ered as estimates of the p roba-
bility of flashover at the k thvoltage level.
In a p revious p ap er [ 1 ] , the author has given a method for
( 2 ) d etermining
best
estimates
of a and
V5 o.
From
these,
an estimate
for the voltage VB at a given " base" p robability of flashover
PB c an easily be found . For ex amp le, the estimate of the 3
sigma w ithstand voltage w ould be found w ith
PB = 0 .0 0 1 3 5
c orresp ond ing to a value of the normal variate
Y B
=
- 3
w hic hy ield s the estimate ( read as " estimate of" )
VB
=
V5 0 + aCY B
= V0 - 3 a.
In the same p ap er, it is show n that the varianc e of the estimate
of V is given by
( 4 )
N
E rk ( Y k
-
Y B) 2
VV' BB= NN N 2
E rk Y k E
rk
-
( E rk Y k )
k =1 k = k =l
( 5 )
1 1 8 6
BROWN: CUM ULATIVE FLASHOVER DISTRIBUTION TESTING1
w here rk is a w eighting fac tor
rnk Z k -
rk =
-
Pk ( I
-
Pk )
Z k and Pk are d efined using ( 2 ) - ( 4 ) :
rY k
Pk J= Z ( y d y
Y k =
Vk
-
V5 0
( 6 )
( 7 a)
( 7 b)
e- Y k 2 /2
Z k
=
- 0
* ( 7 C)
The varianc e VVB is a measure of the error of an estimate of
VB.
WHERE TO TEST: THE PROBLEM
To minimiz e the error in the estimate of a d esired voltage
( e.g., 3 w ithstand , c ritic al flashover ( CFO) , etc .) , the varianc e
VVB ( i.e. ( 5 ) ) must be minimiz ed w ith resp ec t to the loc ation of
the test p oints: V1 , V2 , ..., V, , . Simultaneously , the varianc e
must be minimiz ed w ith resp ec t to the number of voltage ap p lic a-
tions at eac h of the voltage levels: ni at V1 , n2 at V2 , ..., n, at
VN. A fix ed number of total voltage ap p lic ations nt is the c on-
straint
N
nt Z F nk . ( 8 )
k =1
The solution to this op timiz ation p roblem is given in d etail in
the Ap p end ix . The results involve a p artic ular value of Y variate
Y o:
Y o= 1 .5 7 5
This c orresp ond s to the p robabilities
p ( - Y 0 )
=
5 .7 6
p erc ent
p ( + Y 0 )
=
9 4 .2 4 p erc ent.
The solution d evolves into tw o p arts as follow s.
TESTING FOR WITHSTAND
Case 1 : /Y B! > Y o
If the Y variate for the d esired voltage is greater in magnitud e
than Y o ( e.g., Y B = - 3 for the 3 sigma w ithstand voltage) , then
the error in the estimate of VB is minimiz ed w hen
N =2
Y ,
= - Y o c orresp ond ing to p = 5 .7 6 p erc ent
Y 2 = +Y o c orresp ond ing to p = 9 4 .2 4 p erc ent
ni
= 2 ( 1 - ( 9 a)
n2
=2 ( 1
+
) ( 9 b)
In view of the limits of p rac tic al engineering ac c urac y , 5 .7 6
p erc ent and 9 4 .2 4 p erc ent w ill henc eforth be referred to as 5 p er-
c ent and 9 5 p erc ent, resp ec tively . Consid er as an ex amp le esti-
mating the 3 sigma w ithstand voltage w ith 1 2 0 voltage ap p lic a-
tions:
nfl
=
2 - (
-1 - 5 ) 7 9 1
n2 =1 2 0 - 9 2 =2 9 .
TESTING FOR CFO
Case 2 :
/Y BI
< Y o
If the Y variate for the d esired voltage ( e.g., Y B = 0 for the
CFO, V5 0 ) is less than Y o in magnitud e, then the error in the
estimate of VB is minimiz ed if all test p oints are tak en at Y B.
If, for ex amp le, it is d esired to d etermine the CFO V6 o( Y B = 0 )
then the c riterion bec omes to mak e all ap p lic ations at V5 o.
