www.ietdl.

org

Published in IET Science, Measurement and Technology

Received on 26th May 2009
Revised on 11th September 2009
doi: 10.1049/iet-smt.2009.0048

ISSN 1751-8822

Parameters optimisation for surge arrester

circuit models
C.A. Christodoulou1 L. Ekonomou2 G.P. Fotis1 P. Karampelas3
I.A. Stathopulos1
1
School of Electrical and Computer Engineering, High Voltage Laboratory, National Technical University of Athens,
9 Iroon Politechniou Street, Zografou Campus, Athens 157 80, Greece
2
Department of Electrical Engineering Educators, A.S.PE.T.E. School of Pedagogical and Technological Education,
N. Heraklion, Athens 141 21, Greece
3
IT Faculty, Hellenic American University, 12 Kaplanon Street, Athens 106 80, Greece
E-mail: [email protected]

Abstract: The appropriate modelling of metal oxide (MO) surge arresters and the determination of the
equivalent circuit parameters are signicant issues, since the efcient representation of the arresters dynamic
behaviour by the models improves the quality and reliability of the lightning performance studies. In the
current work, an optimisation algorithm for the evaluation of the surge arrester models parameters is
developed. Optimum arrester model equivalent circuit parameter values are selected, minimising the error
between the simulated peak residual voltage value and this given by the manufacturer. The proposed
methodology is applied on a 120 kV MO arrester, presenting its effectiveness, reducing the relative error
between the simulated and manufacturers peak residual voltage value around 50% with the use of the
obtained optimum parameter values.

1 Introduction non-linearity of the V I characteristic. The SiC resistors

non-linearity of the V I characteristic is considerably less
Lightning strokes, switching actions and short-circuits often than this in an MO resistor, thus arresters with silicon
cause temporary overvoltages and high-frequency current carbide resistors need serial spark gaps, which separate the
oscillations. Surge arresters are devices that are connected arrester from the line during continuous operation and
between the phase and the earth and they protect interrupt the power frequency follow the current that ows
distribution and transmission lines and their equipment through the arrester after a discharge operation. MO surge
from external and internal overvoltages, providing a low- arresters, besides the more convenient V I characteristic,
impedance path to ground for the overvoltage current. The have simpler design, improved protective characteristics and
voltage current (V I ) characteristic of a surge arrester is faster time response [2, 3].
non-linear, that is, the voltage at the terminals increases
above a certain limit proportionally less than the increase in The basic parts of a MO surge arresters are the cylindrical
the current. The greater the non-linearity, the narrower the MO resistor blocks, the insulating housing and the electrodes
range of the residual voltage of the arrester [1 3]. (Fig. 1). Between the varistor column and the polymeric
housing there is a glass bre structure that either completely
Surge arresters can be classied as gapped or gapless, encloses the resistor blocks or exerts sufcient force on the
depending on whether or not they have an external series ends of the stack to hold the MO blocks rmly together.
gap. Nowadays, gapped arresters, consisting of SiC varistor
in series with gaps, have been replaced by metal oxide Laboratory results indicate that MO surge arresters have
(MO) arresters without gaps, presenting higher dynamic characteristics and behave differently, depending

86 IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92

& The Institution of Engineering and Technology 2010 doi: 10.1049/iet-smt.2009.0048
 www.ietdl.org

Figure 2 IEEE model [9]

Figure 1 MO surge arresters cut (1, electrodes; 2,

breglass; 3, non-linear resistor; 4, external insulator)

on the form of the transient wave. For fast front surges, with
rise times less than 8 ms, the peak of the voltage wave occurs Figure 3 Pianceti Gianettoni model [10]
before the peak of the current, and the residual voltage across
the arrester increases as the time to crest of the arrester
discharge current decreases. This increase of the residual
voltage could reach approximately 6% when the front time
of the discharge current is reduced from 8 to 1.3 ms. This
frequency-dependent behaviour requires a more
sophisticated model for lightning overvoltages than the
simple static non-linear resistance, which can be used only
for switching surge studies. Figure 4 Fernandez Diaz model [11]

