PERPUSTAKAAN UTHMeprints.uthm.edu.my/id/eprint/817/1/24_Pages_from...analisis bagi model yang berbeza iaitu NN1, NN2, NN3 dan NN4 yang mana model terbaik akan dipilih bagi analisis

PERPUSTAKAAN UTHM

*30000001883623*

KOLEJ UNIVERSITITEKNOLOGI TUN HUSSEIN ONN

PENGESAHAN STATUS LAPORAN PROJEK SARJANA

FORECASTING SUNSPOT NUMBERS USING NEURAL NETWORK: EFFECT TO THE

ELECTRICAL SYSTEM

SESIPENGAJIAN: 2006/2007

Saya REZA EZUAN BIN SAMIN mengaku membenarkan Laporan Projek Sarjana ini disimpan di Perpustakaan dengan syarat-syarat kegunaan seperti berikut:

1. 2 . 3.

Laporan Projek Sarjana adalah hakmilik Kolej Universiti Teknologi Tun Hussein Onn. Perpustakaan dibenarkan membuat salinan untuk tujuan pengajian sahaja. Perpustakaan dibenarkan membuat salinan tesis ini sebagai bahan pertukaran antara institusi pengajian tinggi. ** Sila tandakan (V)

• • HI

SULIT

TERHAD

TIDAK TERHAD

(Mengandungi maklumat yang berdarjah keselamatan atau kepentingan Malaysia seperti yang termaktub di dalam AKTA RAHSIA RASMI 1972)

(Mengandungi maklumat TERHAD yang telah ditentukan oleh organisasi/badan di mana penyelidikan dijalankan

Disahkan oleh

(TANDATANGAN PENULIS)

Alamat Tetap:

7 (TANDATAF JG iN PENYELIA)

63 J ALAN BESAR,

TONGKANG PECAH,

83010 BATU PAHAT,

JOHOR.

Tarikh:

PROF. MADYA SITI HAWA BT RUSLAN

Nama Penyelia

Tarikh:

CATATAN: ** Jika Laporan Projek Sarjana ini SULIT atau TERHAD, sila lampirkan surat

daripada pihak berkuasa/organisasi berkenaan dengan menyatakan sekali sebab dan tempoh laporan ini perlu di kelaskan sebagai SULIT atau TERHAD.

FORECASTING SUNSPOT NUMBERS USING NEURAL NETWORK: EFFECT

TO THE ELECTRICAL SYSTEM

REZA EZUAN BIN SAMIN

A project report submitted in partial fulfillment of the requirements for

the award of the degree of

Master of Electrical Engineering

Faculty of Electrical and Electronics Engineering

Kolej Universiti Teknologi Tun Hussein Onn

NOVEMBER, 2006

"I hereby declare that the work in this report in my own except for quotations and

summaries which have been duly acknowledged"

Student

REZA EZUAN BIN SAMIN

_ ; ^ g / ( ] / ^ o f c

Supervised by

Supervisor I

Supervisor II

DR. AZME BIN KHAMIS

iii

ACKNOWLEDGEMENT

Assalamualaikum. First of all I would like to thank Allah The Almighty for

giving me the strength to complete my research as one of the requirement for my Master

degree.

I would also like to thank Associate Professor Siti Hawa Bt. Ruslan, my project

supervisor for her support and encouragement all the time especially during the

difficulties that I faced during the completion of my research. Not to forget, Dr Azme

Bin Khamis, my project co-supervisor for his guidance and opinion especially in the

area of Neural Network.

I would also like to thank my parents for their support, love and understanding

during the completion of my Master study. I also like to thank my beloved fiance for the

understanding and support all the time.

Last but not least, to all my colleagues that also gave the encouragement and

opinion in making my research a great success. Wassalam.

IV

ABSTRACT

The purpose of this research is to develop the forecasting system of Sunspot

Numbers that highly related to Geomagnetic Induced Current (GIC). This geomagnetic

induced current (GIC) have the effect to the electrical system especially to the

transformers. Sunspot data obtained from the National Geophysical Data Center

(NGDC) ranging from 1700 until 2005 is analyzed using Neural Network (NN) using

the MATLAB 7.0 Graphic User Interface (GUI) method computer program called

"Sunspot Neural Forecaster" so that the analysis and simulation of the sunspot data can

be done easily and more user friendly. First, a comparison analysis between

Feedforward Neural Network (FNN) and Recurrent Neural Network (RNN) is done to

choose the best NN type for the next analysis. The second stage of the analysis involved

the selection of NN training algorithm between Levenberg Marquardt, Resilient

Backpropagation and Gradient Descent. As in the selection of NN type analysis, the

best NN training algorithm is selected for the next analysis. The next analysis involved

the selection of NN models between NN1, NN2, NN3 and NN4 and the best models is

selected for the last analysis which is the transfer function analysis. The NN transfer

function analysis involved Tansig/Purelin and Logsig/Purelin transfer function for the

hidden layer and output layer respectively. Based from the analysis that have been done,

FNN using Levenberg Marquardt training algorithm with NN2 model and

Tansig/Purelin transfer function are used for forecasting the sunspot data. The

forecasting result obtained shows the system managed to forecast the sunspot numbers.