This last c omment p oints up a d iffic ulty w hic h ap p lies for
both c ases 1 and 2 : to d etermine at w hat voltages to test ( and
thus d etermine the d istribution) , the d istribution must first be
k now n. In fac t, the same p roblem ex ists w hen using c onventional
testing p rac tic es ( four or more voltage levels, relatively evenly
sp ac ed betw een, roughly , 1 0 - p erc ent and 9 0 - p erc ent p roba-
bility of flashover) . In both ( c onventional and that of this p ap er)
the test voltages used must be c hosen based on p ast ex p erienc e
and /or on a few c onfirming voltage ap p lic ations; in other w ord s,
based on a " k now led ge" of the d istribution.
TESTING FOR STANDARD DEVIATION
It may be d esired to minimiz e the error in the estimate of a
stand ard d eviation, rather than the voltage at a sp ec ific p roba-
bility . The varianc e in the estimate of sigma is [ 1 ]
N
E: rk
va = U2
N= N ( N 2 ( 1 0 )
E rk Y k 2 E
rk - rk Y k J
k =1 k =1 k k =1
M inimiz ing this w ith resp ec t to the test voltages lead s to the
c onc lusion that eq ual numbers of voltage ap p lic ations should be
mad e at + Y 0 ( 5 - p erc ent and 9 5 - p erc ent p robability of flashover) .
COM PARISON OF M ETHODS
To illustrate the d ifferenc e betw een various method s, c onsid er
testing to d etermine a d istribution for w hic h p ast ex p erienc e
and /or a few voltage ap p lic ations suggest
V5 0 = 1 0 0 0 k V
=
0 .0 5 V5 0 = 5 0 k V.
Three c omp arisons are mad e, all w ith a total of 1 2 0 voltage
ap p lic ations.
M ethod a: Four voltage levels- 2 0 - p erc ent, 4 0 - p erc ent, 6 0 -
p erc ent, and 8 0 - p erc ent p robability of flashover.
M ethod b: Tw o voltage levels- using the test p roc ed ure w hic h
minimiz es the error in the estimate of the 3 sigma w ithstand
voltage.
M ethod c : Tw o voltage levels- ap p rox imately 5 - p erc ent and
9 5 - p erc ent p robability of flashover w ith eq ual number of shots
p er p oint, w hic h minimies the error in the estimate of stand ard
d eviation.
The d ata are given in Table I, in w hic h
Vk voltage for the k thvoltage level k V
X k ex p ec ted number of flashovers at voltage Vk
nk number of voltage ap p lic ations at voltage Vk .
1 1 8 7
IEEE TRANSACTIONS ON POWER APPARATUS AND SY STEM S, JULY /AUGUST 1 9 7 0
TABLE I
DATA FOR COM PARISON OF M ETHODS
M ethod k Vk X k nk
a 1 9 5 7 .9 8 4 0
2 9 8 7 .3 8 2 0
3 1 0 1 2 .7 1 2 2 0
4 1 0 4 2 .1 3 2 4 0
b 1 9 2 0 .1 5 9 1
2 1 0 7 4 .2 2 7 2 9
c 1 9 1 7 .8 3 6 0
2 1 0 8 2 .2 5 7 6 0
TABLE II
COM PARISON OF M ETHODS
9 0 - PERCENT CONFIDENCE LIM ITS
V5 0 3 Sigma Withstand
M ethod ( k V) ( k V)
a 1 0 .2 4 7
b 1 7 .4 3 3
c 1 5 .8 2 9
All p oints at Vs5 o 9 .4
6 !
je0
V5 0 tI0 5 0 K V E a /E |
\ A ~~~~~~~~~~~~~~n ,
n2
1 0
' - M INIM UM FOR~( BSTFOR3 a
tic ,
- 1 .!5 - 2 0 - 2 .5
NORM AL VARIABLE- Y
- 3 .5 - 4 .0
Fig. 2 . 9 0 - p erc ent c onfid enc e limits of voltage estimates for various
test method s.
