MODELING THE CRITICAL FLASHOVER VOLTAGE OF HIGH

VOLTAGE INSULATORS USING ARTIFICIAL INTELLIGENCE

M. Belkheir i1, B. Zegnini1 & D. Mahi1
1
Laboratoire d’études et Développement des Matériaux Semi-conducteurs et Diélectriques, LeDMaScD,
Université Amar Telidji de Laghouat, BP 37G route de Ghardaïa, Laghouat 03000, Algérie

Abstract: This work attempts to apply artificial intelligent techniques in high voltage
applications and especially to estimate the arc parameters for polluted insulators
during the pollution flashover mechanism, using experimental measurements carried
out in an insulator test station according to the IEC norm and a mathematical model
based on the characteristics of the insulator: the diameter, the height, the creepage
distance, the form factor and the equivalent salt deposit density and estimates the
critical flashover voltage. First, Genetic Algorithm approach (GAs) is used to find the
optimal parameters of the mathematical model that describes this phenomenon. Then
two types of artificial neural networks (ANN) are designed to establish a nonlinear
model between the above mentioned characteristics and the critical flashover voltage.
The GAs and ANNs models, algorithms, and tools have been developed using the
software package Matlab. The obtained results are promising and insure that artificial
intelligent techniques can help High voltage Engineers to understand more deeply and
estimate the critical flashover voltage for new designed insulators with different
operating conditions and constitute an indispensable models that can be used in field
simulations of arc parameters for polluted insulators. Further comparative analysis of
the estimated results with the measured data collected from the site measurement
amply demonstrate the effectiveness of the use of Artificial intelligent techniques for
optimization (Gas) and modeling (ANNs) of critical flashover voltage and leakage
current against contamination level in terms of σs the surface conductivity.
Keywords: Critical Flashover Voltage, Polluted Insulators, Artificial intelligent
techniques, Genetic algorithms (GAs), Artificial Neural Networks (ANNs), High
Voltage, Simulation, Measurements.

1. INTRODUCION
The reliability of the power system mainly depends on the environmental and weather
conditions which cause flashover on polluted insulators leading to system outages. It is
generally recognized that the main causes leading to the contamination of insulators are
marine pollution-found in the immediate neighborhood of the coastal regions and solid
pollution-found in the dense industrial areas. A major problem of insulation systems is the
accumulation of airborne pollutants due to natural, industrial or even mixed pollution, during
Corresponding author: b.zegnini@mail.lagh-univ.dz bakzegni@yahoo.fr
 138 Journal of Intelligent Computing and Applications (JICA)

the dry weather period and their subsequent wetting, mainly by high humidity. At the coastal
areas the high voltage insulators are affected by salt particles that settle on the insulators
surfaces. The winds that blow from the sea carry the salt particles. These particles are not
dangerous in its dry condition but with high environmental humidity or drizzle rain conditions
the salt can absorb the water and form a thin film with high conductivity. This layer gives
an ideal path for the leakage current to pass through between the high voltage side and the
ground side. The conductivity of this layer depends on the type of salts which this layer
consists of [1], [2]. High failure rate of polluted insulator due to the flashover has been
found near the coastal areas [3]. This problem was the motivation for the installation of a
test station performs laboratory tests on artificially polluted insulators.
Experiments concerning the critical flashover voltage are time-consuming and have
further obstacles, such as high cost and the need for special equipment. This has resulted in
the development of several approaches for the estimation of the flashover voltage on
polluted insulators. Most are based on circuit models for the calculation of the analytical
mathematical relationship for either dc or ac flashover voltage on polluted insulators.
This work attempts to utilize the available experimental data and the results of a
theoretical approach, in order to construct a model that can estimate the arc parameters of
polluted insulators using some geometrical characteristics of the insulator: the diameter,
the height, the creepage distance, the form factor and the equivalent salt deposit density
and estimates the critical flashover based on Artificial intelligent techniques.
The proposed approach will help in the establishment of maintenance policy and for
addressing an effective solution against pollution flashover of high voltage insulators.