V

ABSTRAK

Tujuan penyelidikan ini adalah untuk membangunkan suaru sistem ramalan

"Sunspot Neural Forecaster" bagi meramal nombor bintik suria (Sunspot Numbers) yang

mempunyai kesan terhadap arus teraruh geomagnetik, {Geomagnetic Induced Current,

GIC). Fenomena GIC ini memberi impak kepada sistem elektrik terutamanya sistem

transformer. Data bagi penyelidikan ini diperolehi daripada National Geophysical Data

Center (NGDC) dari tahun 1700 hingga 2005. Analisis kemudian dijalankan

menggunakan data tersebut dengan menggunakan rangkaian neural menggunakan

perisian MATLAB 7.0 diberi nama "Sunspot Neural Forecaster" menggunakan kaedah

GUI agar analisis bagi nombor bintik suria dapat dibuat dengan lebih mudah serta

mesra pengguna. Pada peringkat awal, analisis perbandingan dibuat antara FNN dan

RNN dan rangkaian neural terbaik dipilih bagi analisis seterusnya. Analisis seterusnya

merupakan analisis bagi algoritma pembelajaran yang berbeza. Tiga jenis analisis

algoritma pembelajaran telah dibuat iaitu Levenberg Marquardt, Resilient

Backpropagtion dan Gradient Descent dan algoritma pembelajaran yang memberikan

prestasi terbaik akan dipilih bagi analisis seterusnya. Analisis seterusnya merupakan

analisis bagi model yang berbeza iaitu NN1, NN2, NN3 dan NN4 yang mana model

terbaik akan dipilih bagi analisis terakhir iaitu analisis bagi fungsi pindah yang berbeza

iaitu Tansig/Purelin dan Logsig/Purelin bagi fungsi pindah pada lapisan tersembunyi

serta lapisan keluaran. Berdasarkan analisis-analisis yang telah dibuat, suatu model

rangkaian neural menggunakan FNN dengan algoritma pembelajaran Levenberg

Marquardt menggunakan model NN2 serta Tansig/Purelin masing-masing sebagai

fungsi pindah pada lapisan tersembunyi dan lapisan keluaran digunakan bagi meramal

nombor bintik suria dengan tepat.

