Printer Friendly

ANN based internal fault diagnosis of HVDC converter transformer.


Fault diagnosis of HVDC converter transformer is necessary for reliable operation of the HVDC system. Proper design of insulation prevents the operational difficulties and also the cost. The purpose of the impulse test is to ascertain the ability of the insulation of the transformer to withstand the application of the definite magnitude of the test voltage. In the event of the failure of the transformer neutral current characteristics undergo changes. In order to assess the integrity of the winding, specific high voltage tests are conducted on the converter transformer. One of the tests comprises the comparison of applied voltage and neutral current signal at reduced and full test voltage level. Any difference in the current wave shape at reduced and full voltage shows the existence of fault in the converter transformer winding. Minor changes in the wave shape do not decisively predict the existence of fault in the converter transformer winding. Such situation arises due to minor nature of faults like inter-turn in the winding. In order to detect such faults, the transfer function technique (FFT) is used widely by several testing agencies. Although this technique is quite effective in detecting major faults in transformer winding, the minor one like interturn becomes difficult to detect with clarity. Transfer function has been applied for evaluating the transient admittance of the winding both at reduced and full voltages. These transient admittances are deconvoluted into frequency domain using Fast Fourier Transform (FFT) method [1,2,3]. The comparison of the frequencies obtained from the transfer function analysis reveals the condition of the winding. The theory of the above method emerges from the fact that the natural frequencies of an electrical circuit comprising parameters like resistance, inductances and capacitances are independent of voltage and shape and can alter only in the event of a change in the parameters. Since a transformer can be represented by an equivalent electrical network using these parameters the frequency spectrum of transient admittance is not altered by the types of surge voltages. Though, the transfer function technique has proved quite effective in diagnosing the fault in the transformer, some of the lower range frequencies are difficult to detect as it has no clarity regarding the occurrence of fault. [4].

Artificial Neural Networks is gaining importance and is a powerful tool in analyzing these natural currents of HVDC converter transformer due to its excellent pattern recognition technique and pattern recognition capability. Back propagation algorithm is used to classify the neutral currents of the HVDC transformer generated after subjecting various faults such as section to section fault, winding to winding fault and winding to ground fault etc., The artificial neural network is trained for a set of normalized inputs and tested.

The present paper deals with analysis of neutral current obtained after applying a standard impulse voltage waveform represented as

V = V0([e.sup.-[alpha]t] - [e.sup.-[beta]t]) (1)

Where [alpha] and [beta] are constants and depend upon the rate of raise and decay of the pulse. For a given value of a and [beta] impulse voltage generated has a raise time of 1.2[micro]s and a tail of 50[micro]s.

Description of the Transformer

The HVDC converter transformer is of power rating 315 MVA [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.] single phase is as shown in Figure 1.


Figure 1 shows the constructional geometry of one limb of the converter transformer comprising of tapping, HVAC and HVDC star windings. For the purpose of the neutral current calculation the tapping winding has been fully earthed as is done in a particular case of impulse tests. The HVAC winding is divided into 8 sections and HVDC winding is divided into 16 sections as shown in Figure 2.


Based on the turns in a section, paper insulation thickness, physical clearance between the windings and discs is provided. The calculation of self and mutual inductances, series and ground capacitance have been carried out. The network is suitably formed. The network is solved for Eigen voltage values and frequencies and finally node voltage and node neutral current are calculated. The equivalent electrical network showing the inductive and capacitive parameters of the winding is shown in Figure 3.


Artificial Neural Networks and Implementation

Artificial Neural Networks are being widely used in the classification problems. For a neural network, if activation and out put functions are chosen, it is completely described by the weights and node thresholds. The training process is the process of finding the weights and thresholds for the network and it is equivalent to finding the unknown Input--output relationship. Thus neural networks are appropriate and especially powerful when they are used to find such relationships that are difficult to describe explicitly[5].

Among all the proposed neural network structure, the feed forward neural network (FFN) is most popular one. It contains an input layer, an output layer and many hidden layers. Each layer can have many processing nodes or neurons as represented in Figure 4. In order for a neural network to learn certain relationship, data sets describing the relationship must be presented to the net and certain learning rules be applied to the network parameters. In this paper, back propagation learning algorithm is used to train the network. This learning rule is also noted as generalized data rule, exploits gradient information of the error function. The individual pattern error ei of pattern i is given by the formula:


Where [t.sub.ik] is the desired output of the pattern I and oik is the actual net output. Summing the error for all the patterns, the error (E) is obtained and minimizing E is the task of the training process.


The objective is to classify the 4 different categories of faults i.e. No Fault, Section Fault, Winding to Winding Fault, Winding to Ground Fault.

Generally, a design procedure to diagnosis faults using Neural Network is based on the following steps:

(1) Selection of Input

(2) Desired Output

(3) Processing the Input / Output

(4) Structural design of the neural Network

(5) Fault simulation to generate training and test patterns.

(6) Training of the neural Network

(7) Evaluation by using test patterns

Determination of the Neutral Current

A standard impulse voltage with a raise time of 1.2[micro]s and a tail of 50[micro]s is applied to one of the limb of the star winding of the HVDC converter transformer. The calculated neutral current is normalized after taking the absolute value of the signal. The neutral current signal is sampled at 0.1 [micro]s of time and 1024 samples taken for 102.4 micro seconds of time. The neutral current signal for various fault conditions is plotted in Figure 5. The neutral current samples of the HVDC converter transformer are divided into 16 sets. The normalized data is given as input to training of artificial neural networks using back propagation algorithm.


