Deep Neural Network Based Fault Classification

and Location Detection in Power Transmission Line

Sheikh Iftekhar Ahmed∗, M. Fahmin Rahman†, Shomen Kundu‡, Raihan Mahmud Chowdhury§,
Auronno Ovid Hussainfj and Munia Ferdoushi∥
Department of Elecrtrical and Electronic Engineering,
Bangladesh University of Engineering and Technology,
Dhaka, Bangladesh
Email: ∗[email protected], †[email protected], ‡[email protected],
§
[email protected], [email protected], ∥[email protected]

Abstract—Transmission lines constitute a significant portion of

any power system as they transmit power from the source to the the values of voltages and currents during fault [3]. In order
load end. Faults in the overhead transmission lines are common to make power system resilient to fault occurrences, fast and
phenomena which can significantly disrupt the effective precise identification of faults is essential for proper system
operation of the power system. Swift and efficient fault restoration and protection. Moreover, the capability to identify
classification along with identification of the exact location of fault occurrence at any point of the transmission line by
fault occurrence to isolate the affected region are extremely
essential for ensuring power system reliability. In this paper, measuring the current and voltages at any one specific
we propose a Deep Neural Network (DNN) based method to location would reduce the need for multiple sophisticated
address the problem of transmission line fault detection, measurement equipment, thereby making the system more
classification, and fault location identification. Our model is economical.
evaluated on a dataset developed through the modeling and
Previously, methods like visual inspection, trial and error
simulation of a four-bus power system with three 100 km
transmission lines. The dataset consists of the RMS values of techniques, and were used for fault identification in transmis-
three-phase currents and voltages due to the occurrence of sion systems [4]. A combination of discrete wavelet trans-
different types of faults on the transmission lines of this form (DWT), multi-class support vector machines (SVM)
system. The model performs satisfactorily in classifying [5] later gained popularity in classifying faults. Hardware-
symmetrical faults and unsymmetrical single line to ground
efficient logic using field programmable gate array (FPGA)
faults with 100% accuracy, yielding an overall 98.775%
accuracy for all four types of faults studied. It also [6] and GSM techniques [7] have also been used to predict
successfully locates the faulty site within at least 2% precision fault types. However, the computational complexity of fault
in more than 80% cases. Index Terms—Deep Neural Network location identification of in the transmission line still remains
(DNN), Mean Absolute challenging. The synchronized phasor measurement technique
Error (MAE), Cross-entropy, Fault Analysis, Power System is capable of estimating fault location in transmission lines,
but installation cost of phasor measurement units is high. [8].
I. INTRODUCTION Deep Neural Networks (DNN) have recently gained pop-
The growing world population and the associated need ularity in the classification and detection of faults due to
for energy have resulted in increased electricity demand and their self-adaptability. Recently, Sahoo et al. [9] used and
consumption in recent years. The efficacy and reliability of Padhy et al. [10] used voltage and current magnitudes to
the power system, therefore, are more important now than classify the fault type in a single transmission line system
ever before. Due to the immense scope and spread of modern with 97% and 97.9% accuracy, respectively. Chawardol et
power systems, the occurrence of faults have become al. [11] additionally used three phase power to classify fault
inevitable. The fault is most likely to occur in the overhead type with 77.6% accuracy and the faulted zone with 95%
transmission lines than in any other part of the power system. accuracy in a three-line transmission system. Elnozahy et
Generally 80-90% short-circuit faults tend to occur on al. [12] used two cycles of FFT-based frequency data and
overhead lines and 10-20% on station equipment and bus bars achieved 100% accuracy in classifying fault and predicted
combined [1]. These faults can be caused by a variety of fault location distance with remarkable precision; however
causes such as lightning, strong wind, the collapse of tree their studied system had only a single 100 km transmission
branches during a natural disaster, and equipment line which ignores interference effects due to the contribution
malfunction. Faults cause high current surges to flow through of neighbouring lines in practical system. In this work, we
the network and thereby disrupt the distribution of electric present a DNN-based approach to classify fault types and
power to consumers.Faults are usually classified as Line-to- identify the location of fault occurrence in a transmission
Ground (LG), Line-to-Line (LL), Double Line-to-Ground system consisting of three 100 km lines by analyzing only
(LLG), and three-phase symmetrical fault (LLL) depending on one cycle of post-fault RMS value data. In this text, sections
the transmission line phase affected by it. Estimating the II-III describe the power system model used in our study,
location of the fault is difficult before isolating the network our developed algorithm, and data generation techniques. The
and it is quite challenging to extract
 obtained results are discussed in section IV and section V and patterns, and the result is conveyed using the output layer.
concludes the paper. The neural network computes the weighted total of the inputs,
II. MODELING OF POWER SYSTEM incorporates a bias and then the weighted total is passed to the
activation function. An activation function determines how the
The simulink model used for fault simulation and dataset
weighted sum of the input is transformed into output from a
generation is shown in Fig. 1. It is constructed using Simscape
node or nodes in a layer of the network. The following figure
Electrical toolbox.
2 depicts a deep neural network.

