Electrical Power and Energy Systems 33 (2011) 721730

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

Transmission lines distance protection using articial neural networks

Universidade Federal do ABC, R. Santa Adlia, 166, Bloco B, sala 906, Santo Andr, SP, CEP 09210-170, Brazil Escola Politcnica da Universidade de So Paulo, Av. Prof. Luciano Gualberto, 158, So Paulo, SP, CEP 05508-900, Brazil

a r t i c l e

i n f o

a b s t r a c t
This work presents the development and implementation of an articial neural network based algorithm for transmission lines distance protection. This algorithm was developed to be used in any transmission line regardless of its conguration or voltage level. The described ANN-based algorithm does not need any topology adaptation or ANN parameters adjustment when applied to different electrical systems. This feature makes this solution unique since all ANN-based solutions presented until now were developed for particular transmission lines, which means that those solutions cannot be implemented in commercial relays. 2011 Elsevier Ltd. All rights reserved.

Article history: Received 13 November 2008 Received in revised form 29 October 2010 Accepted 9 December 2010 Available online 1 February 2011 Keywords: Protective relaying Articial neural networks Transmission line protection Distance protection

1. Introduction Transmission lines are one of the most important power system components, responsible for delivering energy from generating plants to consumers. Those lines may present a high failure rate since, in most cases, they cover large distances and are exposed to different climate conditions. The distance protection method is one of the most used techniques for preventing damages that can be inicted to transmission lines. The conventional distance protection is based on impedance measurement elements that estimate the line impedance from the relay to the fault point using voltage and current phasor quantities. The ANN-based algorithms have been developed as an alternative to conventional methods, since they present very promising results with regard to precision and operating time. Refs. [17] present ANN-based algorithms that act as classiers, indicating whether the fault is located inside or outside the protection zones. Refs. [810] present ANN-based algorithms that act as function approximators which point out, directly or not, the fault distance. The algorithms presented in [4,6,10] are based on the phasor quantities of voltage and current fundamental components, obtained via DFT. Those components are used as inputs for their ANN. The other algorithms but [9] are based on voltage and/or current samples. Those samples are used as inputs for their ANN. The ANN-based transmission line protection methods described in [110,16] surely present limitations once they are trained to cor Corresponding author. Tel.: +55 11 4996 0106.
E-mail address: [email protected] (Ricardo Caneloi dos Santos). 0142-0615/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.ijepes.2010.12.029

rectly respond only to a particular transmission line. Therefore, considering such methods, it shall be necessary to re-create a new simulation environment set containing hundreds or thousands of fault cases for each change made at the transmission line characteristics. Besides that, it shall be also demanded to repeat the whole ANN training process, resuming the major drawback presented by those algorithms. It is proposed then the development and implementation of a new ANN-based algorithm, suitable for protecting any transmission line, regardless of its characteristics (such as tower geometry, type of conductors, length and voltage level). This new ANN-based algorithm does not need any topology adaptation or ANN parameters adjustment when applied to different transmission lines [17 19]. It was developed to exceed the existing limitations of previously developed ANN-based algorithms, allowing it to be implemented in commercial relays. This ANN-based algorithm operates as a function approximator, estimating the line impedance from the relay to the fault point. Similarly to conventional distance protection relays, this algorithm is also based on six impedance measurement elements, which estimate the line impedance, based on current and voltage samples provided by CTs and VTs, respectively.

2. Impedance measurement elements The conventional distance relays have six impedance measurement elements. Three of them (phase elements AB, BC and CA) are responsible for detecting faults that involve more than one phase. The remaining elements (three ground elements AN, BN and CN) are responsible for detecting phase-to-ground faults.

722

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730

Nomenclature ANN VT CT MLP Articial Neural Network Voltage Transformer Current Transformer Multilayer Perceptron DFT EMTP Discrete Fourier Transform Electromagnetic Transients Program

All six impedance measurement elements are responsible for protecting transmission lines against the following fault types: single line-to-ground (AN, BN and CN); double line (AB, BC and CA); double line-to-ground (ABN, BCN and CAN); and three-phase (ABC and ABCN), according to [11,12]. The presented algorithm implements all six measurement impedance elements using ANN. Therefore, this algorithm has the advantages of ANN-based algorithms, plus the advantages of conventional ones. All processing steps performed by the algorithm and presented in Fig. 1 are detailed in the following items. 2.1. Pre-processing step As illustrated in Fig. 1, the pre-processing step consists of generating suitable signals to compose the ANNs input arrays for each impedance measurement element using voltage and current samples from VTs and CTs. The type of the signal generated in this step was dened from the equating provided by the simplied model shown in Fig. 2. Such model describes a three-phase transmission line, ideally transposed, with no capacitive effect. As an example, consider a single line-to-ground fault (AN) in the transmission line depicted in Fig. 2. Eq. (1) represents this fault, assuming no fault resistance.