The appropriate modelling of a surge arrester is circuit models and in Table 1 the equations for the
signicant for insulation coordination studies, in order to parameters computation for each model.
extract reliable estimations. For this reason, several
frequent-dependent models of MO surge arresters, using
The Pinceti Gianettoni [10] and the Fernandez Diaz
the physical and the electrical data given by the
models [11] are simpler forms of the IEEE, using in
manufacturer, have been proposed, in a way that the
addition to the dimensions, the physical characteristics data
model simulation results correspond to the actual
obtained from manufacturers datasheet.
behaviour of the arrester [4 8]. The existing models
differ in the parameters estimation procedure, but all are
efcient to simulate the frequency-depended behaviour of Table 1 Models parameters computation
the arresters. Main issue for the suggested models is the
determination of the parameters of each model, in a way IEEE Pinceti Gianettoni Fernandez
that the simulated curve ts as possible the real recorded Diaz
waveform. To this direction, the current work suggests an R0 (100d )/n V 1 MV 1 MV
optimisation algorithm for the circuit parameter values
evaluation, which uses the initial values obtained by R1 (65d )/n V
adjustment methods dened for each one model, in order L0 (0.2d )/n mH n L00
1 Vr(1=T2 )  Vr(8=20)
to minimise the error between the computed and the  Vn
manufacturers residual voltage curves. 4 Vr(8=20)

L1 (15d )/n mH
1 Vr(1=T2 )  Vr(8=20)
 Vn
12 Vr(8=20)
2 MO surge arrester models
C (100n)/d pF 100/d pF
The most used frequently-depended arrester models that
represent efciently the arresters performance are: the Vn is the arresters rated voltage, Vr(8/20) is the residual
IEEE model [9], the Pinceti Gianettoni [10] and the voltage for a 8/20 10 kA lightning current, Vr(1/T2) is the
Fernandez Diaz model [11]; the last two models are based residual voltage for a 1/T2 10 kA lightning current, n is a
on the IEEE model, but they differ in the parameters scale factor, d is the length of arrester column in metres,
computation. In Figs. 2 4 are shown the three examined L 00 is the inductance obtained from diagrams [11].

IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92 87

doi: 10.1049/iet-smt.2009.0048 & The Institution of Engineering and Technology 2010
www.ietdl.org

3 Problem analysis iteratively updates the worst point by four operations:

reection, expansion, one-dimensional contraction and
The predicted residual voltage for an injected impulse current multiple contraction. Fig. 5 illustrates these operations in
of each model must be as much closer to the measured three-dimensional variable space.
residual voltage (manufacturers datasheet) for the same
current. Considering that the accuracy of the computed It must be mentioned that the downhill simplex method
residual voltage depends on the model parameters, these starts with N 1 points and not just with a single point. If
can be determined by minimising the function [7, 8] one of these points is selected randomly to be the initial
T starting point P0 , then the other N points can be extracted using
e [Vc (t, x)  Vm (t)]2 dt (1)
0 Pi P0 lei (5)

where T is the duration of the impulse current, x is a column where ei are N unit vectors and l is a constant, corresponding to
vector containing the parameters x1 , x2 , . . . , xn of each the problems characteristic length scale.
model, Vc(t, x) is the computed residual voltage and Vm(t)
is the measured by the manufacturer residual voltage. Reection involves the moving of the worst point (vertice)
of the simplex (highest value of the objective function) to a
point reected through the remaining N points. The
4 Formulation of the optimisation method uses a series of steps, most of them have occurred
problem just moving the point of the simplex, where the function is
maximum (highest point), through the opposite face of
The three models presented in this paper are consisted of the simplex to a lower point. If this point is better than the
different parameters whose values may have various values. best point, then the method attempts to expand the
The IEEE model [8] has ve parameters (R0 , R1 , L0 , L1 , simplex along this line. This operation is called expansion.
C ), the Pianceti Gianettoni model [10] and the On the other hand, if the new point is not much better
Fernandes Diaz model [11] have three parameters (R0 , than the previous point, then the simplex is contracted
L0 , L1) and (R0 , L0 , C), respectively. The objective is the along one dimension from the highest point. This
minimisation of the relative error of (1), which is a function procedure is called contraction. Moreover, if the new point
of three or ve variables depending on the model. is worse than the previous points, the simplex is contracted
along all dimensions towards the best point and steps down
The parameters (R, L, C ) of each model form a column the valley. By repeating this series of operations, the
vector x method nds the optimal solution. Termination criteria can
be dened in any multidimensional minimisation routine.
x [x1 , x2 , x3 , x4 , x5 ]T [R0 , R1 , L0 , L1 , C]T Using more than one independent variable, there is no
(2)
(IEEE model [8]) longer the option of requiring a certain tolerance for a
single independent variable, so it can be typically identied
x [x1 , x2 , x3 ]T [R0 , L0 , L1 ]T one step of the multidimensional algorithm. The
(3)
(Pianceti-Gianettoni model [10]) optimisation procedure is then terminated, when the vector
distance is moved in that step and it is fractionally smaller
x [x1 , x2 , x3 ]T [R0 , L0 , C]T in magnitude than some tolerance.
(4)
(Fernandes-Diaz model [11])

