On-line Trained Neuro-Fuzzy Distance Relay with Directional Element

Tamer Abdelazim, Student Member, IEEE, A. H. Osman, Member, IEEE, O. P. Malik, Fellow, IEEE

Abstract-- A new directional element for an on-line trained Neuro-Fuzzy Network (NFN) numerical distance relay is developed. Numerical simulation studies are performed on a differential equation based transmission line model that includes the fault resistance and the remote-end infeed. The fault direction is identified based on the +/- sign of the estimated fault distance. The results obtained show fast estimation of distance to the fault with accurate fault direction assessment

Index Terms-- Distance Protection Relaying, Directional

I. INTRODUCTION ISTANCE protection relays are commonly used for the protection of transmission lines. Since the impedance of a transmission line is proportional to its length, such relays are based on impedance calculation. They ensure the integrity of directional elements to decide the direction of close up faults. Basic distance relaying principles employ steady-state phasor approach. This approach may involve well-known algorithms based on discrete Fourier transform [1,2], symmetrical components [3], Kalman filtering [4], or orthogonal expansion [5]. These algorithms work on the estimation of fundamental frequency components of voltages and currents. Conventional negative sequence directional elements base their measurements on the phase angle between the negative sequence voltage and current. Another technique based on measuring the positive sequence incremental impedance for assessing the direction to the source of a disturbance has been described in [6]. These types of algorithms need at least one cycle of signal to extract the fundamental frequency phasors needed to estimate the distance and direction of the fault and may be much longer when current transformer (CT) saturation occurs [7]. In digital relays, other methods based on the line differential equation or on traveling waves approach have been proposed. The differential equation approaches compute impedance from the differential equation relating voltages and currents. These techniques do not require so much time to detect the fault. In most cases, the tripping decision may
The authors are with the Department of Electrical and Computer Engineering, University of Calgary, AB T2N 1N4, Canada Tamer Abdelazim (e-mail: [email protected]). A.H. Osman (e-mail: [email protected]). O. P. Malik (e-mail: [email protected]).

occur before the CT saturation. The differential equation based techniques also are not affected by the exponentially decaying current when faults occur or by the variation of the fundamental frequency. Conventional differential equation distance relaying algorithms [8,9] are based on a specific line model consisting of a resistance in series with an inductance. The values of these two parameters are obtained by minimizing the meansquare mismatches between the actual current and voltage and those predicted by the assumed line model. However, these formulations are mathematically not well posed since the minimization procedures are performed as two-parameter problems when they should be dealt with as single-parameter minimization problems. This error is due to the fact that the actual resistance and inductance seen by the relay are related to each other by the fault distance. An improved version of the distance relay described in [8,9], that includes the fault resistance and the remote-end infeed, has been proposed in [10]. This version needs the storage of one cycle pre-fault load current and the fault distance is estimated using the recursive least-squares technique. A modification for the transmission line model developed in [10], which does not depend on the storage of the pre-fault load current, is introduced in this paper. Neural networks have been shown to possess good function approximation capability. One potential limitation of conventional neural network models is that they can be difficult to interpret. An approach to improve interpretability is based on decomposing the process operation into a number of local operating regions. Then a reduced order linear model can be used. Takagi-Sugeno Fuzzy systems can provide an appropriate means to define linear operating regions [11,12]. Fuzzy models can be implemented as a Neuro-Fuzzy networks [13]. Unlike the off-line trained Neuro-Fuzzy Networks (NFN) proposed for distance relaying in [14], a new solution based on on-line trained feed forward Neuro-Fuzzy Network (NFN) is descried in this paper. The technique proposed is able to accurately estimate the distance and direction of the fault in less than one cycle of the fundamental frequency after the detection of the fault. Computer simulation studies have been conducted using PSCAD program to generate the voltage and current signals from the simulated network. Numerical simulation results are presented for various types of forward and reverse faults with different fault resistances and under different operating conditions.

