You are on page 1of 4

Non Linear Regression Model to Predict

Flas over of Nonceramic Insulators


S. Venkataraman, Student Member, IEEE, R. S. Gorur, Fellow IEEE
Department ofElectrical Engineering, Arizona State University, Tempe AZ, USA

breakers in the circuit [1], [2].


Abstract-- This work is based on the application of non-linear There are numerous models that have been developed to
regression analysis technique to develop a model for prediction of predict the flashover voltage of ceramic insulators (porcelain,
flashover voltage of non-ceramic insulators (NCIs). The glass). These models are based on the contamination level of
superiority of nonlinear model over a linear model is
demonstrated. Surface resistance measurement and flashover
gqu
experiments were performed with field aged samples. This work Salt Deposit Density). The models developed for ceramic
finds applications in distribution class insulators and can be insulators cannot be directly applied for nonceramic insulators
extended for higher voltage class of insulators. due to factors like location and aging playing a major role in
Index Terms- Flashover, Gauss Newton iteration, Regressor causing flashover in NCIs [1]-[3].
variable, Response variable, Surface resistance The surface of NCIs is observed to change, as it has a
tendency to interact with the contaminant with time due to
aging phenomena. Consequently, the surface conductance of
I. INTRODUCTION NCIs is dependent on these factors apart from conductivity of
JIGH voltage insulators form an essential part of high the contaminant. This necessitates the need for an alternative
lvoltage electric power transmission and distribution approach to characterize NCIs. Measurement of surface
systems. Any failure in the satisfactory performance of resistance by leakage current measurement in laboratory under
insulators would result in considerable loss of capital as there wet conditions has been shown to be a successful approach
are numerous industries such as semiconductors, mining etc and is adapted in this work [4].
that depend upon the availability of an uninterrupted high
quality power supply. The principle dielectrics used for II. REGRESSION ANALYSIS TECHNIQUE
outdoor insulators are either ceramics or polymers. Ceramic Regression analysis is a widely used statistical technique
insulators are made up of either porcelain or glass. However, for modeling the relation between different variables. The
since the late 1960's, polymers have been preferred over applications of regression are in numerous fields such as
porcelain and glass by many utilities for reasons such as sciences engineering, medicine and management. The
economics, light weight, easy handling and installation, and minp
deinapcs'] regression models find itS uses such as in data descriptihon,
.

flxblt in flexibility in design aspects [1].control, and parameterr estimation. Statistical variables can be
Polymers due to their hydrophobic nature (resistance to eithro an reraesr aespon.S variable swr the
water film formation) have a much better flashover regressor variable iS the surface resistance. The response
performance over ceramic insulators. The term flashover can variable is theah rvage Tesstatistical mode
be defined as disruptive electric discharge over or around the velope in thiswor is laied Tpedictinthe aover
insulator. NCIs are observed to lose their hydrophobicity with vol ef ifferen levels ofmsurfaceresistance[5].
time and in these conditions when exposed to moisture in the Regesson models d ar etrflie rn-inear.mo.
els The
Regression are either linear or non-linear. The most
form of dew, fog it gives rise to leakage currents. Leakage generic form of linear regression model is given as (1)
current formation promotes dry band discharges (arcs) in the
insulator. Flashover occurs when the arc propagates and Y /+I x & (1)
ultimately bridges about two thirds of insulator leakage where,
length. The various factors that contribute to flashover can be y - is the response variable
summarized as contamination accumulation, aging (location Po - is the intercept
of insulators, UV radiation) and material composition. The Pl- is the slope
extent of power interruption due to insulator flashover x- regressor variable
depends on the operation of associated relays and circuit - is the error

sraerssac
relationship ob
In this work the sosre between flashover voltage and
o iera tflosa
The authors acknowledge Power Systems Engineering Research Center
(PSERC) for sponsoring this work.
exponential trend. As the model is not linear in the unknown
1-4244-0228-X/06/$20.00 2006 IEEE parameters a nonlinear regression analysis technique is used in
663
this scenario. The generic form of the nonlinear model is measurements for the same level of surface resistance [7].
given by (2). As we have an exponential trend between the
variables the equation takes the formI of (3) [5].

y=f(X, CP)e (2)


ie 02 & (3)
where,
f(x, (P)- is the expectation function
(PI, Z-are unknown parameters
The proposed regression model is true only when certain
assumptions are satisfied. The three assumptions to be
checked are given as follows
* Normal distribution of errors. The residuals should
lie approximately in a straight line. If this is satisfied,
then the data points can be taken to be noirmally
distributed. Fig. 1 Insulators in field a) Ceramic b) Non-ceramic
* Zero mean and constant variance of errors. If the
residual plot is observed to be structureless, the HV power
assumption of constant variance is satisfactory.supyTt
* Uncorrelation of errors. For an ideal case, a insulator
horizontal band would enclose all residuals and these
would randomly fluctuate within the band.
In the models developed in this work the validity of
assumptions were checked and appeared satisfactory [5], [6].
111. EXPERIMENTAL DETAILS Shunt

