You are on page 1of 4

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 16, NO.

2, APRIL 2001

215

A FIRANN as a Differential Relay for Three Phase Power Transformer Protection


ngel L. Orille-Fernndez, Member, IEEE, Nabil Khalil I. Ghonaim, and Jaime A. Valencia
AbstractThis paper presents an application of a Finite Impulse Response Artificial Neural Network (FIRANN) as a differential protection for a real three phase power transformer. Three FIRANNs are designed, trained and tested. The first one has an output, which identify internal faults from any other cases like inrush current and external faults. The two others FIRANNs, each have two outputs that classify between internal and external faults, so that, a backup protection is included. These FIRANNs have six inputs, one for each sampled current signal from both transformer sides. The sample rate selected is 2 kHz for a 50 Hz power frequency. All FIRANNs were trained to have a 3.5 ms fault detection time, which is considered as a very fast protection. The test results show very good behavior of the FIRANN as a differential protection and it is planned to build a prototype. Index TermsArtificial neural network, digital relays, transformer protection.

Fig. 1. Three-phase transformer circuit scheme.

I. INTRODUCTION HE FIRST use of artificial neural networks (ANN) applied on power transformer differential protection was in 1994 [1]. Since that time there are really few papers suggested the application of ANN on power transformer protection and there is a lot of work to do in this subject. The uses of ANN on transmission line protection have been given more consideration. There are different ways how ANN can be applied to a differential transformer protection. The first investigations suggested the use of ANN as an inrush current identifier to be included as a part of a differential protection. In [1], it were proposed to use a Time Delay Artificial Neural Network (TDANN) to process the normalized sampled current signals. Another group [2], suggested the use of the DFT (discrete Fourier transform) to filter the fundamental component and the harmonics from the second order through fifth order of the current signals and apply them to a Multilayer Feed Forward Artificial Neural Network (MFANN). A recent paper kept the same idea and suggested an additional MFANN to reconstruct the distorted secondary current signal of the current transformer due to saturation in order to improve operation of the power transformer protection [3]. Later papers proposed MFANN as a transformer differential protection, one suggested to use harmonic ratios as inputs [4]; while the other used the negative sequence component of current signals and the voltage signals as inputs of the neural network [5].
Manuscript received February 7, 2000. . L. Orille-Fernndez and N. K. I. Ghonaim are with the Electrical Engineering Department, Polytechnic University of Catalonia, Barcelona, Spain. J. A. Valencia is with the Electrical Engineering Department, University of Antioquia, Colombia. Publisher Item Identifier S 0885-8977(01)03412-4.

The behavior of FIRANN (Finite Impulse Response Artificial Neural Network) applied to a transformer differential protection is presented in this paper. This neural network is well known for its ability to manage time variable signals [6]. Three different FIRANNs were trained as differential protection for a threephase power transformer. The first one has an output to classify the internal faults from any other case. The other two FIRANN have two outputs to classify between internal and external faults. This, in fact, is a novel approach to differential protection that improves selectivity of the proposed relay and includes a backup protection for the other elements, which are connected directly to the power transformer. The inputs of the FIRANNs are the normalized sampled current signals from both sides of the transformer. It means that the FIRANNs acts as a signal processor based relay. The FIRANNs structure, training method, training strategy and test results will be reported.

II. SIMULATED SYSTEM The system used in this work to generate the training patterns of the FIRANN was a 15 kVA, 220 V/1300 V Yy solidly grounded power transformer. It has six taps on every phase winding of the high voltage side. The shortest segment between two consecutive taps is 6% and the longest is 26% of the total coil turns. The method used to simulate internal faults is explained in [7]. The Alternative Transient Program (ATP) was used to simulate all cases needed in this work. The simulation model was validated using real laboratory measurements. The scheme of the system is shown in Fig. 1. The transformer is modeled as coupled branches to simulate its coils wind, ings. Iron losses are simulated as three resistive branch , were added to conThree nonlinear induction branches, sider the saturation effect for inrush simulation cases. The sampled current signals are taken from measurement switch type, so there was not included any current transformer model. The

08858977/01$10.00 2001 IEEE

Authorized licensed use limited to: Iran Univ of Science and Tech. Downloaded on May 9, 2009 at 09:29 from IEEE Xplore. Restrictions apply.

