CRITICAL FLASHOVER VOLTAGE ON POLLUTED INSULATORS

ESTIMATED USING CONVENTIONAL AND INTELLIGENT TECHNIQUES

1 2 1 2 1
V. P. Androvitsaneas, E. Karampotsis, I. F. Gonos, G. Dounias, I. A. Stathopulos
1
High Voltage Laboratory, School of Electrical and Computer Engineering, National
Technical University of Athens, 9 Iroon Politechniou Street, Zografou Campus, Athens
15780, Greece
2
Management & Decision Engineering Lab, Department of Financial and Management
Engineering, University of the Aegean, 41 Kountouriotou Street, 82100 Chios, Greece

Abstract: This paper proposes the use of conventional data analysis and computational
intelligence methods for the estimation of critical flashover voltage on polluted Cap & Pin
porcelain insulators. Specifically, modeling using Artificial Neural Networks (ANNs) and
Multiple Linear Regression (MLR) has been attempted, based on related application data.
The database, used for the analysis, consists of 168 cases (i.e. series of related
measurements) represented by six (6) numeric variables, namely, the diameter, the
height, the creepage distance, the manufacturing constant, the pollution of insulators and
the critical flashover voltage. Part of these data derives from a specialized data
generation model corresponding to incidents of flashover voltage on polluted Cap & Pin
insulators (simulation data), while the rest of the data consist of real experimental
observations (real data). The comparison showed that ANNs are proved to be superior
for modeling the estimation problem of the critical flashover voltage on polluted insulators.
Finally, comparison between the results of this work with other similar approaches,
previously existing in literature, shows that the results of the ANNs application are
satisfactory, yet there is certainly still room for further improvement.

chamber by means of a spraying system

1 INTRODUCTION arrangement [9] in accordance with IEC standards
[10]. These experiments were carried out at the
Critical flashover voltage UC is called the applied Research and Standards Testing Center of the
voltage across the insulator when the partial arc is Public Power Corporation, Athens, Greece.
developed into a complete flashover. The main
reason for the flashover phenomenon is the The rest of the data have been derived from the
pollution of insulators (marine or/and industrial), application of a special mathematical model, built
leading to the initiation of partial discharges on by the High Voltage Laboratory of NTUA. This
their surface and to the final breakdown of the air mathematical model is a theoretical approach to
layer across the insulator surface [1–4]. the flashover phenomenon on polluted insulators,
based on the Obenhaus–Alston's model. The form
The purpose of this study is the development of of the specific model for Cap & Pin insulators is
conventional data analysis and computational illustrated in Figure 1 and described by (1)–(4) [5].
intelligence methods for the estimation of the
critical flashover voltage on polluted Cap & Pin
porcelain insulators. Specifically, a modeling of
flashover occurrence is attempted using the
methods of Artificial Neural Networks (ANNs) and
Multiple Linear Regression (MLR) based on related
application data.
The related application data consist of 168 cases
(i.e. series of related measurements) represented
by six (6) numeric variables. Some of these data
(140 cases) are simulation results, derived from a
specialized data generation model corresponding
to flashover occurrence on polluted Cap & Pin
insulators and the rest of the data (28 cases)
consist of real experimental observations, both
presented in preceding research studies [5–8].
They have been collected from experiments on
Cap & Pin porcelain insulators, which previously Figure 1: Critical flashover voltage on a polluted
have been exposed to artificial pollution in a fog insulator (Obenhaus–Alston's model).
Figure 2 depicts a Cap & Pin insulator with its If D(l) is the diameter of the insulator, then the
basic design parameters and Table 1 summarizes Equation 5 is converted as follows [5]:
the problem variables. 𝐿 1
F = ∫0 𝑑𝑙 (6)
𝜋·𝐷(𝑙)

2 STATISTICAL ANALYSIS OF DATA

The purpose of the statistical analysis was to

observe the values distribution of the problem
variables and the identification of possible linear
correlations existing between the input–output
variables. Conclusions drawn from these
observations are used for the selection and pre-
processing of training, validation and test sets. The
statistical analysis was performed with the aid of
Figure 2: Basic design parameters of a Cap & Pin “STATISTICA 7” software defining significance
Insulator. level p<0.05.