R1 Rf L1 Lf Lm R0 Rf 3:Rn L0 Ls 2:Lm 3:Ln

6 7 8 9

v a R1 iar L1
where

d iax dt

iar ia K r i

2 3 r0 r1 r1 l l l
1 0 1

iax ia K x i Kr R0 R1 R
1

where R is the positive seq. resistance of the faulted section; L the positive seq. inductance of the faulted section; R0 the zero seq. resistance of the faulted section; L0 the zero seq. inductance of the faulted section; r1 the positive seq. resistance of the faulted section per km; l1 the positive seq. inductance of the faulted section per km; r0 the zero seq. resistance of the faulted section per km; l0 the zero seq. inductance of the faulted section per km; va, vb, vc the voltage samples at the relay; ia, ib, ic the current samples at the relay; and Kr, Kx is the compensating factors. Eq. (1) shows that the positive sequence resistance (R1) and inductance (L1) for the section of the line between the relay and the fault point can be calculated from the M-samples of each signal va, iar and iax, using, for instance, some parameters estimation method. The solution adopted in this work is to estimate the faulted line section parameters using ANN. It was noticed during the networks training step that the inuence of the signal iax at the estimation of the R1 is very small and the same occurs to the inuence of iar at the estimation of X1. Based on those results, only the samples of the signals va and iar are used as input to the ANN that calculates R1 during a type AN fault. Also in this same case, only the samples of the signals va and iax are utilized as input to the ANN that calculates X1. The signals iar and iax are generated using (2) and (3) at the pre-processing step. Similar equations can be obtained for BN or CN faults by replacing the index a at (1) with b or c respectively. As another example, (10) represent a double line fault (AB) in the transmission line depicted in Fig. 2.

v a v b R1 ia ib L1
5

Kx

L0 L1 L1

d ia ib dt

v ab R1 iab L1

d iab dt

10

Fig. 1. Impedance measurement element: Processing steps.

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730

723

(2) Base values of voltage and current input data windows at kth sample

Vbabk max jv abk j Ibabk max jiabk j

jv abkM1 j jiabkM1 j

13

14

(3) Base impedance for kth sample

Zbabk

Vbabk Ibabk

15

Fig. 2. Transmission line model (Lf, Lm: self and mutual inductances; Rf: self resistance; Rn, Ln: resistance and inductance of the neutral current return path).

(4) Normalized values of voltage and current input data windows

 v ab
It is shown in (10) that the impedance for AB fault can be obtained from samples of vab and iab. Such equation is also valid for double line-to-ground faults ABN and three-phase faults ABC. Other similar equations can be obtained for BC, CA, BCN and CAN faults. Two ANN with the same structures but with different weights and bias are utilized in this work. The rst ANN, known as ANNR, estimates the positive sequence resistance of the faulted transmission line section. The second ANN, named as ANN-X, estimates the positive sequence reactance for the same line section. Table 1 shows the signals provided by the pre-processing step which are utilized as input to those two ANNs in each one of the six measurement elements. The pre-processing step also consists of normalizing each voltage and current input data windows. In order to perform this task, the pre-processing step divides the voltage and current input data windows by the largest voltage and current sample presented in these windows. The largest voltage sample is called Vb (base voltage) and the largest current sample is called Ib (base current). The ratio between Vb and Ib in each new input data window produces the base impedance, known as Zb. The base impedance is then utilized at the post-processing step. It converts the ANN output normalized values (R and X) into resistances and reactances expressed in ohms, as it will be discussed at Section 2.3. This procedure is presented below for input data windows containing M-samples. The k index indicates the most recent sample of the input data window. (1) Input data windows to phase element AB

v ab
Vbabk

 v abk

 v abkM1

16

ab iab abk i i Ibabk

abkM1 i

17

(5) ANN input data array

 v abk 3 7 6 . 7 6 . . 7 6 7 6  v ab ! 6 v abkM1 7 7 6 6  input data array AB  7 7 6 iabk iab 7 6 7 6 . 7 6 . 5 4 . abkM1 i 2

