FAULT DETECTION IN

POWER SYSTEM USING

ARTIFICIAL NEURAL
NETWORK
Hania Zahra
MS20-EE-23
SUPERVISED BY : Dr Umar Farooq
Institute of Electrical, Electronics & Computer Engineering
UNIVERSITY OF THE PUNJAB, LAHORE
Abstract
This project is based upon reliability enhancement of the power transmission system. At present
three-phase AC transmission is extensively used in Power Transmission System. For large scale
Application of three phase AC transmission, safety and stability of the system is the main
consideration. Faults occurring in transmission line interrupt the power flow and it is necessary to
locate them and recover as quickly as possible. The research work proposes an Artificial Intelligence
based strategy Artificial Neural Network (ANN). ANN is designed to locate and classify the faults in
transmission line. The system detects if there is any fault in line or not, classifies and detects the
faults and enables the relay system to operate quickly to avoid any harmful consequence.4-8-4
neural network scheme has been implemented for fault detection. Eight layer artificial neural
network with 4 neurons in input layer, 15 neurons in hidden layer and 4 neurons in output layer used
in MATLAB to detect the fault with high accuracy as compared to other strategies. The inputs of the
Artificial Neural Network are the voltage and current signals. MATLAB Neural networks have been
performed and faults are tested at different locations. Loss values and epochs graphs of the proposed
strategy are also compared with the conventional strategies depicting the peculiarities of the strategy.
Statement of the Problem
A power system is divided into three parts such as generation, transmission, and distribution of
electricity. Power transmission lines carries electricity from generation side to the near and farthest
away load centers. Transmission lines and its total length have been increased with expansion of the
electric power system. Occurrence of fault in transmission lines is quite obvious, such as lightening
strokes, short circuits, defective equipment, human error and ageing. Mechanical damage can also
generate faults in power system; these faults have to be resolved before reconnection of the system.
Restoration procedure can be rapid if the fault location is known or the fault location is estimated
accurately. Short and long time power outages can be caused by faults for customers, which can
result in major losses. Maintaining a reliable power system operation, the capability to identify, cut
off, locate, and repair these faults rapidly is very tough. The faults occurring on transmission lines
can be permanent fault or momentary fault depending upon the reason of it. If these faults cannot
resolve speedily, they can cause destructive consequences. The complete power system can be
destroyed and also interrupt the power supply of a wide portion of the load center.
Objectives of the Project
 In transmission line when any fault occurs, Voltage at that side (fault side) abruptly decreases to
very low value. This unexpected alteration generates the elevated frequency electromagnetic
impulse, a travelling wave. These travelling waves shift out from the fault and move towards the
sides of sending and receiving end at nearly light speed.
 Signal which has to be filtered and subjected to different signal processing approaches are
collected by the instrument transformer in order to scrutinize the defect. These faults are located
by using filtered signal. To exactly locate the fault position, we have to compute some parameters,
i.e. polarity, phase and time delay of the line.
 In this project, we are trying to explore the various approaches, i.e. Artificial Neural Networks
(ANN) for fault detection in transmission lines. The significance of this project is that we are
trying to reduce power outages. Furthermore, compare the power system faults which are not
equipped with ANN to the power system faults equipped with ANN.
Literature Review
To analyze fault in OH is much easier as compared to UG. Fault in overhead lines are quite obvious
and visible, e.g. tree fallen over line, equipment failure and conductors have fallen over the ground.
Modern approaches (fast protection techniques) are used to clear the transmission line faults rapidly.
There are several techniques has been developed for fault classification. Extensively used fault
classification approaches are categorized into three parts as follows:
 Prominent Approaches
 Hybrid Approaches
 Modern Approaches
 Prominent Methods
1: Wavelet Approach
Wavelet approach is actually a mathematical tool which is best for signal processing. This technique selects a
wavelet function which is called mother wavelet. This mother wavelet function can perform several operations
such as evaluation of electric signal and time varying portents. With the use of Multi resolution array, we can
analyze the signal at different values of frequencies with different resolutions. MRA is specially designed for
outstanding frequency resolution and adverse time resolution in case of low frequency level, for high
frequency level it gives adverse frequency resolution and outstanding time resolution.
2: Artificial Neural Network Approach
For the classification of faults, an artificial neural network can be used. It contains 4 different types of layers
such as input layer, hidden layer, softMax layer and output layer. Input data is adjusted in such a way that
when it is given to the input layer it is in the form of 0 and 1. The most important task is performed in hidden