II. RELAY MEASUREMENTS AND FAULT DETECTION A. Relay Measurements In order to investigate the performance of the proposed distance protection algorithm and its associated directional element, a sample power network has been simulated using the electromagnetic transient computer program PSCAD. The simulated power network is shown in Fig. 1. The network models two adjacent 240kV ideally transposed transmission lines. Both lines, AB and BC are 225km long. Thvenin impedance of area 1 is chosen to be higher than that of area 2 to insure a high remote end infeed. The tested directional distance relay is located at bus B, at the beginning of the (BC) line as shown in Fig. 1. The relay at this location will be tested for forward faults on the (BC) line and for reverse faults on the (AB) line. The relay requires the measurement of the three line-to-neutral voltages and the three line currents at the relay location. The measured six analog signals at the relay location are passed through transfer function blocks of the PSCAD program, simulating analog anti-aliasing 2nd order low-pass filters. The output signals of these filters are the input signals to the proposed distance protection algorithm. The analog to digital conversion is modeled as sampling the voltage and current values with 16 samples per fundamental power cycle, corresponding to a sampling frequency of 960Hz.
area 1 240kV B 225km C area 2

nth and (n-3)th samples are compared. If the difference between them is more than a specified threshold value, a fault is assumed to have occurred. However, for security reasons, the final indication of the fault is only given if three successive comparisons show that the difference is more than the specified threshold. For the studies reported in this paper, the threshold is set at two times the peak value calculated using the (n-3)th sample [16]. III. PROPOSED DISTANCE RELAYING ALGORITHM A. Transmission Line Model Suppose that a fault to ground with a resistance R f occurs on one of the three phases (for example phase a) of Fig. 1 at a distance D f from the relay location at bus A. The line is fully transposed and its capacitance is neglected. The wellknown symmetrical components equations of the voltage of phase a at the relay are
Va 0 = Z 0 I a 0 + Vaf0 Va1 = Z1 I a1 + Vaf 1 f Va 2 = Z 2 I a 2 + Va 2

(4-a) (4-b) (4-c)

225km

125Km

where Z 0 , Z1 , Z 2 the sequence impedances of the line up to the point of the fault. Va 0 , Va1 , Va 2 the symmetrical components of the voltage of phase a at the relay. I a 0 , I a1 , I a 2 the symmetrical components of the current of phase a at the relay.
f Vaf0 , Vaf 1 , Va 2 the symmetrical components of the voltage

Fr

Ff

Fig.1. Transmission line model system.

of phase a at the fault. Adding (4-a), (4-b) and (4-c)

(5)
Z1 = R1 + sL1 and

B. Fault Detection Various techniques are available to detect a disturbance or a fault in a power system. It is desired that a fault detector have a fast transient response. This can be achieved by using a short data window (4 samples) algorithm, such as the one described in [15]. For the work presented in this paper, the algorithm of [15] was selected for the fault detection. The algorithm uses the first derivative of the input signal to compute the phasors from the sampled values. If (1) describes the assumed signal, (2) represents its first derivative, then (3) provides the estimate of the peak value of the input signal
i = I p sin ( t + )

where

Z1 = Z 2 . In the s-domain,

Z 0 = R0 + sL0 . Equation (5) becomes

(1) (2) (3)

i ' = I p cos ( t + )

I2 p

i'

+ i2

To detect the presence of a fault, the estimated peak values of the fundamental frequency of the current phasors from the

(6) + Vaf (s) Applying the inverse Laplace transform to (6), the voltage at the relaying point in the time domain can be obtained as di (t ) f (t) + D f r1 ia (t) + l1 a va (t) = va dt (7) di0 (t ) + D f (r0 r1 ) i0 (t ) + (l0 l1) dt where distance between the fault and the relay location. Df r1, r0 positive- and zero-sequence resistances, respectively, per unit length of the line. l1, l0 positive- and zero-sequence inductances, respectively, per unit length of the line.
vaf (t ) voltage to ground of phase a at the fault.