18

2.2. ANN step The ANN step is the main step of the proposed algorithm. It is basically composed by previously trained articial neural networks used to estimate resistance and reactance values from the relay to the fault point. According to Fig. 1 the impedance measurement elements are composed of two different ANNs. These ANNs (ANN-R and ANNX) are trained to estimate resistance and reactance, respectively, from the relay to the fault point. Both ANN-R and ANN-X receive data arrays containing samples of voltage and current signals, and continuously estimate the resistance and reactance from the relay to the fault point. Fig. 3 illustrates one of the ANN (ANN-R or ANN-X) of the phase element AB, which uses a data array composed by a voltage data window of M-samples and another current data window of M-samples. Notice that both ANNs present exactly the same structure, except for their weights and bias values. The resistance and reactance are estimated using normalized values as shown at Section 2.1 and discussed at Section 3. As illustrated in Fig. 3, both the input and output data arrays of the ANNs (ANN-R and ANN-X) are presented in normalized values. There are two main points that justify this option:  Imposition of activation function used in the ANNs neurons (hyperbolic tangent, 1 to +1);  A way to provide a generic behavior of the algorithm making it independent from particular transmission line characteristics. The normalizing process proposed by this method not only properly equates the ANN inputs but also enable the ANN training for different voltage levels, as it will be presented at Section 3.

v ab v ab
n iab iabk

v abkM1
iabkM1 o

11 12

Table 1 Input signals for impedance measurement elements. Impedance measurement elements Phase AB Input signals RNA-R RNA-X Fault type BC CA Ground AN BN CN

vab
iab

vbc
ibc

vca
ica

van
iar

vbn
ibr

vcn
icr

vab
iab AB ABN

vbc
ibc BC BCN

vca
ica CA CAN

van
iax AN

vbn
ibx BN

vcn
icx CN

724

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730

Fig. 3. ANN (ANN-R or ANN-X) of phase element AB.

Fig. 4. Post-processing step: Transforming R and X from normalized values to Ohms.

Fig. 5. Phase element AB.

2.3. Post-processing step This step is responsible for translating the ANNs output values (resistance and reactance) from normalized values to ohms and deciding whether to trip or not. This process consists of multiplying the ANNs output values by Zb (base impedance) calculated for each voltage and current input data windows, as presented in (15). The decision whether to trip or not is based on the resistance and reactance estimated values, in Ohms. This process is illustrated in Fig. 4. The protection zone shown at Fig. 4 is adjustable and can

present either Mho or Quadrilateral characteristic just as it can be found in any conventional distance protection relay. 2.4. Impedance measurement elements implementation The implementation of phase element AB and ground element AN are illustrated in this item. It is important to notice that all phase elements have the same structure presented in Fig. 5, except for the input signals. Analogously, all ground elements have the same structure presented in Fig. 6, except for the input signals. Both measuring elements are now presented using input arrays

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730

725

Fig. 6. Ground element AN.

of data windows containing M-samples of voltage and current signals. Table 1 shows input signals for each measurement impedance elements.

3. Training methodology The ANN-X and ANN-R training sets were prepared considering three different alternatives, as follows: (1) Complete three-phase system model: implemented using EMTP or the Power System Toolbox Matlab. (2) Second order single-phase model: implemented using the transmission line positive sequence parameters. (3) First order single-phase model: implemented using the transmission line positive sequence parameters. At rst sight the Alternative 1 should provide more precise results once the signals utilized at the ANN training come from a distributed parameters three-phase line model and then being closer to a real electrical system. However, a large amount of data is necessary to the train both ANNs to properly operate according to all different transmission lines (length, voltage level, tower geometry, etc.). This approach is rather complex and demands a great amount of time to generate all training sets. Alternative 2 is less complex when compared to the rst one, but also represents some of the high frequency fault transients.

Fig. 7. First order single-phase system.

