Received October 20, 2021, accepted October 30, 2021, date of publication November 4, 2021, date of current version

November 12, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3125379

Artificial Intelligence-Based Power Transformer

Health Index for Handling Data Uncertainty
DHANU REDIANSYAH 1,2 , (Graduate Student Member, IEEE),
RAHMAN AZIS PRASOJO 1,3 , (Graduate Student Member, IEEE),
SUWARNO 1 , (Senior Member, IEEE), AND A. ABU-SIADA 4 , (Senior Member, IEEE)
1 School of Electrical Engineering and Informatics, Institut Teknologi Bandung, Bandung 40132, Indonesia
2 PT. PLN (Persero) Head Office, Jakarta 12160, Indonesia
3 Electrical Engineering Department, Politeknik Negeri Malang, Malang 65141, Indonesia
4 Electrical and Computer Engineering Discipline, Curtin University, Perth, WA 6102, Australia

Corresponding author: Rahman Azis Prasojo ([email protected])

This work was supported in part by PT. PLN (Persero), and in part by the Institut Teknologi Bandung.

ABSTRACT Power transformer is a critical and expensive asset in electric transmission and distribution
networks. It is essential to monitor the health condition of all power transformer fleet in such networks
to avoid unwanted outages. The health index (HI) is a quick and efficient way to assess the condition of
power transformers based on multi-criteria. While Power transformer HI method has been well presented
in the literature, not much attention was given to handle the uncertainty and reliability of this method due
to unavailability of used data. Therefore, this paper aims to tackle this issue through employing Artificial
Intelligence (AI)-based techniques to reveal the health condition of power transformers with high accuracy
and at the same time handling data uncertainty. The proposed HI approach assesses the power transformer
insulation system based on oil quality, dissolved gas analysis (DGA), and paper condition. In this regard,
collected data from 504, 150-kV transformers are used to establish the proposed AI-models. Seven AI
algorithms including k-Nearest Neighbor (kNN), Support Vector Machine (SVM), Random Forest (RF),
Naïve Bayes (NB), Artificial Neural Network (ANN), Adaptive Boosting (AdaBoost), and Decision Tree are
investigated. A performance comparison of the proposed AI-based HI models is carried out using the scoring-
weighting-based HI method as the reference. Results show that RF model provides the best performance in
predicting power transformer HI with an accuracy of 97.3%.

INDEX TERMS Power transformer, health index, insulation system, condition monitoring, artificial
intelligence.

I. INTRODUCTION the aging mechanism of the transformer oil and paper insu-
Power transformers are among the most critical and expen- lation have been published in the literatures [3]–[7]. The
sive assets within the electrical transmission and distribution health condition and remnant life of power transformers have
networks. With the continous increase in the load demand, been assessed through various parameters of the insulation
power transformers are operating close to their nominal rat- system [8], [9].
ings and becoming more prone to failures. Therefore, power Reliable and cost effective condition monitoring and asset
transformers need to be continuously monitored during their management techniques are essential for electric utilities in
entire operational life to avoid sudden and catastrophic fail- preparing an appropriate financial plans to estimate the future
ures. Over the past two decades, several condition monitoring cost of maintenance, repair, and replacement of their power
techniques have been developed for power transformers [1]. transformer fleet. In this regard, much research effort has
Among these techniques, insulation system has been a com- been conducted to help utility companies optimize their asset
mon key component to identify transformer health state and maintenance costs. Transformer asset management (TAM)
estimate its useful rmnant life [2]. Several papers to explain practice has been explained in [10]–[12] in which strategic
plans for future maintenance and replacement activities based
The associate editor coordinating the review of this manuscript and on various diagnostic testing methods are presented. The
approving it for publication was Zhouyang Ren . diagnostic testing methods can be generally categorized into

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 150637
 D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

