Ann Based Protection
Ann Based Protection
Ann Based Protection
ABSTRACT This paper presents an artificial neural network (ANN) based approach for three-zone distance protection of transmission lines. The proposed neural network-based distance relay has multilayer feedforward architecture with two inputs and three trip/(no trip) output signals. The first output is responsible for main protection of the transmission line section, whereas the other two outputs provide back-up protection for the adjacent line sections. The input features of the neural network are the fundamental frequency voltage and current magnitudes extracted by discrete-Fourier transform. In this paper, the back propagation training technique has been used for off-line training of the proposed ANN distance relay. The Input-output patterns were simulated for faults covering the three zones of protection at different locations, operating conditions, and fault inception angles. The simulation results presented in this paper show that the proposed ANN distance relay is very effective in detection and classification of line faults and therefore can be considered as a good tool for main and backup digital distance protection.
relays.
This paper presents a new approach for transmission line protection using ANN. The proposed ANN-based distance relay is a system comprising a multilayer feedforward ANN with two inputs and three outputs. The proposed ANN has been trained to provide Instantaneous as well as back up protection for faults detected within its three protection zones. One output is assigned for instantaneous tripping for faults occurring in the first protection zone, while the other two outputs give delayed tripping signals for faults occurring in the second and the third protection zones.
473
(neurons) connected in parallel to perform u s e l l computational tasks such as pattern recogition or classification as an alternative to conventional [9]. ANN computing computing approaches characteristics are distinguished from conventional pattem recognition by their capability to map complex and highly nonlinear input-output patterns. ANN can be used to classify patterns by selecting the output which best represents an unknown input pattems in cases where an exact input-output relationship is not easily defined. It has been proven that a network with one hidden layer can perform any nonlinear mapping and no more than two hidden layers are needed for most applications [ 9 ] . ANN have attracted attention in the last decades to solve problems related to electric power system engineering problems such as load forecasting, security assessment, economic dispatch, and fault detection and classification [IO].
are calculated recursively backwards from the output layer toward the input layer through the hidden layer(s). Once all weight changes are calculated the weights are updated. The entire process is repeated until a target error is satisfied.
In its basic form a feedforward ANN consists of an input layer, an output layer, and one or mare hidden layers. Each layer consists of a set of neurons or nodes that are fidly connected to the neurons in the next layer. The connections have multiplying weights associated with them. The node receives its input from either other nodes or from outside world. The sum of all weighted inputs represents the node activation h c t i o n . The output of the node is determined by an output function which responds to this activation. Frequently, the so called logsigmoid function is used. A typical ANN with two hidden layers is shown in Fig* 1.
Figure 1 Multilayer Feedforward ANN The number o f neurons and hidden layers is problem dependent. The process of determining the weights is called training process. In the training process, sets of input-output patterns are associated by properly adjusting the weights in the network such that a s u m of squared error function is minimized:
where Ep is the pattern error, rk is the target (desired) output, and ok is the actual output of the neural network. Various training algorithms have been developed to adapt the weights in ANNs to reduce the error defined in (1). The ANN used in this study is trained using the back propagation learning algorithm, Weight changes
474
tone 3
1
Lmd
1
TRAINING AND TESTING OF THE ANN
In this paper, the backpropagation training algorithms has been used for off-line training of the proposed ANN relay. A total of 1800 input-output training patterns simulating faults covering the three zones of protection at different locations, fault inception angles and operating conditions have been generated by digital simulation. The fault locations include zone I : lo%, 20%, 30%, 40%, 50%, 60%, 70%, and 80% of the transmission line section, zone 2: 90%, loo%, 110%, and 120% of the transmission line section, and zone 3: 130% - 220% of the transmission line section (adjacent line). For each of these fault locations, faults were simulated at 10 different fault inception angles 0 90. Moreover, the above fault studies were conducted for three loading conditions loo%, SO%, and 33.3% of fullload conditions. In addition to the training input-output patterns, another different 700 testing patterns were generated to test the generalization capability of the proposed ANN-based relay. The fault studies have been conducted on a single-phase system comprising a generator connected to a load by short transmission line sections each represented by lumped-series impedance R-L model. The fault transient voltage and current signals were obtained by numerical solution of the differential equation of the transmission line:
~
+;zy
equations (3) and (4j using the trapezoidal integration rule:
(3)
(4)
where At is the integration step size. The magnitudes of the voltage and current phasors corresponding to the post-fault fundamental frequency were calculated using the full cycle DFT:
(5
SIMULATION RESULTS The proposed ANN-based distance relay was trained using the Neural Network Toolbox [ 1I]. The criterion for determining the number of hidden layers and hidden neurons was based on a combined consideration of the training error (accuracy) and speed. In this study, several tests were performed to determine the optimum number of hidden layers and hidden neurons based on the mean square error (MSE) and number of training epochs. Moreover, different training functions were examined for convergence including traingd, traingdm, traingdx, and trainrp. After extensive simulations, it was found that an ANN architecture with 40 neurons in each o f first and second hidden layers
v ( t ) = R(x)i(i)+ L ( x ) -
di
dt
where v(t) and i(t) represents the instantaneous voltage and current at the relay terminals. R(x) and L(x) represent the resistance and inductance respectively of the portion of the transmission line between the relay and the fault location. The numerical solutions of v(t) and i(t) were obtained by integrating the differential
475
with logsigmoid functions for the hidden and output layers are capable to minimize the MSE of the AN" to a final value less 0.00001within 943 epochs. The MSE training error convergence diagrams for the ANN using the "trainrp" training function is shown in Fig. 3.