Alternative 3 is simpler than the other two. It does not represent the high frequency fault transients, but represents the exponential component existing in the current signals. There was no success while training the ANN using alternatives 1 and 2 because the errors found at the end of the processes were extremely high and so the trained network have not presented sufce generalization capability. However for the third alternative, the training process presented fast convergence to the maximum specied error and excellent performance on generalization tests. The Alternative 3 was chosen then due to the results obtained at the training process. The third alternative simplies the ANNs training process, once it avoids the complete modeling of a three-phase system to generate the training sets. On the other hand, this alternative does not represent the high frequency fault components, that are common in long distance high voltage transmission lines. The inuence of this restriction in ANNs performance at a real three-phase system is discussed at the Section 4. The rst order single-phase transmission system, used to generate the ANN training signals, is illustrated in Fig. 7. This system is composed by a voltage at bus (A) feeding a load through a transmission line which positive sequence series impedance is represented by a resistance and an inductance. The fault condition is simulated by a switch (K) closing at time t = 0. The estimate of resistance (R1) and reactance (X1) between the relay and the fault point is provided by two distinguished ANNs known respectively as ANN-R and ANN-X, as previously mentioned in Section 2.2. Such estimate shall present enough precision for steady state line operation as well as for fault occurrences. Therefore it is necessary to generate a training set for each one of those scenarios considering also both ANNs. Each training set is composed by the input data vectors described at Section 2.1, and also by their respective ANNs outputs (R1 or X1), generated as described at Section 2.3. The samples that set up the input vectors are obtained from the equations provided by the model depicted at Fig. 7 operating in steady state condition (19) and (20) and also during fault occurrence (21) and (22), considering the fault current much bigger than the pre-fault one.

726

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730

The voltages and currents at the relay during the steady state operation condition:

3.1. Measured steady state impedance

v V cosw t h
i Iload cosw t h /

19 20

Z relay R j X

V cos/ j sen/ Iload

32

The voltages and currents at the relay during fault occurrence:

3.2. Measured fault impedance

v V cosw t h
i Ifault cosh b eat cosw t h b

21 22 Z relay R j X V cosb j senb Ifault 33

where / is the power factor for load and line; h the angle of voltage at the instant of fault occurrence; a the time constant inverse of the rst order system.

b arctg

x:L1
R
1

! 23

As presented at (34) and (35), it is necessary to normalize the impedance estimated by the ANN once the voltage and current vectors are also normalized.

Z relay R j X Z base V base Ibase

Z relay Z base

34 35

The voltage and current values provided by (19)(22) are acquired using N samples per fundamental cycle and grouped at data windows containing M elements. Those windows are normalized using suitable base values shown as follows. It can be seen below the non-normalized data windows utilized to compose the training vector set during steady state operation condition:

The normalized output values of the ANN-R and ANN-X training vectors are provided by (36)(39). Again, it is possible to adopt unitary values for Iload, Ifault and V. It is presented below the normalized output value to settle the training vector of ANN-R: 3.3. Steady state condition

fv k g V cos

  2p kh N   2p kh/ fik g Iload cos N

24 25

V cos/ Iload Z base

36

These are the non-normalized data windows utilized to compose the training vector set during fault condition:

3.4. Fault condition

  2p fv k g V cos kh N fik g Ifault

26 27 R V cosb Ifault Z base 37

 ! 2p T cosh b:ea:kN cos khb N

where k = 0, . . . , M 1 and T is the period of the fundamental frequency. The normalized data windows constitute the ANNs input training vectors and are resultant from the application of (28) and (29) into (24)(27).

It is presented below the normalized output value to settle the training vector of ANN-X: 3.5. Steady state condition

 fv k g

fv k g V base fik g fk g i Ibase

X 28 29

V sin/ Iload Z base

38

3.6. Fault condition

where

V base maxfjv k jg Ibase maxfjik jg

30 31

V sinb Ifault Z base

39

As it can be seen the normalization process brings out some advantage. It is not necessary anymore to train the ANN for each and every line transmission operating voltage and current levels. Using the proposed process, some voltage signal presenting amplitude equal to V in (24) and (26) generates the same normalized training vectors as a voltage signal with amplitude equal to 2 V, given that the latter Vbase worths twice the value of the former one. This shows that the presented method does not need to rely on the voltage and current amplitudes allowing to adopt unitary values, i.e.: V = 1.0 V; Iload = 1.0 A; Ifault = 1.0 A. The correct steady state and fault impedance values measured by a relay that is installed at the beginning of the transmission line (bus A of Fig. 7) are presented as follows:

The procedure used to generate the training sets for the ANNs consisted of implementing a software routine in Matlab programming language, which evaluates (24)(39), considering different values of load power factor (/), relationship between the positive sequence reactance and resistance values of the transmission line (b) and fault instant (h). The voltage and current input data windows created after each iteration allow the calculation of resistance and reactance values for each simulated scenario. It is important to point out that /, b and h values are randomly chosen considering a range of plausible values, avoiding unrealistic situations. The value ranges for those quantities are shown in Table 2. The values of b and a are interdependent, therefore it is necessary to dene just one of those quantities. One shall notice that the range for b considers the condition observed at all existing volt-

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730 Table 2 Range for quantities /, b and h. Operation condition Steady state Fault Quantities h / h b Range () 0360 0360 0360 4588

727

 Desired error (minimum error): 1 109.  Maximum number of iterations: 500,000.  Activation function: Hyperbolic tangent Tansig (all neurons of neural network). The ANNs with different input data windows sizes and different sampling rates were evaluated to dene the ANN-R and ANN-X architectures during the training stage. After detailed analysis it was concluded that ANNs that use sampling rate equal to 32 samples per cycle of 60 Hz and data windows of 16 samples, i.e., N = 32 and M = 16, are more appropriate to implement the proposed algorithm. Both networks have four layers with 32 input neurons, 65 and 8 neurons at two hidden layers and 1 neuron at output layer. During the training process the ANN-R reached an error of 6.42 105 and the ANN-X reached an error of 2.95 105. In this work only the MLP type of ANN was considered. The evaluation of other ANN types such as RBF is available for future researches. The training sets of each ANN are formed by 4500 input/output arrays, which means that each array of (40) and (41) has 4500 columns and 32 rows. Each column represents a different training scenario and the 32 rows represents the 16 voltage and current samples.

age levels for transmission lines. The utilization of the normalizing process and the value range for b allow the ANN to be trained in order to respond correctly to all possible voltage levels for transmission lines. The training sets are created after a pre-determined number of rounds and consist of input and output arrays as illustrated in (40) and (41). The arrays represent the ANN-R and ANN-X training sets. These training sets consist of input data windows containing Msamples of voltage and current signals and output arrays containing resistance and reactance values, R and X , respectively. There is one value of R or X associated to each input data array, i.e., the input data array P is associated to values of RP and XP, while input data array S is associated to values of RS and XS. All values are normalized.

6 . 6 . . 6 6 P  6 v kM1 6 6 P 6 ik 6 . 6 . 4 . P ikM1 2 k vP 6 . 6 . . 6 6 P  6 v kM1 6 6 P 6 ik 6 . 6 . 4 . P ikM1

k vP

7 7 7 S  7 v kM1 7 P S 7 S 7 ) R R ik 7 7 . 7 . 5 . S ikM1 . . . 7 7 7 7  kM1 7 vS P S 7 S 7 ) X X ik 7 7 . 7 . 5 . S ikM1 . . . k vS 3

k vS

4. Results of the proposed algorithm

40

41

The performance of the proposed algorithm is discussed in this section. Such algorithm was submitted to signals generated by the Matlab Power System Toolbox. The considered Matlab model is based on existing voltage levels applied to three-phase transmission lines congured as distributed parameters. It was possible to verify the performance of both ANNs during fault conditions that result in traveling waves, which were not considered in the ANN training process, since the utilized model (rst order single-phase model) does not represent such phenomenon. The assessment considers the response time of the algorithm, i.e., the trip time (Ttrip) and the settling time (Tset), as follows:  Ttrip: time for the algorithm to generate three consecutive responses inside the rst protection zone of the relay (Mho) adjusted to 85% of the transmission line length;  Tset: time for the algorithm to provide continuous responses considering an error smaller than 5% related to correct value of the fault impedance. Fig. 8 illustrates the trip time and settling time concepts. The simulated system is composed by a transmission line and an equiv-

The ANNs used are MLP and use backpropagation supervised training, [1315]. Such option is typical when applying ANN to power system studies [4]. The training parameters used in this work are:  Momentum: 0.9.  Learning rate: 0.01.

Fig. 8. Trip and settling times (Ttrip and Tset) of the implemented algorithm.

