IET Science, Measurement & Technology

Research Article

ISSN 1751-8822
Health index calculation for power Received on 1st May 2016
Revised on 5th June 2016
transformers using technical and economical Accepted on 8th June 2016
doi: 10.1049/iet-smt.2016.0184
parameters www.ietdl.org

Hamed Zeinoddini-Meymand, Behrooz Vahidi

Department of Electrical Engineering, Amirkabir University of Technology (Tehran Polytechnic), 424 Hafez Ave, Tehran 1591634311, Iran
E-mail: [email protected]

Abstract: In this study technical diagnostic tests and economical lifetime assessment of transformers are investigated to
evaluate the overall health condition of working transformers. Two artificial intelligence models including artificial neural
network and adaptive neuro-fuzzy inference system models are presented to determine the health index for transformers.
The technical and economical parameters are used as input parameters to develop the models. Technical parameters are
extracted from oil characteristics and dissolved gas analysis of different transformers. Economical parameters are
constructed with transformer capital investments, maintenance and operating costs. The models are developed using 226
experimental field datasets of transformers technical and economical parameters. The models are trained using 80% of
the experimental datasets. The remaining 20% is used to evaluate the performance and applicability of the models. The
results prove that the models can be used to determine the health condition of transformers with high accuracy.

1 Introduction and electrical tests, and the information of operating and

maintenance costs of transformers. It means that the proposed HI
Power transformer is one of the most critical assets in electricity is calculated using the technical and economical parameters
networks. Comprehensive condition assessment of transformers altogether.
should be developed to achieve sustainable reliability of power ANN and adaptive neuro-fuzzy inference system (ANFIS) models
system. The overall condition of a transformer should be evaluated are used to calculate transformer HI. The mature well-known ANFIS
in a comprehensive method considering technical and economical and ANN methods are used as modelling tools to calculate
parameters. It should be noted that most utilities operators are highly transformer HI with technical and economical parameters. ANN
motivated to assess the overall condition of transformers because [15, 27] and ANFIS [28, 29] methods are used in the literature for
there is an increasing demand for improved economical and fault diagnosis and aging of power transformers, but they have not
technical performance. To achieve the optimal balance among capital been used for transformer HI calculation.
investments, asset maintenance costs, and operating performance, The models are applied on 226 experimental eld datasets of
there is a need to provide economic and technical evaluation for transformers provided by Iran Transformer Research Institute
engineering decisions and capital replacement plans [14]. (ITRI) and their performances are compared. The operating and
Transformer condition assessment is an important issue that loading conditions of transformers are different depending on the
studied in the literature from different aspects. The frequency type of the industrial facilities they are utilised in, such as power
response analysis test as one of the sensitive tools for detecting plants, regional electricity companies, petrochemical and renery
electrical and mechanical faults inside transformers is presented in plants, cement, steel and other industries.
[5, 6]. Degradation of cellulosic insulation of transformer Technical parameters obtained from transformer diagnostic tests
considering the formation of furan products in insulating oil and include breakdown voltage (BDV), dissipation factor (DF at 90
effects of oxygen and water content on transformer oil aging are C), Acidity, interfacial tension (IFT), Water (Water content in oil at
studied in [79]. Investigation on furan compounds in transformer 20C), %WaterPaper (per cent Water in Paper insulation),
oil [7, 10], dissolved gas analysis (DGA) [9], transformer Furfural (2-Furaldehyde content), and DGA factor (DGAF).
dielectric monitoring [11], and moisture content of oil and paper Economical parameters obtained from transformer aging variables
insulation [1214] are presented in the literature. Methods for life and cost functions are PE% [per cent of economical lifetime
assessment of transformers are investigated in [1, 4]. Condition (EL)], and FAA (aging acceleration factor). These 10 parameters
assessment of power transformers, fault diagnosis of power are the inputs of ANN and ANFIS models, and the output
transformers, and prediction of thermal ageing in transformer oil parameter of the models is HI.
using articial neural networks (ANN) are presented in [15]. The main contributions of the paper are: (i) Considering the
Health Index (HI) is a procedure of combining complex condition technical and economical aspects of transformer characteristics
information to give a single numerical value as a comparative (using 10 various technical and economical parameters) at the same
indication of overall condition of transformer. It helps the operator time to calculate an overall HI for transformers. (ii) Considering
to make the distinction between degradation which needs DGAF parameter instead of total dissolved combustible gas, and
maintenance and diagnosis plans, and degradation that indicates proposing the inclusive PE% parameter to provide a comprehensive
approaching end of life. HI provides a methodology of employing economical view. (iii) Investigating on a large experimental dataset
existing engineering knowledge and experience to predict future collected from different power transformers.
performance, failure probabilities and replacement plans.
Transformer HI evaluation using technical parameters are presented
in [1621], and the economical aspects of transformer operation are 2 Modelling
investigated in [2226].
In this paper a quantitative approach is presented to calculate the In this section ANN and ANFIS models are presented to calculate
HI of transformers, which combines the results of various chemical the transformer HI value. Dataset is divided randomly into training