2. MATHEMATICAL MODEL
The simplest model that has been developed by Obenhaus [4] consists of a partial arc
bridging the dry zone and the resistance of the polluted wet zone in series (Figure 1).

HV Arc Residual resistance

Figure 1: Obenhaus Model

The mathematical model for the evaluation of the flashover process of a polluted
insulator consists of a partial arc spanning over a dry zone and the resistance of the
pollution layer in series, as shown in Figure. 2, where Varc is the arcing voltage, Rp the
resistance of the pollution layer and U a stable voltage supply source. The critical voltage
 Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 139

Arc

Varc Rp

Figure 2: Equivalent Cir cuit

Uc (in V), which is the applied voltage across the insulator when the partial arc is
developed into a complete flashover, is given by the following formula [5]:
−𝑛𝑛
𝐴𝐴
𝑈𝑈𝑐𝑐 = (𝐿𝐿 + 𝜋𝜋𝜋𝜋𝐷𝐷𝑚𝑚 𝐹𝐹𝐹𝐹)(𝜋𝜋𝜋𝜋𝐷𝐷𝑚𝑚 𝜎𝜎𝑠𝑠 )𝑛𝑛 +1 (1)
𝑛𝑛+1
where
L is the creepage distance of the insulator (in cm),
Dm the maximum diameter of the insulator disc (in cm) and F is the form factor. A and
n are the arc constants.
The surface conductivity σs (in Ω-1) is given by the following equation:

𝜎𝜎𝑠𝑠 = (369.05𝐶𝐶 + 0.42) × 10−6 (2)

where C is the is the equivalent salt deposit density (in mg/cm2).

The coefficient of the pollution layer resistance K in case of cap-and-pin insulators is

𝑛𝑛 +1 𝐿𝐿
𝐾𝐾 = 1 + ln � � (3)
2𝜋𝜋𝜋𝜋𝜋𝜋 2𝜋𝜋𝜋𝜋𝜋𝜋

where R is the radius of the arc foot (in cm) and is given by
1
𝑅𝑅 = 0.46(𝜋𝜋𝐴𝐴𝐷𝐷𝑚𝑚 𝜎𝜎𝑠𝑠 )2(𝑛𝑛 +1) (4)
The above mathematical model is a result of experiments in specific insulators types
and specific pollutants in their surface. There are many values for the arc constants A and
n in the literature [6] as a result, the above mathematical model could be applied with
satisfactory accuracy in specific insulator types and pollutants.

3. GENETIC ALGORITHMS
According to the literature high voltage transmission line problems are faced using
conventional analytical methods, which include in most cases empirical and/or
approximating equations. Artificial intelligence and more specifically Genetic Algorithm
approach (Gas) are addressed in section, in order to give accurate solutions to high voltage
 140 Journal of Intelligent Computing and Applications (JICA)

transmission line problems using in the calculations only actual field data.
A simple GA relies on the processes of reproduction, crossover, and mutation to reach
the global or “near-global” optimum. To start the search, GAs require the initial set of the
points Ps, which is called population, analogous to the biological system. A random number
generator creates the initial population. This initial set is converted to a binary system and
is considered as chromosomes, actually sequences of “0” and “1”. The next step is to form
pairs of these points that will be considered as parents for a reproduction. Parents come to
reproduction and interchange Np parts of their genetic material.
This is achieved by crossover. After the crossover, there is a very small probability Pm
for mutation. Mutation is the phenomenon where a random “0” becomes “1” or a “1”
becomes “0.” Assume that each pair of “parents” gives rise to Nc children. Thus, the GA
generates the initial layouts and obtains the objective function values. The above operations
are carried out and the next generation with a new population of strings is formed. By the
reproduction, the population of the “parents” is enhanced with the “children,” increasing
the original population since new members are added. The parents always belong to the
considered population. The new population has now Ps+Nc/2 members. Then, the process
of natural selection is applied. According to this process, only Ps members survive out of
the Ps+Nc/2 members. These Ps members are selected as the members with the lower
values of Fg since a minimization problem is solved. Repeating the iterations of
reproduction under crossover and mutation and natural selection, GAs can find the
minimum of Fg. The best values of the population converge at this point. The termination
criterion is fulfilled if either the mean value of Fg in the Ps members population is no
longer improved or the number of iterations is greater than the maximum number of
iterations.
This paper proposes a methodology, which uses the developed GA for the calculation
of the arc parameters for polluted insulators. The operation of GA, which has been developed,
is described in the flowchart of Figure 3.
The experimental data (Uc versus σs) and the geometrical characteristics Dm, L, F of the
insulators of Table I are applied to (Eq.1), thus resulting in a set of 36 simultaneous
equations with two unknowns, the arc constants A and n