Fig. 1. Simulink Model of Power System Fig. 2. Forward and Backward Propagation in a neural network
The model is a 4-bus 50 Hz system. Two of the buses,
bus 1 and 4, are generator buses and the other two are load While forward propagation, the following calculations are
buses. Buses 1 & 2, 2 & 3, 3 & 4 are connected via 11 kV made in each layer:
transmission lines of length 100 km. Static loads of 10, 15,
and 5 MW have been added to buses 2, 3, and 4 respectively. Zi = Wi × Ai−1 + Bi (1)
Fault simulators have been added to each transmission line Ai = σ(Zi) (2)
for simulating LG, LL, LLG and LLL type faults. The fault
resistance is varied to 8 different values of 0.25, 0.5, 5, 10, Here, if we have ‘m’ number of training samples and hi
25, 50, 75, and 100 ohms. Generator specification and trans- is the number of neurons in ith layer, then both Zi and Ai
mission line details are given in Table I and II respectively. have the same shape of (hi × m). Wi is the weight matrix of
The transmission line parameters have been taken from IEEE ith layer having a shape×(hi hi−1) and Bi is the bias vector
9-bus model [13]. of that layer having a shape × (hi 1). σ() denotes the
activation function used in the layer. In backward
TABLE I propagation, the output is compared with true expected output
GENERATOR DETAILS and the produced error is minimized by adjusting the weights
and biases. Depending on the type of problem, different loss
Bus Bus Power R L
Name functions are used. The categorical cross-entropy loss (Eqn.
No Type Generation (ohm) (mH)
G1 (11kV) B1 Swing — 0.8929 16.58 3) is for classification tasks where true labels are one-hot
G2 (11kV) B4 PV 20 MW 0.8929 16.58 encoded with probability 1 and 0.
Σn
Lk = − ykj × log(yˆkj ) (3)
j=1
TABLE II
TRANSMISSION LINE DETAILS where, Lk is the loss for k th sample, yˆkj is the predicted prob-
th
ability of j class of that sample, ykj is the true probability of that class and ‘n’ is the no of classes. Mean Absolute Error
Specifications Values
(MAE) (Eqn.Zero
4) loss is for regression problems,
Sequence Resistance
where yˆk and
0.11241 ohms/km
Zero Sequence Inductance 3.53 mH/km yk are the predicted and true values of the kth sample.
Zero Sequence Capacitance 6.15nF/km m
Positive Sequence Resistance 0.044965 ohms/km 1 Σ
Positive Sequence Inductance 1.414 mH/km MAE = y — yˆ | (4)
k
Positive Sequence Capacitance 10.47 nF/km |
m k=1
k

B. Proposed Algorithm
III. FAULT IDENTIFICATION
We have constructed two DNNs separately, one for fault
A. Deep Neural Network detection and classification and another for fault location iden-
A Deep Neural Network is a machine learning architecture tification. The number of hidden layers, neurons and
that draws inspiration from biological neural cells of human activation functions has been chosen to generate optimum
brains [2]. It has three layers: Input Layer, Hidden Layer, Out- performance on the data set. Both models have 6 neurons
put Layer. The input layer accepts inputs in several different in input layer. The classification model network consists of 3
formats. The hidden layers calculate and find hidden features hidden layers having 60-100-50 neurons (Fig. 3) respectively
with activation
functions (relu-relu-relu). Output layer has 5 neurons with
acti- vation function – softmax. The model compilation has
‘Adam’ optimizer and ‘categorical cross entropy loss’
function. The fault location model has 4 hidden layers
having 60-100-80- 80 neurons (Fig. 4) with activation
functions (relu-tanh-relu- relu). Output layer has single neuron
for regression without activation function. Model compilation
has ‘Adam’ optimizer and ‘mean absolute error’ loss function.
Batch Normalization layers are used between hidden layers to Fig. 5. Training performance of classification model
maintain mean value of data close to 0 and standard
deviation close to 1.