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
& The Institution of Engineering and Technology 2016 823
(80% of dataset) and testing (20% of dataset) subsets which the 2.2 ANFIS model
testing dataset is used to evaluate performance of the models. The
models should be trained in order that predict the HI for unseen ANFIS is an adaptive network, consisting of a number of nodes
data (testing dataset) with possible least deviation (error) from connected through directional links, which uses neural network
experimental eld HI values. learning algorithms and fuzzy reasoning to map inputs into an
One of the strong points of this work is use of a large experimental output. The ANFIS is a strong tool for the prediction and
dataset. In this paper a diverse dataset of 226 test records of power simulation of complex non-linear systems. The hybrid neuro-fuzzy
transformers with different voltage levels and power ranges in approach works by applying neural learning rules to identify and
different weather and operating conditions are used. Diagnostic tune automatically the membership function parameters [31]. The
tests are conducted on power transformers located in different ANFIS structure consists of ve layers (fuzzy layer, product layer,
regions of vast country of Iran with different climates in terms of normalised layer, defuzzify layer, and total output layer) is shown
temperature, humidity and atmospheric pressure. Moreover, in Fig. 2.
operating and loading conditions of transformers are different In this work, the ANFIS model is developed on the basis of the
depending on the type of the industrial facilities they are utilised subtractive clustering algorithm with eight inputs and one output.
in. Therefore, the results of the models will be accurate by using The subtractive clustering method partitions the data into groups
such a diverse and large dataset. The model provides the best called clusters, and generates a fuzzy inference system with the
prediction of the HI and decisivelycan be referred as a reliable minimum number of rules required to distinguish the fuzzy
model trained by the diverse dataset. ANN and ANFIS methods qualities associated with each of the clusters. The advantage of the
are data driven tools and their parameters and weight matrices are subtractive clustering algorithm is the fact that the number of
adjusted using input/output data. clusters does not need to be specied in advance and the
algorithm itself determines the number of clusters. In this method,
the total number of fuzzy rules is only related to the number of
clusters. Hence, it will be a correct choice to use this algorithm for
2.1 ANN model solving the problems with the large number of input dimension.
ANFIS applies a hybrid learning algorithm which is a
ANN is an articial intelligence method that its idea is inspired by combination of least-squares and BP gradient descent methods, in
biological structure of the human brain. Each kind of data with order to train the network according to inputoutput data pairs. In
complicated relations can be modelled via ANN which can operate the hybrid learning algorithm the gradient descent method is used
like a black box model that requires no detailed information about to assign the premise parameters in layer 1, whereas the least-
the investigated system [15, 30]. squares method is employed to identify consequent parameters in
ANN is an efcient model that assesses transformer health layer 4. The hybrid learning procedure is an efcient method to
condition by learning the relationships between inputs (transformer obtain the optimal premise parameters and consequent parameters
technical and economical parameters) and output (HI) based on in the learning process [31].
training data.
To apply the technique, a three-layer (one hidden layer)
feed-forward neural network trained with the Levenberge
Marquardt (LM) back-propagation (BP) algorithm is employed. 3 Input parameters for the models
From a practical perspective, it has been shown through extensive
experiments that single-hidden-layer neural networks are superior The technical and economical input parameters of the models used to
to networks with more than one hidden layer with the same level calculate the output of the models (HI value), are BDV, DF, Acidity,
of complexity and also the latter are more susceptible to fall into IFT, Water, %WaterPaper, Furfural, DGAF, PE%, and FAA. The
poor local minima. In engineering applications, there is a clear experimental eld HI values provided by transformer experts at
tendency toward using neural networks with only one hidden layer ITRI are used as the output for the presented models.
[30]. The schematic diagram of the ANN architecture is illustrated
in Fig. 1.

3.1 Breakdown voltage

BDV is a measure of the ability of the insulating oil to withstand

electric stress and has primary importance for the safe operation of
electrical equipment. Dry and clean oil exhibits an inherently high
breakdown voltage. Free water and solid particles, the latter
particularly in combination with high levels of dissolved water,
tend to migrate to regions of high electric stress and reduce
breakdown voltage dramatically. A low value of breakdown
voltage can indicate that one or more of these are present.
However, a high breakdown voltage does not necessarily indicate
the absence of all contaminants [3235].

3.2 Dissipation factor at 90C

DF is a measure for dielectric losses within the oil. This parameter is

very sensitive to the presence of soluble polar contaminants, ageing
products or colloids in the oil. Changes in the levels of the
contaminants can be monitored by measurement of this parameter
even when contamination is so slight as to be near the limit of
chemical detection. High values of DF may deleteriously affect the
dielectric losses and/or the insulation resistance of the electrical
equipment. Useful additional information can be obtained by
measuring DF at both ambient temperature and a higher
Fig. 1 Schematic diagram of the ANN architecture temperature such as 90C [3235].

