A WAVELET POWER BASED ALGORITHM FOR SYNCHRONOUS GENERATOR PROTECTION

O. Ozgonenel, Member, E. Arisoy, M.A.S.K Khan, Student Member, and M.A. Rahman, Fellow, IEEE
windings/dampers, field windings, and magnetic cores. However, stator-winding faults are considered the most dangerous faults, which can cause severe damage in the generator itself, and consequently require a lot of time and high cost of maintenance [4]. These faults are usually related to insulation failure. In common practice, these are generally known as internal and ground faults. Internal faults can cause catastrophic damage by the large transient short circuit current, while ground faults may be slightly less damaging, less destructive in their initial stages, may, in the long term develop into the internal short circuit faults [5]. There have been many techniques to detect these faults. Online approaches that have been proposed for detecting internal and ground faults in the stator windings of synchronous generators include field current monitoring, spectral analysis of fault generated transient current signals, measurement of third harmonic voltages, and artificial neural networks [6-22]. The exciting current harmonics analysis method was used in references [6-7] to study the effect of internal faults, including short circuit between turns of the stator and rotor windings. An additional harmonic current is induced in the rotor winding when the stator/rotor appears winding inter-turn short circuit fault [7]. The error between the calculated value and practically acquired value of the excitation current was used in reference [8] for detecting rotor winding inter-turn short circuit fault. However, the accurate calculation of the excitation current is very difficult due to the effect of the magnetic saturation on the generator coefficients of self and mutual inductances. Detection of ground faults in synchronous generators has been reported in references [9-13]. The authors have focused mainly on the characteristics of the third harmonic voltages at the generator neutral and terminals by using the equivalent circuits. However, considering the influence of the active load on the third harmonic voltages, a ground fault protection technique based on the ratio of the third harmonic voltages in generator neutral and at the generator terminals, has been proposed in reference [14] for the elimination of irregular generator tripping. Artificial neural networks (ANNs) are employed in references [4, 15-16] for the purpose of generator winding protection. The first scheme [15] utilized two ANNs for detecting the fault occurrence and for classifying the faulted phase, respectively. Three phase currents at both the terminals and neutrals in addition to the field current are used as inputs to these networks. However, the stability and sensitivity limits were not completely covered in the reported schemes. As a result, the whole range of the generator winding is not completely covered for the purpose of protection. In addition, the scheme in reference [4] was concerned only about the last five consecutive samples of the input that could affect the recognition quality due to the

Abstract-In this work, a different synchronous generator protection algorithm is presented. The proposed approach is based on the wavelet coefficients of current and voltage samples acquired from the terminal nodes of the generator. The proposed technique is termed as wavelet power (WP) relating to each phase, and able to distinguish either internal or external faults. Discrete wavelet transform (DWT) is used to get related coefficients. This technique is tested to show its ability to distinguish the faults by using two different synchronous generators in the laboratory environment and a simulation model on ATP-EMTP. MATLAB is used for all required computations. Index Terms- Fault diagnosis, synchronous generators, wavelet transforms, wavelet power.

I. INTRODUCTION

VERY power system is subject from time to timetransient disturbances primarily due to faults and/or switching of major loads. Normally the systems adapts to a new steady state condition with the help of power system control schemes such as generator excitation, turbine governor control systems and so on. For this reason, synchronous generators are one of the most important elements in the power system. Large system disturbances are typically caused by short-circuits of different types. The opening of appropriate high-speed breakers isolates the fault. During the fault, the terminal voltage dips and presents a non-stationary nature. In response, the exciter increases its output voltage to ceiling, which causes the excitation current into the field to increase at a rate determined, by the voltage divided by the inductance of the field [1]. As a result of unwanted switching of the line of the large generators, the other energy-producing units of the power system are overloaded. Therefore, the fault must be cleared as soon as possible and the continuity of the power has to be maintained. One of the lessons learned from the past major disturbances is that some of the functions associated with generator protection may operate during these transient conditions. It is important for the relays to provide protection while optimizing their coordination to avoid undesirable operation during the system disturbances and thereby preserving the integrity of the power grid [2-3]. Intensive research has been conducted to develop and implement reliable techniques for the protection of the salient pole synchronous generators or turbo generators. In general, most of the conducted research concentrates on faults in the four major components of any synchronous generators, which are stator windings, rotor