Fg = ∑36
i=1|Uci − fi (A, n)| (5)

This set of equations must be minimized over A and n. This is why A and n are
converted to the binary system and are considered as parts of a big chromosome.
The search starts with a randomly generated population of such 2Ps chromosomes. Each
constant pair (A, n) is converted to a t bits binary number. 2t bits are required for the
‘chromosome’ of A, n with 0 < A < 500, 0 <n < 1.

The available data in the international literature show that the values of constants A
and n for thin pollution layers are in the range of 50-400 and 0.3-1,
respectively (A=131.5, n=0.374[5]; A=63.0, n=0.76 [7]; A=270- 461, n=0.42-0.66
[8]; A=197.418, n=0.593 [9] ).
Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 141

yes

No

Figur e 3: Flowchar t of Genetic Algor ithm

Table 1
Properties of the Investigated Insulators
Insulator Type (I) Type (II) Type (III) Type (VI)
Maximum diameter DM (mm) 254 254 254 254
Distance between centers H (mm) 146 146 146 146
Creepage distance L (mm) 431 279 279 305
Form factor F 0.916 0.684 0.680 0.696
 142 Journal of Intelligent Computing and Applications (JICA)

4. MODELING USING ARTIFICIAL NEURAL NETWORKS

An artificial neural network (ANN) is made up of a number of simple, and highly

interconnected processing elements called neurons, which processes information by its
dynamic state response to external inputs. ANN models have been applied for different
engineering applications including machine vision [14], speech processing [15] and
system identification, diagnosis and control [16]. In electrical power systems, ANN
have been used for accurate load forecasting, alarm processing, etc. [12,13]. In high
voltage systems, application of ANN has been reported for pattern recognition of partial
discharges [17].
In these applications, the property of universal function approximation of ANN is widely
exploited in which ANN is useful because it acts as a model of a real-world system or
function by adjusting the network parameters (weights). Then, this model can be used for
prediction or forecasting. The useful properties of ANN, like, adaptable and non-linearity
are well suited to many function estimation tasks in many engineering fields [16].

4.1. Multilayer Feed Forward Network

According to the universal approximation theorem, a Neural Network with two layers can
approximate a given function to a desired precision. Figure 4a shows the schematic diagram
of a multilayer feed-forward network used in this paper. The Network is made of
three layers: input layer, output layer, and hidden layers. Each neuron of the output
layer receives a signal from all input via hidden layer neurons along connections with
adjustable weights. The neural network can identify input pattern vectors, once the
connection weights are adjusted by means of the learning process. The back-propagation
learning algorithm which is a generalization of Widrow–Hoff error correction rule is the
most popular method in training the ANN [16] and is employed in this work. This learning
algorithm is presented here in brief. For each neuron in the input layer, the neuron outputs
are given by
Oi = ni (6)
where ni is the input of neuron i and Oi the output of neuron i. Again, for each neuron in
the output layer, the neuron inputs are given by