condition monitoring (CM) and condition assessment (CA) using several factors, industry standards, and utility expert
techniques. CM techniques employ electrical, chemical, and judgment [27]. AI-based general regression neural network
physical tests to be collectively used by the CA tools to (GRNN) was introduced in [28] to calculate the HI based
determine the health condition of the transformer. on four-class transformer conditions; very poor, poor, fair,
The transformer health index (HI), as primarily explained and good. In [29], AI-based HI using Bayesian network was
in [13], is a single factor that utilizes the information from proposed to quantify the parameters contribution through
operating observations, field inspections, and laboratory tests score-probability and population failure statistics. In [30],
to provide a reliable TAM decision. Majority of the data used principal component analysis (PCA) and analytical hierarchy
in the HI model are based on the insulation system testing. process (AHP) were proposed to determine the transformer
Oil insulation tests include dissolved gases analysis (DGA), health condition based on an expert empirical formula.
oil quality analysis (OQA), and furan analysis (FFA). Due to The use of AI methods to define transformer health
the high electric and thermal stress within operating trans- index has been presented in the literatures. For instance,
formers, oil and paper insulation decomposes and releases decision tree, random forest (RF), static vector machine
some gases that dissolve in the oil and decrease its dielectric (SVM), artificial neural network (ANN), and k-nearest neigh-
strength. These gases include hydrogen (H2 ), methane (CH4 ), bor (kNN) methods were used in [27] to automate the assess-
ethylene (C2 H4 ), acetylene (C2 H2 ), ethane (C2 H6 ), carbon ment process. A study in [31] compared several machine
monoxide (CO), and carbon dioxide (CO2 ). DGA test is learning algorithms such as ANN, SVM, and Gaussian
conducted to quantify these gases from which internal trans- bayesian networks (GBN), to evaluate the health condition
former electrical and thermal faults can be identified [14]. of power transformers through probabilistic HI framework.
Considering the guideline in [15], the OQA is determined by An approach is presented in [32] that compared the perfor-
analyzing the oil breakdown voltage (BDV), acidity, water mance of various AI methods such as naïve Bayes (NB),
content, interfacial tension (IFT), dielectric dissipation fac- multinomial logistic regression (MLR), ANN, SVM, kNN,
tor (DDF), and color. FFA is conducted to measure furan one rule (OneR), decision tree (J48), and RF in identifying
compounds that are generated due to cellulose degradation transformer health index. However, regardless the various
and dissolve in the transformer oil [16]. Among the 5 furan AI methods proposed in the literatures to calculate the
compounds, furfuraldehyde (also known as furfural/2FAL) transformer HI, more thorough studies are still required to
dominates the measurements and is correlated to the degree improve the accuracy of such methods, in particular with the
of polymerization of the paper insulation [17]. uncertainty or unavailability of the used data.
A scoring-weighting (also called weighted-sum) is the As discussed in [9], data uncertainty is the main shortcom-
most commonly used method to calculate the HI of power ing of the HI method which affects its reliability and accuracy.
transformers [13], [18]–[23]. The approach starts by com- Data unavailability is the main reason of uncertainty. The
paring each parameter to a scoring table, then weighting study in [33] reported the effect of data unavailability and
each parameter based on its importance. The weights are proposed a remedial approach based on RF to predict the
usually identified by expert personnel. The individual scores missing data and improve the scoring-weighting HI. How-
are combined into a single index that reveals the overall health ever, the development and investigation of AI-based HI model
condition of the transformer. in handling data uncertainty has not yet discussed thoroughly
As the HI is a linear combination of different scores and with proven implementation feasibility.
weighted measurement data, it is a challenging task to deal
with the uncertainty of the used data through the current TABLE 1. Summary of previous researches.
utility practice to identify the HI.
The rapid development of computer science and data
processing has resulted in new HI approaches based on
machine learning algorithms for big data analysis [9]. Arti-
ficial Intelligence (AI)-based HI approach was presented
in [24] using DGA, furan, and oil test data to assess the
health condition of several in-service transformers. However,
due to the limited available data, the prediction accuracy
of the developed neuro-fuzzy model has only been 56.3%.
The study in [25] conducted a transformer HI sensitivity
analysis using self adaptive neuro-fuzzy inference system
(ANFIS) whose parameters were tuned using particle swarm Table 1 shows the summary of the previous researches.
optimizer (PSO). In [26], probabilistic Markov chain models From the above discussion, the gaps in this area of research
were used to predict the future condition of the transformer can be summarized as below:
based on HI calculation using a non-linear optimization tech- • despite the several AI-based methods used to identify the
nique. AI with a fuzzy-based support vector machine (SVM) HI of power transformers, accuracy of such methods is
was employed in transformer HI-based condition assessment still relatively low.

150638 VOLUME 9, 2021

D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

•there is no reliable and widely accepted technique to deal unavailable data on the accuracy of both conventional SW
with data uncertainity which affects the reliability of the and the identified best performing AI-based HI is evaluated
current HI practice. by adjusting the complete data into 50 missing data scenarios.
As such, the main contribution of this paper is to: A preprocessing approach is then proposed to enhance the
• construct an AI-based approach to enrich the current
ability of the AI-based HI model to deal with such unavailable
conventional HI scoring-weighting method for power data and provide an accurate HI value.
transformers.
A. TRANSFORMER POPULATION
• improve the accuracy of the HI calculation using multi-
AI models to process most common routine testing data The data used in this study were collected from PT. PLN
for power transformer insulation system. PERSERO, an Indonesian state electricity company. Data are
• develop an effective correlation for all used parameters
collected from 504 units, all with a specific high voltage
in calculating the transformer HI. of 150 kV with a low voltage of either 70 kV or 20 kV. Major-
• identify the best AI-based combination for dealing with
ity of these transformers use Kraft paper insulation and they
data uncertainty and missing values. are periodically inspected in a condition-based maintenance
scheme by checking their insulation properties through DGA
and dielectric characteristics (oil quality analysis). Dielec-
tric characteristics include oil breakdown voltage (BDV),
water content, interfacial tension (IFT), acidity, and color.
In addition, 2-FAL measurement are provided for some units.
Normally, the oil testing is conducted once a year or more
frequent, if needed. Data used for the analysis in this paper are
the same year data when HI is calculated, therefore omitting
the uncertainty caused by outdated data. The age classifica-
tion of the observed population sample is shown in Figure 2.

FIGURE 2. The age classification of the transformer population sample.