layer by using nonlinear transformation. There is an activation function used for nonlinear problems ,also
transform the problem from linear to nonlinear.
3: Fuzzy Logic Approach
Fuzzy logic-based approach can be used for fault classification/detection. This approach has been utilized to
categorize the faults to classify whether it is L-G, L-L or L-L-G. For this approach, three phase currents at the
end of the transmission line are mandatory to be painstaking for fault classification/detection. We can
implement this strategy on EMTP and MATLAB.
 Hybrid Methods
1: Neuro-fuzzy Approach
Neuro-fuzzy model designed for a transmission line having different line currents under normal and odd
situations at different localities. Fuzzy controller contains three major blocks named as fuzzification, fuzzy
inference and defuzzification. These ANN are very beneficial for determining input and output behavior of
a system, by considering only the weight between its three layers.
2: Wavelet and ANN Approach
Fault has been categorized as pre-fault, fault and post fault. These currents indicate the single phase to
ground fault. The pre-fault is actually a stable state and normal system procedure. Interval of the fault is
also the most important thing to consider, which helps to understand that when the fault has been occurred
and for how long it stays and what are the characteristics of the fault during that whole period. Post fault is
completely depend on the protection system of the transmission and on the installed relays system.
3: Wavelet and Fuzzy logic Approach
Fuzzy logic theory tackle heuristic problems and also provides guidance related to uncertainty issues.
Procedure of fuzzy network contains four parts i.e. fuzzification, inference, composition, and
defuzzification. There are different inputs at the start and then truth values has been applied on it to figure
out the problem. Then we form a fuzzy set in the next step and at the end these values fed back to the
fuzzification level.
Modern Methods
1:Genetic Algorithm
Genetic algorithm and feed forward neural networks has been specially designed to receive and
upgrade the attributes of the input and output. Genetic Algorithm has been designed by considering
genetic evolutionary concepts. This approach can be used to solve optimization difficulties. Only the
best suited features from the complete population lies at the last and create offspring. At the end,
genetic legacy has been transferred to future generations.
2:GSM Approach
GSM module contains phase voltage sensor which is responsible for the continuous sensing of
voltage of the transmission line . Once the fault has been occurred, the values of current and voltage
increases. Then voltage has been fed to the micro-controller. Relay has been installed there to sense
the fault and then show it on the fault display unit. Micro-controller has been used to operate relay.
Micro-controller has been equipped to diagnosis the type of the fault. When fault has been occurred,
GSM approach is used to send SMS to the concerned person. Fault display section is used to display
the type of fault.
Modern Methods
3: Pilot Approach
Major benefit of the pilot approach as compared to other neural networks is that it is setting free. There
is no need to fix the value of the parameters in this approach. Pilot pattern arrangements have to be
implemented on the protection system of the transmission line. But there are no hard and fast rules to
implement the pilot scheme on protection system because pilot approach can easily apply to the existing
protection system. Second advantage of this approach is that it does not bother to identify the fault
starting time and detection of the fault locality. Pilot scheme is used to narrate the signs of reactive
power dignified by different relays.
4: Pattern Recognition Approach
With the implementation of pattern recognition algorithm, we can easily diagnosis the fault and locate
the fault which has been occurred in transmission line. A new methodology related to neural networks
has been used for fault classification by this pattern recognition approach. Inputs should be voltage and
the current signals and generates an output having four signals These four signals helps to differentiates
the different categories of the faults such as phase to ground fault, dual phase to ground fault, dual
phase fault and triple phase faults.
Methodology
Artificial Neural Network
Artificial neural network control and the ANN plant modeling are approximately similar. The only
difference between them is that we do not have a desired output to operate the controller network.
The less information about the function will considerably increase the complexity of the entire
process. The way out to this problem is to use an ANN model of the plant in parallel and utilize it to
estimate the preferred output of the controller.
Methodology
Methodology
Multi-layer Perceptron
Multi-layer approach is basically the extension of forward neural network. In multi-layer approach similar to
forward network data move from input to output side. But in multi-layer approach, we can also perform the
backward procedure. Multi-layer approach is used to solve the problems having nonlinearity in it. Majorly the
frequent use of multi-layer approach is for design formation, calculation and estimation. Multi-layer approach
composed of different number of input, hidden and output layers.
Methodology
Simulation and Results
1: Layer Recurrent Method

