A Novel Application of Gene Expression

Programming in Transformer Diagnostics

A.Abu-Siada, Lai Sin Pin and Syed Islam
Curtin University of Technology, Australia

Abstract-Furans are the major degradation of insulating 7100

paper in transformer oil. Hence the concentration of furans in oil DPV = (1)
can be used as a good indicator of paper deterioration. Furan 8.88 + 2 FAL
concentration in transformer oil is currently measured using where 2FAL is the furfural concentration in mg/kg of oil.
High-Performance Liquid Chromatography (HPLC) or Gas
Chromatography-Mass Spectrometry (GC/MS). Both methods Then the concentration of 2-furfuraldehyde in the oil can be
provide accurate and reliable results in detecting furan used as a good indicator of paper deterioration. It has been
concentration. However, the two methods need very expensive
estimated that new paper, under normal running conditions will
equipments and take long time to get the result for one sample.
Moreover it requires a trained expert to perform and interpret generate furfural at the rate of 1.7 ng/g of paper/h. The rate of
the results. This paper introduces a novel approach for detecting production increases with increasing degree of degradation and
furan concentration in transformer oil through measuring its a total yield of 0.5 mg of furfural/g of paper is expected in
spectral response. The Ultraviolet-to-Visible (UV-Vis) spectral 100000 h of running (15-20 years)[3-5]. Furan level in a
response of transformer oil can be measured instantly with
transformer can be correlated with paper DP, and therefore an
relatively cheap equipment and does not need an expert person to
conduct the test. Results show that there is a good correlation in-service assessment of the mechanical strength of the paper
between oil spectral response and its furan contents. Also, the insulation can be made. Furans concentration is measured by
paper introduces a novel application for Gene Expression High-Performance Liquid Chromatography (HPLC) or Gas
Programming (GEP) to estimate the relationship between furan Chromatography-Mass Spectrometry (GC/MS) based on
concentration and spectral response of transformer oil.
American Society for Testing and Materials (ASTM D5837,
Standard Specifications for Mineral Insulating Oil in Electrical
I. INTRODUCTION
Apparatus)[6]. Both analysis methods provide accurate and
Power transformers are a vital link in a power system. reliable results in detecting the concentration of furan
Monitoring and diagnostic techniques are essential to decrease derivatives; 2- furfural (2-FAL), 2-Furfurol (2-FOL), 5-Hyroxy
maintenance and improve reliability of the equipment. methyl-2-furfural (5-HMF), 5-Methyl-2-furfural (5-MEF), and
Currently there are several of chemical and electrical 2-Acetylfuran (2-ACF). However, these two methods need
diagnostic techniques applied for power transformers[1]. As very expensive equipments and take long time to get the results
the entire energized and high temperature transformer in addition it requires for an expert person to perform and
components are immersed in the transformer oil, the interpret the results. The technique for furan testing is
transformer oil is a key source to detect incipient faults, fast described in IEC, but there is no guideline for interpretations.
developing faults, insulation trending and generally reflects the This paper presents a novel approach for determination of
health condition of the transformer. The electrical windings in furan concentration in transformer oil through measuring its
a power transformer consist of paper insulation immersed in spectral response. The Ultraviolet-to-Visible (UV-Vis) spectral
insulating oil. Paper insulation is composed of approximately response of transformer oil can be measured instantly with
90% of cellulose, 6-7% hemi-cellulose and 3-4% of lignin. relatively cheap equipment and does not need an expert person
Due to electrical and thermal stresses, oil and cellulose to conduct the test. Results show that there is a good
decomposition occurs evolving gases that will decrease the correlation between oil spectral response and its furan contents.
heat dissipation capability and the dielectric strength of the oil.
II. LABORATORY AGED OIL
When degradation of paper insulation occurs, the cellulose
molecular chains get shorter and chemical products such as The study has been performed on in-service as well as
furanic derivatives are produced and dissolve in the oil. Furans laboratory aged transformer oil. Laboratory aged insulating oil
are the major degradation of paper in oil and the 2-furaldehyde is prepared by utilizing the heating process available in IEC
in oil is the most prominent component of paper decomposition. 61125[7]. Section of new craft paper (20mmx280mm) was cut
The increase in furans concentration in the oil corresponds to and wrapped around copper strips (3mmx10mm) then it was
the decrease in the tensile strength and the degree of impregnated in 25ml of new transformer oil (shell Diala B).
polymerization (DP) of the paper. De Pablo reported the All samples were heated up to 100°C in a thermostatically-
following relation between furfural and Degree of controlled aluminum alloy block heater for 7 days. Oxygen
polymerization based on viscosity DPv[2] :