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
824 & The Institution of Engineering and Technology 2016
Fig. 2 Schematic diagram of the ANFIS structure

3.3 Acidity process is slow. Hence, there is a water buildup in the oil, giving
the impression of a wet transformer. Such variations in oil water
The acidity (neutralisation value) of oil is a measure of the acidic content hardly affect water content of paper. This is not surprising
constituents or contaminants in the oil. The acidity of used oil is because more than 99% of the water is in the solid insulation. The
due to the formation of acidic oxidation products. Acids and other water in the oil could be a true indicator of the water in paper,
oxidation products will, in conjunction with water and solid only if the paper and oil are in thermal equilibrium, which almost
contaminants, affect the dielectric and other properties of the oil. is never the case in operating transformers [3235]. In this paper
Acids have an impact on the degradation of cellulosic materials the %WaterPaper parameter is considered as an individual
and may also be responsible for the corrosion of metal parts in a parameter to monitor the insulating paper condition, because the
transformer. The acidity increment rate of the oil in service is condition of transformer solid insulation is very important factor in
a good indicator of the ageing rate. The acidity level is used as a determination of transformer health condition.
general guide for determining when the oil should be replaced or
reclaimed [3235].
3.7 2-Furaldehyde content (Furfural)
3.4 Interfacial tension Furanic compounds are generated by the degradation of cellulosic
materials used in the solid insulation systems of electrical
The IFT between oil and water provides a means of detecting soluble equipment. Furanic compounds that are oil soluble to an
polar contaminants and products of degradation. This characteristic appreciable degree will migrate into the insulating liquid. All of
changes fairly rapidly during the initial stages of ageing but levels these compounds except 2-furaldehyde are not very stable under
off when deterioration is still moderate. With overloaded operating conditions found in transformers. These compounds
transformers, the deterioration of materials is rapid and IFT is a apparently form and then further degrade to 2-furaldehyde over a
tool for detection of deterioration [3235]. time span of a few months. 2-furaldehyde is apparently stable for
several years under the same conditions. The presence of high
3.5 Water content in oil at 20C (Water) concentrations of furanic compounds is signicant in that this may
be an indication of cellulose degradation from aging or incipient
Depending on the amount of water, the temperature of the insulating fault conditions [34, 35].
system and the status of the oil, the water content of insulating oils
inuences the breakdown voltage of the oil, the solid insulation, and
3.8 Dissolved gas analysis factor
the ageing tendency of the liquid and solid insulation. There are two
main sources of water increase in transformer insulation: ingress of In this paper the purpose of using DGA is to assess the overall health
moisture from the atmosphere; and degradation of insulation. condition of the transformer, not to determine the types of faults
For the proper interpretation of moisture content and for trending inside the transformer. Therefore, the DGA parameters (seven
purposes, the analytical result of water content in the oil given at a dissolved gases) are combined to one inclusive DGAF [18]
sample temperature needs to be corrected to that at a dened parameter. In some studies the effect of DGA parameters on
temperature. For practical reasons, the dened temperature is set at transformer HI is considered using dissolved combustible gas
20C, since below 20C the rate of diffusion of water is too slow parameter [17] which is the simple summation of DGA gases
to achieve equilibrium in operational equipment [32]. In this paper except CO and CO2 and has two disadvantages. The rst is that
in order to make the parameters at different oil temperatures CO and CO2 gases which have useful information about paper
comparable, the corrected values (to 20C) are used according to insulation degradation [16, 36, 37], are ignored. The second
standard [3235]. disadvantage is that the importance and weighting of different
gases [16, 18] is not considered.
3.6 Per cent water in paper insulation The DGAF is calculated by considering different dissolved gas
limits recommended in the standards [36, 37] as follows [18].
Monitoring water content in oil is part of a set of routine tests for
transformer oil. However, wet oil does not always mean a wet
7
i=1 Si Wi
paper insulation. As a transformer cools down due to load DGAF =
7
(1)
reduction or shut down, water tends to return to the paper, but this i=1 Wi

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
& The Institution of Engineering and Technology 2016 825
where i denotes seven dissolved gases (H2, CH4, C2H6, C2H4, C2H2, Thus, the estimated OC in year t is calculated as follows [39, 40].
CO and CO2), Si is the score value based on the volume of dissolved
gases, and Wi represents the weighting factor of each individual OC1
dissolved gas. The weighting factors are equal to 1 for both CO OCt,estimate = (10)
(1 + i )k1
and CO2; 3 for CH4, C2H6, and C2H4; 5 for C2H2; and 2 for H2 [18].
where, OCt,estimate is the estimated OC in year t, OC1 is the OC in
3.9 Per cent of economical lifetime (PE%) reference year, i is the increase rate of OC, and k is number of
years to reference year (k for the reference year is 1).
In this paper, the economical lifetime (EL) is calculated by life cycle The year in which the EUAC parameter has the minimum value is
cost (LCC) analysis performed in accordance to standard [38]. LCC the transformer EL. Therefore, per cent of EL (PE%) parameter is
analysis is an essential method of analysis in economic evaluation dened as follows.
which structurally decides and equalises costs within overall life
cycle. Life cycle costing is the process of economic analysis to Age
PE% = 100 (11)
assess the total cost of acquisition, ownership and disposal of a EL
product. It can be applied to the whole life cycle of a product or to
parts or combinations of different life cycle phases. The primary where, Age is the age of the transformer at the date of test, and EL is
objective of life cycle costing is to provide input to decision the transformer EL.
making in any or all phases of a products life cycle [38].
Annual equivalent cost analysis is used to model cost function in
LCC analysis. In this study, the equivalent uniform annual cost 3.10 Aging acceleration factor
(EUAC) method for a cash ow analysis during an operation
Since, the temperature distribution is not uniform; the part that is
period is employed [22, 39, 40].
operating at the highest temperature will ordinarily undergo the
While a transformer ages, the maintenance costs should be
greatest deterioration. Therefore, in aging studies it is usual to
increased to maintain the optimal performance of transformer.
consider the aging effects produced by the highest (hottest-spot)
Therefore, the operating and maintenance costs are considered in
temperature [41]. The parameter of FAA for a given load and
the economic lifetime evaluation by calculating the minimum
temperature is calculated as follows.
EUAC of the transformer [22, 39, 40].
The EUAC is used to determine an EL with consideration of an  
interest rate on investment cost (IC) and operating cost (OC), both 15000 15000
FAA = exp (12)
of which refer to the cash ow during a time period [22, 39, 40]. 383 uH + 273