Table 1: Problem variables. 2.1 STATISTICAL ANALYSIS OF

SIMULATION DATA
Problem Variables
Var. Unit Description According to statistical analysis, performed on the
Dm cm Maximum diameter of insulator simulation data derived from the mathematical
H cm Height of insulator models described above, it was ascertained that all
L cm Creepage distance of insulator the variables follow a normal distribution. The
F - Manufacturing constant of the insulator correlation factor (r) among the input variables and
C mg/cm2 Equivalent Salt Deposit Density (ESDD)
UC kV Critical flashover voltage
between the input–output variables is presented in
Table 2.
n
A −� �
Uc = · (L + π · n · Dm · F · K) · (π · A · Dm · σs ) n+1 (1) Table 2: Correlation among problem variables for
n+1
simulation data.
𝜎𝑠 = (369,05 · 𝐶 + 0.42) · 10−6 (2)
n+1 L
Simulation Data, N=140
K=1+ · ln � � (3) Correlation Factor (r), p<0.05
2·π·F·n 2·π·R·F
Var. Dm H L F C Uc
1
Ι Dm 1.00 0.35 0.51 -0.45 0.00 0.17
R=� = 0,469 · (π · A · Dm · σs ) 2·(n+1) , I in [A] (4) H 0.35 1.00 0.77 0.51 0.00 0.21
π·1,45
L 0.51 0.77 1.00 0.48 0.00 0.26
where: σs = Surface conductivity, in Ω
-1 F -0.45 0.51 0.48 1.00 0.00 0.09
C 0.00 0.00 0.00 0.00 1.00 -0.85
K = Resistance factor of polluted layer UC 0.17 0.21 0.26 0.09 -0.85 1.00

R = Radius of the arc path, in cm According to Table 2, it is observed that, the

correlations among the variables Dm, H, L, and F
A & n = Constants of insulator’s arc are relatively strong, the correlations among the
N = Number of insulator’s caps variable C and variables Dm, H, L, and F are null,
and the correlations among the variable UC and the
Main objective of this work is the development of input variables are weak enough, except for the
methods for the estimation of critical flashover correlation with the variable C.
voltage, based on specific input variables. The
input variables used for the construction of the The mathematical equations expressing the linear
application database were the diameter Dm, the dependence of the variable UC on the other
height H, the creepage distance L the pollution of variables of the problem are the following:
insulators C, whilst the output variable (target UC(Math.) = 4.9781 + 0.33568 ∙ Dm (7)
value) was the critical flashover voltage UC. The
manufacturing constant F is determined by the UC(Math.) = 5.693 + 1.2256 ∙ H (8)
dimensions of the insulator and is calculated
according to the Equation 5 [10]. UC(Math.) = 5.6080 + 0.21975 ∙ L (9)
𝐿 1
F= ∫0 𝑝(𝑙) 𝑑𝑙 (5) UC(Math.) = 11.0 + 3.3643 ∙ F (10)

where: p(l) = Circumference of the insulator as a UC(Math.) = 20.026 − 25.94 ∙ C (11)

function of the partial creepage distance l.
2.2 STATISTICAL ANALYSIS OF REAL DATA 4. MODELING OF THE PROBLEM