216

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 16, NO. 2, APRIL 2001

TABLE I THREE-PHASE TRANSFORMER SIMULATIONS

current sensors used on laboratory measurement were Hall effect elements, which have lineal characteristic. The sampling frequency used is 2 kHz. III. SIMULATED FAULT CASES One of the most important stages during the designing of ANN is the selection of the simulated cases that must be included in the training and testing sets. The transformer model was simulated to cover all possible operating and fault conditions such as the source short circuit level, fault inception instants, fault types and inrush current. The simulated cases are divided into three groups. The first is the training group and its patterns are selected randomly and normally distributed in order to make the FIRANN to generalize and to prevent skew learning. The second group is used to validate the FIRANN during the training process and the last one is the testing group. Table I summarizes the cases simulated with the system explained above. Each row specifies a simulated case and each column designates a different parameter for each case. The IR set includes inrush cases; 8 different switching instants and 27 initial conditions are combined in order to have 216 cases. The EB group has the external fault cases on source side; time instants, loads and type of faults are combined. The EA row includes external fault cases on load side with equal parameter combination as in the previous group. Internal faults are classified in two sets, the IT set that has turns to earth faults and the IE set which has turn to turn faults. All the internal faults are combined with the six taps in each high level side winding. IV. THE ARTIFICIAL NEURAL NETWORK The ANNs used in previous works were the MFANN and TDANN. A neuron model based on Finite Impulse Response (FIR) filter theory is shown in Fig. 2. An ANN that has neurons of this configuration is called the Finite Impulse Response Artificial Neural Network (FIRANN). This type of ANNs is suggested in [6] as good structure for temporal signal processing. The selection of the number of hidden layers, the neuron number in every layer and the time delay on each neuron has been done by trial and error method. The structure of a certain FIRANN depends on the protection functions that were tried to include in it. The number of total time delay is selected to have a detection time of 3.5 msec. A. One Output FIRANN The structure of this FIRANN is shown in Fig. 3. There are 4 layers, the input layer has 6 inputs. The first hidden layer has 6 neurons while the second hidden layer has 4 neurons. All

Fig. 2. Neuron model of a FIRANN.

Fig. 3. FIRANN structure with single output.

Fig. 4.

FIRANN structure with 2 output (FIRANN-1 and FIRANN-2).

neurons have 2 time delay units. There is one output that has been trained to be 1 in case of internal faults and 1 in any other case (i.e., inrush current cases, external faults and healthy cases). B. Two Output FIRANN There are two FIRANNs, FIRANN-1 and FIRANN-2, both have 2 outputs. Both have the same architecture, which is shown on Fig. 4. There are 4 layers, the inputs are the same as in the first case. Each one of the two hidden layers have 8 neurons and 2 time delay units per neuron. These two FIRANNs has two outputs labeled internal fault and external fault. The difference between the two networks

Authorized licensed use limited to: Iran Univ of Science and Tech. Downloaded on May 9, 2009 at 09:29 from IEEE Xplore. Restrictions apply.