1-4244-0493-2/06/$20.00 2006 IEEE.

changes in system configurations. Three ANNs have been used in the second scheme [16] in order to cover 99.71% of the generator winding with higher boundaries of stability and sensitivity limits. However, the scheme could not able to classify the faulted phases. Analysis of armature magneto motive force (mmf) and field current has been reported in [21] for detecting the stator winding faults and rotor circuit deterioration respectively. A non-differential protection of synchronous generator stator winding utilizing fault currents of the generator terminals has been used in [22]. In this work, multi channel filters of different centre frequencies have been implemented at the outputs of the CTs for extracting desired fault transient bands. The spectral energies of these bands have been compared for detecting faults in the synchronous generator. However, space harmonics in the air gap field were neglected and no experimental validation was provided [5]. In this paper, the main objective is to analyze discontinuities in current and voltage signal during fault conditions. The acquired and generated data are first preprocessed with high pass wavelet filter to get the discontinuities of the analyzed signal. Interpreting the DWT coefficients can sometimes be rather difficult because of the way DWT coefficients are presented. Wavelet coefficients are of inner product of the signal and a wavelet basis function. Thus, the cross-correlation function and the similarity between the signal and wavelet basis are only interpreted. The DWT most often employs a Fast Fourier Transform (FFT) and dyadic grid in its computation, and it is not feasible to make a color-magnitude scale versus frequency plot like that produced for the continuous wavelet transform (CWT). To avoid the problem described above, a windowing function such as .e .t is applied to wavelet coefficients to get the importance of the last sampled datum in a discrete current or voltage vector. Afterwards, a statistical method called Maximum Likelihood Parameter Estimation (MLE) is applied for windowed wavelet coefficients. It is observed that wavelet power of the related phases get positive high value during the internal faults and negative high value during the external faults. The overall proposed algorithm is seen in Figure 1.

II. THEORETICAL BACKGROUND OF DWT The proposed technique is mainly based on the discrete wavelet transform of the fault currents by using Daubechies mother wavelet and scaling functions at the first level. The Daubechies wavelets (filters) have three important properties: a) they are orthogonal, b) they are maximally flat and c) they are compactly supported (the impulse response is zero outside a certain time interval). The advantage of analyzing a signal with wavelets is that it enables one to study local features of the signal with a detail matched to their characteristic scale. In the temporal domain such a property allows for an effective representation of transient signals. It can be said that the DWT enables one to make a multiresolution analysis of a signal. The basic notion of the DWT is that this filtering can be done on a dyadic basis, dividing bands into halves. It is possible to have both smooth wavelets with compact support and symmetry of the associated scaling functions, and this avoids bias for the locations of maxima and minima of the signal. Discrete families of orthonormal wavelets are obtained by dilating or contracting and translating 0,0 , with the choice
a = 2l and b = ka with l , k Z (Z is the set of integers). Equation (1) shows definition of orthonormal wavelets.

l ,k ( x ) =

1 2l

x k 2l = 1 2 l x k l 2 2l

(1)

Then, DWT and inverse DWT are given by the following equations in the L2-sense.
W l ,k = f ( x) =
~

f ( x)
~

l , k ( x ) dx

(2)

l =k =

l ,k l , k ( x )

(3)

Fig. 1. Proposed protection algorithm

where f ( x) is the discrete signal, l ,k is the wavelet function [23]. When analyzing signals of a non-stationary nature, it is often beneficial to be able to acquire a correlation between the time and frequency domains of a signal. The Fourier transform provides information about the frequency domain, however time localized information is essentially lost in the process. The problem with this is the inability to associate features in the frequency domain with their location in time, as an alteration in the frequency spectrum may result in changes throughout the time domain. In contrast to the Fourier transform, the wavelet transform allows exceptional localization in both the time domain via translations of the mother wavelet, and the scale (frequency) domain via dilation. The wavelet coefficients are calculated for each wavelet segment, giving a time scale function relating the wavelets correlation to the signal. The dilation function of the wavelet transform can be represented as a tree of low and high pass filters, with each step transforming the low pass filter. In this paper the unpadded zero filter kernel is used for having an impulse response duration of sampling intervals. Therefore, zero padding is unnecessary for filter implementation in MATLAB .