The proposed ANN operates in a static manner. The ANN was trained off-line. Once the desired performance was achieved, the weights o f the ANN were frozen. The next step is to test the generalization capability of the proposed ANN when exposed to test patterns which are different from the training pattems. The testing performance of the ANN was examined using 700 input-output patterns representing faults at I' 5 ' fault inception angles (S, 30",4, 60", and 75') at different fault locations over the three zones and at different load conditions. The results of the testing performance are shown in Fig. 5 .
0.81
100
2bO
3bU
400
5b0
6bo
700
800
s! b
EPOchs
OYIPYl
b?
0%
0.4
1
I
100
Figure 3 MSE Training Convergence of the ANN The training performance of the proposed ANN is depicted in Fig. 4. As can be seen from the plots, the results are very encouraging and match accurately with the target values.
OC
0
ZOO
300
400 500
600
700
Paltern Number
,
200
AOO
.
600
eon
.
1000
iaoo
I
0
100
I
200
300
400
500
600
700
Psllsrn N u m b a t
Pattern Number
NN Duylul
I21
200
400
1101)
800
IBQO
100
200
300
400
500
600
IO
PBtlem Number
Panern Numuer
NN
0"tP"t
PI
zoo
.
100
.
600
.
800
.
1000
.
1200
.
1lu0
MOO
I ieao
Pattern Number
The testing phase simulations of the three outputs of the ANN illustrate the efficiency of the proposed scheme. As one can see that almost 100% of the 700 tested fault pattems of outputs 1 and 3 have been successfully classified into the right zone resulting in correct tripho trip output signal. However among the 700 pattems, only one or two patterns of output 2 were misclassified as shown in Pig. 5-b. This well performance can also be observed in Fig. 6 which shows the error per pattern between the actual output pattern and the desired one or the three outputs of the ANN.
476
1
0 6.
The authors acknowledge the support and facilities of Yarmouk University, Irbid, Jordan. REFERENCES
0.604Error (11
0.20
-0.i-0.4-
-0.6-
-0
lbo
2bO
300
ab0
5bLl
6bO
TOO
o a0.6 0 4-
0.2EIlO,
121
-0.2.
0 4. 0 f-
1
1
100
-0.6.
I
0
200
300
P I tte m
400
500
600
700
04
Errol
(3
-0. -0.l o -
io0
2on
300
400
500
600
700
Pattern
CONCLUSION
In this paper, an ANN for fault classification has been developed to simulate the performance o f a distance relay over three zones of line protection. The ANN has been trained wt BP to generate a trip (1) or no trip (0) ih signal according to the location of the fault. Computer simulations for various fault inception angles, locations and operating conditions show that the proposed ANNbased distance relay scheme is reliable and very encouraging. The ANN was able to classify faults they have not been exposed to during the training phase. The proposed scheme can be implemented for real time operation by using dedicated digital signal processing circuits. It is believed that the proposed ANN-based distance relay can support or replace the conventional three-zone distance protection system.
Elmore, W., Protective relaying theoy and applications Marcel Dekker, NY, 1994. 2. Johns, A. T., and Salman, S . K., Digiiai protection for power systems, Peter Peregrinus Ltd., UK, 1995. 3. Kudu, H., Sasaki, H., Seo, K., Takahashi, V., Youshida, K. and Meda, T., Implementation of a digital distance relay using an interpolated integral solution of a differential equation, IEEE Trans. on Power Delivery, Vol. 3, No. 4, 1988,pp. 1475-1481. 4. Ramanioorty, M., Application of digital computers to power system protection, Journal of Inst. Eng. (India), Vol. 52, No. 10, 1972, pp. 235-238. 5. Sachdev, M. S., and Baribeau, M. A., A new algorithm for digital impedance relays, IEEE Trans. on Power Apparatus and Systems, Vol. 98, No. 6, 1979, pp. 2232-2240. 6. Youssef, 0. A, S., A fundamental digital approach to impedance relays, IEEE Trans. on Power Delivery, Vol. 7, No. 4, 1992, pp. 18611867. 7. Cow, D. V., and Jorge, D. C., Artificial neural network approach to distance protection of transmission lines, IEEE Trans. on Power Delivery, Vol. 13, No.], 1988, pp. 102-108. 8. Sidhu, T. S., Singh, H., and Sachdevm, M, S., Design, implementation, and testing of an artificial neural network based fault direction discrimination for protecting transmission lines, IEEE Trans. on Power Delivery, Vol. 10, N0.2, 1995, pp. 697-706. 9. Haykin, S., Neural networks a comprehensive foundation, Macmillan College Publishing Comnanv. 2dEd. NY. 1999. 10. El-Shariawi, M. and Niebur, D., Artificial neural networks with applications i power o systems, IEEE Service Center, NJ, 1995. 1 1 . Demuth, H. and Beale, M., Neural network toolbox users guide for use wirh M T L A B ,
1.
I ,
2002.
AUTHORS ADDRESS
ACKNOWLEDGEMENTS
477