Fig. 3. Layer structure for Classification model

Fig. 6. Confusion Matrix for Classificaton

Fig. 4. Layer structure for Location model The model was simulated for 1000 epochs and after the 848 th
iteration the optimum accuracy was obtained. Fig. 5 and Fig.
6 illustrate the training performance plot and the confusion
C. Dataset Generation matrix respectively. The classification model identifies the
Our generated dataset contains two types of target labels. correct type of faults for all the test samples except in 5 cases
One is the distance of the fault location and the other is the for LL and LLG faults. 100% accuracy is obtained for LG,
type of fault. The fault type classes are LG, LL, LLG, LLL LLL and no fault case. The sum of the diagonal percentages
and no fault (labeled as None) and the classes are one-hot confirms the accuracy on the test samples to be 98.7745%.
encoded. The fault distance in a transmission line is measured The minimum cross-entropy loss is found to be 0.0351. Since
from it’s corresponding bus. no fault condition is set as an individual class, the model
We have taken RMS values of the three-phase voltages equally performs well for detecting if there is any fault and
(V , V , V ) and currents (I , I , I ) as the input features. For the type
a b c a b c of fault when it occurs. The performance comparison with the
transmission lines 1, 2, and 3, input voltages and currents other related works are summerized in the table III.
are taken from bus 1, 2 and 3 respectively. The sampling For the location detection problem, the no-fault conditions
frequency is set to 2000Hz which is a suitable rate for the are removed from the whole data set and it is split into 3
power system. In a single run, we took data of one electrical parts, each part corresponding to a particular transmission
cycle resulting in 40 samples for each phase of voltages and line. For each transmission line data, 85% of total samples
currents, and then calculated the RMS value of the samples.
In each transmission line, faults are introduced at 17 different
places, each 5 km apart. At each location, 4 types of faults TABLE III
are introduced along with a no-fault condition. For each type COMPARISON OF RECENT WORK ON FAULT CLASSIFICATION