Application of an optimisation algorithm will determine the

5 Optimisation algorithm
optimal values xi . The goal of the optimisation is to The algorithm for the downhill simplex method is applied on
minimise the relative error, which is a function of several (2) (4), in order to minimise (1), which describes the
variables. e(x) is the relative error, where x is the column function (error) that needs optimisation. The algorithm is
vector presented in (2)(4). The minimum cannot be applied on (1) three times, one for each one of the three
reached in a full step, but requires a step analysis, namely an models described in (2) (4).
iteration technique. Many algorithms exist to perform, such
as the gradient methods [12, 13], that is, NewtonRaphson, The downhill simplex method repeats a series of the
LevenbergMarquard and quasi-Newton methods and the following steps:
direct search methods [14], that is, simplex of Nelder and
Mead. The method that will be applied on this paper is the 1. Order and re-label the N1 points so that
downhill simplex method [15], an optimisation algorithm e(xN1) . . . .. e(x2) . e(x1).
that requires only function evaluations, and not calculation of
derivatives. In the N-dimensional space, a simplex is a 2. Generate a trial point xr by reection, such that
polyhedron with N 1 vertices. The N 1 points are xr xg a (xg 2 xN1), where xg is the centroid of the N
chosen and an initial simplex is dened. The method best points in the vertices of the simplex.

88 IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92

& The Institution of Engineering and Technology 2010 doi: 10.1049/iet-smt.2009.0048
 www.ietdl.org

Figure 5 Four basic operations in the downhill simplex method (xh , highest point; xl , lowest point; xg, centroid of N points)

3. If e(x1) , e(xr) , e(xN), replace xN1 by xr . In Tables 35 are shown the initial computed parameters for
each one model, according to the computations described in
4. If e(xr) , e(x1), generate a new point xq by expansion, Section 2, as well as the optimum parameter values obtained
such that xq xr b(xr 2 xg). using the optimisation algorithm described in Section 5.

5. If e(xq) , e(xr), replace xN1 by xq , otherwise replace xN1 Figs. 6 and 7 show the computed residual voltage curves
by xr for each one model in a common diagram, using the initial
and the optimised parameter values, respectively. In
6. If e(xr) . e(xN), generate a new point xc by contraction, Table 6 is given the peak value of the residual voltage for
such that xq xg g(xN1 2 xg). each one model (for initial and optimised values) and the
relative errors (for initial and optimised values) comparing
7. If e(xc) , e(xN1), replace xN1 by xq .
Table 2 Electrical and insulation data of arrester
8. If e(xc) . e(xN1), contract along all dimensions towards
x1 . Evaluate the objective function values at the N new MCOV 120 kV
vertices, xi x1 h(xi 2 x1).
rated voltage 150 kV
Standard (efcient) values of a, b, g and h are a 1.0, maximum residual voltage with 5 kA 367 kV
b 1.0, g 0.5 and h 0.5. lightning current 8/20 ms
10 kA 396 kV
20 kA 449 kV
6 Simulation results height 1330 mm
Simulations were performed for each one model using the
insulation material porcelain
PSCAD software program for a one-column 120 kV MO
arrester, with height 1330 mm. In Table 2 are given the creepage 3320 mm
electrical and insulation characteristic data of the used surge
energy capability 8 kJ/kV
arrester.

IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92 89

doi: 10.1049/iet-smt.2009.0048 & The Institution of Engineering and Technology 2010
www.ietdl.org

Table 3 Parameters for the IEEE model where Vrsimulation is the residual voltage from the simulation
and Vrmanufacturer is the residual voltage given by the
Parameters Initial values Optimised values manufacturer datasheet.
L1(mH) 19.95 21.58
As it can be observed the IEEE model presents the lower
R1(V) 86.45 92.04 error, offering higher accuracy, mainly because of its
L0(mH) 0.267 0.167 complexity in comparison to the other models.
Furthermore, it is obvious that the use of the optimum
R0(V) 133 115.6 parameter values in the equivalent circuit models reduce
C(pF) 75.2 75.8 signicantly the error between the manufacturers and the
simulated residual voltage, something really very important
since the arresters models can be more reliable for the
insulation coordination studies, representing more efcient
Table 4 Parameters for the Pinceti Gianettoni model
the MO arresters behaviour and resulting in more precise
Parameters Initial values Optimised values analysis.
L1(mH) 3.21 4.28 All the examined models reproduce efciently the residual
L0(mH) 1.07 0.32 voltage of the arresters, since the relative errors are low
enough. After algorithm implementation, the relative error
R0(MV) 1 1 for all models is reduced almost to half. The residual
voltage values of the IEEE and the Fernandez Diaz
models are very close, something that means that practically
there is no big difference to the results for an insulation
Table 5 Parameters for the Fernandez Diaz model
coordination study, if the one or the other model is chosen.
Parameters Initial values Optimised values The decision about which model will be used is based on
the available data, the complexity of the system and the
L1(mH) 0.48 1.02 demanded speed of the simulation.
C(pF) 75.18 68.25
R0(MV) 1 1 7 Application to a 150 kV
transmission line
these values to the manufacturers waveform, for the nominal Surge arresters equivalent circuit models are used for
discharge current of the arrester, the peak value of lightning insulation coordination studies in high-voltage electrical
current impulse, which is used to clarify an arrester [16]. networks. In order to show the importance and the
effectiveness of the optimised parameters for each model,
The relative error is computed by the equation the three different models are implemented in a 150 kV
high-voltage transmission lines of the Hellenic system,
using the initial and the optimised parameter values
Vrsimulation  Vrmanufacturer computing the arresters failure probability for each model
e 100% (6)
Vrmanufacturer according to the procedures described in [17, 18]. The

Figure 6 Computed residual voltages using the initial values for each one model

90 IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92

& The Institution of Engineering and Technology 2010 doi: 10.1049/iet-smt.2009.0048
 www.ietdl.org

Figure 7 Computed residual voltages using the optimised values for each one model

Table 6 Residual voltages and relative errors for each one model

Nominal discharge current, 10 kA Residual voltage, kV Relative error, %

Model Initial parameters Optimised parameters Initial parameters Optimised parameters
IEEE 401.30 398.25 1.33 0.71
Pinceti Gianettoni 387.85 391.88 22.05 21.04
Fernandez Diaz 402.75 399.30 1.70 0.83
datasheet 396.00

Table 7 Arresters failure probability

R, V PIEEE,initial, % PIEEE,optimised, % PPinceti,initial, % PPinceti,optimised, % PFernandez,initial, % PFernandez,optimised, %

10 0.56 0.54 0.50 0.52 0.55 0.53
20 0.81 0.77 0.73 0.76 0.79 0.76
40 1.24 1.19 1.12 1.17 1.21 1.16
60 1.63 1.57 1.48 1.53 1.59 1.54
80 2.05 1.97 1.86 1.93 1.94 1.87

arresters are applied on each tower of the line and is parameters selection. Models with optimised parameters
performed sensitivity analysis with the tower footing reproduce more adequately the arresters residual voltage,
resistance. Table 7 presents the arresters failure probability and should be preferred, in order to extract more accurate
obtained using the three dynamic models, before and after results.
the optimisation procedure.