𝒏𝒏𝒌𝒌 = ∑𝑵𝑵𝑵𝑵
𝒊𝒊=𝟎𝟎 𝒘𝒘𝒌𝒌𝒌𝒌 𝑶𝑶𝒋𝒋 𝒌𝒌 = 𝟏𝟏, … , 𝑵𝑵𝒌𝒌 (7)
where wkj is connection weight between neuron j and neuron k, and Nj, Nk are the number
of neurons in the hidden and output layers, respectively; the neuron outputs are given by
1
𝑂𝑂𝑘𝑘 = −�n +θ �
= 𝑓𝑓𝑘𝑘 (nk , θk ) (8)
1+e k k

where θk is the threshold of neuron k, and the sigmoid function fk is usually used as an
activation function.
 Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 143

Output (Uc)

Output Layer

……
Hidden Layers

……

Input Layer

Inputs (D m , H , L, F , σ s )
Figure 4: The Str uctur e of a Multilayer Perceptr on

For the neurons in the hidden layer, the input and the outputs are given by the relationships
similar to those given in the Eqs. (7) and (8), respectively. The connection weights
of the feed-forward network are derived from the input–output patterns in the
training set by the application of generalization delta rule [16]. The algorithm is based
on minimization of the error function of each pattern p by the use of the steepest
descent method. The sum of squared errors which is the error function of each pattern is
given by

1
𝑁𝑁𝑘𝑘 2
𝐸𝐸𝑝𝑝 = ∑k=1 �𝑡𝑡𝑝𝑝𝑝𝑝 − 𝑂𝑂𝑝𝑝𝑝𝑝 � (9)
2

where t pk and Opk are target and calculated outputs for output neuron k, respectively.
The overall measure of the error for all the input–output patterns is

given by
𝑁𝑁
𝑝𝑝
𝐸𝐸𝑝𝑝 = ∑𝑝𝑝=1 𝐸𝐸𝑝𝑝 (10)

where Np is the number of input–output patterns in the training set.

The hidden and output layer weights are adjusted according to the following expressions:

∆𝑊𝑊𝑘𝑘𝑘𝑘 (p) = 𝜂𝜂𝑗𝑗 𝛿𝛿𝑝𝑝𝑝𝑝 𝑂𝑂𝑝𝑝𝑝𝑝 + 𝛼𝛼𝑗𝑗 ∆𝑊𝑊𝑘𝑘𝑘𝑘 (𝑝𝑝 − 1) (11)

 144 Journal of Intelligent Computing and Applications (JICA)

𝑂𝑂𝑝𝑝𝑝𝑝 �𝑡𝑡𝑝𝑝𝑝𝑝 −𝑂𝑂𝑝𝑝𝑝𝑝 ��1−𝑂𝑂𝑝𝑝𝑝𝑝 � for the output layer

𝛿𝛿𝑝𝑝𝑝𝑝 = � 𝑁𝑁𝑘𝑘 (12)
𝑂𝑂𝑝𝑝𝑝𝑝 �1−𝑂𝑂𝑝𝑝𝑝𝑝 � ∑𝑗𝑗 =1 𝛿𝛿𝑝𝑝𝑝𝑝 𝑊𝑊𝑗𝑗𝑗𝑗 for the hidden layer

where ηj, αj are the learning rate and momentum constant, respectively.

Most of the proposed optimization methods for updating the Neural Network weights
are based on Equations (11-12), the difference relies on either fixing the learning rate and
momentum for some variants or using adaptive searching methods.

4.2. The modeling of an insulator by ANN

In this paper, a new approach using ANN as a function approximator is used to model
accurately the relationship between critical flashover voltage Uc (KV) as the output of the
neuronal model and the High voltage insulator parameters the maximum diameter
Dm(cm), the height H (cm), the creepage distance L (cm), the form factor F of the
insulator and the layer conductivity σs (mS).

𝑈𝑈𝑐𝑐 = 𝑓𝑓(𝐷𝐷𝑚𝑚 , 𝐻𝐻, 𝐿𝐿, 𝐹𝐹, 𝜎𝜎𝑠𝑠 ) (13)

The data used for the training, evaluation and testing of the ANN were selected from
various sources for different types of insulators. The aim of the present work is to apply
different Neural Network topologies for flashover modeling to be used for forecasting the
critical flashover voltage for new operating conditions. A multilayer feedforward neural
network is constructed and trained with normalized experimental input/output data pairs
using Levenberg-Marquardt optimization algorithm, a variant based on equations (11,12)
to find the optimal network weights (W*) that minimize equation (10).

