INDIAN INSTITUTE OF TECHNOLOGY, KHARAGPUR 721302, DECEMBER 27-29, 2002
107
Artificial Neural Network Based Improved Protection Scheme for Power Transformers U.J. Shenoy
K. Parthasarathy
H.P. Khincha
Abstract--This paper suggests the solution for possible improvement of power transformer protection using Artificial Neural Network (ANN). A three layer feed forward neural network (FFNN) has been trained to discriminate between internal faults in power transformer and other abnormalities such as magnetizing inrush currents, saturation of CTs and over-excitation of the core. The results obtained demonstrate the power of ANN in classifying the faults and other abnormal conditions. The training algorithm used is resilient backpropagation (`TRAINRP') along with log-sigmoid transfer function. An improved digital transformer scheme has been suggested by adding the proposed waveform classifier with the existing digital protection scheme to improve the reliability of the existing relays. The results of testing show significant gain in both sensitivity and selectivity of the differential relay as compared to traditional approaches. Several novel concepts have been introduced including the formal measure of information carried by the relaying signals and expected consequences of maloperation of a protective relay. Index Terms— Artificial neural networks, digital relaying, Fourier algorithm, power transformer protection. I. INTRODUCTION
The digital technology has brought unquestionable improvements in numerical relay design in terms of criteria signals estimation in short time, better filtering, standardization of hardware, self monitoring features etc. However, it did not make a major breakthrough in power system protection as far as security, dependability and speed of operation is concerned[6]. The relaying task, however can be approached as a pattern recognition problem. To enhance the performance of protection, the artificial neural network (ANN) approach has been perceived by researchers in the field of power system protection. By monitoring the relaying inputs, the relay can be trained using ANN to classify the ongoing signals between faults and all other conditions. During the energization of transformers, due to strong saturation of the transformer core, very high Inrush currents may result. The establishment of inrush in power transformers is becoming unreliable. Current transformers used on either side of the power transformer may have different performance and particularly at high currents they do not transform their primary currents so accurately for a U.J. Shenoy, corresponding author (e-mail:
[email protected]) is senior scientific officer, H.P. Khincha and D. Thukaram are professors and K.G. Sheshadri is the project assistant in the Electrical Engg. Department, Indian Institute of Science, Bangalore- 560 012. K. Parthasarathy is with the Power Research & Development Company, Bangalore.
D. Thukaram
K.G. Sheshadri
short time under transient conditions when short circuit occurs. Also the mismatch of ratios among different current transformers may introduce error in current transformation at the primary and secondary of power transformer. For units with tap changer, mismatch can occur at taps. Inrush magnetizing currents, external faults combined with saturation of CTs and CT ratio mismatch are the most relevant phenomena, which may upset the current balance causing the relay to maloperate. The two common methods used to avoid undesired tripping due to above mentioned abnormalities are: i) Implementation of delays in protective devices ii) Restraining or blocking the relay operation according to the harmonic content of the measured current. The first solution has been used for primary over-current protection and in differential schemes. However this option is undesirable because of the potential danger of delaying the tripping time during real internal fault. The second solution is based on the fact that the second and fifth harmonic component of the inrush current and over-excitation condition respectively, is considerably larger than a typical fault current. As with classical protection, new numeric protection can also lead to unnecessary operation or failures. This is the consequence of the inability of protection algorithms to distinguish transient states from faults. Especially, critical for the operation of digital protection are distortions, which are caused by protection CTs at partial or total saturation of the iron core. This leads to large short circuit currents with high dc offset or large inrush currents with simultaneous short circuit currents. Hence in real circumstances, there may be delayed or even failure to operate condition. In cases of worst distorted waveforms not resembling close to fundamental or harmonic frequency, it is seen that the digital relaying algorithms fail to converge. Hence it is necessary to classify relaying signals based on wave shapes to discriminate fault and no-fault cases. The classification process is a pattern recognition problem where the solution process must be able to handle noisy data and where functional mapping is nonlinear. In many of relaying problems, the principle is to identify and classify the shape of waveform and then take some appropriate action. This is so because the shape of the waveform gives a pointer to the type of abnormality existing in the system. The ANN method presented in this paper detects inrush currents, external fault combined with saturation of CTs and over-excitation of the core based on recognizing the wave shapes more precisely and differentiates these wave shapes from the internal fault wave shapes for power transformer protection applications.
