1778 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 31, NO.

4, AUGUST 2016

Prediction of Transformer Furan Levels

Khaled Bashir Shaban, Senior Member, IEEE, Ayman H. El-Hag, Senior Member, IEEE, and Kamel Benhmed

Abstract—In this letter, the ranges of furan content in oil in measurements such as BDV, acidity and water content are used
power transformers are predicted using measurements of oil tests, to predict the levels of furan content. Initially, the recognition
such as breakdown voltage, acidity, water content, and dissolved rate obtained is around 67% because of the imbalance in class
gas analysis. Predictive models based on machine-learning tech-
niques are trained and tested to estimate the furan level. A predic- distributions. The recognition rate is improved to achieve 80%
tion accuracy of 90% is achieved when using k-nearest neighbors after treating the imbalanced classification problem using an
as the classification model with a wrapper method as the feature over-sampling technique. However, the small data size (only 90
selection technique. transformers) used in the study makes it difficult to generalize
Index Terms—Artificial intelligent, furan, health index, trans- the findings. The objective of the present study is to predict the
former. level of the furan content in transformer oil using a set of 731
transformers.
I. INTRODUCTION
OWER transformers are considered to be one of the most II. MATERIALS AND METHODS
P important elements in the power grid. Unexpected fail-
ures of power transformers may cause large interruptions to
The maintenance history of 731 power transformers from a lo-
cal utility in the Gulf region was collected. The records include
industrial, commercial and residential customers. Hence, esti- the test results of DGA; oil quality tests, such as breakdown
mating the life span of transformers’ is required to minimize voltage, water content and acidity; and furan content. The trans-
the probability of their unexpected failures. Moreover, the age formers were made between 1960 and 2011; their ratings range
of transformers is highly correlated with both the degree of from 12.5 to 40 MVA and the voltage level is 66/11 kV. The
polymerization (DP) and tensile strength (TS) of their paper target feature, also known as the target variable for prediction,
insulation [1]. The measurements to find the DP and TS of cel- is 2-furfuraldehyde (2-FAL).
lulose insulation paper should be conducted on paper samples Our methodology for furan content prediction is based on
taken from transformers in operation. This sampling requires the machine learning (ML) approach. The process of applying
transformers to be out of service, which is costly to utilities. ML consists of three phases: training, testing and deployment.
However, it has been reported that the amount of furan existing In the training phase, data with known target values is collected
in transformer oil can be correlated to the DP, and hence to and a subset of features is selected to be used for constructing a
the paper insulation aging condition [2]. Therefore, the furan predictive model (classifier). The portion of data that is used in
concentration in oil can be used as an important indicator in the classifier training phase is referred to as training data. There
determining the health condition of transformers. are a plethora of algorithms to build ML-based classification
Several attempts have been investigated to predict the furan models that may adapt differently to the given data. In this
content in transformer oil. In [3], stepwise feature selection and study k-NN, J48 and Naı̈ve Bayes are attempted to determine
ANN are used to predict furan content using oil tests as input the range of the furan content. Moreover, several feature
feature vectors. While a high recognition rate is reported, the selection (FS) techniques have been employed to eliminate the
dataset is relatively small, which limits its generalization. A least original dataset from irrelevant and redundant features. Both the
square-SVM model is developed to determine the relationship wrapper and the step wise regression have been used to select
between furan content as a function of gases (CO and CO2) [4]. the needed features. Applying FS also results in reducing data
In [5], an adaptive neuro-fuzzy inference technique is attempted dimensionality. This is a desirable data quality that positively
to correlate furanic compounds obtained by a high performance affects learning algorithm performance.
liquid chromatography spectral test to the remnant life of the
power transformer. Specific wavelength and peak absorbance of
transformer oil are used as inputs. III. RESULTS AND DISCUSSION
Recently, the authors investigated the possibility of predicting The furan values are divided into five intervals as suggested
furan content using machine learning algorithms, [6]. The main- by [7], Table I. The classification results are shown in Table II
tenance records of 90 transformers that include transformer oil where it is evident that using k-NN as the classifier and a
wrapper as the feature selection technique, the recognition rate
Manuscript received August 17, 2015; revised October 15, 2015; accepted achieves 90%. The confusion matrix for this result is shown in
December 18, 2015. Date of publication January 26, 2016; date of current
Table III. It is evident from Table III that both classes A and
version July 21, 2016. Paper no. PESL-00127-2015.
K. B. Shaban and K. Benhmed are with Computer Science and Engi- E have recognition rates greater than 90%. On the other hand,
neering Department, Qatar University, Doha, State of Qatar (e-mail: khaled. the recognition rates for classes B–D are less than 50%. This
[email protected]; [email protected]). is mainly caused by the imbalance in the data between classes
A. H. El-Hag is with Department of Electrical Engineering, American Uni-
versity of Sharjah, Sharjah, United Arab Emirates (e-mail: [email protected]). A and E on one side and classes B to D on the other side. To
Digital Object Identifier 10.1109/TPWRD.2016.2521320 overcome this imbalanced classification problem, the value of

0885-8977 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed
Seeuse limited to: Hitachi Energy. Downloaded
http://www.ieee.org/publications on June 21,2024 at 05:04:10 UTC for
standards/publications/rights/index.html frommore
IEEEinformation.
Xplore. Restrictions apply.
IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 31, NO. 4, AUGUST 2016 1779