Different normalization schemes have been tested for normalizing the input-output
training patterns. They are detailed in Table 2 [18].
Table 2
Normalization Schemes
SchemeNumber Input output
1 MAX MAX
2 MAX MAX MIN
3 MAX MEAN & S.D

4 MAX MIN MAX

5 MAX MIN MAX MIN
6 MAX MIN MEAN & S.D

7 MEAN & S.D MAX

8 MEAN & S.D MAX MIN
9 MEAN & S.D MEAN & S.D

After normalization, the input and output variable ranges are within (0,1) [13]. Each
 Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 145

input–output variable was normalized one by one using scheme number 8 in this work.

𝑛𝑛𝑛𝑛𝑛𝑛 𝑖𝑖 (p)−𝑛𝑛𝑛𝑛𝑛𝑛 𝑖𝑖,av (p)

𝑛𝑛𝑛𝑛𝑛𝑛𝑖𝑖,nor (p) = σi
(14)

where : net i ,nor is the average value of the ith component of the input vector;
σi is the standard deviation value of the ith component of the input vector.

O (p)−𝑂𝑂𝑘𝑘min
𝑂𝑂𝑘𝑘,nor (p) = O 𝑘𝑘 (15)
𝑘𝑘max −𝑂𝑂 𝑘𝑘min

where Ok ,min ,Ok ,max are the minimum and the maximum elements of the ith component
of the output vector respectively.

5. RESULTS AND DISCUSSION

5.1. Genetic Algor ithms Results

The mathematical model parameters of the flashover phenomenon are first obtained by a
genetic algorithm. We started by randomly generated population of 20 chromosomes. It
generates 20 random values for the arc constant A (0 <A < 500) and 20 random values for
the arc constant n (0< n< 1). Each constant, A or n, is converted to a 16-bit binary number.
32 bits are required for the chromosome. Each pair of parents with single point crossover
generates four children. The crossover begins as each chromosome of any parent is divided
into six parts, and the pair of parents interchange their genetic material. After crossover
there is a 1% probability of mutation. The procedure is terminated after 200 generations.
The above parameters of the algorithm are summarized in Table 3.

The application results in pairs of (A, n) values that finally converge to the optimum
values A = 177.22, n = 0.510 as the number of iterations increases (Table 4).

Table 3
Genetic Algorithms Parameters
Population size Bits/chromosome Number of children Mutation probability Number of generations
Ps =20 t=16 Nc=5 Pm =1% Nmax=200

Table 4
Genetic Algorithms Results
Generation 10 20 40 60 80 100 140 200
A 189.96 179.52 177.47 177.75 176.76 177.1 177.43 177.22
N 0.472 0.480 0.499 0.502 0.517 0.512 0.507 0.510
Fg 33380 32056 30978 30963 30960 30959 30958 30957

The simulation results are shown in Figure 5 and Figure 6 where it becomes obvious that
 146 Journal of Intelligent Computing and Applications (JICA)

the algorithm converges rapidly to these values.

-4
x 10
3.6

3.5

cost function F g
3.4

3.3

3.2

3.1

3.0
0 20 40 60 80 100 120 140 160 180 200
generation
Figure 5: The Cost Function in Each Gener ation

Figure 6: Convergence of the Optimum Values of the Arc Constants (A , n)

It is important to verify the validity of the genetic algorithm by applying the optimum
values of the arc constants in the mathematical model to the insulators of Table 1. The
obtained results are plotted in Figure 7. The Figure also shows the results from the
experiments [5], [10]. The computed values agree well with the experiments.
 Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 147

x 104
3
T(I) Model
T(I) exp
T(II) Model
2.5 T(II) exp
T(III) Model
Critical Voltage Uc (Volts)