The arresters failure probability is affected by the 8 Conclusions

residual voltage waveform and its peak voltage; for this In this work, an optimisation method was developed in order
reason the calculated failure probabilities change in optimum surge arrester equivalent circuit model parameters
function with the residual voltage after parameter to be determined. The minimisation of the error between
optimisation. It is obvious that the appropriate choice of the computed and the manufacturers residual voltage for a
the parameters plays important role to installation given injected impulse current was the criterion of the
coordination simulations, since the output voltage of the optimisation. The obtained results prove the efciency of
arresters varies, depended on the used model and its the method and the optimum obtained parameters improve

IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92 91

doi: 10.1049/iet-smt.2009.0048 & The Institution of Engineering and Technology 2010
www.ietdl.org

the behaviour of the models, with the simulated residual [9] IEEE Working Group 3.4.11: Modeling of metal oxide
voltage to t better the manufacturers one. The proposed surge arresters, IEEE Trans. Power Deliv., 1992, 7, (1),
methodology can be proved very valuable in the pp. 302 309
achievement of more accurate and reliable results in
transient analysis and lighting performance studies. [10] PINCETI P., GIANNETTONI M.: A simplied model for zinc
oxide surge arresters, IEEE Trans. Power Deliv., 1999, 14,
(2), pp. 393 398
9 References
[11] FERNANDEZ F., DIAZ R.: Metal oxide surge arrester model
[1] KUFFEL E., ZAENGL W.S., KUFFEL J.: High voltage engineering
for fast transient simulations. Int. Conf. Power System
fundamentals (Butterworth-Heinemann, 2000, 2nd edn.)
Transients IPAT01, Rio De Janeiro, Brazil, 2001, paper 14
[2] IEEE Std 1243-1997: IEEE guide for improving the
[12] FORSYTHE G.F. , MALCOLM M.A., MOLER C.B. : Computer
lightning performance of transmission lines, 1997
methods for mathematical computations (Prentice Hall,
1976)
[3] IEC 60099-4: Surge arresters part 4: metal oxide
surge arresters without gaps for a.c. systems, 2nd edn.,
[13] GILL P.E., MURRAY W., WRIGHT M.H.: Numerical linear algebra
2004 2005
and optimization (Addison Wesley, 1991, 1st edn.)
[4] ITNIK B., BABUDER M.: The ability of different simulation
[14] SHANNO D.F.: Conditioning of quasi-Newton methods for
models to describe the behavior of metal oxide varistors.
function minimization, Math. Comput, 1970, 24,
Twenty-eighth Int. Conf. Lightning Protection, Kazanawa,
pp. 647 656
2006, pp. 1111 1116
[15] MORDECAI A. : Nonlinear programming: analysis and
[5] BAYADI A., HARID N., ZEHAR K., BELKHIAT S.: Simulation of
methods (Dover Publishing, 2003)
metal oxide surge arrester dynamic behavior under fast
transients. Int. Conf. Power Systems Transients, New
[16] IEC 60099-4: Surge arresters part 4: metal oxide
Orleans, USA, 2003
surge arresters without gaps for a.c. systems:
(2004 2005, 2nd edn.)
[6] SULJANOVIC N., MUJCIC A., MURKO V.: Practical issues of
metal oxide varistor modeling for numerical simulations.
[17] MONTANES L., GARCIA-GRACIA M., SANZ M., GARCIA M.A.: An
Twenty-eighth Int. Conf. Lightning Protection, Kazanawa,
improvement for the selection of surge arresters based
Japan, 2006
on the evaluation of the failure probability, IEEE Trans.
Power Deliv., 2002, 17, (1), pp. 123 128
[7] ABDELHAFID B.: Parameter identication of ZnO surge
arrester models based on genetic algorithms, Electr.
[18] CHRISTODOULOU C.A., GONOS I.F., STATHOPULOS I.A.: Lightning
Power Syst. Res., 2008, 78, pp. 1204 1209
performance of high voltage transmission lines protected
by surge arresters: a simulation for the Hellenic
[8] LI H.J., BIRLASEKARAN S. , CHOI S.S.: A parameter
transmission network. Proc. 29th Int. Conf. Lightning
identication technique for metal-oxide surge arrester
Protection (ICLP 08), Uppsala, Sweden, 2008, pp. 6b-7-
models, IEEE Trans. Power Deliv., 2002, 17, (3),
1 6b-7-8
pp. 736 741

92 IET Sci. Meas. Technol., 2010, Vol. 4, Iss. 2, pp. 86 92

& The Institution of Engineering and Technology 2010 doi: 10.1049/iet-smt.2009.0048