According to statistical analysis, held in the set of As already mentioned above, the conventional and
the real data, it was observed that the only intelligent methods, used for the estimation of
variables, which follow normal distribution, are the flashover voltage on polluted Cap & Pin porcelain
variables C and UC. The correlation factor (r) insulators, were Artificial Neural Networks (ANNs)
among the individual input variables and between and Multiple Linear Regression (MLR). Evaluation
input–output variables is presented in Table 3. was performed according to the known measures
According to Table 3, it is observed that the of correlation factor (r), the mean-absolute error
correlations among all variables are potent and (MAE), the root-mean-square error (RMSE), the
range approximately within the same level. root-absolute error (RAE) and the relative-root-
squared error (RRSE), between of the actual value
Table 3: Correlation among problem variables for (Uc(act)) and predicted value (Uc(pr)) of critical
real data. flashover voltage.
Real Data, N=28 4.1 ANN METHOD
Correlation Factor (r), p<0.05
Var. Dm H L F C Uc
Dm 1.0 -1.0 -0.61 -0.95 0.44 -0.51 The ANN model used in this study belongs to the
H -1.0 1.0 0.61 0.95 -0.44 0.51 category of the feed-forward artificial neural
L -0.6 0.6 1.00 0.83 -0.57 0.67 networks (Multilayer Perceptrons–MLPs) [11, 12].
F -0.9 0.9 0.83 1.00 -0.52 0.62
C 0.4 -0.4 -0.57 -0.52 1.00 -0.87 On finding the optimal MLP for the estimation of
UC -0.5 0.5 0.67 0.62 -0.87 1.00 critical flashover voltage, 1895 architectures of
MLPs, in total, have been developed and
The regression equations expressing the linear compared with each other. All MLPs had one
dependence of the variable UC on the other hidden layer and the activation function of hidden
problem variables are in this case as follows: nodes was the logistic function of (17), [11].
UC(Exp.) = 76.749 − 2.550 ∙ Dm (12) 1
φ�uj � = −α∙uj (17)
1+e
UC(Exp.) = −34.58 + 3.1881 ∙ H (13)
The online learning for the supervised training of
UC(Exp.) = −4.366 + 0.50413 ∙ L (14) each MLP into batch mode has been performed
using the back-propagation algorithm (BP) [11, 12],
UC(Exp.) = 2.8882 + 11.719 ∙ F (15) while validation of each MLP has been carried out
using 10-fold cross validation method [12]. The
UC(Exp.) = 19.226 − 20.63 ∙ C (16) training process was divided into four (4) separate
stages, each one of them for the selection of the
optimum value for ANN parameters applying
3 SELECTION AND PRE-PROCESSING OF consecutive iterations. The goal in each stage was
DATASETS to minimize MAE, RMSE, RAE, and RRSE and to
maximize the correlation factor (r) between the
The entire database was divided into a training set UC(actual) and UC(predicted).
and a test set. The training set was used for the
development of the predictive models of critical The aim of the first stage of the training process
flashover voltage and the test set was used to was to find the optimum combinations among the
validate the final predictive models. learning rate (η), the momentum term (α), and the
number of neurons in the hidden layer (Nn) for
The test set is derived from the random selection specific numbers of epochs (e) (Table 4).
of 6 cases of actual data and was used solely for
the final test of predictive models with a view to Table 4: First stage of the training process.
their generalization ability. The training set consists
of the rest 162 data cases (140 cases of simulation First Stage
data and 22 cases of actual data). It is sorted in No e η α Nn
ascending order (according to the values of the 0.1–0.3 0.1–0.9 2-20
variable UC) and it was used in both modeling 1 500
with step 0.1 with step 0.1 with step 2
methods (ANNs and MLR). Before the 2 2500
0.1–0.3 0.1–0.9 2-20
experimentation with the datasets, all variables with step 0.1 with step 0.1 with step 2
(Dm, H, L, F, and C) were normalized in the interval 0.1–0.3 0.1–0.9 2-20
3 4500
with step 0.1 with step 0.1 with step 2
[-1, 1], a usual practice for ANNs.
0.1–0.3 0.1–0.9 2-20
4 6500
with step 0.1 with step 0.1 with step 2
0.1–0.3 0.1–0.9 2-20
5 8500
with step 0.1 with step 0.1 with step 2
0.1–0.3 0.1–0.9 2-20
6 10 500
with step 0.1 with step 0.1 with step 2
The search which took place at the second stage
of the training process concerns the finding of the
best combinations between the number of epochs
and the number of neurons in the hidden layer of
MLPs, for the combinations η-α, resulted from the
first stage (Table 5).
In the third stage, having taken into account the
best combinations of e–η–α–Nn, yielded from the
second stage of the experimental procedure, the
behavior of MLPs was observed by altering the
number of neurons in the hidden layer (Nn) while
keeping fixed the combinations of e–η–α (Table 6).

Table 5: Second stage of the training process. Figure 3: Scatter plot for the training–validation
set.
Second Stage Table 8: ANN model for the estimation of critical
No c η α Nn flashover voltage.
12 500–20 500 2–20
1 0.1 0.8
with step 2000 with step 2
ANN Architecture
12 500–20 500 2–20
2 0.1 0.9
with step 2000 with step 2 Name ANN1
12 500–20 500 2–20 Type MLP
3 0.2 0.5
with step 2000 with step 2 Number of Layers 3
12 500–20 500 2–20 Number of Hidden Layers 1
4 0.2 0.6
with step 2000 with step 2 Number of Hidden Units (Nn) 21
12 500–20 500 2–20 Number of Inputs 6
5 0.2 0.7
with step 2000 with step 2 Number of Outputs 1
12 500–20 500 2–20 Inputs Dm, H, L, F, C
6 0.3 0.5
with step 2000 with step 2 Outputs UC
12 500–20 500 2–20
7 0.3 0.6 Cases of Training-Validation Set 162
with step 2000 with step 2
Training Procedure Supervised
Training Type Batch Mode
Validation Procedure 10-fold cross validation
Table 6: Third stage of the training process.
Training Algorithm Back Propagation
Learning Rate (η) 0.1
Third Stage Momentum Term (α) 0.8
No e η α Nn Number of Epochs (e) 36 500
1 20 500 0.1 0.8 19,20,21,22 Performance
2 20 500 0.2 0.6 13,14,15 Training–validation set Test set
r 0.9994 0.9842
MAE (kV) 0.1370 1.3249
In the last fourth stage, using the combinations of
RMSE (kV) 0.1886 1.6957
the third one, the behavior of MLPs was observed
RAE (%) 2.9136 -
by altering the number of epochs (e) (Table 7). RRSE (%) 3.3936 -