The ap p rox imate 9 0 - p erc ent c onfid enc e limits of an estimate of
voltage VB are [ 1 ]
1 .6 4 5 /\v
k -
for any c orresp ond ing value of Y B ( or p robability PB) . The vari-
anc e v
- VB is c omp uted from ( 5 ) . The resulting c omp arison is
show n in Fig. 2 . Also ind ic ated in Table II is the minimum
p ossible error in V5 0 , w hic h w ould oc c ur ( see c ase 2 for
/Y BI
<
Y O)
if all 1 2 0
ap p lic ations
w ere tak en at Vw i.
It is noted that for
the latter c ase, ( 5 ) d egenerates to
2
vv=- -
2 nt
It is seen from Table II, w hic h is a c ond ensation of the p erti-
nent results, that w hen M ethod a ( w hic h the author c hose as
rep resentative of a c onventional test p roc ed ure) is used instead
of M ethod b, the 9 0 - p erc ent c onfid enc e band on the 3 sigma
w ithstand inc reases from about 5 8 k V to 9 4 k V. On the other
hand , the 9 0 - p erc ent c onfid enc e band in CFO d ec reases from 3 5
k V to 2 0 k V.
CHOICE OF TECHNIQ UE
It is rec ogniz ed that the d esired result of most testing is a good
estimate of the entire d istribution, inc lud ing both CFO and
w ithstand . Henc e, some sort of c omp romise must be mad e.
M ethod c ( testing at 5 p erc ent and 9 5 p erc ent, w ith about eq ual
number of shots p er p oint) is the method w hic h minimiz es the
varianc e of the estimate of sigma, and at the same time y ield s
a q uite ac c urate estimate of CFO ( + 1 6 k V out of a CFO of 1 0 0 0
k V is the 9 0 - p erc ent c onfid enc e band ) . Henc e, it may be a reason-
able c hoic e for a " good " estimate of the entire d istribution. On
the other hand , if CFO is d eemed c ritic al, then most shots should
be tak en as c lose as p ossible to CFO ( but, of c ourse, strad d ling
it) . If the 3 sigma w ithstand is the most imp ortant item then
( 9 a) and ( 9 b) should be used to d etermine test p oints, w hic h
results in
ni
= 0 .7 6 nt at
p i
;
5 p erc ent
n2
= 0 .2 4 nt at P2 9 5 p erc ent.
( lla)
( lIb)
DIFFICULTY OF APPLICATION
The same d iffic ulty is attend ant to any method c hosen; namely ,
that in ord er to d etermine w hat test voltages to ap p ly ( so that,
for ex amp le, ( 1 la) and ( 1 lb) are satisfied ) , the d istribution must
first be k now n. This rather severe req uirement is mitigated by
several fac ts. With " p ast ex p erienc e" and a few c onfirming
shots, an estimate ac c urate enough to begin is relatively easy to
mak e. Furthermore, as testing p roc eed s, and the d istribution be-
c omes more evid ent, the voltage c an be ad justed as d esired .
For ex amp le, sup p ose it is d esired to p lac e 6 0 shots eac h at
5 - p erc ent and 9 5 - p erc ent p robability of flashover, and the first
1 0 shots y ield ed flashovers ( 2 0 p erc ent) . If p ast ex p erienc e sug-
gests a stand ard d eviation of 5 p erc ent ( i.e., a
= 0 .0 5 V5 0 ) , an
immed iate read justment d ow nw ard in voltage is p ossible. Finally ,
it is often the c ase that a series of tests are being run, all mild ly
d ifferent. ( For ex amp le, grad ual variation of gap length or
variations of number of insulators in an insulator string.)
In
suc h c ases, it is relatively easy to d etermine test voltages w hic h
w ill be c lose to the d esired p erc entages.
CONCLUSIONS
1 ) Eq uations giving the test p oints w hic h minimiz e varianc e
in the estimate of the voltage at a given p robability are d evelop ed .
2 ) The results req uire an a p riori k now led ge of the d istribu-
tion. It is ind ic ated that this is not usually an insurmountable
p roblem.