B. TRANSFORMER HEALTH INDEX

FIGURE 1. Flowchart of the proposed methodology.
There are many methods to calculate the transformer HI
score that reflects the overall health condition of the power
II. METHODOLOGY transformer. Generally, there are two main approaches to
The study in this paper is conducted in accordance to the determine the HI:
workflow of Figure 1. Diagnostic data of 504, 150kV-power Scoring-Weighting: This approach starts with a scoring and
transformers including oil characteristics, furan content, and weighting process conducted by expert personnel. As shown
DGA were collected from in-service units of various life in Figure 3, condition data are processed into scores by com-
spans and health conditions. The assessment data are used to paring them to the scoring tables. Then, individual scores are
calculate the power transformer insulation system HI using weighted and aggregated into a single index value revealing
the conventional Scoring-Weighting (SW) method. At the the overall transformer health condition. A final health index
same time, data are used to establish the AI-based HI method. is a linear combination of different scored and weighted mea-
The performance of various AI algorithms is compared to surement data. Therefore it is a challenging task to conduct
identify the best performing AI algorithm through compar- a consistent and reliable HI process in case of some of the
ison with the SW method as a reference. The effect of data are unavailable. In this regard, this paper highlights the

VOLUME 9, 2021 150639

 D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

is divided into three-layers as shown in Figure 5 which is

based on previously proposed method in [20]. The HI is
equipped with scoring tables to score each measurement,
as well as weighting factors based on the reliability and
criticality of each parameter. These tables and weighting fac-
tors can be continuously adapted based on new information,
experience, and recent international guidelines. The layers in
Figure 5 include:
Data layer: this layer consists of frequently measured diag-
nostic data such as key gases of DGA, BDV, water content,
IFT, acidity, oil color, 2FAL, and transformer age.
Factor layer: This layer has three categories that are
derived from data layer/ measurement:
1. Oil Quality Factor (OQF) that consists of oil parameters;
FIGURE 3. The principle of conventional HI scoring-weighting method. BDV, water content, IFT, acidity, and color.
2. Faults factor (FF), which consists of gas evolution, gas
concentration level, and Duval’s pentagon DGA interpreta-
drawback of the conventional scoring-weighting models in
tion method (DPM), which is shown in Figure 6.
handling unavailable data using the certainty level and output
3. Paper condition Factor (PCF): This factor consists of
accuracy as proposed in [33].
the operating age, CO/CO2 ratio, and 2FAL concentration.
Artificial Intelligence: various AI methods have been
Health Index layer: in this layer the HI calculation is
employed to estimate the transformers HI. The typical
conducted to provide a single value that reveals the overall
AI-based HI model development is as shown in Figure 4.
health condition of the investigated transformer.
In this method, training database including various diagnostic
measurements conducted on several transformers at differ-
ent health condition levels must be carefully prepared. The
training dataset is used to establish the AI model to estimate
the overall transformer health condition. As discussed above,
Various artificial intelligence algorithms have been used to
establish such model including ANN, ANFIS, SVM, GRNN,
decision trees [34]. The use of AI in power transformer
condition assessment is useful especially for analysing large
transformer datasets [35].

FIGURE 5. Layers of the proposed AI method.

FIGURE 4. The principle of AI-based HI method.

D. SCORING-WEIGHTING-BASED HI
All of the data used to calculate the transformer HI are
C. POWER TRANSFORMER INSULATION SYSTEM HI determined by international standards limits [37], [38]. The
The HI is developed based on the transformer reliability scoring-weighting based HI for each factor is calculated as
survey published by CIGRE TB 642 [36] that is discussing in (1).
the failure modes of high voltage power transformers. The Pn
survey results show that the most common fault location i=1 Si Wi
HI each factor = P n (1)
for power transformer is the winding, with a dominant fault 1 Wi
mode within the dielectric system. The main cause of failures where n is the number of parameters used in every factor. The
includes aging and external short circuit. Therefore, this study value of Si is based on the scoring for each parameter based
focuses mainly on the integrity of the power transformer on standard scoring tables, and Wi is the weighting factor that
insulation system. The structure of the proposed HI method describes the importance of every parameter.

150640 VOLUME 9, 2021

D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

TABLE 2. The scoring-weighting for oil quality. TABLE 6. The scoring-weighting paper condition.

TABLE 3. The level scoring for DGA.

FIGURE 6. Duval’s pentagon DGA interpretation method.

TABLE 4. The rate scoring for DGA.

TABLE 7. The scoring for health index factor.

TABLE 5. The matrix of fault factor category.

TABLE 8. The health index categories.

The assessment methods are based on Tables 2 through 7,

along with Figure 6. The final HI value is calculated using (2).
Pn
j=1 SF j Wj
HI final = Pn x100% (2)
j=1 4W j

where SFj is the parameter’s scoring factor and Wj is the condition monitoring, fault diagnosis and asset manage-
weighting factor. ment [39], [40].
The weighting parameters and factors are obtained by The health index category and description is classified as
using AHP. This technique is built based on the judgment shown in Table 8. A more detailed explanation of the scoring-
of five experts with in-depth experience in transformer weighting method can be found in [20].

VOLUME 9, 2021 150641

 D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

satisfactorily accuracy, additional tuning process is conducted

through more training. On the other hand, if the accuracy has
met the criteria, the final HI is calculated as per the process
shown in Figure 7. Various AI algorithms are employed to
establish the proposed HI model; this includes kNN, SVM,
RF, NB, ANN, Adaptive Boosting (AdaBoost), and Decision
Tree (Tree). The proposed method to calculate the trans-
former HI using multi-AI approaches is shown in Figure 8.
The performance of these approaches is compared based on:
AI-SW: which is based on factors category stage (OQF, FF,
and PCF), then scoring-weighting-based in HI category stage.
AI-Ful: which is based on all classification stages (OQF,
FF, PCF, and HI category) of the AI model.