T(III) exp
T(VI) Model
2 T(VI)exp

1.5

0.5

0 0.5
0 1 1.5 2 x 10-4
−1
Surface conductivity σ(Ω )
Figure 7: Cr itical Voltage Against the Sur face Conductivity

Data reported in [11] are used to validate the model for the obtained optimal arc
characteristic numbers (A, n) (Table 5); the surface conductivity is calculated using
equation (2).
Figure 8 and Figure 9 present the computed critical flashover voltage and critical
leakage current against the surface conductivity for type T (V) and type T (IV) insulators
using the optimum values of the arc constants in the mathematical model.
Experimental tests on these types of insulators show very good agreement between
the measured values and the computed ones.

Table 5
Experimental Values for Validation
InsulatorType C (mg/cm2 ) Uc (kV)[10]
T(V) D=254 mm 0 .13 12.0
H=146 mm F=.68 0.16 11.1
0.23 8.7
0.28 9.1
0.34 7.5
0.37 7.8
0.49 6.2
0.52 6.8
0.55 6.1
T(IV) D=254 mm 0.02 22.0
H=146 mm F=.68 0.05 16.0
0.10 13.0
0.16 11.0
0.22 10.0
0.30 8.5
148 Journal of Intelligent Computing and Applications (JICA)

x 104
T(V) Model
2.5 T(V) exp
T(IV) Model

Critical Voltage Uc (Volts)

T(IV) exp

2
A=177.22, n=0.510

1.5

0.5

0
0 0.5 1 1.5 2 x 10-4
Surface conductivity σ(Ω−1)
Figure 8: Cr itical Flashover Voltage Against the Sur face Conductivity for
Type T(V) and Type T(IV) Insulator s

T(V)
T(IV)
2

A=177.22,n=0.510
I (A)

1.5
Critical Current

0.5

0
0 0.5 1 1.5 2 x 10-4
Surface conductivity σ(Ω−1)

Figure 9: Critical Leakage Current versus the Surface Conductivity for Type T (V) and Type T (IV)
Insulators
 Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 149

5.2. Neural Networ k Results

Modeling using Neural Networks is a not straightforward especially in selecting the

number of hidden layers, the number of neurons per layer and the activation functions for
each layer. In this paper we have worked on Two and Three Layer Feedforward Neural
Networks.
Our methodology of designing a neural network, as depicted in figure 10, is summarized as
follows:
- Among the available experimental data, we have constructed two sets, a
training set to find the optimal Network parameters and weights, and a validation
set.
- To validate the obtained model for generalization.
- For all the patterns in the training set, the respective parameters of the network are
optimised through a set of trials where we have tried different number of neurons
per layer (2-12 neurons) and we have tested different activation functions the linear,
the logistic or hyperbolic).
- For each combination of the parameters, the ANN training process is updated
using the respective training set. After the respective convergence, the
Absolute Mean square Error is used as a comparison index between the experimental
and estimated values of the critical flashover voltage for the evaluation set is
calculated and the lowest index is chosen as the best one with the respective
parameters.
The effect of different learning rates and momenta on the convergence property of the
learning process is extensively studied and the best combination is identified. It has been
verified that using two hidden layers has better effect on the learning accuracy than the use
of one hidden layer, with the same number of hidden layer nodes. The effect of different
learning rates and momenta in different layers on the convergence property of the learning
process is also presented in details.
Figure 11 shows the impact of changing the number of neurons in a One Hidden Layer
Network, for a smaller number of neurons, the error between the network output and the
target is higher and as the number of neurons increased this error is reduced until we reach
a certain number where generalization is no longer achieved.
Figure 12 illustrates the impact of changing the number of neurons in a two hidden layer
Neural Network. It is clear that a medium number of neurons is better where the training
error is small as well as the validation error.
The same data used in determining the arc constants for the mathematical model using
GAs are used to train a Neural Network model, Figure 13 illustrates the best obtained
black box model using a 2 Hidden Layer Network (4,7,7,1) whereas Figure 14 shows the
ANN capability of predicting the Critical Flashover Voltage for New types of insulators.
150 Journal of Intelligent Computing and Applications (JICA)

