oornputl:ers
i~ial ongineorlng PERGAMON
Computers & Industrial Engineering 37 (1999) 399-402
A TRANSFORMER DIFFERENTIAL PROTECTION BASED ON FINITE IMPULSE RESPONSE ARTIFICIAL NEURAL NETWORK A. L. ORILLE
NABIL KHALIL
J.A. VALENCIA V.
Professor Electrical Engineering Department Polytechnic University of Catulonia
Ph. D. Student Electrical Engineering department Polytechnic University of Catalonia
Ph.D. Student-UPC Electrical Engineering department University of Antioquia - Colombia
ABSTRACT This paper presents the application of a finite impulse response artificial neural network (F1RANN) on digital differential protection design for a three-phase transformer. The neural network inputs are normalized sampled current dsta~ Any pre-processing signal as in other neural network applications is not needed. The network was trained to identify external fault on load side besides internal fault as in the other differential protection. The FIRANN has 6 inputs and 2 outputs. The first output goes on when there is an internal fault while the second output goes on in case of external fault. The simulated system used to get data for training and testing the neural network is presented. The neural network architecture and some of the obtained results are reported. © 1999 Elsevier Science Ltd. All rights reserved. KEYWORDS Protection, differential, transformer, digital, neural, network. INTRODUCTION The first paper published on the subject of Artificial Neural Network (ANN) in power transformer differential protection is Perez et al (1994). Since that time few papers suggesting application of ANN on power transformer protection were written. The use of ANN on transmission line protection has been given more consideration. There are different ways to use ANN's on transformer protection. In early investigations there was tried to use ANN as inrush current identifier to be included as a part of a differential protection. In these papers, one used a timedelayed feed forward neural network (TDFFNN) to process the normalized sampled current signal (Perez et al, 1994), the other one included a DFT (Discrete Fourier Transform) before a feed forward neural network (FFNN) whose inputs were the normalized magnitudes of fundamental and the second through fifth harmonic of current signals (Nagpal e~ al., 1995). A recent paper keeps the same idea of using a ANN as just one part of the differential protection and suggests a FFNN to reconstruct the distorted secondary CT signal due to saturation in order to improve operation of power transformer protection besides inrush current identification (Pihler et al., 1997). Probably, the fLrst paper proposing an ANN as differential protection for transformer, not only as part of it, is from Bastard et al. (1995). The authors suggest a FFNN with four inputs that need some harmonic components. Yongli et al. (1995) presented a one-layer-FFNN with 12 inputs that include negative sequence current and voltage samples. The behavior o f a FIRANN (Finite Impulse Response Artificial Neural Network) applied to a transformer differential protection is presented in this paper. This kind of ANN is well known for its ability to manage time variable signals (Haykin, 1994) and this is the special characteristic that we take advantage of for development of a differential protection based on ANN, without any necessity of pre-processing signals. The first reported work used the FIRANN in power protection is OriUe et al. (1996 a), where FIRANN is applied to transmission line protection. Orille et al. (1996 b, 1997 a) also reported other ways to use FIRANN on transmission line protection. The latest and the first one paper using FIRANN on differential protection of power transformer is Orille et al. (1997 b). In the present work, a FIRANN was trained for differential protection of a three-phase power transformer. The FIRANN has two outputs. One output goes high in case of internal fault, while the other goes high in case of external fault on load side. The latter output is intended as a backup protection. The inputs of these FIRANN's are normalized sampled currents taken from the transformer. The sample rate selected is 2 kHz for a 50 Hz power wave. The FIRANN was trained to have a 3.5 ms tripping time which is considered as a very fast protection. The test results show very well behavior of FIRANN as a differential protection and it is planned to build a prototype. 0360-8352/99 - see front matter © 1999 Elsevier Science Ltd. All rights reserved. PII: S0360-8352(99)00103-5
400
Proceedings of the 24th International Conference on Computers and Industrial Engineering
SIMULATED SYSTEM The system simulated in order to get the pattern to train the FIRANN was a three-phase transformer, 15 kVA, 220 V / 1300 V and Yy0 connection. It has 6-tap derivation on every phase winding of high voltage side. The shortest segment between two consecutive taps is 6% of total turns and the largest is 26% of total turns. The methods used to simulate internal faults are explained in Bastard et al. (1994), and ATP-EMTP was used to compute all cases needed in this work.