Results and Discussions

A separate training data is prepared after taking the local average values and maximum for every 6.4[micro]s of time. Out of the recorded neutral current data of 102.4[micro]s, the normalized local average data and the local maximum data is given as an inputs to the neural networks and the networks are trained till the error is reduced. The training data for all the fault conditions is given after normalization is given in Figure 6(a) and (b). The error plot of neural network after training with the local average values is given in Figure 7(a).




The error plot of neural network after training with the local maximum values is given in Figure 7(b).


From Figure 7(a) it can be stated that the neural network has taken more than 180 iterations to reduce the error to 4.64276x10-16 in case when average data is taken as the input sets and from Figure 7(b) it has taken less than 130 iterations to reduce the error to 3.76898x10-16 when the local maximum values have been taken as the training and testing patterns.

A total number of 34 sets of data is prepared and out of this 20 sets of data is used for training and 14 sets of data is used for testing the Artificial Neural Network. Table 1 is showing the various test results with both average test patterns and Maximum test patterns.

From the test results as listed in Table 1 it can be stated that the local maximum value based training patterns are providing better diagnosis characteristics when compared to that of the local average value based data sets. The artificial neural network based fault diagnosis of HVDC converter transformer is providing an efficiency of 96.97%.



The article presents the diagnosis of various faults using the neutral current in the DC winding of HVDC converter transformer. The neutral current for healthy winding, inter turn faults, inter winding faults and ground faults are analyzed using back propagation algorithm. From testing results summary as in Table 1 and Figure 8 it can striated that 97% of efficiency can be obtained in identifying the type of fault in the HVDC transformer windings by using the training and testing patterns from the local maximum values of the neutral current trace when compared to that of the local average values. Hence the results indicates that the HVDC converter transformer faults can be identified and analyzed more efficiently by using artificial neural networks.


[1] B P Singh, N K Kishore, K S R Sheriff and A Bhoomaiah, "Adoption of "Transfer Function Technique for Failure Analysis of Transformer Winding". Conference on Electrical Insulation and Di-electrical Phenomena, Ausitn, Texas, USA oct 17-21,1999.

[2] Kardey, George G, Reta-Hemandez Manual, Amarh Felix, Mc.Culla, Gary, "Improved Technique for "Fault Detection Sensitivity in Transformer Impulse Test", Proceedings of IEEE Power Engineering Society, Transmission and Distribution Conference, vol 4 2000, pp 2412-2416.

[3] Prasanth Babu, C.; Surya Kalavathi, M.; Singh, D.B.P., "Use of Wavelet and Neural Network (BPFN) for Transformer Fault Diagnosis" Electrical Insulation and Dielectric Phenomena, 2006 IEEE Conference on Volume, Issue, 15-18 Oct. 2006 Page(s):93-96.

[4] M Surya Kalavathi et al "Transient analyisi and Neural Network method of fault detection in High voltage transmission system" NPSC Proceedings of 13th National Power System conference, Vol 1, Dec. 2004. PP 453-457.

[5] M. Surya Klavathi et al "Neural Network method of identification of faults within a power transformer based on computer studies" 14th international symposium on high voltage Engineering Tsinghua University Beijing, Chaina, Aug: 2005 F46.

(1) Pannala Krishna Murthy (2), J. Amarnath (3), B.P. Singh (4) and S. Kamakshaiah

(1) Swarna Bharathi Institute of Science and Technology, Khammam,-507002, E-mail:

(2) Jawaharlal Nehru Technological University, Kukatpally, Hyderabad--500084, E-mail:

(3) Jyotishmathi college of engineering and Technology, Turakapally, Shameerpet, E-mail:

(4) CVR Engineering College, Ibrahimpatnam, Hyderabad. E-mail:
Table 1: Test result for the network for both average
and maximum data sets.

                                        Diagnosis by
Fault           Location       Local Average   Local Maximum
                                   value           value

        Nornal Operation             YES            YES

                 12th Disk           YES            YES
                 14th Disk           YES            YES
                 15th Disk           YES            YES
Turn to Turn     17th Disk           YES            YES
Faults           20th Disk           YES            YES
                 21st Disk           YES            YES
                 22nd Disk           YES            YES
                 24th Disk           YES            YES
                 17th to 4th         YES            YES
                 17th to 5th         YES            YES
                 18th to 4th         YES            YES
                 18th to 5th         YES            YES
                 19th to 4th         YES            YES
                 19th to 5th         YES            YES
Winding to       20th to 4th         YES            YES
Winding          20th to 5th         YES            YES
                 20th to 6th         YES            YES
                 21st to 5th         YES            YES
                 21nd to 6th         YES            YES
                 21nd to 7th         NO             NO
                 22nd to 6th         YES            YES
                 22rd to 7th         NO             YES
                 17th                NO             YES
                 18th                NO             YES
                 19th                YES            YES
Turn to Ground   Ground
                 20th                YES            NO
                 21st Ground         YES            YES
                 22nd                YES            YES
COPYRIGHT 2009 Research India Publications
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2009 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:artificial neural network and high-voltage direct current
Author:Murthy, Pannala Krishna; Amarnath, J.; Singh, B.P.; Kamakshaiah, S.
Publication:International Journal of Applied Engineering Research
Article Type:Report
Geographic Code:1USA
Date:Jun 1, 2009
Previous Article:Performance evaluation and optimization for urea crystallization system in a fertilizer plant using genetic algorithm technique.
Next Article:Simple diversity scheme for IDMA communication system.

Related Articles
Features selection of SVM and ANN using particle swarm optimization for power transformers incipient fault symptom diagnosis.
Mechanical fault diagnositics for electromechanical devices using ANFIS.
Multiple feedback loop control strategy for UPS inverters using artificial neural networks.
Application of neural network structure in voltage vector selection of direct torque control induction motor.

Terms of use | Privacy policy | Copyright © 2018 Farlex, Inc. | Feedback | For webmasters