III. THE PROPOSED FAULT DETECTION TECHNIQUE The proposed technique is based on the correlation of wavelet coefficients of both current and voltage signals obtained from terminal nodes. In this study, faults in the generator protection region are called internal and outside the generator protection region are called external. Fig. 2 shows experimental set-up in the laboratory environment. The DC motor is considered as the prime mover for the test. The following equation is used to get detail coefficients of the analyzed signal: 1 m d m, n =< f (t ), mn (t ) >= m / 2 f (t ) (a0 t n) (4) a0 Both time and time scale parameters are discrete in discrete wavelet transform. As far as the structure of computations is concerned, the DWT is in fact the same as an octave band filter bank.

seen in Figure 3. The aim of this procedure is to get the significance of the latest sample in detail coefficients. It is essential because when the fault begins, last samples are multiplied with a high value and afterwards it changes the mean value calculated by MLE.

Fig. 3. An exponential function applied to detail coefficients

Fig. 2. Experimental set-up in the laboratory environment

This technique requires both current and voltage samples for one cycle. Therefore, the computation time of the proposed algorithm needs 20ms (for 50 Hz) for initial stage and after acquiring the one period date, a sliding window calculation needs only 0.5msec time delay to acquire the current and voltage samples. To acquire signals NI-DAQ PCI 16MIO-E is used and sampling frequency is set to 1 kHz, which corresponds 20 samples per cycle. After having samples, wavelet decomposition is performed for the current and voltage samples. It can be observed the detail signal (D) clearly shows distinctive features of the transients, i.e., immediately after the fault occurrence several sharp spikes appear in the detail signal indicating the occurrence of a disturbance because of including the high frequency components ranging from Fs / 2 Fs . Detail coefficients of current and voltage signals are multiplied one by one and instant wavelet power of each phase are calculated. Calculation procedure is as follows:
Power _ R =

It is observed that average value of the new distribution function has distinctive features for normal operating condition and different kinds of faults. This procedure is performed by using mle(unid, Power_R) in MATLAB environment. Since a windowing function has been already applied to detail coefficient, discrete distribution function is applied to calculate MLE value of Power_R. If the sign of calculated MLE values is negative, it is concluded that there is a fault (disturbance) outside the generator protection region. If the sign of calculated MLE values is positive, it is concluded that there is a fault inside the generator protection region [24]. Figure 4 shows possible fault generations in the laboratory environment. To simulate high impedance faults, fault resistances are used.

Fig. 4. Possible fault generations

cCD .cVD
k k =1

IV. RESULTS
k

(5)

where N gives the length of the discrete vector, cCD is detail coefficients of current and cVD is detail coefficients of voltage samples, Power_R is wavelet power of R phase. Similarly, wavelet Power_S and Power_T are calculated by using equation 5. Then, a special sliding windowing function is applied to each calculated wavelet power of the phases as

A lot of experiments are performed to show accuracy of the proposed algorithm in the laboratory environment by using two different synchronous generators (SG1 and SG2) and an ATP-EMTP simulation of a generator, TYPE59. These experiments include normal operating conditions at different load changes, single phase to ground fault and phase-phase faults. Threshold value is calculated by using rated values. This is simply mean value of the wavelet powers of the phases

under normal condition. Two different laboratory synchronous generators are rated at 3-phase, 220V, 0.8kVA and 3-phase 220V, 1.1 kVA for SG1 and SG2, respectively. The wavelet powers for each phase are given in Table-I and Table-II for SG1 and SG2, respectively for normal operating (unfaulted) conditions.
TABLE-I Normal operating condition for SG1 (0.8kVA, 220V) WP_R WP_S WP_T 567 524 442

Variation of wavelet power (WP) for Table-VI is shown in Figure 5.