vi

TABLE OF CONTENTS

CHAPTER TITLE PAGE

ACKNOWLEDGEMENT iii

ABSTRACT iv

ABSTRAK v

TABLE OF CONTENTS vi

LIST OF TABLES ix

LIST OF FIGURES x

LIST OF SYMBOLS xiii

LIST OF ABBREVIATION xiv

LIST OF APPENDICES xvi

I INTRODUCTION 1

1.1. Research Background 1

1.2. Problem Statement 2

1.3. Importance of Study 3

1.4. Research Objective 3

1.5. Scope of Project 4

1.6. Thesis Outline 5

vii

II LITERATURE STUDY 7

2.1 Introduction 7

2.2 What Is Sunspot Numbers 8

2.3 Sunspot Numbers & Geomagnetic Induced Current 11

2.4 GIC and Its Effect to Electrical System 12

2.5 Sunspot Forecasting 20

2.6 The Role of Forecast 24

ffl RESEARCH METHODOLOGY 27

3.1 Introduction 27

3.2 Introduction to Neural Network 27

3.2.1 Feedforward Neural Network 31

3.2.2 Recurrent Neural Network 33

3.2.3 Training Algorithm 34

3.2.4 Transfer Function 35

3.2.5 Improving Generalization 38

3.2.6 Neural Network Application in Forecasting 39

3.3 Development of NN System 40

3.3.1 Data Collection 41

3.3.2 Preparing of Input and Output Data 41

3.3.3 Design of Neural Network Model 42

3.3.4 Network Training 44

3.4 Summary 44

IV PROGRAMMING & GRAPHIC USER INTERFACE (GUI) 45

4.1 Introduction 45

4.2 GUI for Sunspot Neural Forecaster 46

4.2.1 Simulation for MSE and Correlation Analysis 50

viii

V RESULT AND DISCUSSION 53

5.1 Introduction 53

5.2 Selection of NN Type 54

5.3 Feedforward Neural Network Analysis 5 6

5.3.1 Selection of Training Algorithm 56

5.3.2 Selection of Model 60

5.3.3 Selection of Transfer Function 73

5.3.4 Optimized NN parameters analysis 77

5.4 Sunspot numbers forecasting 79

VI CONCLUSION 82

6.1 Conclusion 82

6.2 Recommendation for Future Works 83

REFERENCES 86

APPENDIX 91

ix

LIST OF TABLES

3.1 Combination of transfer function 43

5.1 MSE performance analysis for different NN type 54

5.2 MSE Performance analysis for different training

algorithm. 57

5.3 NN1 MSE and correlation analysis 61




5.7 Average MSE performance analysis for different

models 71

5.8 MSE performance analysis for different transfer

function 74

5.9 Forecast value of sunspot numbers 80

X

LIST OF FIGURES

NO OF FIGURE TITLE PAGE

1.1 Time series plot for Sunspot number from 1700 until 2005 4

2.1 Sunspots observed from the sun 9

2.2 Sunspot and geomagnetic activity 12

2.3 Six steps of sunspot chain from the Sun to the ground 13

2.4 Ejection from the sun travels to earth and distorts earth

magnetic field 13

2.5 Half cycle saturation of power transformers due to GIC 15

2.6 Relationship between sunspot numbers and major

transformer breakdown due to GIC 16

2.7 Observed Regional GIC Index (RGI) as measured at

the Ottawa observatory on 12-14 March 1989 18

2.8 Disturbance environments observed by region on

13 March 1989 19

2.9 The Salem nuclear plant transformer damage due to GIC

half cycle saturation of transformer on 13-14 March 1989 20

3.1 Main components of neurons 28

3.2 Neural network main components 29

3.3 Feedforward Neural Network (FNN) 32

3.4 Recurrent Neural Network (RNN) 33

3.5 Linear transfer function 36

3.6 Log Sigmoid transfer function 37

xi

3.7 Tangent Sigmoid transfer function 37

3.8 NN system development flow 40

4.1 GUI for "Sunspot Neural Forecaster" 46

4.2 Actual vs NN prediction 47

4.3 Multiple hidden nodes analysis at MATLAB command

window 49

4.4 GUI simulation for MSE and correlation analysis 50

4.5 Flow chart manual for "Sunspot Neural Forecaster" 52

5.1 FNN MSE performance analysis 55

5.2 RNN MSE performance analysis 55

5.3 MSE performance for Resilient Backpropagation algorithm 58

5.4 MSE performance for Gradient Descent 58

5.5 MSE performance for Levenberg Marquardt 59

5.6 NN1 MSE performance analysis 62

5.7 NN1 correlation analysis 62







5.14 MSE training performance for different models 69

5.15 MSE validation performance for different models 70

5.16 MSE testing performance for different models 70

5.17 Average MSE performance for different models 72

5.18 NN2 Tansig/Purelin MSE performance analysis 75

5.19 NN2 Logsig/Purelin MSE performance analysis 75

5.20 NN2 average MSE performance analysis for different

transfer function 76

5.21 Actual vs. NN prediction for optimized NN parameters 78

5.22 Current & forecast value of sunspot number (1700-2025) 79

Forecast value of sunspot numbers

De-rectification of the sunspot numbers

LIST OF SYMBOLS

Transfer function

Connection matrix from input layer to hidden layer

Bias vector

Connection matrix from hidden layer to output layer

Function

Nonlinear mapping

Function argument

Input

Time lag

Number of observation

Actual input

Actual inputs at their maximum

Actual inputs at their minimum

Scaled input

Actual observation

Output of model

Output vector

Observation with m input

Observation with m+1 input

Observation with N-m patterns

Observation at time t

LIST OF ABBREVIATIONS

ARV Average Relative Variance

CME Coronal Mass Ejection

CT Current Transformer

DSF Disappearing Filaments

EHV Extra High Voltage

EPvNN Elman Recurrent Neural Network

FFN Feedforward Neural Network

GA Genetic Algorithm

GEA Genetic and Evolutionary Algorithm

GIC Geomagnetic Induced Current

GMDH Group Method of Data Handling

GRNN General Regression Neural Network

GUI Graphic User Interface

MAE Mean Absolute Error

MAPE Mean Absolute Percentage Error

MLP Multi Layer Perceptron

MSE Mean Square Error

NGDC National Geophysical Data Center.

NN Neural Network

RGI Regional GIC Index

RNN . Recurrent Neural Network

SETAR Self Exciting Threshold Autoregressive

SSN Sunspot Numbers

Self Organizing Map

Time delay Added Evolutionary Forecasting

LIST OF APPENDICES

APPENDIX TITLE

A. 1 Sunspot Numbers Data (1700-2005)

CHAPTER I

INTRODUCTION

1.1. Research Background

Solar activity forecasting is an important topic for various scientific and

technological areas like space activities related to operation of low earth orbiting

satellites, electric power transmission line, high frequency radio communications and

geophysical applications. The particles and electromagnetic radiations flowing from

solar activity outbursts are also important for long term climate variations and thus it is

very important to know in advance the phase and amplitude of the next solar and

geomagnetic cycles.