ORILLE et al.: A FIRANN AS A DIFFERENTIAL RELAY FOR THREE PHASE POWER TRANSFORMER PROTECTION

217

TABLE II OUTPUT DESIGN OF FIRANN-1

TABLE III OUTPUT DESIGN OF FIRANN-2 Fig. 5. Statistic results of the test.

current almost in phase and with the same magnitude. This will make the FIRANN to generalize for different windings connections, sizes and ratings. The time consumed in the training cycle of these FIRANNs was about 26 hours in a PC Pentium II. Number of iterations needed to get good results was in general less than 1500. only exists in the designing of the outputs response. The first output works as internal faults detector in both networks while the second output detects external faults in the forward direction only in the case of the FIRANN-1. In the case of the FIRANN-2 the second output detects external faults in both forward and backward directions. This second output is very useful to increase the selectivity of the relay and make it works as backup protection for the elements connected to the transformer such as the transmission lines and generators. The training patterns used to train both FIRANN-1 and FIRANN-2 was included all types of internal and external shunt faults except in the case of three-phase fault on source side. In this case transformer currents go to zero which looks like disconnecting the transformer and/or the source. Tables II and III show output design for FIRANN-1 and FIRANN-2. V. TRAINING METHOD The training method used to train the FIRANN is called temporal back-propagation [2]. A program based on temporal backpropagation was developed by authors to train the proposed networks. This program was developed in C language and runs under the Linux operative system. The patterns used to train each FIRANN 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 to generalize. The one output FIRANN needed 882 simulated fault patterns, the FIRANN-1 needed 630 and the FIRANN-2 used 1260 fault patterns. Each simulated case consists of 87 samples, 40 samples before fault inception and 47 samples after fault inception. The ideal value of the FIRANN output goes from 1 to 1 in 3.5 msec exponentially to avoid false detection of faults during heavy transients and load swing. One of the training targets is to search about the minimum training set, which is sufficient to make the FIRANN to generalize. Therefore, we used less than 4% of all simulated cases as training data showing the great ability of this FIRANN to generalize. The inputs of the FIRANN must be normalized and adjusted to have the waveforms of the transformer primary and secondary VI. TEST AND RESULTS The test of the FIRANNs is done using a set of patterns, which completely differs from the training ones. Fig. 5 summarizes test statistics. On training stage the FIRANN was tested periodically using about 10% of the total simulated cases including fault and nonfault events as a validation patterns. When it was considered that training was over, a sample of 50% of total simulated cases was chosen for each FIRANN in order to test its behavior. Those samples included all kind of event and did not include cases used on the training file. We consider a tested case as Good when the FIRANN actual response is equal or faster than ideal one. The Short Retard SR means a delay less than 2 ms and Long Retard LR means a delay greater than 2 ms and less 10 ms. The Bad means a delay more than 10 ms. Note that the SR and LR are caused by light fault currents and the Bad is caused by very small fault currents. The FIRANNs show Good behavior on more than 95% of the tested cases. The SR and LR classifications give more than 4.6%. Leaving BAD classification less than 0.4%. Also it can be noticed that FIRANN-2 has less Good cases and greater retard cases than the others do. This difference is caused by the source side external faults, which make the transformer current to be less than the nominal values in both sides. In order to overcome this problem, the training patterns that correspond to this case must be increased and the FIRANN-2 must be trained again. Some test examples are shown in Figs. 68. The actual network answer and input signals are displayed in the figures. Fig. 6 shows a typical behavior of the single output FIRANN. The diagram on upper left corner shows the actual answer of the FIRANN with marks. The other three diagrams show the three-phase input current signals to the FIRANN. Figs. 7 and 8 show the novel approach to a differential protection. The first figure case shows the response of the single output FIRANN-1 for an external fault while the second one shows the FIRANN-2 response for an internal fault. The IF and EF denote the internal fault and external fault outputs respectively.

Authorized licensed use limited to: Iran Univ of Science and Tech. Downloaded on May 9, 2009 at 09:29 from IEEE Xplore. Restrictions apply.

218

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 16, NO. 2, APRIL 2001

connection types of power transformers by choosing a suitable current normalization values in order to match the primary and secondary current magnitudes and to reduce the current transformer errors. In addition a suitable method must be applied to eliminate the phase shift between the primary and the secondary currents in order to make the relay general for all three phase transformer connections. A prototype of this FIRANNs will be created using a nonexpensive Digital Signal Processing (DSP) card. Finally it is clear that these FIRANN had verified the four protection bases, i.e., reliability, selectivity, speed and economy.
Fig. 6. Single output FIRANN response.