EUAC(n) = IC(n) + OC(n) (2) where, H is the winding hottest-spot temperature in C.

where, n is the year of study.

IC is generally composed of initial purchase cost (PC) and salvage 4 Results and discussion
value (SV). The IC decreases annually [39, 40].
The dataset includes 226 sets of technical and economical data of
   transformers provided by ITRI. This dataset divided randomly
A A into training (80% of dataset) and testing (20% of dataset) subsets.
IC(n) = PC , i, n SVn , i, n (3)
P F Therefore, 181 sets of training data have been used for developing
the models whereas the remaining 45 sets of testing data are
  used to evaluate and demonstrate the performance of the trained
F
SVn = SV1 , i, n (4) model in the HI prediction with the transformer parameters. The
P
results of HI calculation with ANN and ANFIS models are
  illustrated in the following.
A i(1 + i)n Performance of ANN is generally based on parameters of its
, i, n = (5)
P (1 + i)n 1 architecture and setting. Appropriate designation of the initial
amounts of weights and biases is very effective on the
 
A i performance of the network. The BP LM training algorithm is a
, i, n = (6) non-linear optimisation method which may not necessarily lead to
F (1 + i)n 1
a unique solution at each run. One of the most difcult tasks in
  studying ANN is nding an appropriate architecture. This task is
F
, i, n = (1 + i)n (7) performed via trial and error and the optimum number of neurons
P in hidden layer is identied. Increasing the number of the hidden
layer neurons leads to improvement of the estimation ability of
where, i is the interest rate, n is the time period (year) of study, and ANN, but when it exceeds an optimum number, the over-tting
SVn is the SV at year n. problem may occur. It means that the network has memorised the
OC is composed of overhaul cost, preventative maintenance cost, training examples, but it has not learned to generalise for the new
the cost of diagnostic tests, labour cost and so on. The OC increases situations (unseen data). The way of identifying suitable
annually [39, 40]. architecture of ANN that is very time-consuming is the trial and
error.
    In this study ANN has 10 input neurons, one hidden layer

n
P A
OC(n) = OCt , i, t , i, n (8) containing S neurons and one output layer with one neuron.
t=1
F P Therefore, the numbers of ANN modiable parameters (weights
and biases) are 12S + 1. Since the size of training dataset should
  be several times the number of modiable parameters, the number
P 1
, i, n = (9) of hidden neurons should not be high. Moreover, as a rule of
F (1 + i)n thumb the number of hidden neurons should be between the size
of the input layer and output layer [42]. Therefore, in this study it
Due to the long-time operation of some transformers and lack of is not considered more than eight neurons in hidden layer. The
data, it is difcult to obtain data of the OC in the previous years. optimal number of hidden neurons can be determined by nding

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
826 & The Institution of Engineering and Technology 2016
Table 1 Comparison of average RMSE for different ANN configurations

Bias (bk)

2.4225
No. of Best Worst Average Standard
hidden deviation

Output layer
neurons

1 0.1704 0.3816 0.2311 0.0802

Weights (wjk)
2 0.1398 0.2076 0.1804 0.0211

6.4284
0.4293
3 0.1726 0.2560 0.2112 0.0263

HI
4 0.1622 0.3280 0.2321 0.0499
5 0.1569 0.3056 0.2324 0.0365
6 0.1634 0.2356 0.2033 0.0262
7 0.1861 0.3019 0.2200 0.0348
8 0.1748 0.3659 0.2436 0.0551

Biases (bj )

0.4338
4.5269
the network through comparison among average calculated root
mean squared error (RMSE).
In Table 1 the average and standard deviation of RMSE for testing
dataset is presented for different number of hidden neurons for 100

0.2501
31.6857
FAA
trails.
It can be concluded from Table 1 that optimal ANN conguration
has two neurons in hidden layer, because it results minimum average
and standard deviation of RMSE.
The weight and bias values of the optimal ANN conguration

0.1259
15.5378
PE%
(with two hidden neurons) have been given in Table 2.
In this paper the ANFIS model with subtractive clustering method
and hybrid learning algorithm is developed. The premise and
consequent parameters for the optimum ANFIS model are given in
Tables 3 and 4.