Table 7: Fourth stage of the training process. Finally, Figures 3 and 4 present the correlation (r)
between the actual and estimated values of UC.
Fourth Stage
No e η α Nn
1 20 500–38 500 with step 2000 0.1 0.8 21
2 20 500–28 500 with step 2000 0.2 0.6 14

The final ANN model and its performance on the

training and test set are presented in Table 8.

Figure 4: Scatter plot for the test set.

4.2 THE MLR METHOD 4.3 COMPARISON AND RESULTS

The application of multiple linear regression (MLR) The comparison between the models of this paper
method [14] has been performed on the same (ANN1 and MLR) shows that ANNs are superior for
datasets with those used in the ANN methodology the estimation of flashover voltage on polluted Cap
(i.e. training and test set). The MLR model for the & Pin insulators. The final performance results for
estimation of flashover voltage is described by both models are tabulated in Table 10.
(18).
Table 10: Comparison between ANN1 and MLR.
Therefore, Table 9 presents the performance of
the MLR model and Figures 5 and 6 give the Comparison ANN1–MLR
corresponding scatter diagrams.
Modeling ANN1 MLR
UC = W + Q + 12.6931 (18) r 0.9994 0.8748
MAE (kV) 0.1370 2.2036
where: RMSE (kV) 0.1886 2.6868
RAE (%) 2.9136 46.8765
W = 0.7663 ∙ D𝑚 + 0.6931 ∙ H (19) RRSE (%) 3.3936 48.3468
Test ANN1 MLR
Q = 2.0396 ∙ L − 0.1124 ∙ F − 6.7013 ∙ C (20) r 0.9842 0.9322
MAE (kV) 1.3249 2.7646
Table 9: Performance of the MLR model. RMSE (kV) 1.6957 3.2932

MLR Model Moreover, the comparison of ANN1 with

Modeling Test corresponding models of other relative studies
r 0.8748 0.9322 (Table 11–Table 13) show that current results are
MAE (kV) 2.2036 2.7646 quite satisfactory, yet, there is room for further
RMSE (kV) 2.6868 3.2932 improvement.
RAE (%) 46.8765 -
RRSE (%) 48.3468 - Table 11: Comparison between ANN1 and the
ANN presented in [7].

Comparison ANN1–ΑΝΝ
Training–validation ANN1 ANN
r 0.9994 0.9972
Test ANN1 ANN
r 0.9842 0.9853
MAPE (%) 7.48 3.84

Table 12: Comparison between ANN1 and the

ANN models presented in [8].