3 ) If sole interest is to k now the voltage V at a p robability
PB outsid e the range of 5 - p erc ent to 9 5 - p erc ent p robability of
flashover ( suc h as the 3 sigma w ithstand at p
= 0 .1 3 p erc ent) ,
then testing should be d one as c lose as p ossible to 5 p erc ent and
9 5 p erc ent, w ith " most" shots at the p robability nearest PB ( e.g.,
most shots at 5 p erc ent for the 3 sigma w ithstand ) .
" - M IlNIM OUM FOR
ALL SHOTS AT 5 0 %
- .5 - 1 .0
1 1 8 8
BROWN: CUM ULATIVE FLASHOVER DISTRIBUTION TESTING
4 ) If sole interest is in a voltage VB at a p robability PB insid e
the range of 5 p erc ent to 9 5 p erc ent, then all shots should be
mad e as c lose as p ossible to the d esired p robability . For ex amp le,
if CFO w ere the sole interest, all testing should be d one as c lose
as p ossible to 5 0 - p erc ent p robability of flashover.
5 ) Usually , interest is in both the CFO and sigma, req uiring a
c omp romise. The p roc ed ure w hic h minimiz es error in the estimate
of the stand ard d eviation a is to p lac e an eq ual number of shots
at 5 - p erc ent and 9 5 - p erc ent p robability of flashover. This
p roc ed ure results in only a small error in CFO above that
ac hieved by p lac ing all test p oints at 5 0 - p erc ent p robability of
flashover.
6 ) While eac h test req uires its ow n d ec ision as to aims and
p roc ed ures, the author suggests that c onsid eration be given to
ap p ly ing voltages near to 5 - p erc ent and 9 5 - p erc ent p robability
of flashover, in an effort to red uc e errors, p artic ularly in the
stand ard d eviation.
APPENDIX
M INIM IZ ATION OF VARIANCE WITH RESPECT TO TEST POINTS
Eq uation ( 5 ) tak en from [ 1 ] is
N
E rk ( Y k
-
Y B) 2
VVB
k
2 /=1
V
V- a k =N N
N
2
E rk Y k 2 E rk - f rk Y k
k =1 k =1 \k =1 /
w here rk , Y k , and Y B are as d efined p reviously [ ( 5 ) - ( 7 ) and sur-
round ing tex t] . It is d esired that this varianc e be minimiz ed
w ith resp ec t to 1 ) the test voltages: V1 , * . * , Vk , * * * , Vn and 2 )
the number of voltage ap p lic ations at the test voltages: n1 ,
n2 , * * l* , nk , * * * , nN. Furthermore, the testing is restric ted to the
c onstraint of a given total number nt of voltage ap p lic ations:
N
E nk =nt
k 1 =1
Using the method of Lagrange multip liers [ 2 ] set
F =VEBn+X
-
( 1 2 )
w e have
0 .2 r' 8
=
- f- - ( Y ' S- Y B)
2
D
n,
N N
E
rk ( Y k - Y B) 2
E
_ k =1 k =1
D
( Y k - Y S)
2
w here
N N N 2
D =
E rk Y k E rk - E rk Y k )
k =l1 = =
It is easily verified that ( 1 7 ) may be w ritten
N N N
D =
E
rk ( Y k - Y B) 2
E rk - E rk (
k
k =l k =1 k =(
Y 2
Y B) )
w hic h is used in rew riting ( 1 6 ) to
O2 r'
x =
-
rS2 ( Y s - R) 2
D
n,
N
S
=
E rk ( y k
-
Y B)
i
N
R = -
E
rk Y k ( Y k
-
Y B) .
S k =1
( 1 6 )
( 1 7 )
( 1 8 )
( 1 9 a)
( 1 9 b)
( 1 9 c )
Nex t, aF/0 Y 8 = 0 is c onsid ered . After a great d eal of rela-
tively simp le d ifferentiation and algebraic manip ulation, the
follow ing results:
OF a
2
f
_,
_
=
_ S2
( Y s
-
R) 2
iY + 2 r8 ( Y ,
-
R) }
0 lY 8 D
2
a
fs
If R is treated as a c onstant, this may be w ritten
OF = 2 a
~~2 - r8 , ( Y ., - R) 2 =0 .