Table 1 . Set of simulations
Sources
Time
L C.
Bil3
BH3
LOAD
IR
8
27
1
1
1
1
216
EB
8
I
1
1
27
10
2160
EA
8
I
1
1
27
10
2160
IT
8
I
18
I
27
1
3888
IE
8
I
1
45
27
1
9720
Total
FAULT # SlM
18144
Laid
Fig. 1. Simulated system.
The scheme of the system is shown in Fig. 1. The transformer is modeled as R-L coupled branches to simulate the winding and iron losses. 3 non-linear induction branches include the saturation effect on inrush cases. The sampled current signals are taken from measurement type switch, so there was not included any CT model as the real system in this work. The time step for simulations was 10"s sec and signals were sampled at 2 kHz, which means 40 samples on 50 Hz power frequency.
Fault Cases Simulated Table 1 summarizes the cases simulated with the system explained above. Each row is a simulated set and columns designate the different parameter for every case. The IR sets are inrush cases; 8 different connection instants and 27 initial conditions are combined in order to have 216 cases. EB groups are the external fault cases on source side; time instants, loads and type o f faults are combined. The EA sets are external fault cases on load side and there are equal parameter combinations as in the previous group. Internal faults are classified in two sets, IT sets that have turns to earth faults and IE set which have turn to turn faults. It is always assumed that internal faults are in one phase, so the number of the cases is defined by combining 6 taps derivation in each high level side winding. It is possible to simulate as many cases as you want, but we thought that it was sufficient. It was tried to choose a wide range of cases but we know that it is impossible to include all possible cases of this simple system. THE NEURAL NETWORK The ANN used in a previous work was the FFNN and TDFFNN. The TDFFNN is just a FFNN that has time delay inputs in order to adapt the architecture to manage time variable signals. A FFNN with neuron based on FIR filter like the one shown on Fig. 2. is suggested in Haykin (1994) as a good structure for temporal processing. This kind of ANN, that we call FIRANN, has shown good performance in time series prediction, so we have decided to apply it on power system protection. The hidden layers number, neuron number in every layer and time delay on each neuron has been selected by trial and error method. It does not mean that the FIRANN proposed is optimal. It only means that it performs well as differential protection and it could be improved. The number of total time delays is intended to have a 3.5 ms tripping signal response.
Proceedings of the 24th International Conference on Computers and Industrial Engineering
401
x,
L~Ia Ib
•
•
ylc
°
•
I8
•
•
iib
*
.
•
•
IF EF
X2
x~
xa~2) t
C
*
I .~(3) J----/
Fig.2. Neuron modelofa FIRANN g.3. FIRANNstructure.
F I R / ~ Str.ct.re The scheme of the FIRANN structure used is shown in Fig. 3. There are 4 layers. The first layer has 6 inputs, which are the sampled line current on both sides of transformer. The two hidden layers have 8 neurons each and there are two outputs. Each neuron has two time delay units on every input. Table 2. Programmingoutputs INTglR,'qALFAULTS EXTERNALFAULTS IE
+1
IT EA
+I
-I
-I -1 -1
+1 .1 -1
EB IR
-I
This FIRANN has two outputs that have been labeled IF (internal fault) and EF (external fault). The programming of those outputs is shown on table 2. As labels suggest, each output will turn on in case of internal or external fault respectively. TRAINING METHOD Haykin (1994) suggests two methods to train a FIRANN. The first one is an instantaneous gradient appmacl,., the oth~ one is temporal beck-propagation learning. We decided to use the last one, which is recommended to overcome some problems associated with the fLrStone. A program based on temporal back-propagation was developed by us to train the networks proposed, The pattern to train each ANN is a subset of all simulated cases. This subset must be properly selected in order to include a representative of all cases and let the FIRANN generalize. This task has no rules clearly stated, so try and error method was used to select simulated cases that were piled up to form a pattern file. The FIRANN was trained with 540 cases. The pattern file pile alternate between fault and healthy cases. TEST AND RESULTS A sample of simulated cases was chosen in order to test its behavior. We consider a tested case as Good when the real answer was equal or faster than ideal, Short Retard (SR) when the real answer has a delay less than 2 ms, Long Retard (LR) when the real answer delay is longer than 2 ms but shorter than 10 ms, and any other answer was considered Bad. Finally, 86% of tested cases was Good, 6% was SR, 3% LR and 5% was considered Bad. This means that this FIRANN shows good behavior on 95% of tested cases, considering SR and LR as good answers. This was really the target point in training process. As an example for the FIRANN behavior, Fig. 4 depicts current signals on bottom graphics and outputs signals on top in case of internal fault. Fig. 5 shows an external fault case.