Va (s) = (R1 + sL1) I a (s) + (R0 + sL0 R1 sL1) I 0 (s)

va (t ) voltage to ground of phase a at the relay. ia (t ) current of phase a at the relay. i0 (t ) zero-sequence current at the relay. Equation (7) is the desired transmission line model. When

bolted faults are considered vaf (t ) = 0 and (7) reduces to the same model used in [8]. In [9] an extension to this model to include the effects of the fault resistance and the remote end infeed has been proposed. The transmission line model obtained required the storage of one full cycle of the pre-fault instantaneous load current in order to solve for the current through the fault resistance. The use of this stored instantaneous pre-fault load current might affect the accuracy of the fault distance estimation. In this paper a new modification is proposed in order to eliminate the need to store the instantaneous pre-fault load current and to work with the instantaneous measured signals. The voltage at the fault in (7) can be expressed in terms of the fault resistance R f and the fault current iaf (t ) as
vaf (t ) = R f iaf (t )

Substituting (15) into (7) yields the required differential equation that models the transmission line during line-toground faults. This equation is given by di (t) va (t) = D f r1 ia (t) + Df l1 a + D f (r0 r1) i0 (t) dt (17) di0 (t) ' + D f (l0 l1) + R f i0 (t) dt To simplify the notation, the terms multiplied by D f and
R f can be denoted by f1a (t ) and f 2a (t ) , respectively.

Equation (17) becomes

(18-a)

where
f1a (t ) = r1 ia (t ) + l1 dia (t ) + (r0 r1 ) i0 (t ) dt di (t ) + (l0 l1 ) 0 dt

(18-b)

and (8)
f 2a (t ) = i0 (t ) (18-c) Equations similar to (18) can be easily written for phase b to ground, and phase c to ground faults. For phase-to-phase faults the fault resistance is normally negligible, and there is no zero sequence current component. Therefore, the differential equation that describes these types of faults can be given as (example: phase a to phase b fault) di (t ) vab (t ) = D f r1 ( ia (t ) ib (t )) + D f l1 a dt (19) dib (t ) D f l1 dt where vab (t ) = va (t ) vb (t ) represents the line-to-line voltage between phase a and phase b. ia (t ) the current of phase a at the relay. ib (t ) the current of phase b at the relay.

The zero sequence component of the fault current is known to be

(9)

(10)

The relation between the zero-sequence current at the relay with the zero-sequence fault current can be approximated as
i0 (t) = i0f (t)

RE x0 + (L D f )x0

SE ( x0

+ D f x0 ) +

RE x0 + (L D f )x0

(11)

SE RE and x0 are the zero-sequence reactances behind where x0 the sending-end and receiving-end, respectively, xo is the zero-sequence reactance per unit length of the line and L is the length of the protected line.

If

K0 =

SE x0 + D f x0 RE x0 + ( L D f ) x0

then, (12) (13) (14)

Equation (19) can also be simplified using the f (t ) notation as done before in the line-to-ground case, vab (t ) = f1ab (t ) D f + f 2 ab (t ) R f (20-a) where
f1ab (t ) = r1 ( ia (t ) ib (t )) + l1 dia (t ) di (t ) l1 b dt dt

(t ) = (1 + K 0 ) i0 (t ) Substituting (12) into (10) yields vaf (t ) = 3R f (1 + K 0 ) i0 (t ) R 'f = 3R f (1 + K 0 )

i0f

(20-b)

and
f 2ab (t ) = 0 (20-c) Equations similar to (20) can be easily written for phase b to phase c, and phase c to phase a faults. The linear differential dynamic operator, based on the use of numerical derivative techniques, is applied to (18) and (20) to obtain a form that can be used with digital signals. For example, (18) transformed with a first-order derivative becomes Va (n) = F1a (n) D f + F2 a (n) R 'f

Let then

vaf (t ) = R 'f i0 (t ) (15) The relation between the zero-sequence current at the relay and the measured line currents is known to be 1 i0 (t ) = ( ia (t ) + ib (t ) + ic (t ) ) (16) 3

(21-a)

where
l F1a (n) = r1 (ia (n)) + 1 (ia (n) ia (n 1)) T (21-b) l l + (r0 r1 ) (i0 (n)) + ( 0 1 ) (i0 (n) i0 (n 1)) T F2 a (n) = i0 (n) (21-c) and T is the sampling period. In the same way (20) can be transformed with a first-order derivative. Left-hand side of the transformed differential equations, for example in (21-a), are the actual measured voltage samples whereas those on the right-hand side are the predicted voltage samples as estimated using the current samples and the assumed line model. The differences between the actual and predicted voltage samples, due to the approximations made in the derivation, are considered as errors to be minimized by properly selecting the parameters D f and R 'f .