VIII. LIST SYMBOLS ANN FIRANN MFANN TDANN Artificial Neural Network Finite Impulse Response Artificial Neural Network Feed Forward Artificial Neural Network Time Delay Feed Forward Artificial Neural Network REFERENCES
[1] L. G. Perez, A. J. Flechsig, J. L. Meador, and Z. Obradovic, Training an artificial neural network to discriminate between magnetizing inrush and internal faults, IEEE Trans. on Power Delivery, vol. 9, no. 1, pp. 434441, Jan. 1994. [2] M. Nagpal, M. S. Sachdev, K. Ning, and L. M. Wedephol, Using a neural network for transformer protection, in Proc. of the International Conference on Energy, Management and Power Delivery, vol. 2, Singapore, Nov. 2123, 1995, pp. 674679. [3] P. Pihler, B. Grcar, and D. Dolinar, Improved operation of power transformer protection using artificial neural network, IEEE Trans. on Power Delivery, vol. 12, no. 3, pp. 11281136, July 1997. [4] P. Bastard, M. Meunier, and H. Regal, Neural network-based algorithm for power transformer differential relays, IIE Proc.Gener. Transm. Distrib., vol. 142, no. 4, pp. 386392, July 1995. [5] A. L. Orille, N. Khalil, and J. A. Valencia, A transformer differential protection based on finite impulse response artificial neural network, in 24th International Conf. on Computers & Industrial Engineering. Uxbridge, UK: Brunel University, Sept. 911, 1998. [6] P. Bastard, P. Bertrand, and M. Meunier, A transformer model for winding fault studies, IEEE Trans. on Power Delivery, vol. 9, no. 2, pp. 690699, Apr. 1994. [7] S. Haykin, NEURAL NETWORKS A Comprehensive Foundation: Macmillan College Publishing Company, 1994, p. 510.

Fig. 7.

The FIRANN-1 response for a load side external fault.

Fig. 8. The FIRANN-2 response for internal fault.

VII. CONCLUSION A new approach of power transformer differential protection is proposed in this paper. The three FIRANNs performance is very good. Their response to a fault case takes less than 3.5 ms on more than 95% of the tested cases and less than 10 ms in 4.5% of the tested cases. The prolonged delay is due to very weak faults, which is a logic behavior of the FIRANN and could be improved by increasing the FIRANN size. The response time of the proposed relays is inverse proportional to the fault strength which makes them insensitive to load swings and heavy transients. This demonstrates that the relays are fast and reliable in general and more selective in the cases of FIRANN-1 and FIRANN-2 which can be used to backup other relays. The proposed relays can be generalized for all

ngel L. Orille-Ferndez received the Dr. Ing. degree in electrical engineering from Polytechnic University of Catalonia, Spain, in 1988. Since 1989, he is Full Professor of the Department of Electrical Engineering of Polytechnic University of Catalonia and Head of this department since 1995.

Nabil Khalil I. Ghonaim was born in Cairo, Egypt on October 1965. He received the B.Sc. degree with distinction first class honors and the M.Sc. degree in electrical engineering from Helwan University, Cairo, Egypt, in 1987 and 1993, respectively. From 1994 to the present, he has started his studies toward the Ph.D. degree from Polytechnic University of Catalonia, Barcelona, Spain in the field of power system protection. His areas of interest are power system simulation, planning and protection, signal processing and neural networks.

Jaime A. Valencia was born in Medellin Colombia. Received the electrical engineer degree in 1982 from National University of Colombia and the M.Sc. degree on mathematics in 1988. He has been Teacher Assistance at Antioquia University since 1990 and a Ph.D. student at Polytechnic University of Catalonia, Spain since 1995.

Authorized licensed use limited to: Iran Univ of Science and Tech. Downloaded on May 9, 2009 at 09:29 from IEEE Xplore. Restrictions apply.

You might also like