TABLE I TABLE VI
FURFURAL TEST RATING [7] DETAILED CONFUSION MATRIX FOR THE BEST RESULT USING k-NN WITH 8
FEATURES SELECTED WITH THE WRAPPER METHOD

Rating Code 2-Furfuraldehyde [ppm] Available samples

Classified as Recall
A 0–0.1 512
B 0.1–0.25 26 A B C
C 0.25–0.5 31
D 0.5–1 43 Actual A 508 3 1 99.2
E >1 119 B 15 72 13 72
C 0 15 104 87.4
TABLE II Precision 97 80 88 93.6
MODEL ACCURACIES OF FURAN CLASSIFICATION WITH FEATURE SELECTION

Classifiers Accuracy
model were CO2 and acidity. This could be due to the fact that
CO2 is one of the main outcomes of thermal degradation of
No features Stepwise (4 Wrapper transformer paper insulation. The correlation between furan
selected features)
content and acidity comes from the fact that acidity is the main
Decision Tree (J48) 85.91% 87.68% 88.37% 7 features cause of hydrolytic aging of transformer paper insulation.
k-NN (k = 5) 87.14% 87.55% 90.01% 9 features
Random Forest (100 trees) 89.19% 87.68% 89.46%11 features
IV. CONCLUSION
TABLE III In this paper, a number of machine learning algorithms
DETAILED CONFUSION MATRIX FOR THE BEST RESULT USING k-NN WITH
NINE FEATURES SELECTED WITH THE WRAPPER METHOD are proposed for predicting furan (2-FAL) content. More
than 90% is achieved when predicting classes with a high
number of data points. On the other hand, classes with a low
Classified as Recall
number of data points achieve a prediction accuracy of less
A B C D E than 50%. To improve the classification accuracy, the number
Actual A 508 1 0 1 2 99.2
of classes is reduced from five to three, and hence overall, a
B 7 12 1 4 2 46.2 93.6% prediction accuracy is achieved. Reducing the number
C 5 6 13 2 5 41.9 of classes to three exemplifies the ability of the technique to
D 3 2 1 17 20 39.5
predict more correctly the classes if the number of data points
E 0 0 5 6 108 90.8
are higher and balanced between the classes. The proposed
Precision 97 57 56 57 79 90
approach can be applied to achieve greater accuracy for a more
TABLE IV efficient health index for power transformers.
FURFURAL TEST RATING REDUCED TO THREE LEVELS
ACKNOWLEDGMENT
Rating Code 2-Furfuraldehyde [ppm] Available samples
This work was made possible by the NPRP 5-044-2-016 grant
A 0–0.1 512 from Qatar National Research Fund (a member of Qatar Foun-
B 0.1–1 100 dation). The statements made herein are solely the responsibility
C >1 119
of the authors.
TABLE V
MODEL ACCURACIES OF FURAN CLASSIFICATION WITH FEATURE SELECTION REFERENCES
[1] D. H. Shroff and A. W. Stannett, “A review of paper aging in power
Classifier Tuned parameter Accuracy transformers,” Proc. Inst. Elect. Eng., vol. 132, no. 6 C, pp. 312–319, Nov.
1985.
k-NN k =3 92.47% [2] M. Dong, G. J. Zhang, and Z. Yan, “Synthetic furfural analysis for trans-
k= 5 93.02% former aging,” in Proc. Annu. Rep. Conf. Elect. Insul. Dielect. Phenomena,
J48 decision tree pruned 90.97% 2004, pp. 245–247.
unpruned 92.47% [3] R. A. Ghunem, K. Assaleh, and A. H. El-Hag, “Artificial neural net-
Random Forest 100 trees 93.57% works with stepwise regression for predicting transformer oil furan con-
tent,” IEEE Trans. Dielectr. Elect. Insul., vol. 19, no. 2, pp. 414–420,
Apr. 2012.
the target variable is limited to only three classes by combining [4] A. Mahdjoubi, B. Zegnini, and M. Belkheiri, “A new approach using least
squares support vector machines (LS-SVM) to predict furan in power trans-
classes B–D as depicted in Table IV. The results of the formers,” Przeglad
˛ Elektrotechniczny, pp. 142–145, 2014.
recognition rate using various search techniques and classifiers [5] P. K. Pandey, Z. Husain, and R. K. Jarial, “ANFIS based approach
are shown in Table V. The first observation when comparing Ta- to estimate remnant life of power transformer by predicting furan
contents,” Int. J. Elect. Comput. Eng., vol. 4, no. 4, pp. 463–470,
bles II and V is that the different classifiers exhibited consistent
Aug. 2014.
performance, with better results when the number of classes was [6] K. Benhmed, K. B. Shaban, and A. El-Hag, “Cost effective assessment of
reduced from 5 to 3. Additionally, it is evident from Table V that transformers using machine learning approach,” in Proc. Innovat. Smart
using the wrapper and random forest resulted in the best recog- Grid Technol.–Asia, 2014, pp. 328–332.
[7] A. N. Jahromi, R. Piercy, S. Cress, J. R. R. Service, and W. Fan, “An
nition rate, with around 4% improvement compared to Table II. approach to power transformer asset management using health index,” IEEE
The most repeated selected features for the best classification Elect. Insul. Mag., vol. 25, no. 2, pp. 20–34, Mar./Apr. 2009.

Authorized licensed use limited to: Hitachi Energy. Downloaded on June 21,2024 at 05:04:10 UTC from IEEE Xplore. Restrictions apply.