Comparison ANN1–case study 1

Training–validation ANN1 case study 1
Figure 5: Scatter plot for the training–validation r 0.9994 0.9960
set. RMSE (kV) 0.1886 0.070
MAPE (%) 1.1867 1.300
Test ANN1 case study 1
r 0.9842 0.9861
MAPE (%) 7.480 4.040
Comparison ANN1–case study 2
Training–validation ANN1 case study 2
RMSE (kV) 0.1886 0.178
MAPE (%) 1.1867 4.151
Test ANN1 case study 2
MAPE (%) 7.480 4.225
Comparison ANN1–case study 3
Training–validation ANN1 case study 3
RMSE (kV) 0.1886 0.165
MAPE (%) 1.1867 4.004
Test ANN1 case study 3
Figure 6: Scatter plot for the test set. MAPE (%) 7.480 2.152
Table 13: Comparison between ANN1 and the [3] R. Boudissa, A. Haddad, Z. Sahli, A. Mekhaldi
ANN presented in [6]. and R. Baersch, “Performance of outdoor
insulators under non-uniform pollution
th
Comparison ANN1–ΑΝΝ conditions”, 14 International Symposium on
High Voltage Engineering, Beijing, China,
Training–validation ANN1 ΑΝΝ
r 0.9994 0.95305
August 2005.
RMSE (kV) 0.1886 0.310 [4] Y. Cheng, C. H. Li, C. H. Niu and F. Zhang,
“Porcelain insulators detection by two
dimensions electric field on high voltage
th
5 CONCLUSIONS transmission lines”, 15 International
Symposium on High Voltage Engineering,
The proposed modeling of the flashover Ljubljana, Slovenia, August 2007.
phenomenon on polluted Cap & Pin porcelain [5] F. V. Topalis, I. F. Gonos and I. A.
insulators using Artificial Neural Networks (ANNs) Stathopulos, “Dielectric behaviour of polluted
and Multiple Linear Regression (MLR), proves the porcelain insulators”, IEE Proc.-Gener.
superiority of ANNs. Transm. Distrib., Vol. 148, No. 4, Jul. 2001.
[6] A. A. Gialketsi, V. T. Kontargyri, I. F. Gonos
The proposed ANN model (named ANN1), which is and I. A. Stathopulos, “Estimation of the
a multilayer perceptron (MLP) with 1 hidden layer, flashover voltage on insulators using Artificial
identified and estimated the value of UC from input Neural Networks”, World Scientific and
variables with great success giving results of Engineering Academy and Society,
r=0.9942 and RMSE=0.1886kV. Transactions on Circuits and Systems, Vol. 4,
No. 5, pp. 514-518, May 2005.
In addition, the performance of ANN1 on unknown [7] G. E. Asimakopoulou, V. T. Kontargyri, G. J.
and unseen data during the training, i.e. on the test Tsekouras, F. E. Asimakopoulou, I. F. Gonos
set, (r=0.9842 and RMSE=1.6957kV) and the and I. A. Stathopulos, “Artificial neural network
comparison with similar ANNs models from related optimisation methodology for the estimation of
literature, shows that the model generalizes well the critical flashover voltage on insulators”,
on the estimation of critical flashover voltage on IET Science, Measurement & Technology,
polluted Cap & Pin porcelain insulators, yet, there Vol. 3, No. 1, pp. 90–104, 2009.
might still room for further development and [8] V. T. Kontargyri, A. A. Gialketsi, G. J.
improvement. Tsekouras, I. F. Gonos and I. A. Stathopulos,
“Design of an artificial neural network for the
Future research on the domain is directed towards
estimation of the flashover voltage on
different areas, such as:
insulators”, Electric Power Systems Research,
(a) Use of different ANN architectures, as well as Vol. 77, No. 12, pp. 1532–1540, Oct. 2007.
of other computational intelligence and nature [9] K. Ikonomou, G. Katsibokis, G. Kravaritis and
inspired intelligence techniques. I. Stathopulos, “Cool fog tests on artificially
th
polluted insulators”, 5 International
(b) Enrichment of data collection with more real Symposium on High Voltage Engineering,
data sets from different insulator types. Braunschweig, Germany, 1987.
[10] IEC 60507-04:1991, “Artificial pollution tests
(c) Experimentation with different validation on high voltage insulators to be used on a.c.
methods than the 10-fold cross validation method
systems”, 1991.
used in this work (e.g. leave-1-out approach), and
[11] S. Haykin, “Neural Networks and Learning
(d) Design and application of proper hybrid Machines”, Pearson International Edition
intelligent techniques on the domain. (Prentice Hall-Pearson Education Inc.), 2009,
rd
3 edn.
[12] J. Han, M. Kamber and J. Pei, “Data Mining-
Concepts and Techniques”, Morgan
REFERENCES Kaufmann Publishers, Elsevier, 2012, 3 edn.
rd

[1] C. S. Engelbrecht, R. Hartings, H. Tunell, B. [13] R. Kohavi, “A study of cross-validation and

Engström, H. Janssen and R. Hennings, bootstrap for accuracy estimation and model
th
“Pollution tests for coastal conditions on an selection”, 14 International Joint Conference
800kV composite bushing”, IEEE TPWRD, on Artificial Intelligence, Montreal, Quebec,
Vol. 18, No. 3, pp. 953–959, Jul. 2003. Canada, August 1995.
[2] Z. Aydogmus and M. Cebeci, “A new [14] R. Peck, C. Olsen and J. L. Devore,
flashover dynamic model of polluted HV “Introduction to Statistics and Data Analysis”,
th
insulators”, IEEE TDEI, Vol. 11, No. 4, pp. Cengage Learning, USA, 2014, 5 edn.
577–584, Aug. 2004.