0 Y 8 D2 0 9 Y 8
( 2 0 )
( 2 1 )
It is noted that ( 1 9 ) - ( 2 1 ) are valid only if S i' , 0 . If S = 0 , the
( 1 3 a) results are
( 1 3 b)
and for eac h of the p arameters ns, V8 , set
OlF OiF
- - On - O=V8
, s = 1
2 ,- - ,
c ln, d lVs
( 1 4 )
Sinc e
y =
Vs- V5 0
( 1 4 ) c an be w ritten
OiF O1 F
- =0 ,
=0 , 8 ( 1 5 )
Noting that, in v - B, n, only oc c urs in
r,
and that
c , p / ao
N
E nk
=
1
c ln, d 9 ns k =1
N N
D
E rk E rk ( Y k
-
Y B) 2
k =1 k =1
2
E
rk T( Y /C
- Y B) 2
oa
r
k =l
Dn, N
E rk
k =1
OF
=_
ar.
N
)
aF
- E
rk ( Y k
-
Y B) 2
w hic h y ield s
( 2 2 a)
( 2 2 b)
( 2 2 c )
ar, ~i8 = 0 .
aY s
This is true only at
Y ,
= 0 , c . Henc e, this is a d egenerate
c ase.
The analy sis for S $ 0 is as follow s. Eq uation ( 1 9 a) is w ritten
X D
C =
ts( Y s
-
R)
2 =q 2 5
( 2 3 a)
1 1 8 9
IEEE TRANSACTIONS ON POWER APPARATUS AND SY STEM S, JULY /AUGUST 1 9 7 0
0
z
0
hi
It
w
x
1 -
z
s
- 3 _ - 2 _I___ 1 __3
0 .7
OA ~ ~ ~
"
0 .!
0 .5
0 .4
- 4 - 3 - 2 - I 0 . 3 4
NORM AL VARIABLE- Y
Fig. 3 . Weighting func tion t =
Z
-
p ) .
1 .2
I.0
0 .i
OAG
0 .2
- 4 - 3 - 0! 0 1 2 3 4
NORM AL VARIABLE- Y
Fig. 4 . Illustration for d erivation in tex t ( Y - 0 .5 ) 2 t.
w here
rs Z s2
ns p s( l- Ps)
( 2 3 b)
The p arameter t8 is show n in Fig. 3 . Note that
t, ( Y ,
-
R) 2
d ep end s only on
Y ,
and is the same for all s. Fig. 4 show s a p lot
of
t, ( Y ,
-
R) 2 , w ith R = 0 .5 for p urp oses
of illustration. It is
seen that at this p oint there are at most four p ossible values of
Y 8 s
at a given value of C. Eq uation ( 2 1 ) , how ever, restric ts the
slop e of the c urve to be z ero ( i.e., at the p eak s) . Then if R $
0 , there is only one solution at a given value of C.
If R =0 , then
d Y ( tlY s2 ) =
0 .
( 2 4 )
t, Y 8 2
is show n in Fig. 5 . Consid ering R = 0 first, ( 2 4 ) is satisfied
for
Y 2 = Y o
= 1 .5 7 5 0 3 0 1 , Y 1 = - Y 0 . ( 2 5 )
Q
z
2
I.-
IL
w
s
9 .,
0 .!
0 .4 I
0 .2
- 4 - 3 - 2 - I 0 1 2 3 4
NORM AL VARIABLE- Y
Fig. 5 . Illustration for d erivation in tex t Y 2 t.
From R
= 0 and ( 1 9 c )
N
E rk Y k ( y k
-
Y B)
= 0
from w hic h, w ithN
=
2 , Y 1 and Y 2 as above,
n
n
= t ( 1 -
2 Y B/
n2 =2 t 1 +
2 Y B,
( 2 6 a)
( 2 6 b)
Note that ( 2 6 ) c an only be satisfied ( i.e., nli > 0 , n2 > 0 ) for
Y BI
> Y o. For this c ase the varianc e
( 5 )
bec omes
A
2
1 Y B2
V
VB
=
- -
nt to Y 0 2
( 2 7 )
Here to is as in ( 2 3 b) evaluated at Y o.
