Training Network

New Design of a Neural Network Algorithm for

Detecting and Classifying Transmission Line

Faults

S. Vasilic, Student Member, IEEE, M. Kezunovic, Fellow, IEEE

forward networks. Training of these networks is very slow,

needs much larger training sets, and very easily converges on

Abstract--This paper introduces a new artificial intelligent

based approach for detecting and classifying the faults in power local minima, whenever input patterns with large

system networks. This approach utilizes unique type of neural dimensionality are present as in this particular case.

network specially developed to deal with large amount of input

Furthermore, retraining of this type of network with new

data. A model of an actual power network is implemented in ATP

training data is rather difficult.

program and used for simulating the fault scenarios on

Instead of using multilayer neural network, a unique type of

transmission lines. Protective algorithm is implemented in

neural network may be used for fault classification [3-6]. This

MATLAB and interacts with the network model simulations in

ATP. Procedures of generating training and testing patterns are network is based on ISODATA clustering algorithm [7] and

performed carefully to ensure covering of all possible events. belongs to a group of special neural networks named Self-

Training and testing phases of the neural network algorithm are

Organizing Maps [3]. The adaptive behavior of the neural

optimized to improve classification of a variety of previously

network is described by Adaptive Resonance Theory [8]. The

unseen patterns.

main aim of this work is to enhance that neural network based

clustering algorithm. In the previous version of the algorithm

Keywords--clustering methods, electromagnetic transients,

neural networks, pattern classification, power system faults, relatively small number of training and testing patterns was

protective relaying, testing, training. used. Training patterns did not cover different values of fault

angles and significant performance deterioration became

I. INTRODUCTION obvious, due to insufficient network training. Number of

T HIS paper introduces artificial neural network based passes through the stabilization phase was limited (to speed-up

technique for detecting and classifying faults on the training) and this prevented establishing optimal clustering

transmission lines. Transmission line faults happen randomly, structure. Classification of test patterns was done based on

and they are outcome of unpredictable conditions. Several predetermined number of nearest clusters, instead of selecting

varying parameters: type of fault, fault location, fault optimal number of neighbors for each implementation. Also,

impedance, and fault incident time determine the fault location classification was performed in three cycles

corresponding transient current and voltage waveforms making it challenging for an on-line classification of the fault

detected by the relays at line ends. The main role of the zone in one cycle.

relaying principle is detecting and classifying the faults, based The new algorithm has to overcome observed deficiencies.

on three phase voltage and current samples. The new detection It has to show the importance of proper generation of the sets

and classification approach has to reliably conclude, in a very of training and testing patterns and apply extended set of

short time (1-2 cycles), whether and which type of fault occurs scenarios for improved algorithm design and evaluation.

under a variety of time-changing operating conditions [1]. Moreover, some steps in the algorithm training and testing

Protective relay accordingly performs action, usually phases may be improved assuring smaller classification error.

disconnecting faulted phase/line and/or initiating some alarm Additional task is establishing specially devoted simulation

and control signals. environment, where training and testing procedures can be

Various applications of neural networks were used in the easily controlled and performed. This is done through

past to improve the distance relaying of transmission lines [2]. interfacing different programs, exchanging simulation

These applications are mainly based on multilayer feed- parameters and results, and using appropriate graphical

interface.

A specially developed power network model has been

This study was supported by an Army/EPRI contract between EPRI and

implemented in the electromagnetic transient program ATP

Carnegie Mellon University, and has been carried out by Texas A&M

University under the subcontract titled "Self-Evolving Agents for Monitoring, [9]. That model has been interfaced to MATLAB package for

Control and Protection of Large, Complex Dynamic Systems".

automatic simulation of a large number of scenarios [10-12].

S. Vasilic and M. Kezunovic are with the Department of Electrical

Simulation outputs are used as a signal generator for the neural

Engineering, Texas A&M University, College Station, TX 77843-3128 USA

(e-mails respectively: ********@**.****.***, *******@**.****.***).