F. DEALING WITH DATA UNCERTAINTY

All assessments of transformers include a non-avoidable
level of uncertainty. Unavailability of the data, erroneous,
or obsolete data will adversely affect the assessment results
and increase the uncertainty of the obtained outcomes. The
FIGURE 7. A flowchart for the AI-based HI model’s development.
uncertainty within available data is due to different reasons
such as incorrect data entry, erroneous or questionable test
results, uncertainty in the condition assessment, and obsolete
data [9].
A solution to manage the unavailability of the data is by
employing machine learning to estimate the missing param-
eters based on historical data and well training process.
Following the characteristics of the data, several methods can
be used to preprocess the missing values. In this study, the

FIGURE 8. Flowchart of the proposed multi-AI approaches to calculate HI.

E. AI-BASED HI DEVELOPMENT
To establish the AI-based HI model, the collected transformer
data have been divided into 354 training and 150 testing data
sets. Each set comprises data of transformers of different
aging, operating and health conditions.
The output of the AI model is the overall HI which
is validated using the corresponding HI calculated by the
scoring-weighting method. If the AI model does not provide FIGURE 9. Dealing the uncertainty with AI-based HI.

150642 VOLUME 9, 2021

D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

average/most frequent value is used to handle the issue of

unavailable data as shown in Figure 9.
Three methods are used to implement the proposed
approach namely; scoring weighting (SW), random forest
(RF), and Random forest with preprocessing-based (RFwP).
The certainty level (CL) is calculated as below [33]:
CL i = WPi xWf i (3)
P
(AvailableCLi)
CL = x100 (4)
Max.CL
CLi is the certainty level for each parameter (i), WPi is the FIGURE 10. HI categories by scoring-weighting method.

weighting parameter, and WFi is the weighting factor.

The other categories are good condition (G) for 104 units
G. PROPOSED RANDOM FOREST (21%), caution condition (C) for 134 units (27%), poor con-
Random forest algorithm is a method consisting of a collec- dition (P) for 82 units (16%), and very poor condition (VP)
tion of tree-structured classifiers {h(x, 2 k), k = 1, . . .} for 10 units (2%). This HI assessment is used to highlight
where {2k} represent independent identically distributed faulty transformers and to construct strategies for mainte-
random vectors. Each tree casts a unit vote for choosing nance plans.
the most popular class at input x [41]. An ensemble of the
classifiers h1 (x), h2 (x). . . hK (x), is given, and the training B. DATA MANAGEMENT
data are drawn randomly from the distribution of the random The data used to develop the AI-based HI models are the same
vector X, Y and the margin function is defined as: data classified in the above section to ensure the correctness
mg(X , Y ) = avk I (hk (X ) = Y ) − max j6=Y avk I (hk(X ) = j) and availability of all data. The distributions of the used
parameters are shown in Figure 11 and a matrix correlat-
(5) ing these parameters is shown in Figure 12. As shown in
The generalization error is given by: Figure 12, the IFT and color have the highest correlation
value of -0.69. Sequentially, the significant correlation is
PE ∗ = PX ,Y (mg(X ,Y ) < 0) (6) found between C2 H4 and CH4 (0.65), 2FAL and color (0.65),
In RF, hk (X ) = h(X ,2k ). Almost all sequences 21 . . . PE∗ IFT and acidity (−0.63), color and acidity (0.55), and age and
converge to: color (0.53). These results reveal that several parameters have
a dependency relationship and influences each other.
PX ,Y (P2 (h(X ,2) = Y )−max j6=Y P2 (h(X , 2) = j) < 0) (7) Furthermore, the correlation of parameters with HI score,
This explains why random forests do not overfit as more as shown in Figure 13, can be described from the highest to
trees are added and produce a limiting value for the general- the lowest by applying a threshold limit above 0.2, which are
ization error [41].

H. EVALUATION
The evaluation of the model is based on the accuracy of the
prediction of the transformer HI category. The classification
accuracy (CA) is the proportion of correctly classified data as
given by (8).
(TP + TN )
Accuracy = (8)
(TP + TN + FP + FN )
where TP is true positive, TN is true negative, FP is false
positive, and FN is false negative. This calculation is based
on the confusion matrix of binary classification in which
the evaluation is considered as a multi-class classification
problem.

III. RESULTS AND DISCUSSION

A. TRANSFORMER HEALTH INDEX
The HI is calculated for the investigated 504 units using
the scoring-weighting method. As shown in Figure 10, the
most frequent category of HI transformer is very good condi-
tion (VG) that represents 174 units or 35% of the population. FIGURE 11. Histogram of data collected parameters.

VOLUME 9, 2021 150643

 D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

FIGURE 14. Performance comparison of data trainning multi-AI-based.

When evaluated using the testing data, as shown in Figure 14,

random forest excels with the PCF classification accuracy of
97%. The decision tree excels with the OQF classification
accuracy of 96%, and the FF is 100% accurate. The AdaBoost
FIGURE 12. The correlation matrix of observed parameters. excels at predicting HI categories with about 100% of CA.
The results show that the best prediction models are the
tree-based classifiers such as random forest, decision tree and
AdaBoost.

TABLE 9. The confusion matrix of random forest AI-full-based.