Then, threshold value is

2 (567 + 524 + 442) = 1022 . 3

TABLE-II Normal operating condition for SG2 (1.1kVA, 220V) WP_R WP_S WP_T 415 288 321

Then, threshold value is 2 / 3(415 + 288 + 321) = 683 . The ratio of 2/3 is simply chosen to catch the possible high impedance fault either internal or external. A. Phase To Ground (L-G) Internal Faults In this case study, the generators are loaded with star connected resistors (300 ) and connected to the infinite bus system. An internal fault as R phase to ground fault is generated with a fault resistance of 50 . Faulty phase is typed bold. The wavelet powers for each phase are given in Table-III to Table-VI for SG1 and SG2 for stator phase to ground (L-G) internal fault conditions.
TABLE-III R phase to ground fault for SG1 WP_R WP_S WP_T 5504 717 357

Fig. 5. Voltages, currents and variation of wavelet power (WP) of S phase for the case of phase S to ground fault (internal). The first three groups are voltages of R, S and T phases, while the second three groups are currents of R, S and T phases and the last one is variation of wavelet power (WP) of phase S.

B. Phase-To-Phase (L-L) Internal Faults Similarly, stator phase-to-phase faults are generated for SG1 and SG2. The wavelet powers for each phase are given in Table-VII to Table-X for SG1 and SG2 for stator phase to phase (L-L) internal faults.
TABLE-VII Internal short circuit fault between phase R and phase S for SG1 WP_R WP_S WP_T 6475 11186 716

TABLE-VIII Internal short circuit fault between phase R and phase S for SG2 WP_R WP_S WP_T 3670 6098 630

TABLE-IV R phase to ground fault for SG2 WP_R WP_S WP_T 10002 628 596

TABLE-IX Internal short circuit fault between phase S and phase T for SG1 WP_R WP_S WP_T 696 2700 3508

TABLE-V S phase to ground fault for SG1 WP_R WP_S WP_T 610 3197 496

TABLE-X Internal short circuit fault between phase S and phase T for SG2 WP_R WP_S WP_T 345 7851 24935

TABLE-VI S phase to ground fault for SG2 WP_R WP_S WP_T 646 2929 641

Variation of wavelet power (WP) for Table-VIII is also shown in Figure 6.

Fig. 6. Voltages, currents and variations of wavelet power (WP) of phase R and phase S for the case of phase R and phase S short circuit fault (internal). The first three groups are voltages of R, S and T phases, while the second three groups are currents of R, S and T phases and the last one is variation of wavelet power (WP) of phase R and phase S.

Fig. 8. An internal fault condition between R and S phases in ATP-EMTP environment.

In Figure 7, an external fault is generated on TYPE-59 in ATP-EMTP. An external fault between S and T phases begins at 2 sec and ends at 3 sec. Total simulation time is 5 sec. Note that wavelet power (WP) values of S and T phases are negative. The results confirm that the L-L fault is external.

In the ATP-EMTP studies, the simulation runs about 5 sec. and a total of 4546 samples are acquired. The beginning moment of the fault is 2 sec. and it ends at 3 sec. Therefore, the fault instant at 2 sec. comes across the 1818th sample in the stored vector array. Similarly, the ending time of the fault (at 3 sec.) comes across 2727th sample. Since a sliding window calculation with 20 samples is performed, the WPs of the faulted phases begins at the fault instant, 91st sample, ( 1818 / 20 91 ) and ends when the fault is cleared ( 2727 / 20 136 ). V. CONCLUSIONS In this paper, a novel wavelet power based generator protection algorithm is presented. This approach is based on current and voltage samples acquired from the terminal nodes of the generator and their wavelet decomposition at the first level. Daubechies 4 type mother and scaling vectors are used for the wavelet analysis. It has been observed that MLE values of the calculated instantaneous wavelet powers of each phase have distinctive features during the power system disturbances. During any fault condition, wavelet power of the related phase(s) has higher values than the others. This approach uses 20 current and voltage samples at the rate of 1000Hz sampling frequency.

Fig. 7. An external fault condition between S and T phases in ATP-EMTP environment.