R ( l ) : If F1a (n) is Ai
then

and

F2 a (n) is B j

V l ( n ) = w l
where Ai is a fuzzy set

(22)

F1a ( The group of fuzzy sets defined

in the universe of discourse of F1a (n) ) , Bj is a fuzzy set F2 a y ( The group of fuzzy sets defined in the universe of discourse of F2 a ( n) ),
'l wl = [D l f R f ] and

= [ F1a (n) F2 a (n)]T .

Many techniques for estimating the distance between the relay and the fault point have been proposed using differential equations to model transmission lines [8-10]. These proposals mainly employ the Least-Squares (LS) or the Recursive LeastSquares techniques (RLS) for estimating the fault distance. B. Proposed NFN Directional Distance Relay The proposed technique for estimating the fault distance is based on using an on-line trained neuro-fuzzy network (NFN). A NFN is based on a fuzzy system trained by a learning algorithm derived from neural network theory. In this paper the NFN is trained by Levenberg-Marquardt algorithm[17]. The learning procedure operates on local regions, and causes only local parameters modifications in the underlying fuzzy system. The neuro-fuzzy network can be viewed as a 3-layer feedforward neural network. The first layer represents input variables, the second (hidden) layer represents fuzzy rules and the third layer represents output variables. This NFN can be interpreted as a system of fuzzy rules. By on-line learning, It is proposed to create the system out of the measurements data, although an initial model has been made by prior knowledge using orthogonal least squares (OLS) algorithm to reduce the set of fuzzy rules [11]. For example in (21-a), the input to the NFN should have two inputs F1a (n) and F2 a (n) and a single output Va (n) . The NFN needs to estimate the two parameters D f and R . The NFN structure [18] is shown in Fig. 2. On-line training of the proposed NFN is basically the process of estimating the optimum weight values, D f and
' f

Fig. 2. Proposed NFN structure.

The output of NFN using the weighting average method [18] is

R
l =1

( F1 a ( n )) * l ( F 2 a ( n )) * V
R
l

^ l

a(n)

l =1 l

( F1 a ( n )) * l ( F 2 a ( n ))

(F =
R
l =1

1a

( n )) * l ( F 2 a ( n )) * w l T ( 23 )

l =1

l ( F1 a ( n )) * l ( F 2 a ( n ))

The parameters of the NFN can be written as

( F (n)) * ( F (n)) * w w= ( F (n)) * ( F (n))

1a

2a

(24)

l =1

1a

2a

where

l ( F1a (n)) and l ( F2 a (n)) are

the membership values

R 'f , that can generate the measured output voltage, Va (n) ,

for the inputs F1a ( n) and F2 a ( n) respectively and The described technique is designated to estimate the fault distance D f and the fault resistance R f from R 'f , if desired, using (14). The NFN estimation process starts once the fault detector

indicates a fault. Adaptive input data windows for the inputs, F1a (n) and
F2a (n) , and for the output, Va (n) , are proposed here for accurate and fast estimation. The adaptive data window technique is based on starting to collect input and output samples at the instant of fault detection and progressively increasing the window length as new input and output samples become available. Once the size of the input/output data windows reaches 16, one cycle of the fundamental frequency, the oldest samples are discarded as new samples become available. The on-line training process will then continue its estimation with the 16 samples data windows.

Tests have been performed for different fault resistance values, namely, 0, 100, and 150. All cases have been studied for a fixed fault distance D f of 180 km, and for a fixed loading angle of 20 (power flow is from area 1 to area 2). The estimated fault direction and distance for the various fault resistance values are shown in Fig. 4. It can be seen that the error in estimating the fault distance is slightly increased by increasing the fault resistance.