The performance of the RF classifier for the HI category

FIGURE 13. Plot of the correlation between observed parameters.
is shown in the confusion matrix. The comparison between
actual and predicted values is used to evaluate the classi-
as follows: color (−0.778), IFT (+0.638), 2FAL (−0.568), fication accuracy of AI-based HI. The classification accu-
age (−0.51), acidity (−0.45), CO (−0.413), CO2 (−0.335), racy for RF is 97.3%. Furthermore, the random forest model
water content (−0.278), BDV (+0.268), and C2 H2 (−0.218). only has four incorrect classifications, as shown in Table 9.
The HI score has a limit value of 1, which means HI category The accuracy of other models are: decision tree (96%),
has an equal relationship with the HI score. Therefore HI AdaBoost (94.8%), neural network (91.3%), SVM (89.3%),
scores and maximum level are derived from the value of Naïve Bayes (70.7%), and kNN (70%). In addition, an
the parameter (data layer), then in the subsequent discus- AI-SW-based model, which uses AI in the factor level, and
sion, HI maximum score and LvMax will be ignored. The weight-sum calculation in the HI level has been also devel-
correlation of parameters helps to reduce the number of fea- oped. The model provides a classification accuracy results of
tures without reducing the accuracy significantly as will be 98%. Figure 15 shows a comparison of the performance of
elaborated below. all developed models in terms of the HI calculation accuracy.
Results attest that AI-based tree-structured classifiers pro-
C. EVALUATION OF DEVELOPED AI-BASED HI MODELS duces better HI calculation accuracy than other algorithms.
To evaluate the performance of the developed AI-based HI
models, the data were divided into training and testing sets. D. EFFECTS OF UNAVAILABLE DATA
Using the training data, random forest algorithm was found to In this study, a comparison is done to evaluate the effect of
produce the highest classification accuracy (CA) of 91.2% in uncertainty as a result of unavailable data to the accuracy of
OQF and 92.1% in PCF. The tree decision excels at predicting the HI calculation. Three main methods were compared: SW,
FF categories with about 100% of CA. In the HI-layer model, RF, and RFwP. The missing data scenarios in Table 10 were
the AdaBoost algorithm produces the highest CA of 97.2%. assumed to conduct this analysis.

150644 VOLUME 9, 2021

D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

FIGURE 16. The accuracy with the certainty level of parameters.

FIGURE 15. Accuracy of the AI method to predict HI.

TABLE 11. Scenarios of the factor availability.
TABLE 10. Missing value scenarios of transformer health index.

FIGURE 17. The accuracy with the certainty level of factor.

Figure 16 shows the accuracy with the certainty level of
parameters. The results show that the RF is significantly
influenced by a low certainty level due to particular missing investigate various scenario of unavailable data. The corre-
parameters. With a low certainty level, the accuracy of RF lation of the certainty level and the classification accuracy of
also decreased. In contrast, RF with preprocessing (RFwP), multi-parameters is constructed, as shown in Appendix 1. The
due to missing value using the average or most frequent value, aim is to measure the causal effect between certainty level
can maintain the classification accuracy at an acceptable level and resulting HI accuracy for each of the three developed
of at least 95%. model: SW-based HI, RF-based HI, and RFwP-based HI.
Then, the effect of missing factor on the HI calculation This analysis is helping to define efficient and precise ways
accuracy is examined according to the scenarios shown in to deal with the data uncertainty.
Table 11. The uncertainty due to missing factor significantly
affects the HI classification accuracy as shown in Figure 17. 1) SCORING-WEIGHTING (SW)-BASED HEALTH INDEX
In this case, the observed factor conditions were divided into When evaluated using the testing data in Appendix 1, the
OQF, FF, and PCF. The most significant effect of missing SW-based HI is affected by the uncertainty due to unavailable
value is observed in the RF-based paper condition factor data as shown in Figure 18. This is indicated by the coeffi-
(PCF). The lowest accuracy value is 60.7% using RF based cient of determination (R2 ) of 0.917. From Figure 18, The
on the missing of the paper condition factor. Using RFwP, the correlation of the accuracy (y) and certainty level (x) can be
classification accuracy can be increased to 82%. Therefore, approximated using (9).
it can be concluded that the RFwP model can maintain the y = 0.6554x + 0.2573 (9)
classification accuracy of HI categories with missing factor.
After evaluating the effect of missing each parameter and The challenges of this method is that if there are many
factor on the HI accuracy, the next analysis is conducted to missing values, the classification accuracy of the HI category

VOLUME 9, 2021 150645

 D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

TABLE 12. Scenarios of data unavailability.

FIGURE 18. Effect of certainty level due to unavailable data on the

accuracy of SW-based HI.

FIGURE 19. Effect of certainty level due to unavailable data on the

accuracy of RF-based HI.

FIGURE 20. Effect of certainty level due to unavailable data on the

accuracy of RFwP-based HI.

will be decreased significantly. The range of the classification The color categories in terms of the accuracy indicator are
accuracy in HI categories will have a value above 80% if the green for high (H) accuracy level, yellow for medium (M)
certainty level is above 85%. accuracy level, and red for low (L) accuracy level.