Nevertheless, the solar cycle or sunspot numbers is very difficult to predict on

the basis of time series of various proposed indicators, due to high frequency content,

noise contamination, high dispersion level and high variability both in phase and

amplitude, with intermittent behavior at different scales. This topic is also complicated

by the lack of a quantitative theoretical model of the Sun's magnetic cycle. Many

attempts to predict the future behavior of the solar or sunspot activity are well

2

documented in the literature. Numerous techniques for forecasting are developed to

accurately predict phase and amplitude of future solar cycles, but with limited success.

Depending on the nature of the prediction methods, five classes can be distinguish: 1)

Curve fitting; 2) Precursor; 3) Spectral; 4) Neural Networks; 5) Climatology.

Several method of forecasting the sunspot numbers have been developed by M.

Salvatore and C. Francesco (2006), Dmitriev A.V et.al (1999), Fessant, F, Bengio, S and

Collobert, D. (2000) and L. Ming (1990). All of the researchers have used Neural

Network (NN) in the forecasting system. In term of the NN method, there are many NN

type such as Feedforward Neural Network (FNN) and Recurrent Neural Network (RNN)

that have been used by the researchers and each one have their own reason in selecting

the NN type that they have chosen.

1.2. Problem Statement

A model is to be extracted from the Sunspot Number (SSN) ranging back from

1700 until 2005. This model will be used to forecast the next sunspot number from year

2006 until 2025. In order to forecast the model, Neural Network (NN) system will be

used using the MATLAB 7.0 software.

3

1.3. Importance of Study

It is hope that by forecasting the Sunspot Number, it will help as a preventive

action in protecting our electrical system due to the effect of Geomagnetic Induced

Current (GIC). This is due to that sunspot numbers is highly related with the GIC

phenomena.

1.4. Research Objective

The objectives of this research are as follows:

1. To develop a prototype forecast system for predicting the solar activity using

MATLAB software. Instead of using the ordinary and less user friendly command

window in MATLAB, more user friendly graphic user interface (GUI) is used. By

forecasting the related data, it is hope that it will help in preventing the Geomagnetic

Induced Current (GIC) from affecting the electrical system.

2. To determine the effect of NN parameters such as number of hidden nodes, transfer

function and learning algorithm to the performance of the system.

3. To determine the optimum NN parameters in order to forecast the sunspot numbers.

4

1.5. Scope of Project

This project presents the NN applications for the development of expert system

for forecasting the solar activity based on the sunspot data that strongly affect the earth

communication operation. For the analysis and development of the system, MATLAB

7.0 will be used.

The sunspot data ranging from 1700 until 2005 that was used in this research was

obtained from the National Geophysical Data Center (NGDC) through the ftp server:

ftp://ftp.ngdc.noaa.gov/STP/SOLAR DATA/SUNSPOT NUMBERS/. Figure 1.1

indicates the time series plot for Sunspot numbers from 1700 until 2005. The complete

sunspot data can be seen in the Appendix.

200 180

Year

Figure 1.1: Time series plot for Sunspot number from 1700 until 2005.

ftp://ftp.ngdc.noaa.gov/STP/SOLAR

5

1.6 Thesis Outline

The next chapter will focus on the literature study and brief explanation about the

effect of sunspot numbers to the electrical system. In term of the literature study, it will

not only discuss about the NN method in forecasting the sunspot numbers but also other

methods such as time series and genetic algorithm.

Chapter 3 will discuss on the research methodology which were used in this

research. Brief explanation about NN and the NN parameters that will be used in the

analysis will be made. Furthermore, the procedure of the NN development in this

research will also be discussed.

Chapter 4 will discuss about the programming and the graphic user interface

(GUI) that have been developed using MATLAB 7.0 software. In this chapter also, brief

explanation about how to use the "Sunspot Neural Forecaster" interface will also be

highlighted.

Results and discussion about all the NN analysis are in chapter 5. In this chapter,

the analysis begins with comparison analysis between Feedforward Neural Network

(FNN) and Recurrent Neural Network (RNN). Then the analysis proceeds to the

training algorithm analysis where different training algorithm will be compared in order

to get the best training algorithm. The next stage of the analysis involved analysis for

different models. The models are NN1, NN2, NN3 and NN4. As in the previous

analysis, only the best models will be selected for the next analysis. Finally, the analysis

for different transfer function was done. The combination transfer functions that

involved were Tansig/Purelin and Logsig/Purelin for hidden layer and output layer

respectively. After all the analyses have been done, the optimized NN parameters were

Documents

PERPUSTAKAAN UTHMeprints.uthm.edu.my/id/eprint/817/1/24_Pages_from...analisis bagi model yang berbeza iaitu NN1, NN2, NN3 dan NN4 yang mana model terbaik akan dipilih bagi analisis