VI. REFERENCES
[1] Working Group 16 Report of Rotating Machinery Protection, Power System Relaying Committee, Performance of generator protection during major system disturbances, IEEE Trans. Power Delivery, vol. 19, pp. 1650-1662, Oct. 2004. [2] A.G. Phadke and J.S. Thorp, Computer Relaying for Power Systems, John Wiley & Sons Inc, 1988. [3] IEEE Committee, 1988, Survey of Experience with Generator Protection and Prospects for Improvements using Digital Computers, pp. 1511-1522. [4] H.A. Darwish, A-M.I. Taalab, and T.A. Kawady, Development and implementation of an ANN-based fault diagnosis scheme for generator winding protection, IEEE Trans. Power Delivery, vol. 16, no. 2, pp. 208-214, Apr. 2001. [5] D. Bi, X. Wang, W. Wang, Z.Q. Zhu, and D. Howe, Improved transient simulation of salient-pole synchronous generators with internal and ground faults in the stator winding, IEEE Trans. Energy Convers., vol. 20, no. 1, pp. 128-134, Mar. 2005.

In Figure 8, an internal fault is generated on TYPE-59 in ATP-EMTP. An internal fault between R and S phases begins at 2 sec and ends at 3 sec. Total simulation time is 5 sec. Note that wavelet power (WP) values of R and S phases are positive. The results confirm that the L-L fault is internal.

[6] J. Penman and H. Jiang, The detection of stator and rotor winding short circuits in synchronous machines by analyzing excitation current harmonics, Proc. IEE Int. Conf. Opportunities and Advances in International Power Generation, no. 419, 1996, pp. 137-142. [7] W. Shuting, L. Heming, L. Yonggang, and W. Yi, The diagnosis method of generator rotor winding inter-turn short circuit fault based on exciting current harmonics, Conf. Rec. IEEE 5th Int. Conf. Power Electronics and Drive Systems, 17-20 Nov. 2003, vol. 2, pp. 1669-1673. [8] L. Yonggang, Z. Hua, and L. Heming, The new method on rotor winding inter-turn short circuit fault measure of turbine generator, Conf. Rec. IEEE Int. Conf. Electric Machines and Drives, 1-4 June 2003, vol. 3, pp. 1483-1487. [9] J. Basilesco and J. Taylor, Report on methods for earthing of generator step-up transformer and generator winding neutrals as practiced throughout the word, CIGRE, no. 121, pp. 86101, 1988. [10] E. M. Gulachenski and E. W. Courville, New England electrics 39 years of experience with resonant neutral grounding of unit-connected generators, IEEE Trans. Power Delivery, vol. 6, no. 3, pp. 10161024, Jul. 1991. [11] M. E. Mandour, Design and commission of protective scheme for generator stator winding based on the third harmonics, Conf. Rec. 32nd Universities Power Engineering Conf., vol. 2, 1997, pp. 713-717. [12] R. L. Schlake, G.W. Buckley, and G. McPherson, Performance of third harmonic ground fault protection schemes for generator stator windings, IEEE Trans. on Power Apparatus and Systems, PAS 100, no.7, pp. 3195-3202, 1981. [13] G.W. Buckley and R. L. Schlake, Performance of third-harmonic ground fault protection schemes for generator stator windings, Proc. IEE Int. Conf. Electrical Machines Design and Applications, no. 213, 1982. pp. 225-230. [14] M. Fulczyk and R. Mydlikowski, Influence of the generator load conditions on third harmonic voltages in generator stator windings, IEEE Trans. on Energy Convers., vol. 20, no. 1, pp. 158-165, Mar. 2005. [15] A. I. Megahed and O. P. Malik, An artificial neural network based digital differential protection scheme for synchronous generator stator winding protection, IEEE Trans. on Power Delivery, vol. 14, no. 1, pp. 86-93, 1999. [16] A. I. Taalab, H. A. Darwish, and T. A. Kawady, ANN-based novel fault detector algorithm for generator windings protection, IEEE Trans. on Power Delivery, vol. 14, no. 3, July 1999. [17] P. K. Dash, O. P. Malik, and G. S. Hope, Fast generator protection against internal asymmetrical faults, IEEE Trans. on Power Apparatus and Systems, PAS-96, no. 5, pp. 1498-1506, 1977. [18] T. S. Sidhu, B. Sunga, and M. S. Sachdev A digital technique for stator winding protection of synchronous generators, in Electric Power Systems Research, vol. 36, no.1, pp. 45-55, 1996. [19] V. A. Kinitsky, Calculation of internal fault currents in synchronous machines, IEEE Trans. on Power Apparatus and Systems, PAS-84, no. 5, pp. 381-389, 1965. [20] P. J. Tavner and J. Penman, Condition monitoring of electrical machines, Research Studies Press Ltd., 1987. [21] J. Sottile, F.C. Trutt, and A.W. Leedy, Condition monitoring of brushless three-phase synchronous generators with stator winding or rotor circuit deterioration, Conf. Rec. IEEE 36th IAS Annual Meeting, Chicago, IL, 30 Sept. - 4 Oct. 2001, pp. 1587-1594. [22] Z.Q. Bo, M.A. Redfern, S. Potts, S. Weller, N.F. Chin, and F. Jianq, Non-differential protection of a generators stator utilizing fault transients, IEE 7th Int. Conf. Developments in Power System Protection, Amsterdam, 9-12 Apr. 2001, pp. 503-506. [23] J.C. Goswami and A.K. Chan, Fundamentals of Wavelets, John Wiley & Sons Inc, 1999. [24] Erdinc Arisoy, Wavelet Based Generator Protection Algorithm, Ondokuz Mayis University, Institute of Basic and Applied Sciences, MSc Thesis, 2004.