728

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730 Table 5 Medium time for settling and tripping in milliseconds. Voltage (kV) Ttrip (ms) Medium 138 230 345 500 765 Mean time 7.31 8.57 10.56 6.84 7.05 8.06 Max 16.05 12.64 15.44 11.44 12.66 Min 4.50 4.45 5.62 3.43 3.44 Tset (ms) Medium 12.70 16.12 17.59 38.76 37.91 24.61 Max 19.62 29.62 31.71 45.30 65.47 Min 5.00 9.33 12.78 12.50 10.52

Fig. 9. Simulated system.

Table 3 Short-circuit power for each voltage level. Voltage (kV) Terminal 1 Scc1/ (GVA) 138 230 345 500 765 1.41 5.86 1.37 12.59 12.99 Scc3/ (GVA) 2.06 4.85 1.09 15.90 20.03 Terminal 2 Scc1/ (GVA) 4.69 1.79 2.18 15.24 18.67 Scc3/ (GVA) 3.92 1.98 1.87 14.93 22.25

Table 4 Lines parameters. Parameter R0 (X/km) R1 (X/km) L0 (mH/km) L1 (mH/km) C0 (nF/km) C1 (nF/km) 138 kV 0.550907 0.219275 4.495941 1.313819 5.540583 8.847506 230 kV 0.464812 0.115898 4.239230 1.347695 6.105862 8.642304 345 kV 0.400583 0.029559 3.910398 1.247315 6.455623 9.395225 500 kV 0.382196 0.012106 3.204005 0.830045 7.722758 13.88204 765 kV 0.360164 0.012163 3.009576 0.878713 8.875915 13.35893

Fig. 10. Behavior of impedance measurement elements: AB fault.

alent attached to each terminal as shown in Fig. 9. The terminals short-circuit powers for each voltage level are presented in Table 3. The considered lines are typical lines of the Brazilian system at voltage levels of 138 kV, 230 kV, 345 kV, 500 kV and 765 kV, with compatible length and load for those levels. The operating frequency is 60 Hz. The lines characteristics and parameters are shown in Table 4. Each one of those lines were simulated using three different lengths (138 kV: 50/74/150 km; 230 kV: 80/100/ 200 km; 345 kV: 100/145/300 km; 500 kV: 181/300/400 km; 765 kV: 265/350/500 km). The diverse electrical network operating situations were generated considering the following variations: fault type (phase-toground, phase-to-phase, double-phase-to-ground and threephase); fault distance (20%, 50% and 80% of the total transmission line length); fault resistance (0 X, 10 X and 20 X); equivalent impedance at the relay terminal (30%, 100% and 120% of the real value); voltage angle at fault instant (0, 45and 90). Amongst the many possible combinations of those ve parameters only the 32 most representative operating situations were selected. Those 32 operating situations were simulated using Matlab Power System Toolbox at 5 considered voltage levels and for 3 line lengths for each voltage level, resulting in 480 different simulation scenarios. The variations at transmission line length can provide answers to the behavior of the proposed algorithm for short, medium and long lines. The algorithm performance is evaluated by the indexes Ttrip and Tset for each simulated case. Table 5 presents the medium, maximum and minimum times observed for those two indexes in each voltage level. As it can be observed the medium time for the algorithm tripping is about half of a cycle. Regarding the Tset, it is possible to notice that its value increases as the line voltage level rises. This can be explained by the inuence of the traveling waves which present higher amplitudes for higher voltage levels. It shall

Fig. 11. Behavior of impedance measurement elements: AN fault.

be emphasized that such high frequency phenomenon is responsible for decreasing the algorithm precision once the ANN training was based on a rst order model. In order to present some results, there are three typical cases to be shown in detail. Figs. 1012 present the impedance at RX plane calculated by the measuring elements of proposed algorithm for faults at a 345 kV and 145 km transmission line. The fault occurs in half distance of transmission line and the voltage angle is 45 in all three presented cases. In the rst case (Fig. 10), Ttrip and Tset signals were generated, respectively, after 12.0 and 14.0 ms, for a

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730 Table 7 Medium time for settling in milliseconds. Level 138 kV Algorithm ANN Fourier % of cases 50 km 12.28 17.97 94 80 km 230 kV ANN Fourier % of cases 17.28 17.22 72 100 km 345 kV ANN Fourier % of cases 17.15 19.50 75 181 km 500 kV ANN Fourier % of cases 36.85 16.93 41 265 km 765 kV ANN Fourier % of cases 36.70 17.65 41 74 km 13.65 18.75 97 100 km 17.04 17.18 72 145 km 17.57 18.30 69 300 km 35.03 17.00 47 350 km 35.54 17.59 38