There is a feedback loop in layer recurrent method which enables the data to be reprocessed back into the
input. Biases and weights have been updated. This update is based on cost function gradient at several
dimensional values. Variations in weights and biases can enable the network to strengthen the connections
with neurons. The design of the neural network contains 4, 8 and 4 layers and containing 15 neurons in
each hidden layer.
1: Layer Recurrent Method
2: Feed forward Back propagation Method
There is no feedback loop in feed forward method. This approach is preferably used where the
information is neither time-dependent nor consecutive. This network has four inputs, four outputs
and 6 hidden layers containing 20 neurons in each layer.
2: Feed forward Back propagation Method
3: Elman Back-Propagation Method
There is a feedback loop in Elman method, which is based on the back-propagation. In this method
hidden layer contains another commence layer which operates as a delay operator. The basic purpose
of this layer is to memorize the information so the system stability increases. This network has four
inputs and outputs and 6 hidden layers containing 20 neurons in each layer.
3: Elman Back-Propagation Method
4: Cascade-Forward Back Propagation Method

Cascade-Forward method is similar to the Feed-Forward method. The only difference between them
is that cascade-forward method includes a link from input and every preceding layer to the next
layer. Output is directly connected to input. This network has four inputs and outputs and 6 hidden
layers containing 20 neurons in each layer.
4: Cascade-Forward Back Propagation
Method
5: Perceptron Method
A perceptron is also a neural network that performs certain calculations to analyze the input data. If a
function has input “x” then it is multiplied by weight and biases and provides the output as f(x). This
network has four inputs and outputs and 1 hidden layer containing 4 neurons.
5: Perceptron Method
6: Feed-Forward Distributed Time Delay
Method
Feed-Forward Distributed Time Delay is similar to the Feed-Forward network. The only difference is
that every input and weights of each and every layer has a delay track connected with it. This
procedure enables the network to have limited dynamic reaction to input which is in time series. This
network has four inputs and outputs and 6 hidden layer containing 15 neurons in each layer.
6: Feed-Forward Distributed Time Delay
Method
7: NARX (Nonlinear autoregressive system
having exogenous inputs)
In a cost-effective model, an exogenous variable is who which is measured outside the network and
then applied the measured value on the model. This network has four inputs and outputs and 4
hidden layer containing 20 neurons in each layer.
7: NARX (Nonlinear autoregressive system
having exogenous inputs)
RMS values of given Methodologies
APPROCHES RMS VALUES
Artificial Neural PHASE A PHASE B PHASE C GROUND
Networks