0.0342
12.2466
DGAF
For example, Rule 1 is as follows:

If x1 is A11 and x2 is A12 and x3 is A13 and x4 is A14 and x5 is A15 and
x6 is A16 and x7 is A17 and x8 is A18 and x9 is A19 and x10 is A1,10

0.0351
2.1099
Furfural
Then y1 = 0.006527x1 0.3871x2 + 0.6645x3 + 0.01995x4
0.03556x5 0.1573x6 0.5607x7 1.347x8 + 0.00004034x9
+ 0.4428x10 + 4.374
%WaterPaper

0.0165
12.9325
where x1, x2, x3, x4, x5, x6, x7, x8, x9, and x10 are BDV, DF, Acidity,
IFT, Water, %WaterPaper, Furfural, DGAF, PE%, and FAA
Weights (wij )
Hidden layer

parameters, respectively.
Unlike the ANN model, the ANFIS model is robust that gives the
same result in each run. The error criteria RMSE and R 2 of the HI
evaluation for the presented ANN and ANFIS models with
0.2942
1.0531
Water

training, testing and total dataset are given in Table 5. First the
models are constructed with training dataset and then HI values
are evaluated with trained model, and the RMSE is calculated. The
purpose of doing training with 80% of dataset and remaining 20%
of dataset for testing is for verication of the performance of the
Table 2 Weight and bias values of the optimal ANN configuration

0.3690
8.2666

models. It can be seen from Table 5 that the error values for test
IFT

dataset are close to the errors for train dataset. It proves the
performance and accuracy of the models confronted with unseen
test dataset.
The low values of the RMSE and the R 2 values close to unity in
0.1479
14.2057
Acidity

Table 5 show the accuracy of the presented models to calculate the

transformer HI. It can be seen from Table 5 that ANFIS method
produces superior results for train, test and total dataset. This
better performance can be explained by the fact that ANFIS
combines the learning capabilities of neural network and reasoning
0.3776
7.2802

capabilities of fuzzy logic. Hence it has an extended prediction

DF

capability compared with ANN and fuzzy logic.

To map the HI quantitative values into ve qualitative condition
categories [18, 43] to determine the overall health condition of
each transformer, the HI values are normalised onto scale of 0
0.1514
0.4849
BDV

(completely degraded transformer) to 1 (perfect condition).

Table 6 provides categories of HI values and correlates them to
the failure probability and overall health condition of transformers.
HI values are grouped into condition categories from very good
to very poor.
Neuron

The comparison of experimental normalised HI values provided

by ITRI and those predicted by presented ANN and ANFIS
1
2

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
& The Institution of Engineering and Technology 2016 827
Table 3 Premise parameters of the optimum ANFIS model
BDV DF Acidity IFT Water %WaterPaper Furfural DGAF PE% FAA
[i1,ci1] [i2,ci2] [i3,ci3] [i4,ci4] [i5,ci5] [i6,ci6] [i7,ci7] [i8,ci8] [i9,ci9] [i10,ci10]

rule 1 [8.96 75] [0.1730.0011] [0.0517 0.0629] [4.043 33.8] [5.582 2.3] [0.6889 1.908] [0.4121 0.116] [0.3125 0.993] [124.1 63.16] [0.01780.0056]
rule 2 [8.961 72.5] [0.145 0.0147] [0.0431 0.0564] [4.056 32.8] [5.582 5.9] [0.6671 3.193] [0.4413 0.184] [0.3711 1.118] [124.1 73.68] [0.0024 0.0103]

Table 4 Consequent parameters of the optimum ANFIS structure for predicting HI

pi1 pi2 pi3 pi4 pi5 pi6 pi7 pi8 pi9 pi10 pi0

rule 1 0.006527 0.3871 0.6645 0.01995 0.03556 0.1573 0.5607 1.347 0.00004034 0.4428 4.374
rule 2 0.01076 2.698 0.6902 0.01893 0.001488 0.09745 0.1325 0.7119 0.0001259 15.5 2.926