729

150 km 12.17 18.12 100 200 km 14.05 17.59 88 300 km 18.46 18.53 72 400 km 44.42 16.93 38 500 km 41.48 20.53 44

Fig. 12. Behavior of impedance measurement elements: ABC fault.

solid fault between phases A and B. In the second case (Fig. 11), Ttrip and Tset signals were generated, respectively, after 9.5 and 13.5 ms, for a single-phase fault (phase A) and a fault resistance equal to 10 X. In the third case (Fig. 12), Ttrip and Tset signals were generated, respectively, after 8.0 and 13.5 ms, for a solid three-phase fault. These facts conrm that the proposed system (using pre-processing, ANN and post-processing) performed as expected, once in all cases the measuring elements calculated the fault impedance with precision of less than 5%. The performance of the proposed ANN-based algorithm is now compared to conventional method based on the DFT, in order to measure its applicability. It is important to point out that the Fourier-based method implemented for comparison, uses a digital lter to reduce the inuence of the exponential component, which considerably improves its performance. It shall be observed that half-cycle Fourier algorithm was not considered in this analysis because it presents an unsatisfactory behavior when facing high frequency oscillations and exponential component, both normally included into the voltage and current signals during fault. Although such algorithm has the advantage of using a smaller window data (half of a cycle), the above menTable 6 Medium time for tripping in milliseconds. Level 138 kV Algorithm ANN Fourier % of cases 50 km 06.88 14.20 97 80 km 230 kV ANN Fourier % of cases 09.00 13.30 93 100 km 345 kV ANN Fourier % of cases 11.40 16.70 100 181 km 500 kV ANN Fourier % of cases 07.60 13.06 97 265 km 765 kV ANN Fourier % of cases 07.01 13.46 100 74 km 08.16 15.63 100 100 km 08.83 13.03 93 145 km 10.55 16.20 100 300 km 06.77 12.43 97 350 km 06.75 13.10 97 150 km 06.90 13.70 100 200 km 07.90 12.67 87 300 km 09.75 15.40 100 400 km 06.17 12.56 97 500 km 07.40 12.93 87

tioned conditions make its performance to be overcome by the one-cycle Fourier algorithm, keeping that out of commercial relays. Tables 6 and 7 present both the tripping and settling average time of the algorithms in the following conditions: 32 cases simulated for each line length for each voltage level. The last line of each voltage level presents the percentage of cases where the ANNbased algorithm is faster than the Fourier. As it can be noticed in Table 6, the ANN-based algorithm presents better performance than the conventional Fourier algorithm for tripping times once the developed algorithm presents signicantly lower medium times for tripping no matter the transmission line length. As it can be noticed in Table 7, the ANN-based algorithm presents better performance than the conventional Fourier algorithm for transmission lines with voltage levels up to 345 kV. Nevertheless, the Fourier algorithm is more adequate than the ANN-based algorithm for cases where the voltage level of transmission lines is greater than 345 kV. This fact occurs because the traveling waves effects are strongly present in those lines and the training of ANN-R and ANN-X was made based upon a rst order model (Fig. 7) which are not appropriate to represent such phenomenon, as mentioned before. Another important point to analyze is the impact of the fault resistance to the precision of the measured impedance. It is known that the fault resistance affects not only the resistance but also the reactance both measured by the relay. The effects of the fault resistance on the reactance depends on the line pre-fault load. The inuence of fault resistance can be observed in Table 8 which presents the results of the simulations made for the system in Fig. 9 with a 230 kV line and length of 80 km. A line-to-ground AN fault was applied at the central point of this line (40 km from the relay)

Table 8 Impedance measured by relay. Load (MVA) Fault resistance 0X 11.4 j4.8 76.4 j6.8 ANN Fourier ANN Fourier 4.64 + j20.36 4.65 + j20.34 4.39 + j20.27 4.65 + j20.34 20 X 21.35 + j19.83 23.93 + j20.51 19.78 + j18.11 22.37 + j18.73