150646 VOLUME 9, 2021

D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

2) RANDOM FOREST-BASED by considering the missing values with the average or the
The AI-based RF algorithm also suffers when predicting most frequent value. The impact of missing data on the
transformers HI with many unavailable data as confirmed by model has been evaluated, and the proposed RFwP performed
Figure 19 in which R2 is 0.7614. The correlation of the accu- better than other investigated methods and resulted in a
racy and certainty level for this model can be approximated satisfactory classification accuracy even with low certainity
as below. levels. This solves the current common issue of calculating
the transformer HI with unavailable data. Results in this
y = 0.7542x + 0.1232 (10)
paper pave the way to adopt AI-based methods in calculating
In some cases, RF model can predict the HI with an accu- the transformer HI. However, implementing such techniques
racy above 80% even though the certainty level is only 70%. calls for more validation and feasibility studies on large
Nevertheless, the accuracy will be better if the certainty level transformer population with several scenarios of unavailable
is above 80%. data.

3) RANDOM FOREST WITH PREPROCESSING APPENDIX

The RFwP is slightly affected by unavailable data. In this case SCENARIOS OF UNCERTAINTY MULTI-PARAMETER
R2 is 0.8079 and the accuracy-certainty level correlation can See Table 12.
be approximated from Figure 19 as in (11).
REFERENCES
y = 0.5407x + 0.4904 (11)
[1] A. Abu-Siada, Power Transformer Condition Monitoring and Diagnosis:
The RFwP model has performed better than SW and RF, Concepts and Challenges. London, U.K.: IET, 2018.
[2] N. A. Bakar and A. Abu-Siada, ‘‘Fuzzy logic approach for transformer
resulting in the highest accuracy value in classifying the HI remnant life prediction and asset management decision,’’ IEEE Trans.
category even at a low certainty level. When the certainty Dielectr. Electr. Insul., vol. 23, no. 5, pp. 3199–3208, Oct. 2016.
level is, in the worst-case below 10%, the accuracy is still [3] L. E. Lundgaard, W. Hansen, D. Linhjell, and T. J. Painter, ‘‘Aging of
oil-impregnated paper in power transformers,’’ IEEE Trans. Power Del.,
above 50%. In some instances, the accuracy can reach a value vol. 19, no. 1, pp. 230–239, Jan. 2004.
above 80% even with a certainty level of only 50%. The best [4] M. Wang, A. J. Vandermaar, and K. D. Srivastava, ‘‘Review of condition
condition is when the certainty level is above 70% which assessment of power transformers in service,’’ IEEE Elect. Insul. Mag.,
vol. 18, no. 6, pp. 12–25, Nov./Dec. 2002.
results in an accuracy reaches 90%. [5] D. Martin, Y. Cui, C. Ekanayake, H. Ma, and T. Saha, ‘‘An updated model
For instance for scenario HI-13 in the Appendix, furan to determine the life remaining of transformer insulation,’’ IEEE Trans.
and IFT parameters are unavailable. Based on the data Power Del., vol. 30, no. 1, pp. 395–402, Feb. 2015.
[6] M. Ali, C. Eley, A. M. Emsley, R. Heywood, and X. dan Xaio, ‘‘Measuring
observation, furan and IFT have the highest frequency of and understanding the ageing of Kraft insulating paper in power transform-
data unavailability. The assessment value of the certainty ers,’’ IEEE Elect. Insul. Mag., vol. 12, no. 3, pp. 28–34, May/Jun. 1996.
level for HI-13 scenario is 79%, and when evaluated based [7] M. Arshad and S. M. Islam, ‘‘Significance of cellulose power transformer
condition assessment,’’ IEEE Trans. Dielectr. Electr. Insul., vol. 18, no. 15,
on 150 transformers data, the classification accuracy using pp. 1591–1598, Oct. 2011.
the scoring-weighting method is 78%. It can be observed that [8] S. Forouhari and A. Abu-Siada, ‘‘Remnant life estimation of power trans-
the RF model can achieve 80.7% accuracy while the RFwP former based on IFT and acidity number of transformer oil,’’ in Proc. IEEE
11st Int. Conf. Properties Appl. Dielectric Mater. (ICPADM), Jul. 2015,
classification accuracy is 96%. pp. 552–555.
As shown in Figure 20, the RFwP model achieves medium [9] Condition Assessment of Power Transformers, CIGRE, document 761,
to high accuracy levels in almost all instances. This reveals Mar. 2019.
[10] X. E. Zhang dan Gockenbach, ‘‘Asset-management of transformers based
the ability of RFwP to solve the issue of data uncertainties in on condition monitoring and standard diagnosis [feature article],’’ IEEE
estimating the transformer HI. Elect. Insul. Mag., vol. 24, no. 4, pp. 26–40, Jul./Aug. 2008.
[11] A. E. B. Abu-Elanien and M. M. A. Salama, ‘‘Asset management tech-
niques for transformers,’’ Electr. Pow. Syst. Res., vol. 80, pp. 456–464,
IV. CONCLUSION
Apr. 2010.
The development and implementation of artificial [12] A. Azmi, J. Jasni, N. Azis, and M. Z. A. A. Kadir, ‘‘Evolution of trans-
intelligence-based health index model to assess the power former health index in the form of mathematical equation,’’ Renew. Sustain.
transformer insulation system conditions has been presented. Energy Rev., vol. 76, pp. 687–700, Sep. 2017.
[13] A. Jahromi, R. Piercy, S. Cress, J. Service, and W. Fan, ‘‘An approach
The proposed method is aimed to simplify, speed up, and to power transformer asset management using health index,’’ IEEE Elect.
reduce error due to data uncertainty. A comparison of various Insul. Mag., vol. 25, no. 2, pp. 20–34, Mar. 2009.
AI algorithms is carried out and evaluated using the classifi- [14] N. Bakar, A. Abu-Siada, and S. Islam, ‘‘A review of dissolved gas analysis
measurement and interpretation techniques,’’ IEEE Elect. Insul. Mag.,
cation accuracy in reference to the scoring-weighting-based vol. 30, no. 3, pp. 39–49, May/Jun. 2014.
HI calculation that is widely used by worldwide utilities. [15] IEEE Guide for Acceptance and Maintenance of Insulating Mineral Oil in
Based on this comparison, the random forest-based model Electrical Equipment, IEEE Standard C57.106-2015, Jun. 2015.
[16] A. J. Kachler and I. Höhlein, ‘‘Aging of cellulose at transformer service
is chosen as the proposed method with the highest accuracy temperatures. Part 1: Influence of type of oil and air on the degree of
to predict the transformer HI. Results show that random polymerization of pressboard, dissolved gases, and furanic compounds in
forest does not overfit as it produces a limiting value of oil,’’ IEEE Electr. Insul. Mag., vol. 21, no. 2, pp. 15–21, Mar./Apr. 2005.
[17] A. Abu-Siada, S. P. Lai, and S. M. Islam, ‘‘A novel fuzzy-logic approach
generalization error. As such, random forest with prepro- for furan estimation in transformer oil,’’ IEEE Trans. Power Del., vol. 27,
cessing has been proposed to handle the data uncertainty no. 2, pp. 469–474, Apr. 2012.