Begin

Read Data

Normalise data

Construct
training set

Construct
Evaluation set

Choose ANN
topology and size

Select a training
Algorithm

WN+1=WN+∆W

Evaluation

No Stopping yes
criteria
reached
Evaluate the Model for
generalization

No
Is model
good ?

yes

End

Figure 10: Flowchart of the ANN Optimisation Methodology for the Estimation of the Critical
Flashover Voltage of Insulators
Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 151

Network with One Hidden Layer : Logsig Activation

Parameters: Scheme 8: Lr=.3; alpha=.8; #Epochs=500

4
Training
3.5 Validation

2.5
RMSE

1.5

0.5

0
0 5 10 15 20 25 30 35 40 45 50
Number of Neurons
Figure 11: The Impact of Number of Neurons in Designing a One Hidden Layer ANN

Network with Two Hidden Layer

Parameters: Scheme 8: Lr=.3; alpha=.8; #Epochs=500
3
2T
2V
4T
2.5 4V
7T
7V
10T
10V
2
RMSE

1.5

0.5

0
2 3 4 5 6 7 8 9 10
Number of neurons

Figure 12: The Impact of Number of Neurons in Designing a Two Hidden Layer ANN
152 Journal of Intelligent Computing and Applications (JICA)

x 104
3

ANN T(I)
Exp T(I)
ANN T(II)
2.5
Exp T(II)
ANN T(III)
Critical Flashover Voltage Uc (Volts)

Exp T(III)
ANN T(IV)
2 Exp T(IV)

1.5

0.5
0 0.5 1 1.5 2 2.5 x10-4
-4

Surface conductivity σ(Ω−1)

Figure 13: Trained ANN-Critical Voltage Against the Surface Conductivity

x 104

ANN T(V)
2.5 Exp T(V)
ANN T(VI)
Exp T(VI)
Critical Voltage Uc (Volts)

1.5

0.5
0 0.5 1 1.5 2 2.5
x 10-4
Surface conductivity σ(Ω−1)

Figure 14: ANNs in Estimating Critical Voltage Against the Surface Conductivity
 Modeling the Critical Flashover Voltage of High Voltage Insulators Using... 153

6. CONCLUSION
In this paper, two techniques of artificial intelligence are successfully applied to solve the
problem of critical flashover voltage modeling and estimation. First, a Genetic Algorithm
(GA) has been used for the estimation of arc parameters for polluted insulators based on a
physical model that fairly represents this phenomenon. The fundamental advantage of this
proposed GA is the ability to find the optimized characteristics of the discharge (A, n)
developing along the insulator surface when given some of the insulator’s geometrical
types. Then, ANNs are designed as black-box models to estimate the unknown relationship
between the insulator parameters and the critical flashover voltage.
Therefore, finding a model either based on GAs or ANNs helps in computing the critical
conditions for flashover which leads to better understanding of the transient phenomena in
polluted insulators. Simulation results were obtained from experimental studies and from
application of a mathematical model for the estimation of the flashover voltage on polluted
insulators.
The results prove the validity of Artificial Intelligence for modeling phenomena in
High Voltage Engineering. The method proposed in the paper can be used to eliminate the
flashover fault and in the establishment of maintenance policy and for addressing an effective
solution against pollution flashover of high voltage insulators.