730

Ricardo Caneloi dos Santos, E.C. Senger / Electrical Power and Energy Systems 33 (2011) 721730 [2] Kezunovic M, Rikalo I. Detect and classify faults using neural nets. IEEE Comput Appl Power 1996:427. [3] Vzquez E, Altuve H, Chacn O. Neural network approach to fault detection in electric power systems. In: IEEE international conference on neural networks, vol. 4; June 36, 1996. p. 20905. [4] Jiali H, Yuqian D, Yongli L, Gang W, Shanshan L. Distance relay protection based on ANN. In: Proceedings of the 4th international conference on advances in power system control, operation and management (APSCOM97), Hong Kong; November 1114, 1997. p. 51520. [5] Oleskovicz M, Coury D, Carvalho A. Articial neural network applied to power system protection. In: 5th Brazilian symposium on neural networks (SBRN98), Belo Horizonte, Brazil; December 911, 1998. [6] Li Y, Li B, Zhang X, He J. An ANN-based distance protective relays of transmission lines. IEE Develop Power Syst Protect 2001;479:3114. [7] Vasilic S, Kezunovic M. An improved neural network algorithm for classifying the transmission line faults. In: IEEE Power Engineering Society winter meeting, vol. 2; January 2731, 2002. p. 91823. [8] Salat R, Osowski S. Fault location in transmission line using self-organizing neural network. In: 5th international conference on signal processing (ICSP00), vol. 2, Beijing, China; August 2125, 2000. p. 15858. [9] Tawk M, Morcos M. ANN-based techniques for estimating fault location on transmission lines using Prony method. IEEE Trans Power Delivery 2001;16(2):21924. [10] Bouthiba T. Fault location in EHV transmission lines using articial neural networks. Int J Appl Math Comput Sci 2004;14(1):6978. [11] Horowitz SH, Phadke AG. Power system relaying. 2nd ed. England: Research Studies Press; 1996. [12] Blackburn JL. Protective relaying principles and applications. 3rd ed. New York: Marcel Dekker; 1998. [13] Wasserman PD. Neural computing theory and practice. New York: Van Nostrand Reinhold; 1989. [14] Demuth H, Beale M. Neural network toolbox for use with Matlab. 3rd ed. Berlin (Germany): The Math Works; 1994. [15] Beale R, Jackson T. Neural computing: an introduction. 2nd ed. Great Britain: JW Arrow Smith Ltd.; 2001. [16] Garcia J, Mazn AJ, Zamora I. Best ANN structures for fault location in single and double circuit transmission lines. IEEE Trans Power Delivery 2005;20(4):238995. [17] Samantaray SR, Dash PK, Upadhyay SK. Distance relaying for transmission line using support vector machine and radial basis function neural network. Int J Electr Power Energy Syst 2009;31(4):16772. [18] Samantaray SR, Dash PK, Panda G. Distance relaying for transmission line using support vector machine and radial basis function neural network. Int J Electr Power Energy Syst 2007;29(7):5516. [19] Misra RK, Singh SP. Efcient ANN method for post-contingency status evaluation. Int J Electr Power Energy Syst 2010;32(1):5462.

considering two pre-fault load situations (light and heavy load) and two different fault resistances (0 and 20 X). The correct parameters for positive sequence impedance for the 40 km section of such line are: R1 = 4.63 X and X1 = 20.33 X. Table 8 also presents the impedance that should be measured by a relay based on the traditional Fourier method serving as base for comparison. The results presented in such table show that the impact of the fault resistance at the proposed solution precision (ANN based) is quite similar to that observed at the distance relay implemented using the Fourier method. 5. Conclusion The main contribution of the proposed algorithm is the possibility of its application to any transmission line, no matter the line conguration, with no need for additional adjustments at ANN parameters nor its re-training. The pre-processing and post-processing steps make possible to the developed algorithm to be independent from the line voltage level and allow the ANN to be trained by a rst order single-phase model which simplies and optimize the generation of training sets. It is important to notice that during the tests the ANNs responses were not affected by the exponential component, thereby the performance was maintained. Regarding the response times, the proposed algorithm present advantages over the conventional Fourier algorithm when the tripping time is discussed for any kind of transmission line and the settling time is discussed for transmission lines until 345 kV level. However, the conventional Fourier algorithm presents better performance than the ANN-based algorithm for 500 kV and 765 kV lines, with regards to the settling times. References
[1] Dalstein T, Friedrich T, Kulicke B, Sobajic D. Multi neural network based fault area estimation for high speed protective relaying. IEEE Trans Power Delivery 1996;11(2):7407.