If R 5 0 , there is only one solution. That is, only one test
p oint. Clearly this is at Y B. For this c ase, ( 5 ) red uc es ( w hen
treated as a limit as all test p oints ap p roac h Y B) to
U2
1
VVB
= - - -
nt tB
( 2 8 )
Eq uation ( 2 8 ) must be c omp ared w ith ( 2 7 ) to d etermine w hic h
lead s to the absolute minimum. The varianc e for one test p oint
( 2 8 ) is less than that for tw o test p oints ( 2 7 ) if
tBY 2 0 B2 > tY 0 2
w hic h is never true. Henc e, if tw o p oints y ield a p ossibility
( I Y BI > Y O) , the test p oints are as in ( 2 5 ) and ( 2 6 ) . Otherw ise
( IY BI
<
Y o) ,
one test
p oint,
at
Y B y ield s
minimum varianc e.
REFERENCES
[ 1 ] G. W. Brow n, " M ethod of max imum lik elihood ap p lied to the
analy sis of flashover d ata, " IEEE Trans. Pow er Ap p aratus and
Sy stems, vol. PAS- 8 8 , p p . 1 8 2 3 - 1 8 3 0 , Dec ember 1 9 6 9 .
[ 2 ] R. Weinstoc k , Calc ulus of Variations. New Y ork : M c Graw -
Hill, 1 9 5 2 .
1 1 9 0
BROWN: CUM ULATIVE FLASHOVER DISTRIBUTION TESTING
Disc ussion
J. H. M oran ( Interp ac e Corp oration, Leroy , N.Y .) : In a high- voltage
laboratory suc h as is op erated by the Lap p Insulator Division of
Interp ac e Corp oration, there are tw o sep arate and d istinc t func tions
p erformed . At any given time, the laboratory may be engaged in
either d evelop mental w ork or so- c alled " c ommerc ial" testing. In the
d evelop mental ty p e of w ork it has c ome to be the c ommonly ac -
c ep ted p roc ed ure to d etermine the c ritic al flashover values of the
test p iec e. As a rule, it is not nec essary to d etermine the w ithstand
values sinc e, fox r p rac tic al p urp oses, an assumed sigma of 5 p erc ent
p rod uc es a w ithstand value of suffic ient ac c urac y for the w ork in-
volved . There are only a few oc c asions on w hic h the w ithstand must
be d etermined d irec tly from test d ata.
For the c ommerc ial testing, how ever, the situation is q uite d iffer-
ent. If, for instanc e, a high- voltage sw itc h is the test objec t, the
p urc haser of the sw itc h w ill usually req uire that the manufac turer
show that the d esign of the sw itc h w ill meet c ertain minimum re-
q uirements. For the most p art, these minimum req uirements are
sp ec ified as a " w ithstand k V" at some given p erc ent of p robability
of flashover. In the p ast, the four- p oint method has been c ommonly
used . On a number of oc c asions, d ue to the vagaries of w eather and
other variables, it is nec essary to obtain more than four p oints from
w hic h to ad eq uately obtain the w ithstand values. The results are not
alw ay s satisfac tory .
In this resp ec t it is w orth rec alling that [ 3 1 , w hic h w as p ublished
for trial use, is the only d oc ument w hic h d efines a test p roc ed ure and
method of obtaining the w ithstand values for ex tra- high- voltage
sw itc hes. In this tec hnic al rep ort a method of mod ified linear regres-
sion is given by w hic h a w ithstand value c an be c omp uted . The major
fault in the d oc ument is the req uirement that the value of stand ard
d eviation ( sigma) d etermined by the d ata obtained is to be inc reased
by a given fac tor of 2 p erc ent. This fac tQ r is c omp letely arbitrary and
is d efined as " an allow anc e for unc ertainty in d etermination of d is-
p ersion." There have been a number of oc c asions on w hic h the use of
[ 3 1 has resulted in embarrassment to a sw itc h maniifac turer and to
the ultimate p urc haser of the sw itc h bec ause w hile the - d ata itself as
obtained w ould give w ithstand values ac c ep table to the p urc haser,
the use of the arbitrary fac tors c ontained in [ 3 1 w ould inc rease the
slop e of the p robability c urve to the p oint w here the w ithstand values
w ere below the ac c ep table limits.