IV. SIMULATION RESULTS The directional distance relay under test is located at bus B, at the beginning of the BC line as shown in Fig. 1. The measured three line-to-neutral voltage signals and the three line current signals at the relay location constitute the input to the proposed NFN directional distance-relaying algorithm. Performance of the on-line trained NFN directional distance relay has been investigated extensively for different forward and reverse fault locations, different fault resistances, and different fault types. A. Fault location The proposed relay has been tested for single line to ground (SLG) faults, phase a to ground, at different forward and reverse locations on the transmission line. The forward faults for the relay were simulated on the BC section and the reverse faults were simulated on the AB section as shown in Fig. 1. The forward fault locations were simulated at 5, 100, 200km and the reverse fault locations were simulated at -5, 100, -200km. All fault studies have been performed for a fixed fault resistance R f of 20, and for a fixed loading angle of 20 (power flow is from area 1 to area 2). The estimated fault distance in km versus time in ms is shown in Fig. 3. The time equal to 0 on the time axis represents the moment of fault detection. The proposed technique estimated the direction and distance for the 5 and 100km forward and reverse faults with an error less than 1.7%, while for the forward and reverse faults at 200km the error was 3.5% due to the high remote end infeed and the fault resistance. As aforementioned the positive and negative distance values represent the forward and reverse fault directions, respectively. The fault direction and distance have been estimated in less than 10 ms after the fault detection in all cases. B. Fault Resistance The proposed relay is tested for the effect of the existence of fault resistance R f in ground faults. The test is performed as a SLG fault on phase a at a relatively long distance from the relay in order to test for limiting conditions.
Fig. 3. Estimated fault distance for SLG different fault locations.

Fig. 4. Estimated fault distance for different fault resistances.

The error in the highest fault resistance case for the forward faults is less than 3% and the highest error in the reverse faults is 10%. The reason for the higher error fault distance estimation in the case of the reverse faults is that their direction is opposite to the power flow direction in addition to the high ground fault resistance values. This study illustrates the accuracy of the transmission line model used and the accurate and fast direction and distance estimation obtained by using the on-line trained NFN. C. Fault Type

The relay performance has been tested for different types

of fault. In the previous tests it is already tested for forward and reverse SLG faults. Therefore, in this test the relay is tested for forward and reverse line-to-line, double line-toground, and three phase faults. The estimated fault direction and distance for different forward and reverse fault types at 150 km from the relay location, and for a fixed loading angle of 20 (power flow is from area 1 to area 2) are shown in Fig. 5. The fault types are three-line fault (3L), double line-to-ground fault (DLG) with a ground fault resistance of 50, and line-to-line fault (LL) in the forward and reverse directions. It can be seen that the relaying algorithm successfully estimated the fault direction and fault location with an error of less than 3% for all fault types and within one cycle of the fundamental frequency after the detection of the fault.

M.S. Sachdev and M.A. Baribeau, A new algorithm for digital impedance relays, IEEE Trans. on Power Apparatus and Systems, Vol. 98, 1979, pp.2232-2240. [6] P.G. McLaren et al. A new directional element for numerical distance relays, IEEE Trans. on Power Delivery, Vol. 10, No. 2, April 1995, pp. 666-675. [7] T. Segui and P. Bertrand, Fundamental basis for distance relaying with parametrical estimation, IEEE Trans. on Power Delivery, Vol. 16, No. 1, Jan. 2001, pp. 99-104. [8] R. Poncelet, The use of digital computers for network protection, CIGRE, Paper No. 32-08, 1972, Paris, France, pp. 1-15. [9] P. Bornard and J.C. Bastide, A prototype of multiprocessor based distance relay, IEEE on Power Apparatus and Systems, vol. PAS-101, no. 2, Feb. 1982, pp. 491498. [10] A.S. AlFuhaid and M.A. El-Sayed, A recursive least-squares digital distance relaying algorithm, IEEE Trans. on Power Delivery, Vol. 14, No. 4, October 1999, pp. 1257-1262. [11] J. Abonyi, J.A. Roubos, M. Oosterom, F. Szeifert Compact TS-Fuzzy Models through Clustering and OLS plus FIS Model Reduction, Proc. of IEEE international conference on fuzzy systems, Sydney, Australia, 2001. [12] T. Abdelazim and O. P. Malik, "Analysis of Two Designed Fuzzy Logic Based Adaptive Power System Stabilizers", 35th North American Power Symposium ( NAPS 2003 ), Rolla, Missouri, October 20-21, 2003. [13] Zhang J., and A. J. Morris, Recurrent Neuro-Fuzzy Networks for Nonlinear Process Modelling, IEEE Transactions on Neural Networks. Vol. 10, No. 2, March 1999, pp. 313-326. [14] P.K. Dash, A.K. Pardhan and G. Panda, A novel fuzzy neural network based distance relaying scheme, IEEE Trans. on Power Delivery, Vol. 15, No. 3, July 2000, pp. 902-907. [15] B.J. Mann and I.F. Morrison,