2008 Australasian Universities Power Engineering Conference (AUPEC'08) Paper P-054 page 1
flow at a rate of 1 l/hr was supplied into each dry tube to Where Aλ is the absorbance, S is the sample intensity at
further accelerate the aging process. wavelength λ , D is the dark intensity at wavelength λ , R is
III. FURAN ANALYSIS the reference intensity at wavelength λ .
Same oil samples used for furan concentration measurement
All samples were prepared in accordance to standard ASTM using GC/MS were tested using a laboratory grade
D 5837 and tested using GC/MS system for furan derivatives spectrophotometer for absorbance spectrophotometry. The
identification and quantification. Table I shows furan experiment procedure was set up in reference to ASTM
derivative concentration in particle per million (ppm) for E275[10]. Figure 1 shows the lab set up for measuring spectral
different oil samples using GC/MS. Results show that the 2- response for one oil sample. Figure 2 shows the spectral
furaldehyde (2-FAL) is the most prominent component of response (absorbance) for different oil samples with different
paper decomposition. Therefore, the level of 2-furaldehyde in furan concentration.
transformer oil can be used as an indicator for paper
deterioration.
Table I
Furan concentration result by GC/MS
Test Sample 2-FAL 2-FOL 2-ACF 5-MEF 5-HMF

New Oil <0.01 <0.01 <0.01 <0.01 <0.01

Sample 1 3.1 <0.01 0.01 0.01 <0.01
Sample 2 5.1 <0.01 0.02 0.01 <0.01
Sample 3 10.0 <0.01 0.03 0.03 <0.01
Sample 4 15.0 0.01 0.05 0.05 <0.01

The correlation between 2-furaldehyde and DP with respect Figure 1. Lab set up for measuring the spectral response of transformer oil
to the solid insulation extent of damage is given in table II
depicting insulation paper dielectric and mechanical properties.
When DP test reveals a value of 250 or less, the paper is
considered to have lost all its mechanical strength, and the
transformer has reached its end of life [8].
Table II
DP and 2-Furfuraldehyde (2-FAL) Correlation
2-FAL (ppm) DP Value Significance

0-0.1 1200-700 Healthy Insulation

0.1-1.0 700-450 Moderate Deterioration
1-10 450-250 Extensive Deterioration
> 10 < 250 End of Life Criteria

IV. UV-VIS SPECTRAL RESPONSE Figure 2. UV/Vis Spectrum (Absorbance) for different oil samples with
different furan concentration
UV-Spectrophotometry is a non-intrusive test used to
determine the transformer’s integrity. UV-Spectrophotometry It can be shown from Fig. 2 that the new oil exhibits its
is an accurate and sensitive method to analyze impurities in the characteristics between 200 and 350nm uv-spectrum with
transformer oil using light absorbing properties of a sample. maximum absorbance at 250nm wavelength. However, in-
Light transmitted through the oil sample containing various service and laboratory aged oil samples exhibit their respective
characteristics in the range of 200 and 470nm wavelength uv-
contaminations is decreased by that fraction being absorbed
spectrum. Results show that absorbance as well as bandwidth
and is detected as a function of wavelength[9]. A
for maximum absorbance increases by a significant and easily
spectrophotometer measures the transmission, absorption or
observable margin with oil deterioration and contaminations
reflection of the light spectrum for a given wavelength. which are reflected by the furan concentration level in the oil.
Absorption spectroscopy provides a measure of how much UV spectrum shows considerable noise for contaminated oil
light is absorbed by the oil sample which can be calculated which can be attributed to the variety of contaminations
as[2]: including very high carbon and water content in the oil. Figure
 S − Dλ  2 shows a good correlation between the furan concentration in
Aλ = − log λ  (2)
transformer oil and its spectral response. To prove this
 Rλ − Dλ  correlation in more details, Figure 3 shows the relationship
between spectral response parameters (maximum absorbance

2008 Australasian Universities Power Engineering Conference (AUPEC'08) Paper P-054 page 2
and bandwidth wavelength) of transformer oil and its furan Ct
concentration level; the more furan concentration the more fi = ∑ ( M − Ci , j − T j ) (3)
absorbance and more wavelength. Then the spectral response j =1
parameters can be used as an alternative method to GC/MS to Where M is the range of selection, Ci,j is the value returned by
determine the furan concentration in transformer oil. Spectral the individual chromosome i for fitness case j (out of Ct fitness
response of transformer oil can be measured instantly with cases), and Tj is the target value for fitness case j. Figure 4
relatively cheap equipment and does not need an expert person illustrates the flowchart of the typical GEP algorithm[12].
to conduct the test. In the next section, Gene Expression
Programming (GEP) software tool is used to find a
mathematical correlation between furan concentration in
transformer oil and its spectral response.