NATIONAL POWER SYSTEMS CONFERENCE, NPSC 2002
108
II. TRAINING AND TESTING OF ANN BASED WAVEFORM CLASSIFIER
A. Proposed ANN architecture A three-layer feed forward neural network (FFNN) used for the proposed classification problem is as shown in Fig. 1. The FFNN can form any arbitrary complex decision regions and separate different classes. The number of Neurons in the input layer and hidden layer are selected to be equal to 10 and 20 respectively. Since each column of the input data matrix represents one condition of the system, one output neuron is used in the output layer. The output of the neural network must be in the range of 0 and 1 to indicate FAULT (1) or NOFAULT (0) condition. For the waveform classification of power transformers, output value should be close to ZERO for the input current of inrush and CT saturation and should be close to ONE for internal fault conditions.
positive. The TRAINRP algorithm generally provides faster convergence than most other algorithms[8]. The training is carried out in batch mode for various patterns. In batch mode, weights and biases are updated only after the entire set has been applied to the network. The transfer function chosen is log-sigmoid since commonly this transfer function is used in back propagation networks because it is differentiable. The log-sigmoid transfer function is expressed as in (1) f ( x) =
1 1 + exp( − x )
(1)
For nonlinear multilayer network, resilient back propagation along with log-sigmoid transfer function is used. The sigmoid function generally compresses infinity input range into finite range of outputs. Due to this, the gradient can have a very small change in weights and biases even though the weights and biases are far from optimum values. The program is written using Neural Network Toolbox 3.0 for MATLAB version 5.3. The command ‘newff’ creates a feed forward back propagation network. It accepts the arguments such as input matrix of minimum and maximum values for inputs, size of the layers, transfer function of the layers ('log-sig') and the training function ('TRAINRP'). The other fields to be initialized are momentum, epochs, learning rate, goal etc. After the network has been trained SIM command simulates the network and returns the network outputs. The INITWB is a weight-bias initialization function for each layer. The relaying data for training and testing have been generated using C programs and samples were stored as ‘.m’ files. Typical waveforms simulated for training and testing of ANN based classifier are shown in Fig. 2,Fig. 3 and Fig. 4 along with the responses of the full cycle Fourier digital relaying algorithm. Eight cycles of input data for each pattern has been used to train the network. The sampling rate of 32 samples per cycle has been chosen for this purpose. B. Comparison of conventional digital relaying algorithm (Full cycle Fourier) and neural network based approach
Fig. 1. A three-layer feed forward neural network The training function chosen is TRAINRP. This function updates the weights and biases values in accordance with resilient back propagation algorithm. In contrast to other training algorithms TRAINRP function does not use the magnitude of the gradient ∂E / ∂w (derivative of the error function with respect to weight) but uses only the sign of ∂E / ∂w to determine the direction of weight update. The size of the weight change is determined by a separate update value. The update value for each weight and bias is increased by a factor delta_inc whenever the current gradient has same sign as that of the previous, but decreases this value by a factor delta_dec if the gradient has the opposite sign. If the gradient is zero, then the update value remains the same. This update value is then added to the weight if the gradient is negative and subtracted from the weight if the gradient is
The most commonly used FIR filter in digital differential relay designs is based on Fourier full cycle algorithm. The performance of the Fourier algorithm for typical input current signals during CT saturation condition, internal faults and magnetizing inrush are as shown in Fig. 2, Fig. 3 and Fig. 4 respectively. It has been observed that both the conventional algorithm and the proposed feed forward neural network based approach differentiate between steady state inrush, over excitation, which largely contains fifth harmonic and internal faults accurately. However, for the cases of CT saturation and transients in inrush currents the ANN based classification of waveforms is more efficient as compared to conventional digital relaying algorithm. III.
RESULTS AND DISCUSSIONS
Typically 100 different patterns have been generated and used for training the network. The trained network has been
INDIAN INSTITUTE OF TECHNOLOGY, KHARAGPUR 721302, DECEMBER 27-29, 2002
tested with a set of independent test patterns. The results obtained during training and testing of the simulated network are tabulated in Table I. The second harmonic component less than 15% has been considered to rule out magnetizing condition and less than 30% of fifth harmonic component has been chosen to rule out over-excitation condition[2]. The fundamental component of differential current above 1.0 p.u. has been classified as fault. It has been observed that TRAINRP function for ANN gave satisfactory results as compared to other training functions. Also resilient back propagation training is much faster compared to standard steepest descent and conjugate gradient methods. To achieve efficient training of the network momentum, mc=0.9 and learning rate, lr=0.1 are used.