VOLUME 9, 2021 150647

 D. Rediansyah et al.: Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty

[18] A. Naderian, S. Cress, R. Piercy, F. Wang, and J. Service, ‘‘An approach [41] L. Breiman, ‘‘Random forests–random features,’’ Statistic Dept., Univ.
to determine the health index of power transformers,’’ in Proc. Conf. Rec. California, Berkeley, CA, USA, Tech. Rep. 567, 1999, pp. 1–29. [Online].
IEEE Int. Symp. Electr. Insul., Jun. 2008, pp. 192–196. Available: https://www.stat.berkeley.edu/~breiman/random-forests.pdf
[19] I. G. N. S. Hernanda, A. C. Mulyana, D. A. Asfani, I. M. Y. Negara, and
D. Fahmi, ‘‘Application of health index method for transformer condition
assessment,’’ in Proc. IEEE Region 10 Conf. (TENCON), Oct. 2014,
pp. 1–6. DHANU REDIANSYAH (Graduate Student
[20] W. R. Tamma, R. A. Prasojo, and Suwarno, ‘‘High voltage power trans-
Member, IEEE) received the B.Sc. degree from
former condition assessment considering the health index value and its
the Department of Mechanical Engineering, Insti-
decreasing rate,’’ High Voltage, vol. 6, no. 2, pp. 314–327, Apr. 2021.
tut Teknologi Indonesia, in 2009, and the M.Sc.
[21] W. R. Tamma, R. Azis Prasojo, and S. Suwarno, ‘‘Assessment of high
voltage power transformer aging condition based on health index value degree from the School of Electrical Engineer-
considering its apparent and actual age,’’ in Proc. 12nd Int. Conf. Inf. ing and Informatics, Institut Teknologi Bandung,
Technol. Electr. Eng. (ICITEE), Oct. 2020, pp. 292–296. Indonesia, in 2021. Since 2011, he has been
[22] L. En-Wen and S. Bin, ‘‘Transformer health status evaluation model working at PT. PLN (Persero), the Indonesian state
based on multi-feature factors,’’ in Proc. Int. Conf. Power Syst. Technol., of electricity company. He has experienced at a
Oct. 2014, pp. 1417–1422. generation power plant, distribution, and trans-
[23] K. Taengko and P. Damrongkulkamjorn, ‘‘Risk assessment for power trans- mission. His research interests include high voltage equipment condition
formers in PEA substations using health index,’’ in Proc. 10th Int. Conf. monitoring and diagnostic, renewable energy power plant, and power quality.
Electr. Eng./Electron., Comput., Telecommun. Inf. Technol., May 2013,
pp. 1–6.
[24] E. Kadim, N. Azis, J. Jasni, S. Ahmad, and M. Talib, ‘‘Transformers health
index assessment based on neural-fuzzy network,’’ Energies, vol. 11, no. 4, RAHMAN AZIS PRASOJO (Graduate Student
p. 710, Mar. 2018. Member, IEEE) received the B.Sc. degree from the
[25] K. Ibrahim, R. M. Sharkawy, H. K. Temraz, and M. M. A. dan Salama, Department of Electrical Engineering, Politeknik
‘‘Transformer health index sensitivity analysis using neuro-fuzzy mod- Negeri Malang, Malang, Indonesia, in 2015, and
elling,’’ in Proc. 2nd Int. Conf. Adv. Technol. Appl. Sci. (ICaTAS), 2017,
the M.Sc. degree from the School of Electrical
pp. 1–7.
Engineering and Informatics, Institut Teknologi
[26] M. Yahaya, N. Azis, M. Ab Kadir, J. Jasni, M. Hairi, and M. Talib,
‘‘Estimation of transformers health index based on the Markov chain,’’ Bandung, Bandung, Indonesia, in 2017, where he
Energies, vol. 10, no. 11, p. 1824, Nov. 2017. is currently pursuing the Ph.D. degree. Since 2019,
[27] A. D. Ashkezari, H. Ma, T. K. Saha, and C. Ekanayake, ‘‘Application he has been an Assistant Professor with the Depart-
of fuzzy support vector machine for determining the health index of the ment of Electrical Engineering, Politeknik Negeri
insulation system of in-service power transformers,’’ IEEE Trans. Dielectr. Malang. He has published various conference papers and journal articles
Electr. Insul., vol. 20, no. 3, pp. 965–973, Jun. 2013. regarding high voltage power transformer condition monitoring and diag-
[28] M. M. Islam, G. Lee, and S. N. Hettiwatte, ‘‘Application of a general regres- nostics. He has reviewed several reputable journals.
sion neural network for health index calculation of power transformers,’’