402
Proceedings of the 24th International Conference on Computers and Industrial Engineering IF output
Itow
IFoutl~
EF O~gg
Itush
'I
Fig. 4. Internal fault.
EFoul~t
i ii i/I
Fig. 5. Externalfault.
CONCLUSIONS The FIRANN trained as a differential protection with backup has shown good performance. It is remarkable that with 3% of simulated cases used to train the networks there were 95% good answers on realized tests. It is possible to use all simulated cases to train de network but we consider more important the ability of ANN to 'generalize' because one of the great concern on protection is how the device would perform on cases that were not taken on account. Another important feature is the ability to manage white noise. On test with 5% white noise on every signal, the answer of FIRANN was similar. It is possible to develop other digital protection with this ANN and different kind of differential protections. It is planned to develop a prototype using DSP to test this new algorithms. REFERENCES Bastard, P., P. Bertrand and M. Meunier (1994). A transformer model for winding fault studies. IEEE trans, on power Delivery, Vol. 9, n ° 2, pp. 690-699. Bastard, P., M. Meunier and H. Regal (1995). Neural network-based algorithm for power transformer differential relays. IEE Proc.-Gener. Transm. Distrib., Vol. 142, n ° 4, pp. 386-392. Haykin, S. (1994). NEURAL NETWORKS a comprehensivefoundatior~ Macmillan College Publishing Company. Nagpal, M., M. S. Sachdev, K. Ning and L. M. Wedephol (1995). Using a neural network for transformer protection. Proc. of the international conference on Energy, Management and Power Delivery, Vol. 2, pp. 674-679. Orille, A. L., N. Khalil and J. A. Valencia (1996 a). Using the FIR model of the perceptron for transmission line high speed protection. Seminario anual de automatiea y electronica industrial- Zaragoza, tomo I, pp. 103-108. Orille, A. L., N. Khalil and J. A. Valencia (1996 b). Uso del modeio FIR de1 perceptron pata la determinacion de la direccion del fallo en una proteccion de distancio de alta velocidad. 4a Jornadas MATELEC 96-CIRED de Madrid, sesion 3, pp. 194-203. Orille, A. L., N. Khalil and J. A. Valencia (1997 a). Using the FIR model of the perceptron for pre-fault and postfault current direction estimation for transmission line high speed protection. Fifth International middle east power system conference MEPCON'97-Alexandria, Vol. 2, pp. 597-601. Orille, A. L , N. Khalil and J. A. Valencia (1997 b). Uso de la red neuronal FIR en proteccion diferencial de un transfonv.ador de potencia. 5" Jomadas Hispano-Lusas de Ingenieria Electdea-Salmanea, Tomo I, pp. 127-134. Perez, L. G., A. J. Flechsig, J. L. Meador and Z. Obradovic (1994). Training an artificial neural network to discriminate between magnetizing inrush and internal faults. IEEE trans, on Power Delivery, Vol. 9, n ° 1, pp. 434441. Pihier, P., B. Great and D. Dolinar (1997). Improved operation of power transformer protection using artificial neural network. IEEE trans, on Power Delivery, Vol. 12, n ° 3, pp. 1128-1136. Yongli, L., H. Jiali and D. Yuqian (1995). Application of neural network to microprocessor-based transformer protective relaying. Proc. of Energy Management and Power Delivery International Conference, pp. 680-683.