[5]

Digital calculation of impedance for

transmission line protection . , IEEE Trans. on Power Apparatus and Systems, Vol. PAS-90, No. 2, March/April 1971. [16] T.S. Sidhu, D.S. Ghorta and M.S. Sachdev, An adaptive distance relay and its performance comparison with a fixed data window distance relay, IEEE Trans. on Power Delivery, Vol. 17, No. 3, July 2002, pp. 691-697. [17] H. Simon, "Neural Networks", 1994 Macmillan College Publishing Company. [18] J.-S.R Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE Transactions on Systems, Man and Cybernetics, Vol. 23, No. 3, May-June 1993, pp.665 685

Fig. 5. Estimated fault distance for different fault types.

V. CONCLUSIONS A new digital distance relaying technique based on a modified transmission line differential equation model and an on-line trained NFN has been developed. The relay has been extensively tested for high ground fault resistances, for different fault locations and for different fault types. The proposed relaying technique has the ability to provide accurate estimation for the fault distance and direction within one cycle of the fundamental frequency after the detection of the fault. VI. REFERENCES
[1] [2] [3] W.A. Lewis and L.S. Tippett, Fundamental basis for distance relaying on 3-phase systems, AIEE Trans., vol. 66, pp. 694709, 1947. P.G. Maclaren and M.A. Redfern, Fourier-series techniques applied to distance protection, IEE Proceedings, Vol. 122, 1975, pp. 1301-1305. A.G. Phadke, T. Hibka, M. Ibrahim and M.G. Adamiak, A microprocessor based symmetrical component distance relay, IEEE Power Industry Computer Applications Conference Proceedings, 1979, PICA-79, pp. 47-55. A. Girgis and E. B. Makram, Application of adaptive Kalman filtering in fault classification, distance protection, and fault location using microprocessors, IEEE Trans. on Power Delivery, Vol. 3, No. 1, Jan. 1988, pp. 301309.

VII. BIOGRAPHIES
Tamer Abdelazim (SM03) received his B.Sc. and M.Sc. in electrical engineering from Suez Canal University, Port Said, Egypt in 1995 and 1999 respectively. Currently, he is working towards a Ph.D. degree at the University of Calgary, Canada. His areas of interest include power system control, system identification and AI applications in power system. A. H. Osman (SM01) received his B.Sc. and M.Sc. in electrical engineering from Helwan University, Cairo, Egypt in 1991 and 1996 respectively. In 2003 he received his PhD. degree in electrical engineering from the University of Calgary, Canada. Currently, he is a research associate at the University of Calgary, Canada. His areas of interest include power system engineering, digital protection relaying, and power electronics. O. P. Malik (M66-SM69-F87) graduated in electrical engineering from Delhi Polytechnic, India, in 1952, and obtained the M.E. degree in electrical machine design from the University of Roorkee, India, in 1962. In 1965 he received the Ph.D. degree from the University of London, England, and D.I.C. from the Imperial College of Science and Technology, London. He became professor at the University of Calgary in 1974 and is a faculty professor emeritus. He is a Fellow of the Institution of Electrical Engineers (London), and a registered professional Engineer in the provinces of Alberta and Ontario, Canada.

[4]