Figure 3. Correlation between furan concentration in transformer oil and its

spectrum response parameters (wavelength and peak absorbance) Figure 4. The flowchart of GEP algorithm

GEP genes are composed of head and tail. The head contains
V. GENE EXPRESSION PROGRAMMING
symbols that represent both functions (elements from the set
Gene Expression Programming (GEP) is a new technique for function F) and terminals (elements from the terminal set T),
data analysis was first invented by Candida Ferreira in 1999[11, whereas the tail contains only terminals. Therefore two
12]. GEP is a learning algorithm that can find relationships different alphabets occur at different regions within a gene. For
between variables in sets of data and builds models to explain each problem, the length of the head h is chosen, whereas the
these relationships. GEP is similar to genetic algorithm (GA) length of the tail t is a function of h and the number of
and genetic programming (GP) as it uses populations of arguments of the function with the most arguments n, and is
individuals, selects them according to fitness, and introduces evaluated by the equation:
genetic variation using one or more genetic operators. The T=h(n-1)+1 (4)
nature of individuals is the fundamental difference between the TABLE III
three algorithms. In GA the individuals are linear strings of GEP Model results
fixed length (chromosomes); in GP the individuals are Target Model Residual
nonlinear entities of different sizes and shapes (parse tree); and
0.5 0.539963 4.00E-02
in GEP the individuals are encoded as linear strings of fixed
length expressed as nonlinear entities of different sizes and 1 0.886286 0.113714
shapes so that GEP combines the advantage of both GA and 2 1.963771 3.62E-02
GP while overcomes their some shortcomings[13]. One
3 3.139409 0.139409
important application of GEP is symbolic regression or
function finding, where the goal is to find expression that 4 4.014475 1.45E-02
performs well for all fitness cases within a certain error of the 5 4.912101 8.79E-02
correct value. In GEP individuals are selected according to its 7 7.12213 0.12213
fitness by roulette-wheel sampling with elitism and the best
11 11.13847 0.138472
individual is preserved. Mathematically, the fitness f i of an
12 11.85625 0.143752
individual program i is expressed by the following
equation[14]: 13 12.98926 1.07E-02
15 14.94548 5.45E-02

2008 Australasian Universities Power Engineering Conference (AUPEC'08) Paper P-054 page 3
 To find the correlation between furan concentration in represents the two independent input variables (maximum
transformer oil and its spectral response parameters, the absorbance and bandwidth). From table III and Fig. 5, it can be
bandwidth and the maximum absorbance for each oil sample concluded that the GEP model is very accurate and efficient in
are provided as the input (independent) parameters to the GEP modeling the correlation between furan concentration in
while the corresponding furan concentration is provided as the transformer oil and its spectral response. The mean square
target. Some other parameters should be specified to GEP such error between the model and target is only 0.065 and the
as number of Genes, number of chromosomes, length of head correlation coefficient is 0.9987. The model comes with an
and the set functions. Table III shows the GEP model results Expression Tree (ET) (shown in Figure 6) that representing the
compared to the actual experimental measurements. The input/output relationship. The ET is translated to a
absolute error between GEP model and the actual mathematical expression using the same program which
measurements is given in the third column of table III. The provides the model in many computer languages including
maximum error between the GEP model and the actual Matlab, FORTRAN, C++, Visual basic etc.
measurements is 0.143752. .
To check the capability of the model to estimate the furan
concentration in the transformer oil using its spectrum response
parameters; different random spectrum parameters have been
provided to the model as shown in table IV. The spectrum
parameters have been chosen to be inside and outside the
spectrum response range shown in Figure 2. The estimated
furan concentration corresponding to each parameter coincides
with the practical results shown in Figure 2. This proves the
capability of the model to interpolate/extrapolate the given
parameters to estimate a precise furan concentration in the
transformer oil.
Table IV
Figure 5. GEP model and actual correlation between Furan and spectral Capability of GEP model to interpolate/extrapolate results
response parameters
BW (nm) Max. Absorbance Furan (ppm)
400 1.6 1.98714443
417 1.75 3.75018108
440 1.82 9.20123901
450 1.85 11.0897414
460 1.9 13.8267452
475 1.95 17.4311133