2

network algorithm design, implemented in MATLAB. entire data set. Then Euclidean distances between each pattern

The paper is organized as follows. Description of the neural and the center are calculated and sorted in an increasing order.

network detection and classification algorithm is given in Now, the first cluster is formed by taking pattern closest to the

section II. Section III through its subsections shows the center of data set, and with the radius equal to actual value of

selected model of an actual power network, design the vigilance parameter. Furthermore, all the remaining

implementation steps (devoted to pattern generation, algorithm patterns are presented, in order of the sorted distances.

training and testing), and provides the classification results. Distances between the pattern and existing clusters are

The conclusion is given at the end. calculated. The minimum distance and corresponding (nearest)

cluster are found. If the minimum distance is less or equal to

II. NEURAL NETWORK CLASSIFICATION ALGORITHM the vigilance parameter, then the actual pattern is classified

into the nearest cluster and the cluster center is updated by

Neural networks try to produce a concise representation of

adding new pattern to the cluster. Otherwise, if the minimum

system's behavior through identifying natural groupings of data

distance is greater then the vigilance parameter, the actual

from large data sets. The aim of this procedure, called

pattern forms a new cluster.

clustering, is to partition a given set of input data (patterns)

During stabilization phase the clustering algorithm is

into several groups or clusters, so that each pattern is assigned

reiterated until a stable cluster structure occurs and there are

to a unique cluster. Patterns that belong to the same cluster

no patterns changing their cluster membership during the

should be as similar as possible, while patterns that belong to

iterations. All patterns are presented again. Distance between

different clusters should be as different as possible. Class label

each pattern and existing clusters are calculated, and minimum

is assigned to each cluster, where class symbolizes a group of

distance and corresponding (nearest) cluster are found. Also, a

patterns with a common characteristic.

cluster where the pattern was previously classified is found. If

Self-organizing maps are special type of neural networks,

pattern was classified into the actual nearest cluster and

and they map input patterns with similar features into

minimum distance is less or equal to the vigilance parameter

contiguous clusters after enough input patterns have been

then there is no learning (no changes in the cluster structure).

presented. The similarity between patterns is usually measured

If the pattern was not classified into the actual nearest cluster

by calculating the Euclidean distance between two n-

and minimum distance is less or equal to the vigilance

dimensional vectors. After training, self-organized clusters

parameter then the pattern is moved to the actual nearest

represent prototypes of classes of input patterns.

cluster and that cluster as well as the cluster where pattern was

Adaptive Resonance Theory defines forming a new cluster

previously classified are updated. If minimum distance is

whenever a pattern, sufficiently different from all previously

greater then the vigilance parameter, a new cluster is formed

presented patterns, appears. Adaptive resonance architectures

and the cluster where the pattern was previously classified is

are capable of continuos training with non-stationary inputs.

updated. After processing all patterns, clusters remained

This neural network is without hidden layer and its self-

without patterns are discarded, because their patterns have

organized structure depends only on the presented input data

been moved to other clusters. Stabilization phase is repeated

set. The neural network training consists of unsupervised and

many times until no pattern changes its cluster membership.

supervised learning phases. In the unsupervised learning,

Supervised learning separates non-homogenous clusters

patterns are presented without their class labels, and this

from the homogeneous ones. It assigns class labels to the

procedure tries to identify prototypes that can serve as cluster

homogeneous clusters, and these clusters and their patterns are

centers. In the supervised learning the class label is associated

extracted from further iterative training process. Set of

with each data point. Vigilance parameter is a confidence

remaining patterns (patterns in non-homogeneous clusters) is

measure and is being tuned, consecutively decreasing during

transformed into new, reduced data set of training patterns. If

iterations. It controls the number and size of generated

new set of training patterns is not empty, vigilance parameter

clusters. The large values allow large deviations from the

is decreased, and unsupervised and supervised learning

cluster centers and hence lead to a small set of clusters, while

procedures are repeated. Otherwise, if either all actual training

small values lead to a large number of tight clusters.

patterns are members of only homogeneous clusters, or current

The initial data set, containing all the patterns, is firstly

value of vigilance parameter is less then specified value,

processed using unsupervised learning. The outcome of

learning is completed.

unsupervised learning is a stable family of clusters, defined as

During the testing phase Euclidean distances between test

hyperspheres in an n dimensional space, where n denotes the

pattern and established clusters (prototypes) are calculated,

number of input features. Unsupervised learning forms stable

and k-nearest neighbor rule [13] is used to classify the pattern.

family of both homogenous (having patterns with the same

Given a set of classified data, the k-nearest neighbor rule

class label) and non-homogenous (having patterns with two or

determines the classification of a new pattern based on the

more class labels) clusters. It does not require either the initial

most represented class label amongst the k nearest clusters,

guess of the number of cluster, or the initial cluster center

retrieved from the cluster structure adopted during training.

coordinates. It consists of two steps: initialization and

The outcome of the testing phase are class labels assigned to

stabilization.

testing patterns.