In the p ast, there has been no p rop erly d oc umented method to
remove the imp asse c reated by this c ond ition. I am very hop eful
that as these oc c asions arise in the future, the information p resented
by Dr. Brow n, c an be emp loy ed for the benefit of all p arties c on-
c erned . In p artic ular, the c urves in Fig. 2 c an be used to aid in
mak ing the final d ec ision as to the ac c ep tability of the test ap p aratus,
to the great benefit of the ind ustry .
REFERENCES
[ 3 ] " Tec hnic al rep ort on sw itc hing surge testing of ex tra- high- voltage
sw itc hes, " IEEE Publ. 2 7 1 , Dec ember 1 9 6 6 .
N. Hy lten- Cavallius ( Hy d ro- Q uebec , M ontreal, P.Q ., Canad a) :
I just have tw o q uestions to the author of this interesting and w ell
w ritten p ap er.
In some c ases, the flashover d istribution c urve c an be very far
from Gaussian. How c an the method s be ad ap ted to suc h c ases?
For some test objec ts, suc h as those w ithboth internal and ex ternal
insulation, the manufac turers are very reluc tant to p ermit rep eated
ex ternal flashovers. Can the method s be mod ified to fac ilitate tests
on eq uip ment subjec t to suc h restric tions?
M anusc rip t rec eived February 9 , 1 9 7 0 .
Gord on W. Brow n: I thank the d isc ussers for their k ind remark s.
Regard ing M r. M oran' s c omments, I too have had oc c asion to stud y
[ 3 ] . I am not q uite as c onc erned w ith the method of c urve fitting,
how ever. With any reasonable amount of d ata, most c urve- fitting
tec hniq ues ( w ith some reasonable w eighting) w ill result in very
nearly the same fit. The 2 - p erc ent c orrec tion fac tor is, I agree, an
" imp asse" w hic h need s overc oming. M y c onc ern w ith [ 3 1 is the
restric tion that testing lie w ithin the band of 1 5 to 8 5 p erc ent. With
only 2 0 shots p er p oint, this severely limits the ac c urac y available for
a given number of shots. Consid er the ex amp le of [ 3 ] w ith test
voltages as follow s:
k V Flashovers/Shots
1 2 2 0 3 /2 0
1 2 5 0 4 /2 0
1 3 0 0 9 /2 0
1 3 5 0 1 6 /2 0
For this series, the 9 0 - p erc ent c onfid enc e limits on the 3 sigma w ith-
stand w ould be 7 0 k V. If, on the other hand , voltage had been
ap p lied near the 5 1 and 9 5 - p erc ent levels ( 2 /4 0 at 1 1 7 0 , 3 8 /4 0 at
1 4 3 0 ) , the 9 0 - p erc ent c onfid enc e limits on the 3 sigma w ithstand
w ould bec ome +5 0 k V for the same total number of ap p lic ations, a
substantial imp rovement. As a final c omment on M r. M oran' s
d isc ussion, I antic ip ate the nec essity for a more ac c urate d etermina-
tion of stand ard d eviation in d evelop mental as w ell as c ommerc ial
testing, p artic ularly as sy stem d esigns are based inc reasingly on
thoroughp robability stud ies.
In resp onse to Dr. Hy lten- Cavallius, the method c an be ap p lied to
flashover d istributions other than Gaussian. This has not been d one,
sinc e it req uires use of the method s outlined in [ 1 ] . For most d is-
tributions, this w ould req uire lengthy mathematic s, but the p ro-
c ed ure w ould be straightforw ard . Regard ing nond estruc tive testing,
the method of the p ap er is, unfortunately , not d irec tly ap p lic able.
It is relatively easy to formulate an ap p rox imation to the p roblem;
namely , to ad d an ad d itional c onstraint [ see ( 8 ) ] suc h that the total
number of flashovers is less than some fix ed number.
M anusc rip t rec eived February 5 , 1 9 7 0 .
1 1 9 1
M antisc rip t
rec eived M arc h
3 1 ,
1 9 7 0 .

Vous aimerez peut-être aussi