Table 5 RMSE of ANN and ANFIS models ranges (of Table 6). It means that the experimental HI indicates
Error criteria RMSE R2 that the transformer is at the end of one condition range, and the
predicted HI indicates that the transformer is in the beginning of
ANN train dataset 0.1657 0.9525 the adjacent condition range and vice versa. For example, for one
test dataset 0.1804 0.9422 of points of Grid 1, the predicted HI value is 0.708 (good
total dataset 0.1688 0.9503 condition) and experimental HI value is 0.672 (fair condition). It
ANFIS train dataset 0.1648 0.9529
test dataset 0.1552 0.9592 is clear that both values are close to border value (0.7) of two
total dataset 0.1629 0.9536 condition ranges. Therefore, it is proved that for the mentioned 45
transformers, the predicted and experimental HI values are not
widely divergent.
To demonstrate the accurate performance of the HI prediction for
models for 181 transformers of training datasets and 45 testing
the 45 unseen transformer datasets (test dataset), the train and test
datasets are shown in Fig. 3. The horizontal dashed lines of Fig. 3
datasets have been separated in Fig. 4. As it can be seen from
(which separate the HI ranges of Table 6) help to identify the
Fig. 4, the most of points of test dataset are also located at
health condition for each transformer of dataset (from very good
diagonal grids as well as train dataset points.
to very poor).
The input parameters, the output parameter HI and the condition
The normalised HI values predicted by the ANFIS model (the
of some sample transformers are shown in Table 7. The HI values
more accurate and robust model), versus experimental ones for
of Table 7 are normalised values.
total 226 datasets is shown in Fig. 4.
The differences in the HI condition of the transformers with the
Data points located in diagonal grids of Fig. 4 are the datasets that
same age or worse condition of a younger transformer compared
the health condition (mentioned in Table 6) of experimental and
with an old one may be the result of the loading pattern, the
ANFIS HI values is the same. However, datasets in non-diagonal
harmonic content of the load, a varying number of through faults,
grids show that the prediction is not the same as experiment. For
differences in maintenance plans, partial-discharge activities, or
example, for points of Grid 1, the ANFIS predicted and
problems in the forced cooling system.
experimental HI values have good and fair health conditions,
respectively.
It can be seen from Fig. 4 that the condition of predicted HI and
experimental HI are exactly the same for 80% of total 226 datasets 5 Conclusion
(located at ve diagonal grids of Fig. 4). Nevertheless, the results
for the remaining 20% of dataset (45 transformers) which are not In this paper, two articial intelligence models ANN and ANFIS are
located at diagonal grids (the condition of predicted and developed to determine a HI for transformers. The input parameters of
experimental HI are not the same for these transformers), are not the models are technical and economical parameters including oil
away from the diagonal grids. Therefore, the HI prediction characteristics and DGA related parameters and also maintenance
accuracy is really high for such a large and diverse dataset because and OCs and aging related parameters. The output parameter of the
all points of Fig. 4 are close to the diagonal line of y = x. The models is HI value. The proposed HI approach improves the quality
correlation coefcient between predicted and experimental HI is of condition assessment of the transformer by combining results
equal to 0.9765 which proves high agreement between predicted from various chemical and electrical tests, onsite inspections, and
and experimental HI values. economical information regarding the transformers maintenance
The condition of 45 transformers which have not the same and OCs.
predicted and experimental HI condition, are located at adjacent An experimental eld dataset with technical test records and
grids of the ve diagonal grids and at the border of two condition economical information for 226 transformers with different voltage

Table 6 Transformer health condition based on the normalised HI value [18, 43]

HI Condition Probability of failure Overall health condition

0.851 very good low (0%) satisfactory condition of transformer for continuous operation
0.70.85 good low but slightly increasing normal operation together with specific monitoring
(less than 1.6%)
0.50.7 fair rapidly increasing but lower than increase diagnostic testing with strict overall monitoring
probability at mean age
(between 1.6% and 6.9%)
0.30.5 poor higher than probability at mean restricted operation, increased interval sampling,
age and increasing and detailed diagnostics
(between 6.9% and 14.2%)
00.3 very poor very high, more than double the critical and immediate planning for emergency
probability at mean age major refurbishment or replacement
(more than 14.2%)

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
828 & The Institution of Engineering and Technology 2016
Fig. 3 Comparison of experimental normalised HI with ANN and ANFIS models

levels and power ranges in different weather and operating

conditions is provided by ITRI. The case study demonstrates the
applicability of the developed intelligent models in determining
the HI of the transformers. At rst each model is trained with 181
datasets and then the performance of the model is tested on other
45 sets. The results show that ANFIS model provides more
accurate and robust results comparing with the ANN model. By
the way, the two models can give satisfactory results, but the
ANFIS model is somehow superior.
The condition of predicted HI and experimental HI for ANFIS and
ANN are exactly the same for 80% of total 226 datasets.
Nevertheless, the results for the remaining 20% of dataset are not
away from the diagonal grids (matched points) and they are
located at the border of two adjacent conditions ranges.
The developed transformer condition assessment strategies result
in nancial benets with increased reliability, maximise the
transformers availability, and allow the transformers to be in
service beyond the expected design age.

6 Acknowledgments

Fig. 4 Predicted HI by ANFIS vs. experimental HI for all 226 datasets The authors gratefully acknowledge Iran Transformer Research Institute
(ITRI) for supporting this research. The authors thank Salar Askari,

Table 7 HI value and condition of sample transformers

Voltage ratio, 410 33/6.6 400/15.75 20/0.072 400/15.75 400/15.75 410/13.8 20/0.4 230 10.6/145
kV
Power rating, 100,000 10,000 94,000 2500 94,000 94,000 160,000 1250 84,000 22,000
kVA
Age (year) 2 5 7 9 12 14 14 31 35 41