Initialization phase begins with calculating the center of the

3

III. DESIGN IMPLEMENTATION

A. Power Network Modeling

A typical 345 kV power system section, from Reliant

Energy (RE) HL&P company, was modeled for the testing and

simulation studies. The reduced network equivalent was

obtained by using the load flow and short circuit data, and

verified using both the steady state and transient state results.

Fig. 1 shows one-line diagram of the reduced equivalent for

the used section. STP-SKY section model has nine buses and

contains both short and long transmission lines. This reduced

system is convenient for producing fault waveforms to be used

for transient testing of protective algorithms.

B. Generation of Training and Testing Patterns

Model of the given power network (Fig. 1) is implemented Fig. 2. RE HL&P STP-SKY network model implemented in ATP program

in Alternative Transient Program (ATP) program, and shown

in Fig. 2. This model is used for simulating various fault The testing patterns might be very heterogeneous and quite

different from the training patterns since there are many

operating states and possible events in the power network.

HILL E9

Z9

E1

SKY

They are classified according to their similarity to prototypes

Z1

adopted during training.

This implementation includes all 11 types of fault (AG, BG,

MARION E8

Z8

CG, AB, BC, CA, ABG, BCG, CAG, ABC, ABCG) and the

SPRUCE

normal state. Possible values for fault distance from the SKY

bus are anywhere between 0-100 percents of the total line.

HOLMAN E7

Z7

Fault resistance might be theoretically anywhere between 0

Z2

and Ohms. Fault angle is between 0 and 360 deg. Fault

LHILL STP WAP E6

angle is transformed into corresponding fault incident time,

Z6

E2

where 0 deg is equal to the initial fault incident time, and 360

deg is equal to the initial fault incident time increased for 1

E5

Z3 Z4 Z5

cycle. Parameters are also the initial fault incident time (for

angle 0 deg), and simulation step and end times.

DOW

E3 E4

Combinations of selected values for fault parameters define

total number of simulation cases. After each simulation, output

Fig. 1. RE HL&P STP-SKY Power Network Model

data in specific ATP format containing time and three phase

scenarios on one of its transmission lines (STP-SKY), by

voltage and current variables as well as characteristics of the

varying fault parameters. An intelligent, neural network based,

implemented fault are converted into MATLAB format. The

algorithm is located at the bus SKY, at one end of the selected

example of the simulation output data for one specific case,

line (notation AB1 on the scheme in Fig. 2). It takes voltage

phases A to B to ground fault (ABG), is shown in Fig. 3.

and current measurements from that end of the line and has to

Examples of training patterns for different values of fault

be trained and simulated to recognize the fault on that line.

parameters are shown in Figs. 4 to 8. The algorithm uses

Current and voltage samples obtained through simulations are

training patterns formed by sampling all three phase voltage

used for forming training and testing patterns for protective

and/or current measurements in the selected time window.

algorithm learning and evaluation.

Parameters used for algorithm training in this particular

The MATLAB program interacts with ATP simulations.

simulation example are: three phase currents selected as the

Generation of patterns may be deterministic or random. The

data for training; starting time for taking patterns is 0 seconds

classification algorithm requires deterministic generation of

after the fault incident time; time window is 16.7 ms or 1

training patterns, by specifying several values for each of the

cycle; sampling frequency is 2 kHz (33 samples per cycle).

four fault parameters and combining these values to cover

Features of the training patterns are extracted by using

diversity of fault cases. This forms prototypes that represent

simulation data obtained in a desired time window and with

the space of possible events. The number of training patterns is

selected sampling frequency. Phase A, B, C currents sampled

limited and has to roughly cover all important situations in the

during one cycle after the fault occurs are extracted and placed

power network. Random generation is based on the random

together in one row (Fig. 9) to form feature vector of 99

setting of all fault parameters only constrained by the user.

components (3 phases with 33 samples each). Then all training

This is the requirement for generating testing patterns for

patterns are normalized by scaling all features of all patterns to

algorithm evaluation in heuristic, previously unseen situations.

4

Fig. 3. Voltage and current samples for ABG fault, fault distance 50%, fault

Fig. 5. Example of training patterns for different fault distances. Type of fault

impedance 0 Ohm, and fault incident angle 0 deg.

is ABG and other fault parameters are constant.

have the mean zero, and variance one, and this scaling value is

used later for normalization of the testing patterns. Different

values of fault parameters have to be taken into account for

training, to avoid misclassification later. Figs. 4-7 show

responses during varying type of fault, fault distance,

impedance and angle, respectively. Fig. 8 shows how

combination of varying parameters may cause

misclassification of faults in Zone I and Zone II, supposing

that the bound between zones is established at 80% of the line

length, counting from SKY bus.