Okan Ozgonenel was born in Samsun, Turkey in 1967. He received M.Sc degree in electrical education from Marmara University in 1992, Ph.D degree in electrical engineering from Sakarya University in 2001. He has been with Ondokuz Mayis University since 1991 where he is a lecturer at Electrical & Electronics Engineering Department. His main research interests are digital algorithms, digital signal processing, simulation methods for power transformers, power system control and protection, and wavelet techniques.

Erdinc Arisoy was born in Yozgat, Turkey on March 01, 1979. He graduated from Ondokuz Mayis University, Electrical & Electronics Eng. Department in 2001. He received MSc degree at the same institute in 2004. His main research topics are wavelets, power system protection and numerical techniques. He is currently working for Ondokuz Mayis University, Electrical & Electronics Eng. Department as a research assistant.

M.A.S.K Khan (S05) was born in Bangladesh in 1978. He received the B. Sc. degree in electrical and electronic engineering from Bangladesh University of Engineering and Technology (BUET), Dhaka, Bangladesh in 2004. He is currently working toward the Masters degree in electrical engineering at Memorial University of Newfoundland, St. Johns, NF, Canada. His current research interests include application of wavelet transforms and its real-time implementation for the protection of disturbances in electrical machines and drives.

M. Azizur Rahman (S67-M68-SM73-F88) of Memorial University has been awarded the IEEE William E. Newell Power Electronics Award for outstanding achievement in Aachen, Germany on June 24, 2004. The award is named after Dr. William E. Newell of Westinghouse. It has been awarded annually since 1977. The award recipient is judged to be exceptional in the multidisciplinary field of power electronics. The recipient receives a plaque and cash prize of US $5000. Dr. Rahman is the 2nd Canadian recipient of Newell award; the first Canadian winner was from University of Toronto in 1979. Aziz has received his B.Sc.Eng, M.A.Sc and PhD degrees from the BUET, Dhaka, University of Toronto and Carleton University in 1962, 1965 and 1968 respectively. In addition to his teaching for 42 years, Dr. Rahman has over 10 years of concurrent industrial experiences. He is a professional engineer in Newfoundland and Ontario. In 1993, Dr. Rahman was the first Canadian to receive the IEEE Power Engineering Societys Cyril Veinott Electromechanical Conversion Award. He also received the IEEE Industry Applications Society Outstanding Achievement award in 1992 and is one of the few Canadian scholars, who received the highest achievement awards from the three IEEE Societies.