input parameters BDV (kV) 75 81.8 74.8 81.7 75.5 70.5 75 76.4 75 72.6
DF 0.0015 0.0115 0.0072 0.0165 0.0016 0.0017 0.0014 0.0255 0.0155 0.1048
Acidity 0.04 0.0403 0.07 0.0203 0.06 0.06 0.06 0.0403 0.12 0.0016
(mgKOH/goil )
IFT (mN/m) 35.4 35.5 31.8 36.8 33.8 34.4 35.1 21.7 20.2 13.77
Water (ppm) 3 2.9 2.3 1.6 1 2.5 2.2 5.6 10.5 1.8
%WaterPaper 2.2 3.1 1.9 1.9 1 2 1.9 4.8 4.6 1.5
Furfural (ppm) 0.09 0.32 0.11 0.09 0.11 0.14 0.12 0.65 0.46 1.91
DGAF 1 1 1 1 1 1.111 1.111 1.278 2.056 1.556
PE% 10.526 8.475 36.842 19.149 63.158 73.684 73.684 134.783 269.231 585.714
FAA 5.25 105 4.47 105 1.57 104 1.41 105 1.15 104 4.47 105 4.47 105 1.41 105 3.23 105 5.44 103
output parameter experimental HI 1 0.9219 0.7656 0.8281 0.5625 0.6719 0.4062 0.4688 0.1719 0.2656
(HI) condition very good very good good good fair fair poor poor very poor very poor
ANFIS HI 1 0.9118 0.8296 0.8335 0.6007 0.6425 0.4177 0.4952 0.2269 0.2894
condition very good very good good good fair fair poor poor very poor very poor
ANN HI 0.9927 0.9187 0.8200 0.8067 0.5765 0.6302 0.3859 0.4696 0.2250 0.2467
condition very good very good good good fair fair poor poor very poor very poor