In this work field aged silicone rubber and EPDM samples Ultrasonic
(5 years of exposure in midwest USA) were used. All the nebulizers
experiments were carried out in a fog chamber with
dimensions of about 4mX4mX3m high. Ultrasonic nebulizers
were used for fog generation as these proved to have
considerable advantages over conventional fog generators like
easy to maintain, more efficient and corrosion free. The IV. PARAMETER ESTIMATION IN NONLINEAR SYSTEM THROUGH
surface resistance measurement and flashover experiments LINEARIZATION AND GAUSS NEWTON ITERATION TECHNIQUE
were performed after ensuring 100 00 relative humidity [7].
In order to estimate the parameters of nonlinear system the
To measure surface resistance an ac voltage of the range of first step is to linearize the function and perform Gauss
2-4 kV was used. The leakage distance of the insulators Newton iteration procedure. Linearization is performed by a
between the electrodes was chosen as 15 cm. The voltage
applied was just adequate to establish a measurable leakage T
current and not high enough for discharges to be initiated. It is
advisable to have a vertical orientation of insulators and was
Aol 030.....0pO)
00 = 0020 and retaining the linear terms

followed in this work. The leakage current was measured as a


voltage drop across the resistor which was connected in series (x, 0) )(4)
with the insulator (lkQ). A steady value of surface resistance
was obtained in an hour after starting the test. Fig. 1 shows the
ii
If we assign the following
'0 a-j 0 = Oo
picture of insulators in field. Fig. 2 shows the schematic for 0
surface resistance measurement [4, 7]. fj = f(xi n 00 (5)
After surface resistance was evaluated, in order to estimate 0 = - ) (6)
the flashover voltage (for all NCIs), the voltage was raised to
8000 of the probable flashover voltage (which was obtained Fo q5)(X
, o17
from previous trials) for 15 minutes. If there was no flashover Zj -
the voltage was raised in steps of 100%, and each step was -L (9S (7)0
maintained for 5 minutes until flashover of insulator. The Here fS0 is the initial parameter. From above assignments
flashover voltage reported is the average of three th nolna moe .slnaie n ipiida 5 8
664
Yo =Z0l0
+'6A (8) 12
The estimate of A6o is also termed as vector of increments. ,T
It is also thatf,o = 0- 00 and in general for the kth iteration 10 N
we can generalize as [5]
+1= O+ (9)
k k + (Zk Zk ) Zk (y fk) (10) I
where, 7

Zk =z ] 6

fk 1k
=4 . fkj5
fi,1'1f2k,,, , 5s

i [Olk4 02k . ''..... pk] 0 20 40 60 80 100 120 140 160 180 200

The steps from (4) to (10) were repeated iteratively until


convergence which is when the difference of two successive Iterations
/k iS < 10-6. A matlab program was developed to solve the Fig.4 Convergence of (,
non-linear parameters (DI, (D2. The flow chart is given in Fig. 26
3. 2.4

2.2

1.8

Input values of surface resistance


as x, and flashover voltage as y. qD2 1.4

1.2 i_

Initialize :DI, ')2 to random values ] 1.2

Compute the following 0.61 20 0 40 60 80 100 120 140 160 180 200
f In ((P1) + (P2 X;
,Ln (y)-f; Iterations
l |z,<l := 1/ 01 ; Fig. 5 Convergence of (D2

zo12 -= x TABLE 1. ITERATIONS TAKEN FOR DIFFERENT INITIAL VALUES

IIIIIIIIIIIIILiIIIIIIIIIIII Initial values Iterations Final values


Compute P, Pf, (D], CD2- using (9) L(, | '_2 l to converge (DI (2 |
and (1O) 5 0.1 80 11.56 0.662
I 20 1 75 11.56 0.662
12 0.6 60 11.56 0.662

V. STATISTICAL MODELING

No Yes Based on the above values of (DI, (D2, Minitab was used for
Stop
further statistical analysis. Models with and without intercept
Stop are provided as the constant of intercept model was not
Fig.3 Flow chart of matlab program [5] significant. The results are given as follows
Depending upon the initial values chosen the number of
iterations to converge to obtain <Al, CD2 vary. The sample
number of iterations taken is shown in Table I. Figs 4 and 5
show the convergence <>l, D22