Int. J. Electr. Power Energy Syst., vol. 93, pp. 308–315, Dec. 2017.
[29] S. Li, G. Wu, H. Dong, L. Yang, and X. Zhen, ‘‘Probabilistic health index-
based apparent age estimation for power transformers,’’ IEEE Access, SUWARNO (Senior Member, IEEE) received
vol. 8, pp. 9692–9701, 2020. the B.Sc. and M.Sc. degrees from the Depart-
[30] S. Tee, Q. Liu, and Z. Wang, ‘‘Insulation condition ranking of transformers ment of Electrical Engineering, Institut Teknologi
through principal component analysis and analytic hierarchy process,’’ IET Bandung, Indonesia, in 1988 and 1991, respec-
Gener., Transmiss. Distrib., vol. 11, no. 1, pp. 110–117, Jan. 2017. tively, and the Ph.D. degree from Nagoya
[31] J. I. Aizpurua, B. G. Stewart, S. D. J. McArthur, B. Lambert, J. G. Cross,
University, Japan, in 1996. He is currently a Pro-
and V. M. Catterson, ‘‘Improved power transformer condition monitoring
fessor and the Emeritus Dean of the School of
under uncertainty through soft computing and probabilistic health index,’’
Appl. Soft Comput., vol. 85, Dec. 2019, Art. no. 105530.
Electrical Engineering and Informatics, Institut
[32] A. A. Alqudsi dan El-Hag, ‘‘Application of machine learning in trans- Teknologi Bandung. He serves as the Head for
former health index prediction,’’ Energies, vol. 12, no. 14, p. 2694, 2019. the Electrical Power Engineering Research Group.
[33] R. A. Prasojo and A. Abu-Siada, ‘‘Dealing with data uncertainty for He has published over 200 international journal articles or conference papers.
transformer insulation system health index,’’ IEEE Access, vol. 9, His research interests include high voltage insulating materials and technol-
pp. 74703–74712, 2021. ogy and diagnostics of HV equipment. He was the General Chairperson of
[34] Z. Li, Y. Zhang, A. Abu-Siada, X. Chen, Z. Li, Y. Xu, L. Zhang, and several international conferences, such as ICPADM 2006, ICEEI 2007, CMD
Y. Tong, ‘‘Fault diagnosis of transformer windings based on decision tree 2012, ICHVEPS 2017, and ICHVEPS 2019. He serves as the Editor-in-Chief
and fully connected neural network,’’ Energies, vol. 14, no. 6, p. 1531, for the International Journal on Electrical Engineering and Informatics.
Mar. 2021.
[35] A. Abu-Siada, M. Arshad, and S. Islam, ‘‘Fuzzy logic approach to identify
transformer criticality using dissolved gas analysis,’’ in Proc. IEEE PES
Gen. Meeting, Jul. 2010, pp. 1–5. A. ABU-SIADA (Senior Member, IEEE) received
[36] Transformer Reliability Survey, CIGRE WG, document A2.37, TB642, the B.Sc. and M.Sc. degrees in electrical engineer-
Dec. 2015. ing from Ain Shams University, Egypt, in 1998,
[37] IEEE Guide for the Interpretation of Gases Generated in Mineral Oil- and the Ph.D. degree in electrical engineering
Immersed Transformers, IEEE Standard C57.104-2019, 2019.
from Curtin University, Australia, in 2004. He is
[38] Mineral Insulating Oils in Electrical Equipment—Supervision and Main-
currently the Discipline Lead of electrical and
tenance Guidance, IEC Standard 60422:2013, 2013.
computer engineering with Curtin University.
[39] R. A. Prasojo, A. Setiawan, Suwarno, N. U. Maulidevi, and B. Anggoro
Soedjarno, ‘‘Development of analytic hierarchy process technique in deter- His research interests include power electron-
mining weighting factor for power transformer health index,’’ in Proc. 2nd ics, power system stability, condition monitor-
Int. Conf. High Voltage Eng. Power Syst. (ICHVEPS), Oct. 2019. ing, and power quality. He is the Vice Chair of
[40] R. Azis Prasojo, Suwarno, N. Ulfa Maulidevi, and B. Anggoro Soedjarno, the IEEE Computation Intelligence Society, WA Chapter. He is also the
‘‘A multiple expert consensus model for transformer assessment index Editor-in-Chief for the International Journal of Electrical and Electronic
weighting factor determination,’’ in Proc. 8th Int. Conf. Condition Monitor. Engineering and a regular reviewer for various IEEE TRANSACTIONS.
Diagnosis (CMD), Oct. 2020, pp. 234–237.

150648 VOLUME 9, 2021