References
[1] W. Ricardo S. Garcia, R. Bosignli and E. Jr. Gomes, “Influence of the Nonuniformity of Pollution
Distribution on the Electrical Behavior of Insulators”, Proceedings of the 3rd International
Conference on Properties and Applications of Dielectric materials, July 8- 12, 1991, Tokyo,
Japan, pp. 342-345.
[2] K. Chrzan, Z. Pohl and T. ,Kowalak, “Hygroscopic Properties of Pollutants on HV Insulators”,
IEEE Transactions on Electrical Insulation, February, 1989, Vol. 24, No. 1, pp. 107-112.
[3] L. Qisheng, W. Lai, S. Zhiyi, L. Yansheng, K. Morita, R. Matsuoka, and S. Ito, “Natural
Contamination Test Results of Various Insulators Under DC Voltage in an Inland Area in China”,
Proceedings of the 3rd International Conference on Properties and Applications of Dielectric
Materials, July 8-12, Japan 1991, pp. 350-353.
[4] F. Obenhaus, “Fremdschichtueberschlag und Kriechweglaenge”, Deutsche Elektrotechnik,
1958, H. 4, pp. 135-137
[5] F. V. Topalis, I. F. Gonos, I. A. Stathopoulos, “Dielectric behaviour of polluted porcelain insulators”, IEE Proc. Gen.
Transm. Distribution 148 (4), 2001, 269–274.
[6] D. C. Chaurasia, “Scintillation Modelling for Insulator Strings Under Polluted Conditions”,
in: 11th International Symposium on HighVoltage Engineering, London, UK, 1999.
[7] F. A. M Rizk. “Mathematical models for pollution flashover”. Electra, October, 78, 1981,
pp.71-103
[8] P. S. Ghosh, and N.Chatterjee: “Polluted Insulators Flashover Model for AC Voltage”, IEEE
Trans. Dieiectr. Electr. Insul., 2(l), 1995, pp. 128-136
[9] B. Zegnini , D. Mahi , A. Chaker, “Determination of the E-I Characteristics of the AC Arcs
Developing Along Electrolytic Surface using Regression on Test Results”, ICEEE’2004
Proceeding of the 2nd International Conference on Electrical and Electronics Engineering ; April
 154 Journal of Intelligent Computing and Applications (JICA)

24-26 , Laghouat, Algerie, pp. 179-183. ISSN 1112-4652

[10] G. Zhicheng, and Z. Renyu: “Calculation of DC and AC flashover voltage of polluted
insulators”, IEEE Trans. Electr. Insul., 25, (4), 1990, pp. 723-729
[11] V. T. Kontargyri , A. A. Gialketsi, G. J. Tsekouras, I. F. Gonos, I. A. Stathopulos, “Design of
an Artificial Neural Network for the Estimation of the Flashover Voltage on Insulators”, Electric
Power Systems Research, 77, 2007, 1532–1540 ELSEVIER
[12] M. T. Gençoglu , M. Cebeci, “Investigation of Pollution Flashover on High Voltage
Insulators Using Artificial Neural Network”, Expert Systems with Applications 36, 2009,
7338–7345, ELSEVIER
[13] G. E. Asimakopoulou V.T. Kontargyri G.J. Tsekouras, F.E. Asimakopoulou I.F. Gonos
I.A. Stathopulos, “Artificial Neural Network Optimization Methodology for the Estimation of the
Critical Flashover Voltage on Insulators”, IET Science, Measurement and Technology, Vol. 3, No.
1, 2009, pp. 90–104
[14] K. Fukishama, “Neocognition: A Hierarchical Neural Network Capable of Visual
Pattern Recognition”, Neural Networks, 1, 1988, 119–130.
[15] K. Kohonen, The Neural Phonetic Typewriter. IEEE Computer, 1988, 11–22.
[16] S. Haykin, Neural Networks: Comprehensive Foundations, Prentice Hall, 1995.
[17] A. S. Farag, “Estimation of Polluted Insulators Flashover Time Using Artificial Neural
Networks”, International IEEE/IAS Conference on Industrial Automation and Control: Emerging
Technologies, 1995, 22-27 Mai, Taipei, Taiwan
[18] P. S.Ghosh, S.Chakravorti and N.Chatterjee, “Estimation of Time to Flashover Characteristics
of Contaminated Electrolytic Surfaces Using a Neural Network”, IEEE Trans. on DEIS.,
Vol. 2, No. 6, 1995, pp. 1064-1074.