109
enhance the performance of digital protection scheme. The simplified block diagram of the proposed power transformer protection scheme is shown in Fig. 5. By AND gating the output of the digital relaying scheme and ANN output, the reliability of the existing digital relaying scheme can be improved. The future scope of implementation of ANN classifier along with the existing relay designs requires modification of FFNN. This type of network is called Time Delay Neural Network (TDNN). TDNN may be used to implement the moving window concept as in digital relaying algorithm. Since ANN has to be implemented in real time, it has to make use of samples, which are available after every intersample delay. IV. CONCLUSION
TABLE I TYPICAL ANN OUTPUTS DURING TRAINING AND TESTING
Sl. no
1 2 3 4 5 6 7 8 9 10
Training patterns of current inputs P11 P21 P22 P31 P32 P41 P51 P61 P71 P81
Target fixed for training/ testing patterns 1.0000 1.0000 1.0000 1.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Output during training 0.9986 0.9899 0.9842 0.9853 0.9867 0.0057 0.0012 0.0017 0.0085 0.0150
Testing patterns of current inputs PP11 PP21 PP22 PP31 PP32 PP41 PP51 PP61 PP71 PP81
Output during testing 0.9887 0.9789 0.9632 0.9771 0.9541 0.0165 0.0177 0.0821 0.1096 0.1350
P11,PP11: Fault current vectors (fundamental frequency component >1.0 p.u.) P21,P22,PP21,PP22: Current vectors for magnetizing inrush ruled out ( 2nd harmonic component 30%) P61,PP61: Current vectors for steady state CT saturation condition P71,PP71: Current vectors for CT saturation with transients P81,PP81: Current vectors for transients in magnetizing inrush A. Application of ANN to existing digital transformer protection relay The FFNN network and its training process can be implemented with the existing digital relaying scheme to
It has been shown that a feed forward neural network can be used to discriminate between faults and other abnormalities in power transformer protection applications. The neural network can be trained using resilient back propagation training rule with log- sigmoid transfer function. It has been observed that if the network is trained with large number of patterns representing various cases of abnormalities, ANN classifies fault and no-fault condition accurately. The FFNN network and its training process can be implemented to enhance the performance of digital protection scheme. The proposed study may contribute to the evaluation of the efficiency of neural networks in power system relaying domain. V.
REFERENCES
[1] T.S. Sidhu, M.S.Sachdev, "Online Identification of magnetizing inrush and internal faults in three phase transformers", IEEE Transactions on Power Delivery, vol.7, No.4, pp.1885-1891, October 1992. [2] C.H. Einvall and J.R. Linders, “A three phase differential relay for transformer protection”, IEEE Transactions on Power Apparatus and Systems, vol. PAS-94, No.6, Nov./Dec. 1975, pp. 1971-1980. [3] Luis G. Perez., Alfred J. Flechsig., Jack L. Meador, Zoranobradovic, "Training an Artificial Neural Network to Discriminate between Magnetizing Inrush and Internal faults", IEEE Transactions on Power Delivery, vol.9, No.1, January 1994. [4] P. Bastard, M. Meunier and H. Regal, “Neural network based algorithm for power transformer differential relays”, IEE Proceedings-Generation Transmission Distribution, vol. 142, No.4, July 1995. [5] J.R. Lucas, P.G. McLaren, "Improved simulation models for current and voltage transformers in relay studies", IEEE Transactions on power systems, vol. 1, no.1, pp. 152-159, Jan 1992. [6] B. Kasztenny, E. Rosolowski, M.M. Saha and B. Hillstorm, “ A self organizing fuzzy logic based protective relay-an application to power transformer protection”, IEEE Transactions on Power Delivery, vol.12, No.3, July 1997, pp. 1119-1127. [7] M.A. Rahman, B. Jeyasurya, “ A state-of-the-art review of transformer protection algorithms”, IEEE Transactions on power delivery, vol. 3, No. 2, April 1988. [8] H.Demuth, M.Beale, “Neural Network Toolbox User’s Guide”, January 1998.
110
NATIONAL POWER SYSTEMS CONFERENCE, NPSC 2002
Fig .2. Current waveforms and Fourier algorithm output during CT saturation in steady state and transient condition
Fig. 3. Current waveforms and Fourier algorithm output during internal fault in transformer
INDIAN INSTITUTE OF TECHNOLOGY, KHARAGPUR 721302, DECEMBER 27-29, 2002
Fig. 4. Current waveforms and Fourier algorithm output during inrush in steady state and transient condition
Fig. 5. Simplified block diagram for implementation of ANN based Classifier with digital differential relay
111