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
& The Institution of Engineering and Technology 2016 829
Majid Bahmanabadi (transformer expert at ITRI), and Hormatollah 20 Mohamadeen, K.I., Sharkawy, R.M., Salama, M.M.: Binary cat swarm
Firoozi (Managing director of Rastak-Tavan-Khavarmianeh) for their optimization versus binary particle swarm optimization for transformer health
index determination. Int. Conf. Engineering Technology ICET, Cairo, April
valuable suggestions for improving the quality of this paper. 2014, pp. 15
21 Dominelli, N., Lau, M., Olan, D., et al.: Equipment health rating of power
transformers. IEEE Int. Symp. Electrical Insulation, Indianapolis, IN, September
7 References 2004, pp. 163168
22 Abu-Elanien, A.E.B., Salama, M.M.A., Bartnikas, R.: A techno-economic method
for replacing transformers, IEEE Trans. Power Deliv., 2011, 26, (2), pp. 817829
1 Allan, D.M.: Practical life-assessment technique for aged transformer insulation,
IEE Proc. Sci. Meas. Technol. A, 1993, 140, (5), pp. 404408 23 Cao, W., He, B., Shen, W., et al.: An approach for economic assessment on
2 Brown, R.E., Willis, H.L.: The economics of aging infrastructure, IEEE Power oil-paper insulation diagnosis through accelerated aging experiments, IEEE
Energy Mag., 2006, 4, (3), pp. 3643 Trans. Dielectr. Electr. Insul., 2014, 21, (4), pp. 18421850
3 Wang, M., Vandermaar, A.J., Srivastava, K.D.: Review of condition assessment of 24 Wang, J., Liao, R., Zhang, Y., et al.: Economic life assessment of power
power transformers in service, IEEE Electr. Insul. Mag., 2002, 18, (6), pp. 1225 transformers using an improved model, CSEE J. Power Energy Syst., 2015, 1,
4 Qian, Z., Yan, Z.: Fuzzy synthetic method for life assessment of power (3), pp. 6875
transformer, IEE Proc. Sci. Meas. Technol., 2004, 151, (3), pp. 175180 25 Kawamura, T., Fushimi, Y., Shimato, T., et al.: Improvement in Maintenance and
5 Nirgude1, P.M., Ashokraju1, D., Rajkumar, A.D., et al.: Application of numerical Inspection and Pursuit of Economical Effectiveness of Transformers in Japan.
evaluation techniques for interpreting frequency response measurements in power Report 12-107, Session 2002, CIGRE SC12, Paris
transformers, IET Sci. Meas. Technol., 2008, 2, (5), pp. 275285 26 Sobocki, R., Kazmierski, M., Olech, W.: Technical and economic assessment of
6 Behjat, V., Mahvi, M.: Statistical approach for interpretation of power power transformers, the polish practice, Report 12-104, Session 2002, CIGRE
transformers frequency response analysis results, IET Sci. Meas. Technol., 2015, SC12, Paris
9, (3), pp. 367375 27 Bhalla, D., Bansal, R.K., Gupta, H.O.: Function analysis based rule extraction
7 Emsley, A.M., Xiao, X., Heywood, R.J., et al.: Degradation of cellulosic from articial neural networks for transformer incipient fault diagnosis, Electr.
insulation in power transformers. Part 2: formation of furan products in Power Energy Syst., 2012, 43, (1), pp. 11961203
insulating oil, IEE Proc. Sci. Meas. Technol., 2000, 147, (3), pp. 110114 28 Hooshmand, R.A., Parastegari, M., Forghani, Z.: Adaptive neuro-fuzzy inference
8 Emsley, A.M., Xiao, X., Heywood, R.J., et al.: Degradation of cellulosic system approach for simultaneous diagnosis of the type and location of faults in
insulation in power transformers. Part 3: effects of oxygen and water on ageing power transformers, IEEE Electr. Insul. Mag., 2012, 28, (5), pp. 3242
in oil, IEE Proc. Sci. Meas. Technol., 2000, 147, (3), pp. 115119 29 Khan, S.A., Equbal, Md.D., Islam, T.: A comprehensive comparative study of
9 Wang, X., Li, Q., Yang, R., et al.: Diagnosis of solid insulation deterioration for DGA based transformer fault diagnosis using fuzzy logic and ANFIS models,
power transformers with dissolved gas analysis-based time series correlation, IET IEEE Trans. Dielectr. Electr. Insul., 2015, 22, (1), pp. 590596
Sci. Meas. Technol., 2015, 9, (4), pp. 393399 30 Villiers, J., Barnard, E.: Backpropagation neural nets with one and two hidden
10 Abu-Siada, A.: Correlation of furan concentration and spectral response of transformer layers, IEEE Trans. Neural Netw., 1993, 4, (1), pp. 136141
oil-using expert systems, IET Sci. Meas. Technol., 2011, 5, (5), pp. 183188 31 Jang, J-S.R.: ANFIS: adaptive-network-based fuzzy inference system, IEEE
11 Yang, X., Nielsen, S., Ledwich, G.: Investigations of dielectric monitoring on an Trans. Syst. Man Cybern., 1993, 23, (3), pp. 665685
energised transformer oilpaper insulation system, IET Sci. Meas. Technol., 2015, 32 IEC 60422: Mineral insulating oils in electrical equipment-supervision and
9, (1), pp. 102112 maintenance guidance, 2013
12 Smith, D.J., McMeekin, S.G., Stewart, B.G., et al.: A dielectric frequency 33 IEC 60296: Fluids for electrotechnical applications-unused mineral insulating oils
response model to evaluate the moisture content within an oil impregnated paper for transformers and switchgear, 2012
condenser bushing, IET Sci. Meas. Technol., 2013, 7, (4), pp. 223231 34 IEEE Std. C57.106: IEEE guide for acceptance and maintenance of insulating oil
13 Toudja, T., Moulai, H., Nacer, A., et al.: Moisture and electrical discharges effect on in equipment, 2006
naphthenic mineral oil properties, IET Sci. Meas. Technol., 2014, 8, (6), pp. 588594
35 IEEE Std. C57.152: IEEE guide for diagnostic eld testing of uid-lled power
14 Cui, Y., Ma, H., Saha, T., et al.: Particle tracing modelling on moisture dynamics
transformers, regulators, and reactors, 2013
of oil-impregnated transformer, IET Sci. Meas. Technol., 2016, 10, (4), pp. 335343
36 IEEE Std. C57.104: IEEE guide for the interpretation of gases generated in
15 Xu, W., Wang, D., Zhou, Z., et al.: Fault diagnosis of power transformers:
oil-immersed transformers, 2008
application of fuzzy set theory, expert systems and articial neural networks,
IEE Proc. Sci. Meas. Technol., 1997, 144, (1), pp. 3944 37 IEC 60599: Mineral oil-impregnated electrical equipment in service-guide to the
16 Dehghani-Ashkezari, A., Ma, H., Saha, T.K., et al.: Application of fuzzy support interpretation of dissolved and free gases analysis, 2007
vector machine for determining the health index of the insulation system of 38 IEC 60300-3-3: Application guide-Life cycle costing, 2004
in-service power transformers, IEEE Trans. Dielectr. Electr. Insul., 2013, 20, 39 Sullivan, W.G., Wicks, E.M., Koelling, C.P.: Engineering economy (Prentice
(3), pp. 965973 Hall Press, 2012, 15th edn.)
17 Abu-Elanien, A.E.B., Salama, M.M.A., Ibrahim, M.: Calculation of a health index 40 Blank, L., Tarquin, A.: Engineering economy (McGraw-Hill Press, 2012, 7th
for oil-immersed transformers rated under 69 kV using fuzzy logic, IEEE Trans. edn.)
Power Deliv., 2012, 27, (4), pp. 20292036 41 IEEE Std. C57.91: IEEE guide for loading mineral-oil-immersed transformers and
18 Naderian-Jahromi, A., Piercy, R., Cress, S., et al.: An approach to power step-voltage regulators, 2011
transformer asset management using health index, IEEE Electr. Insul. Mag., 42 Blum, A.: Neural networks in C++: an object-oriented framework for building
2009, 25, (2), pp. 2034 connectionist systems (John Wiley & Sons, Inc., 1992)
19 Cui, Y., Ma, H., Saha, T.K.: Improvement of power transformer insulation 43 Hjartarson, T., Otal, S.: Predicting future asset condition based on current health
diagnosis using oil characteristics data preprocessed by SMOTEBoost index and maintenance level. 11th IEEE Conf. Transmision Distribution
technique, IEEE Trans. Dielectr. Electr. Insul., 2014, 21, (5), pp. 23632373 Construction Operation Live-Line Maintenance, Albuquerque, NM, October 2006

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 7, pp. 823830
830 & The Institution of Engineering and Technology 2016