You are on page 1of 7

APPLICATION OF ARTIFICIAL NEURAL NETWORK

FOR FAULT DETECTION IN POWER SYSTEMS

Rahul Kumar Jha


Tribhuwan University

Abstract: Due to a variety of unpredictable factors, electrical power systems frequently


experience unexpected breakdowns. It is necessary to stop unanticipated faults from spreading to
other parts of the protective system when they occur in power systems. The roles of the
protective systems are to identify, categorize, and then pinpoint the location of the
malfunctioning voltage and/or current lines. Finally, a signal from the protective relay to the
circuit breaker is required to isolate the defective line. Pattern classifiers are effective uses of NN
employed as an intelligent way of detection due to its capacity to learn, generalize, and parallel
process. In order to identify faults in a transmission line power system, this research provides
neural network NN architecture. It seeks to put into action a comprehensive plan for distance
protection that is separated into several neural network zones. Faults from single phase to
ground, double phase, and double phase to ground are all taken into account. As a result, the
NNBP architecture may be used to provide a protective relaying system for power transmission
line systems.

Keywords: - Power system protection, Fault Identification, Neural Network Architecture,


Transmission Lines Protection, . Pattern classifiers

1. INTRODUCTION

The greatest threat to the continuous supply of power is electrical system breakdowns. Electric
power system malfunctions are an inherent issue. As a result, a well-coordinated protection
system must be installed to quickly identify and isolate problems in order to limit the harm and
disruption caused to the electrical system. Devices that can detect faults, respond rapidly, and
disconnect the problematic part are typically used to clear them. Similarly, Active and passive
faults are the two basic categories into which faults may be generically categorized.
An artificial neural network (ANN) is a representation of the human brain in its most basic form.
A natural brain can pick up new skills and adjust to a changing environment. The brain has the
most extraordinary ability to interpret ambiguous, hazy, and incomplete information and draw its
own conclusions from it. For instance, even though the handwriting of another person may be
radically different from our own, we are nevertheless able to read it. A young toddler can
recognize that an orange and a ball both have a round form. A baby as young as a few days old
can identify its mother by touch, speech, and smell. Even in grainy photos, we can recognize a
familiar face.
The brain, which governs the entire body, is a very complicated organ. Even the most
rudimentary animal's brain is more powerful than the most sophisticated machine. Its role
extends beyond simply regulating the body's basic functions to include more sophisticated tasks
like thinking, visualizing, dreaming, envisioning, learning, etc., which are not physically

1
measurable. Even the most sophisticated supercomputers cannot simulate the capabilities of an
artificially intelligent machine.
Basically, we can create and train neural networks to handle specific issues that are challenging
for humans or traditional computer techniques to address. The weight modifications, which are
the main components of the ANN, are what give the training its computational significance. One
of the main distinctions between the neural network method of problem solving and traditional
computer techniques is this.
Thus, the method for locating the fault type and its zone in the line is presented in this work. The
backpropagation neural network method is investigated and used. These three jobs are seen to be
carried out by the line's voltage and current signals. As criteria for fault classification, the
detailed coefficients of all phase current signals, which are only gathered at the transmitting end
of a transmission line, are chosen. The Power System Block Set that comes with MATLAB
(2021a) and the transmission line models are used to simulate and build the information that is
subsequently routed and to provides back-propagation neural networks.

II. TRANSMISSION LINE MODEL

The 110 kV transmission line system connecting two cities is modeled using the AC overhead
transmission lines Pi model network. The recommended designs and methods for this challenge
are developed and implemented over a 320 Kilometers line. The technology used to develop and
evaluate neural networks is depicted in a single-line graphic in Fig. 1. Two Substations form up
the entire system, one in each city at the transmission line's start and finish.
NEA Dhakdhai 33/11 Sub Station, Dhakdhai, 132/33KV Grid Sub Station Phulparas,
Unnamed Road, Basant Pur, Bihar

A B
Zone 1 (50 K m)

Zone 2 (176 K m)

Zone 3 (244 K m)

Zone 4 (286 K m)

320 k m

Fig. 1 Dhakdhai- Phulparas, Transmission Line

The three-phase voltages and currents at substation A are shown in Fig. 1 as V = [Va Vb Vc]T and
I = [Ia Ib Ic]T. Phase to ground faults, phase to phase faults, and double-phase to ground faults are
the three simulation types that are described.

2
III. CONVENTIONAL BACKPROPAGATION ALGORITH

Figure 1: Neural Network Processing

In theory, a network produces an output via forward propagating activation and determines
weight changes by backward propagating error. The supplied values are mediated in both ways
by the weights on the connections between neurons. The backpropagation algorithm may be
defined as a method that trains a certain feed-forward neural network for a specific input pattern
where the classifications are known to us. To determine the weights of a multilayered neural
network with a fixed design, the backpropagation algorithm is used. In order to reduce the total
squared error between the network's output values and the specified target values, gradient
descent is used i.e. (The network examines its yield response to the example input pattern after
each passage of the example set has been shown to it. The comparison between the output
response and the anticipated output is then made, and the error value is measured.) The training
patterns are spread across the network to produce the output of every hidden layer and the output

layer.

3
Figure 1: Neural Network Components

4
In figure 2: Error (shown as a lined pattern) at the output node and activation (shown as a solid
pattern) at the hidden node determine how much of a hidden to output weight changes. While
activation at the input node and error at the hidden node (which in turn depends on error at all the
output nodes) both affect the change in an input to hidden weight.

Notations
• wkj denotes a weight from the hidden to the output layer.
• wji denotes a weight from the input to the hidden layer.
• a denotes an activation value.
• t denotes a target value.
• net denotes the net input.

Review of Calculus Rules


d(eu ) / dx = eu * (du/dx)
d(g + h) / dx = (dg/dx) + (dh/dx)
d(gn )/dx = ngn−1 (dg/dx)

The total error in a network is given by the following equation

1
E = 2 ∑ ( t k −ak ) ……………………………(i)
2

k
∂E
∆ W ∝− ………………………………….(ii)
∂W
∂E
∆ wkj ∝−
∂ w kj
………………………………(iii)
We expand this as

(iv) Derivative of the error with respect to the activation

(v) Derivative of the activation with respect to the net input

Using this fact, we can rewrite the result of the partial derivative as:
ak(1 − ak)
5
6
(vi) Derivative of the net input with respect to a weight

(vii) Weight change rule for a hidden to output weight

∆wkj = ε (tk − ak)ak(1 − ak) aj = ε δk aj

This equation is typically simplified as shown below where the δ term repesents the product of
the error with the derivative of the activation function.
∆wkj = ε δk aj

(viii) Weight change rule for an input to hidden weight

Final Equation is ∆wji = ε δj ai

IV. DESIGN OF ANN FOR TRANSMISSION LINE

You might also like