of fault, fault resistance is set to 8 different values, one at a

time,thus a total of 2040 sets of training data is generated.
IV. PERFORMANCE ANALYSIS
For detection and classification problem, the entire dataset
of 2040 samples has been utilized. Firstly, each input feature
data was normalized for removing the mean and scaling to
unit variance. The data set was split into 80% and 20%
portions for the purpose of training and testing respectively.
The model checkpoint was used which monitors the
categorical accuracy of the testing samples and saves the best
weights of the model for maximum accuracy while training in
a specified local path.
Number of
Data Accuracy
Paper Transmission
Used (%)
Lines (Zones)
Chawardol RMS Volatge, Current and
3 77.6
et al. [11] Three phase Power
Padhy Voltage and Current
2 97.9
et al. [10] Magnitude
Sahoo Pre fault and
2 >97
et al. [9] post fault Current
Elnozahy
1 2 full cycles of FFT 100
et al. [12]
One cycle of post fault
Our work 3 98.775
RMS Voltage and Current
were taken for training the location detection model and the TABLE V
rest of the samples were used for performance evaluation. COMPARISON BETWEEN ACTUAL AND PREDICTED DISTANCE
Considering the MAE between predicted distance and actual
Line Disatnce Actual Predicted
distance as the evaluation parameter, the resulting optimum No From Bus Distance (km) Distance (km)
outcomes for each line are summarized in table (IV). Figure 01 B1 20 19.44476
7 shows the training performance and some of the model 01 B1 40 39.62964
01 B1 55 55.17698
predictions are listed in Table V. The neural model performs 01 B1 85 86.37507
better for line 1 and took more epochs for lines 2 and 3. The 02 B2 25 25.03100
number of correct predictions within ±2 km is considerably 02 B2 45 44.64050
02 B2 65 65.83645
acceptable. Previously, Chawardol et al. [11] attempted to 02 B2 70 68.82495
identify the zone of fault occurence in a three-line 03 B3 10 8.962578
transmission system, however their model could not predict 03 B3 25 25.92330
03 B3 50 51.15203
fault-occurrence distance. With same number of transmission 03 B3 90 88.92009
lines, our model estimates the distance of faulted area in
kilometers form respective buses. The model developed by
Elnozahy et al. isolation of affected portion. Especially for power systems in
[12] also give location in kilometers, however their model had areas that are prone to frequent power disruption due to
only a single transmission line and hence failed to account for natural disasters will significantly benefit from our approach.
the possible contribution of voltage and current components
from neighbouring lines and buses. Our work has presented REFERENCES
a more comprehensive analysis through more accurate and [1] J. Schlabbach and K.-H. Rofalski, Power system engineering : planning,
practical modeling of power system considering contributions design, and operation of power systems and equipment. Weinheim:
Wiley-Vch, 2008.
of surrounding networks while giving precious fault location [2] L. V. Fausett, Fundamentals of Neural Networks: Architectures, algo-
information. rithms and applications. Pearson, 1993.
[3] J. -D. Park, J. Candelaria, L. Ma and K. Dunn, ”DC Ring-Bus Microgrid
TABLE IV Fault Protection and Identification of Fault Location,” in IEEE Transac-
LINE-WISE PERFORMANCE OF LOCATION MODEL tions on Power Delivery, vol. 28, no. 4, pp. 2574-2584, Oct. 2013.
[4] S. S. Gururajapathy, H. Mokhlis, and H. A. Illias, “Fault location
Line Epoch Minimum Distance Distance and detection techniques in power distribution systems with distributed
No No MAE (km) Error <1.5 km Error <2 km generation: A Review,” Renewable and Sustainable Energy Reviews,
01 2151 0.6715 89.0244% 93.9024% vol. 74, pp. 949–958, 2017.
02 4802 1.6058 76.8293% 79.7468% [5] V. Malathi and N. S. Marimuthu, ”Multi-class Support Vector Machine
03 5727 2.1413 74.3902% 81.25% approach for fault classification in power transmission line,” 2008 IEEE
International Conference on Sustainable Energy Technologies, 2008, pp.
67-71.
[6] S. P. Valsan and K. S. Swarup, ”High-Speed Fault Classification in
Power Lines: Theory and FPGA-Based Implementation,” in IEEE
Trans- actions on Industrial Electronics, vol. 56, no. 5, pp. 1793-1800,
May 2009.
[7] M.S. Sujatha, M. V. Kumar, On-line monitoring and analysis of faults
in transmission and distribution lines using GSM, J. Theor. Appl. Inf.,
(November (2)) (2011), pp. 258-265.
[8] A. Rahideh, M. Gitizadeh, and S. Mohammadi, “A Fault Location
Technique for Transmission Lines Using Phasor Measurements,” Inter-
national Journal of Engineering and Advanced Technology (IJEAT),
vol. 3, no. 1, pp. 241–248, Oct. 2013.
[9] B. K. Sahoo, S. Pradhan, B. K. Panigrahi, B. Biswal, N. C. Patel
and S. Das, ”Fault detection in Electrical Power Transmission System
using Artificial Neural Network,” 2020 International Conference on
Computational Intelligence for Smart Power System and Sustainable
Fig. 7. Training performance of location detection model Energy (CISPSSE), 2020, pp. 1-4.
[10] S. K. Padhy, B. K. Panigrahi, P. K. Ray, A. K. Satpathy, R. P.
Nanda and A. Nayak, ”Classification of Faults in a Transmission Line
V. CONCLUSION using Artificial Neural Network,” 2018 International Conference on
Information Technology (ICIT), 2018, pp. 239-243
The proposed DNN algorithm has successfully [11] P. S. Chawardol and H. S. Sheikh, “Transmission Line Fault Classifica-
classificatied various faults in transmission lines with tion and fault zone identification using back-algorithm based Neural
Networkpropagation,” International Journal of Engineering Research
98.775% accuracy. The distance of fault location from and, vol. V5, no. 12, 2016.
reference bus has also been evaluated within 2% of the line [12] A. Elnozahy, K. Sayed and M. Bahyeldin, ”Artificial Neural Net-
length for an average of 85% cases. This approach is capable work Based Fault Classification and Location for Transmission Lines,”
2019 IEEE Conference on Power Electronics and Renewable Energy
of identifying fault class and fault location by observing only (CPERE), 2019, pp. 140-144
one electrical cycle of three phase currents and voltages. The [13] J. Pettikkattil (2022). IEEE 9 Bus, MATLAB Central File Ex-
results obtained indicate quite impressive accuracy of DNN change (https://www.mathworks.com/matlabcentral/fileexchange/45936-
ieee-9-bus). Retrieved September 17, 2022.
algorithm-based approach. When implemented in real power
systems, our model can significantly alter the current
approach to fault detection and