Layer Recurrent

Feed-Forward Back 0.02376 0.14404 0.08858 0.00599

Propagation

Elman Back-Propagation 0.18204 0.16464 0.15883 0.16356

Cascade-Forward Back- 0.09146 0.16033 0.12234 0.198766

Propagation

Perceptron 0.0667 0.0667 0.0667 0.0333

Feed-Forward 8.138516e-07 0.068110 0.012062 1.2188e-07
Distributed Time Delay

NARX 0.003547 0.148761 0.014579 0.003579

Data Analysis using Multi-layer Perceptron
Five folds have been collected and fault analysis is performed on transmission line at 4 different
places. These four cases are:
Case-I: Fault at sending end.
Case-II: Sudden Loss of Line
Case-III: Fault in Middle of Line
Case-IV: Fault in Double circuit Line.
Data Analysis using Multi-layer Perceptron
Fold Class Correct Prediction Accuracy Fold Class Correct Prediction Accuracy
Case-1 (6/7) 85% Case-1 (7/7) 100%
Fold-1 Case-2 (5/6) 83% Fold-1 Case-2 (6/6) 100%
Case-3 (7/7) 100% Case-3 (7/7) 100%
Case-4 (6/6) 100% Case-4 (6/6) 100%
Total Accuracy (Fold-1) (24/26) 92% Total Accuracy (Fold-1) (26/26) 100%
Fold-2 Case-1 (6/7) 85% Fold-2 Case-1 (7/7) 100%
Case-2 (5/6) 83% Case-2 (6/6) 100%
Case-3 (7/7) 100% Case-3 (6/6) 100%
Case-4 (6/6) 100% Case-4 (7/7) 100%
Total Accuracy (Fold-2) (24/26) 92% Total Accuracy (Fold-2) (26/26) 100%
Fold-3 Case-1 (6/6) 100% Fold-3 Case-1 (6/6) 100%
Case-2 (6/7) 85% Case-2 (7/7) 100%
Case-3 (7/7) 100% Case-3 (6/6) 100%
Case-4 (6/6) 100% Case-4 (7/7) 100%
Total Accuracy (Fold-3) (25/26) 96% Total Accuracy (Fold-3) (26/26) 100%
Fold-4 Case-1 (6/6) 100% Fold-4 Case-1 (6/6) 100%
Case-2 (4/7) 57% Case-2 (7/7) 100%
Case-3 (4/6) 66% Case-3 (5/6) 83%
Case-4 (6/6) 100% Case-4 (6/6) 100%
Total Accuracy (Fold-4) (20/25) 80% Total Accuracy (Fold-4) (24/25) 96%
Fold-5 Case-1 (6/6) 100% Fold-5 Case-1 (6/6) 100%
Case-2 (1/6) 16% Case-2 (6/6) 100%
Case-3 (6/7) 85% Case-3 (7/7) 100%
Case-4 (6/6) 100% Case-4 (6/6) 100%
Total Accuracy (Fold-5) (24/26) 76% Total Accuracy (Fold-5) (26/26) 100%
Total Accuracy (112/128) 87% Total Accuracy (127/128) 99%
Results without MLP
Results equipped with MLP
Conclusion
 Different Artificial Neural Networks approaches have been developed to solve problems in power
systems and power plants i.e. load identification, load modeling, forecasting and fault analysis, in
power systems, and sensor validation, monitoring and fault analysis, in power plants.
 This project tried to resolve these losses by implementing seven different artificial neural networks
and MLP as well. These artificial neural networks and MLP have been equipped with modern
learning techniques known as forward propagation and back propagation, and its multiple layers
with multiple neurons in it and capability of tackling nonlinear functions distinguished these
approaches from other approaches. By implementing ANN in this project, system accuracy is
nearly approaches to 100%. Faults can easily detect and resolved by using these approaches.
Future Work
 In Upcoming times, Neural Networks can broadly implement in power transmission system to
improve the steadiness and reliability of the system. The vision of ANN is broad enough and can
be further explored.
 The approach discussed in this project (MLP-ANN) can also be used in future for system better
accuracy. Problem approximation has been done by RBF-ANN. Peculiarity of this approach is its
universal approximation and rapid learning speed. For the solution of complicated problems
related to neural networks can be possible with the implementation of modular NN. The major
benefit of this approach is its uncomplicatedness.