VI. CONCLUSIONS
Results show that there is a good correlation between oil
spectral response and its furan contents. Gene Expression
Programming has been used to find this correlation
mathematically. The great insight of GEP consisted in the
invention of chromosomes capable of representing any
expression tree. The structural and functional organisation of
GEP genes and their interplay with expression trees always
guarantee the production of valid programs. The paper has
introduced a novel application for GEP in transformer
diagnostics. Results show that GEP model can estimate
precisely the furan concentration of transformer oil using the
spectral response parameters (bandwidth and maximum
absorbance). The UV-Vis spectral response of transformer oil
Figure 6. Expression tree for the GEP model can be measured instantly with relatively cheap equipment and
does not need an expert person to conduct the test or to
Figure 5 shows the relationship between furan concentration interpret the results or to interpret the results. Results show that
and spectral response parameters for experimental the model is very efficient and can replace the current
measurements and GEP model. In Fig. 5, the vertical axis expensive and time consuming techniques to detect furan
represents the output (furan concentration in ppm), However, it concentration in transformer oil.
is difficult to represent the horizontal axis by real values as it

2008 Australasian Universities Power Engineering Conference (AUPEC'08) Paper P-054 page 4
 REFERENCES
[1] T. K. Saha, "Review of Modern Diagnostic techniques for [8] M. Wang, A. J. Vandermaar, and K. D. Srivastava, "Review of
Assessing Insulation Condition Aged Transformers," IEEE condition assessment of power transformers in service," Electrical
Transaction on Dielectrics and Electrical Insulation, vol. 10, pp. Insulation Magazine, IEEE, vol. 18, pp. 12-25, 2002.
903-917, 2003. [9] J. A. Palmer, W. Xianghui, R. A. Shoureshi, A. Mander, D.
[2] M. Arshad, "Remnant Life Estimation Model Using Fuzzy Logic Torgerson, and C. Rich, "Effect of aging on the spectral response of
for Power Transformer Asset Management," PhD thesis, Curtin transformer oil," in Electrical Insulation, 2000. Conference Record
University of Technology, 2005. of the 2000 IEEE International Symposium on, 2000, pp. 460-464.
[3] Y. Shang, L. Yang, Z. J. Guo, and Z. Yan, "Assessing aging of [10] ASTM, "Standard Practice for Describing and Measuring
large transformers by furfural investigation," in Solid Dielectrics, Performance of Ultraviolet, Visible, and Near-Infrared
2001. ICSD '01. Proceedings of the 2001 IEEE 7th International Spectrophotometers,," ASTM E275-01, vol. 03.06, pp. 72-81, 2001.
Conference on, 2001, pp. 272-274. [11] W. Chuan-Sheng, H. Li, and K. Li-Shan, "The automatic modeling
[4] C. H. Zhang and J. M. K. Macalpine, "Furfural Concentration in of complex functions based on gene expression programming," in
Transformer Oil as an Indicator of Paper Ageing, Part 1: A Machine Learning and Cybernetics, 2005. Proceedings of 2005
Review," in Power Systems Conference and Exposition, 2006. International Conference on, 2005, pp. 2870-2873 Vol. 5.
PSCE '06. 2006 IEEE PES, 2006, pp. 1088-1091. [12] G. Zhaohui, L. Gaobin, Y. Zhenkun, and J. Min, "Automatic
[5] J. M. K. MacAlpine and C. H. Zhang, "Observations from Modeling of Complex Functions with Clonal Selection-Based Gene
measurements of the furfural content of oil samples from Expression Programming," in Natural Computation, 2007. ICNC
transmission transformers," in Advances in Power System Control, 2007. Third International Conference on, 2007, pp. 228-232.
Operation and Management, 2000. APSCOM-00. 2000 [13] C. Ferreira, "Gene Expression Programming: Mathematical
International Conference on, 2000, pp. 317-321 vol.2. Modeling by an Artificial Intelligence," Springer, 2006.
[6] ASTM, "Standard Test Method for Furanic Compounds in [14] C. Ferreira, "Gene Expression Programming: a New Adaptive
Electrical Insulating Liquids by High-Performance Liquid Algorithm for solving Problems," Complex Systems, vol. 13, pp. 87-
Chromatography (HPLC)," D5837-05, 2005. 129, 2001.
[7] ITC, "Unused Hydrocarbon-Based Insulating Liquids-Test Methods
for Evaluating The Oxidation Stability " IEC61125, 1992.

2008 Australasian Universities Power Engineering Conference (AUPEC'08) Paper P-054 page 5