665
TABLE II. COMPARISON OF LINEAR AND EXPONENTIAL MODEL
Intercept model for FOV vs SR
FOV - 0.01 + 11.6 exp (0.662)*SR Parameter Linear Exponential % Improvement

Predictor Coef SE Coef T P 2.52 146 42.1


Constant -0.014 1.5 -0.01 0.99 S 0.51 0.44 13.7
exp(0.662)*SR 11.6 1.1 10.5 0.00 Average width of Cl 1.5 0.9 40

S = 0.49 R-Sq = 96.5% R-Sq(adj) = 95.7% Average width of Pi 1 3.8 | 2.6 31.6
PRESS = 2.4 R-Sq(pred) = 91.45%
25
Analysis of Variance UL at 95 % PI
Source DF SS MS F P 20 Meanof UL at 95 % Cl
Regression 1 27.3 27.3 110.9 0.000 LL,
Residual Error 4 0.98 0.25
Total 5 28.3 -5-L t 5%P

0 LL at 95 % C~~~~~LLa95%P
No intercept model for FOV vs SR 10LL at95%
FOV = 11.6 exp (0.662)*SR 10
PredictedI
validated
Predictor Coef SE Coef T P 5 Experimental values region
Noconstant
exp(0.662)*SR 11.6 0.13 86.6 0.00 0
S = 0.44; PRESS = 1.46 0 0.2 0.4 0.6 0.8 1 1.2 1.4
Analysis of Variance Surface resistance (Mohm/cm)
Source DF SS MS F P Fig.6 Plot of surface resistance vs FOV; CI - Confidence Interval, PI -
Regression 1 1474.0 1474.0 7497.3 0.000 Prediction Interval, LL- Lower Limit, UL - Upper Limit
Residual Error 5 1.0 0.2
Total 6 1475.0 VI. CONCLUSION

Where, In this work non-linear regression has been used to predict


SE coef- Standard error coefficient flashover voltage of (NCIs) from surface resistance
T -Standard "T" Statistic measurement and flashover experiments performed with field
P- Probability of testing the significance of null hypothesis aged samples. The superiority of nonlinear model compared to
S - Standard deviation a linear model is demonstrated through improved statistical
F- Standard "F" statistic parameters. This work was performed for distribution class
R - Residual sum of squares insulators and can be extended for higher voltage class of
R2 (adj) -Adjusted residual sum of squares insulators in future.
- (pred) - Predicted residual sum of squares

PRESS - Prediction error sum of squares REFERENCES


DF - Degrees of
SS Degrees
- of freedom
freedom
[1] J.T. Burnham, P.S. Givens, T.M. Grisham, "High strength polymer post
SSSum of squares insulators enable economical transmission lines with low environmental
MS - Mean sum of squares impact," Proc. of IEEE Pow. Engg. Society, pp. 494 -503, April 1994.
In general a high value of R2 (adj), indicates the model can [2] B. Hampton, "Flashover mechanism of polluted insulation", Proc. IEEE,
explain the variability in a wide range. A high value of R2 Vol. 111, pp 985-990, 1964.
[3] J. P. Holtzhausen "A critical evaluation of AC pollution flashover models
(pred) with a low value of PRESS indicates the high capability for HV insulators having hydrophilic surfaces," Ph.D. Dissertation, Dept of
in predicting the variability in new observations. A high "F" Elec Engg, Univ of Stellenbosch, S. Africa, 1997
ratio and low "p" value means the model is highly significant. [4] R.S. Gorur, etal "Surface resistance measurements on non-ceramic
[5], [6]. insulators," IEEE Trans. Pow Del, V.16, pp. 801-805, 2001.
[5] D. C. Montgomery, E. A. Peck, G. G. Vining, "Introduction to linear
The developed model is validated by performing an regression analysis," 3rd edition, New York, John Wiley and Sons, 2001
experiment in the region of projection. Fig. 6 shows the graph [6] D. C. Montgomery "Design and analysis of experiments," 6th edition,
of surface resistance to flashover voltage. The superiority of New York, John Wiley and Sons, 2004
choosing exponential model over linear model is demonstrated [7] 5. Venkataraman, R. S Gorur, "Prediction of flashover voltage of non
Table~~~~~~ ~ ~IIitcnb.bevdta contaminated conditions," # 1158, Accepted for
through as h

parameters (ideally to be as low as possible) for exponential


ttsia
~ ~ceramic
~ ~ ~ ~ ~ ~ publicinsulators
ations in IEEEunder
Trans. on DEIS, 2006.

model is lower than linear model.

666

You might also like