Parameters used for generation of the training patterns in the

simulation example given to illustrate the whole algorithm

design are: all 11 types of fault and normal state; fault

distances 5 to 95 % in increments of 10%; fault resistance (for

ground faults) 0, 10, 20 Ohms; fault angle 0 to 330 degrees, in

increments of 30 degrees. Total number of training patterns by

combining all parameters is 2652, and all training patterns are

shown in Fig. 9. Fig. 6 Example of training patterns for different fault impedances. Type of

Parameters used for generation of testing patterns are fault is ABG and other fault parameters are constant.

Fig. 4. Example of training patterns for all 11 types of fault. Other fault Fig. 7 Example of training patterns for different fault incident time. Type of

parameters are constant. fault is ABG and other fault parameters are constant.

5

The first step in algorithm training is to extract the training

patterns from generated patterns and it is described in the

previous section. Two types of classifications were

implemented. Training I was performed for establishing the

cluster structure capable of recognizing only the type of fault.

Training II was performed for establishing the cluster structure

capable of recognizing type of fault and zone of fault.

Boundary distance between the first and second zone is 80%

of the line length. After several hours of iterations, both

training procedures terminated successfully. Simulation output

of Training I is the cluster structure containing 269 clusters,

and of Training II is the cluster structure containing 706

clusters.

D. Algorithm Testing and Classification Results

In the testing phase, input to the neural network is in the

form of the sliding data window containing samples of phase

currents and/or voltages. Classification of testing patterns is

Fig. 8 Example of training patterns for combined different values of fault

parameters.

performed by using cluster structure established during

training and applying the k-nearest neighbor rule. Input

uniformly random selection of fault type, distance between 0

parameter for algorithm testing is only the number k of the

and 100%, angle between 0 and 360 deg, and normally random

nearest neighbors for the rule. Testing patterns are extracted

selection of fault resistance, with mean 0 Ohms and variance

from generated patterns using the same procedure as for

10 Ohms (taking only positive values). Total number of testing

training patterns and have equal number of features. Testing

patterns is 5000.

patterns are normalized by scaling all features of all testing

patterns with the same factor used for scaling of training

patterns. For each testing pattern, distances to all clusters are

computed and sorted in an increasing order. The most frequent

class label of k nearest clusters is computed and assigned to

actual pattern. If the input pattern belongs to any of the

normal-state clusters then the input window is "moved" for

one sample and the comparison is performed again. If the input

pattern does not belong to the normal-state clusters then

fault is detected and execution of the fault classification logic

is initiated. The parameter used to force the neural network to

make the final decision is the time. After decision time has

expired, if pattern still does not come back to the normal state,

neural network will classify fault event according to the fault

type detected in that instance.

Average classification error for the entire testing set of

patterns for selected values of the nearest neighbors from 1 to

Fig. 9 All training patterns. 12, is established by comparing the true and computed class

labels. It is shown in Fig. 10 for both training cases. The

C. Algorithm Training graphic helps in finding optimal values for parameter k in both

Selection of data for training includes either three phase cases. Optimal value k for Training I is 1, and classification

currents, or three phase voltages, or both the three phase error for optimal k is 0.48%. Optimal value k for Training II is

currents and voltages. Vigilance parameter (cluster radius) is 1, and classification error for that k is 6.34%. Obviously,

defined with its initial (maximal) and minimal values, as well classifying the zone of fault is much more difficult task then

as with the decreasing factor during iterations. Type of classifying the type of fault.

classification might be based on detection of fault type

IV. CONCLUSION

(Normal, AG, BG, CG, AB/ABG, BC/BCG, CA/CAG,

ABC/ABCG), fault zone (Normal, Zone I, Zone II), fault The most important aspect of this research is to show that

resistance (Normal, Low, High), or any combination among the proposed neural network approach enables better detection

them. Boundary distances between zones I and II of the fault, of faults in power system networks and proper action to

and between low and high fault resistance, also have to be protect the network. An example of an actual power network

specified. was modeled in ATP program and used to simulate various

6

V. REFERENCES

[1] Power System Relaying Committee, Working Group D5 of the Line

Protection Subcommittee, Proposed statistical performance measures

for microprocessor-based transmission line protective relays, Part I and

II, IEEE Trans. Power Delivery, vol. 12, no. 1, pp. 134-156, Jan. 1997.

[2] M. Kezunovic, "A Survey of Neural Net Applications to Protective

Relaying and Fault Analysis", Engineering Intelligent Systems, vol. 5,

no. 4, pp. 185-192, Dec. 1997.

[3] Y. H. Pao, Adaptive Pattern Recognition and Neural Networks,

Reading: Addison Wesley, 1989, p. 309.

[4] Y. H. Pao and D. J. Sobajic, "Combined Use of Unsupervised and

Supervised Learning for Dynamic Security Assessment", IEEE Trans.

Power Systems, vol. 7, no 2, pp. 878-884, 1992.

[5] M. Kezunovic M., I. Rikalo, and D. Sobajic, "High-speed Fault

Detection and Classification with Neural Nets", Electric Power Systems

Research, vol. 34, pp. 109-116, 1995.

[6] M. Kezunovic and I. Rikalo, "Detect and Classify Faults Using Neural

Nets", IEEE Computer Applications in Power, vol. 9, no. 4, pp. 42-47,

1996.

[7] G. H. Ball and D. J. Hall, "A Clustering Technique for Summarizing

Fig. 10. Results of classification error for Training I and Training II.

Multivariate Data", Behavioral Science, vol. 12, pp. 345-370, 1967.

[8] G. A. Carpenter and S. Grossberg, "ART2: self-organization of stable

fault events in the network. Training and testing patterns are category recognition codes for analog input patterns", Applied Optics,

extracted from the measurements. Neural network based vol. 26, no. 23, pp. 4919-4930, Dec. 1987.

[9] CanAm EMTP User Group, Alternative Transient Program (ATP) Rule

clustering algorithm, implemented in MATLAB, is used to

Book, Portland, 1992.

form pattern prototypes, homogenous structure of clusters [10] The MathWorks, Inc., Using MATLAB, Natick, Jan. 1999.

representing various classes of input data set. Testing patterns [11] M. Kezunovic and S. Vasilic, "Advanced Software Environment for

Evaluating Protection Performance During Power System Disturbances

are classified by combining cluster structure and k-nearest

Using Relay Models", submitted to CIGRE SC 34 Colloquium,

neighbor rule. Romania, Sep. 2001.

This advanced algorithm has several important benefits [12] M. Kezunovic and S. Vasilic, "Design and Evaluation of Context-

Dependent Protective Relaying Approach", submitted to IEEE Porto

comparing to the previous version of the algorithm. New

Power Tech' 2001 Conference, Portugal, Sep. 2001.

algorithm offers easy selection of desired scenarios and [13] T. M. Cover and P. E. Hart, "Nearest Neighbor Pattern Classification",

algorithm parameters by using MATLAB. Various types of the IEEE Trans. Information Theory, vol. IT-13, pp. 21-27, 1967.

classification may be selected and combined. Bounds between

zones of fault may be easily changed. Libraries of the training VI. BIOGRAPHIES

and testing patterns, and cluster structures might be generated

and combined to achieve better algorithm training and Slavko Vasilic (S'00) received his B.S. and M.S. degrees in electrical

validation. Extended sets of training and testing patterns have engineering from University of Belgrade in 1993. and 1999., respectively, and

currently is a Ph.D. candidate in electrical engineering at Texas A&M

been implemented. Since training patterns are generated

University. His research interests are neural networks, fuzzy logic, genetic

uniformly, testing patterns are generated randomly to ensure algorithms, multivariable, robust and adaptive systems, and their

heuristic covering of all possible events. Also, previous implementation in process control and pattern recognition, and especially in

power systems control, protection and monitoring.

version was trained only for particular values of fault angle (0

and 90 deg), while the new algorithm is trained for all possible

Mladen Kezunovic (S'77, M'80, SM'85, F'99) received his Dipl. Ing. degree

values of fault angle (0-360 deg). Fault location classification from the University of Sarajevo, the M.S. and Ph.D. degrees from the

is now performed in one cycle, instead of in three cycles as it University of Kansas, all in electrical engineering, in 1974, 1977 and 1980,

respectively. He has been with Texas A&M University since 1987 where he is

was done earlier. Number of passes through stabilization phase

the Eugene E. Webb Professor and Director of Electric Power and Power

is now unlimited and enables forming more realistic Electronics Institute. His main research interests are digital simulators and

prototypes. Tuning of the new algorithm finds optimal value simulation methods for equipment evaluation and testing as well as

application of intelligent methods to control, protection and power quality

for number of neighbors in k-nearest neighbor rule, while in

monitoring. Dr. Kezunovic is a registered professional engineer in Texas, and

the previous version only predetermined number of three a Fellow of IEEE.

nearest neighbors was used.

Contact this candidate