Lb /DOM / tf o /o 7 
NEURAL NETWORK BASED PREDICTION 
FOR OPTIMUM POWER SYSTEM 
OPERATION 
A dissertation submitted to the 
Department of Electrical Engineering, University of Moratuwa 
in partial fulfillment of the requirement for the 
Degree of Master of Science 
2005/2006 
By 
DON SAMAN RIENZIE ALAHAKOON 
L I B R A R Y 
UNIVERSITY OF MORATUWA, SRI LANKA 
M O R A T U W A 
Supervised by 
Dr. Lanka Udawatta 
Department of Electrical Engineering, 
University of Moratuwa 
Sri Lanka 
December 2006 
University of Moratuwa 
87312 
3 ( 0 * 3 ) 
DECLARATION 
The work submitted in this dissertation is the result of my own investigation, 
except where otherwise stated. 
It has not already been accepted for any degree, and is also not being 
concurrently submitted for any other degree. 
D.S.R. Alahakoon 
Date: December 11, 2006 
I endorse the declaration by the candidate. 
Dr. Lanka Udawatta 
i 
CONTENTS 
Declaration i 
Abstract iv 
Acknowledgement v 
List of Figures vi 
List of Tables vii 
Attachments vii 
1 Introduction to the working environment 1 
1.1 Background 1 
1.2 Present status of the Electrical Power Sector 2 
1.3 Generation 3 
1.3.1 Mahaweli Complex 3 
1.3.2 Laxapana Complex 4 
1.3.3 Other Hydro 5 
1.3.4 Thermal Plants-CEB 5 
1.3.5 Thermal Plants-IPP 6 
1.4 Loads 6 
1.5 Goal 7 
2 Problem Identification 8 
2.1 Properties of the load curve 8 
2.1.1 Introduction 8 
2.1.2 Load factor 8 
2.2 Leading Methods for Forecasting Demand 9 
2.2.1 Time Series Analysis 10 
2.2.2 Regression-based Approaches 10 
2.2.3 Spline or Piecewise Regression 11 
2.2.4 Similar Day Lookup 11 
2.2.5 Artificial Neural Networks 11 
2.2.5.1 Introduction 11 
2.2.5.2 Human Neurones to Artificial Neurones 14 
2.2.5.3 A simple neuron 14 
2.2.5.4 Neural networks in load forecasting 18 
2.2.5.5 Multi-Layer Perceptrori network (MLP) 19 
2.2.5.5.1 Description of the network 19 
2.2.5.5.2 Learning 20 
3 Development of the Software model 21 
3.1 Program structure 21 
3.2 Half hour forecasting 21 
3.2.1 Input variables 21 
3.2.2 Architecture 22 
3.2.3 Algorithm 23 
3.2.4 Training 24 
3.2.5 Application 24 
3.2.6 Accuracy improvement 25 
3.3 24 hour forecasting 26 
3.3.1 Input variables 26 
3.3.2 Variation of temperature, Humidity & Electrical energy 27 
3.3.3 Selection of Architecture 29 
3.3.4 Application 33 
3.3.4.1 Comparison after few months 33 
. » 
ii 
3.3.4.2 Load curve for deferent scenarios 35 
3.3.4.3 Special day forecasting 43 
3.3.5 Test results 44 
3.3.6 Error messages 45 
4 Economic dispatch 46 
4.1 Objective 46 
4.2 Background 46 
4.3 Unit cost 47 
4.3.1 Thermal plants 47 
4.3.2 Hydro plants 47 
4.4 Generation Models 47 
4.4.1 Introduction 47 
4.4.2 Formulation of the LaGrangian 48 
4.4.3 KK.T Conditions for a 2-unit system 49 
4.4.4 Graphical representation of Economic Dispatching 49 
4.4.5 The Lamda-iteration procedure 50 
4.5 Application 51 
4.5.1 Machine availability 51 
4.5.2 Operational policy 51 
4.5.3 Comparison of real & ECD 52 
4.5.4 Comparison of the combination of Load prediction and ECD 55 
5 Discussion, Conclusions & Future improvements 59 
5.1 Discussion and Conclusion 59 
5.2 Future improvements 60 
6 References 62 
7 Appendices 63 
iii 
Abstract 
Neural network techniques are widely use for Load forecasting and accuracy 
depends on the No. of past data, Network structure & influencing factors to 
Electricity demand, such as Day of the week, Month of the year (reflect 
whether, sun rise/set times - monthly cyclic patterns), Temperature, Humidity, 
Wind, Public Holidays etc. Western province of Sri Lanka consumes major 
part of Electricity generation, than other areas. So any whether pattern change 
in other areas would not be affected to the demand pattern considerably. But 
night peak this is not true. 
By examining the past load curve patterns, it is revealed that the major 
influencing factors are time of the day, Day No., Month No., Public Holiday 
status & School day or not others are minor factors. But however temperature 
& Humidity also contribute to some extent, so these two factors also 
considered. Running pattern of Mini-Hydro plants has not been monitoring by 
the System control Centre, Therefore the loading pattern of those plants is not 
considered. But it is understood that the running pattern depends on the rain 
fall of particular area. These all plants are run of river plants, so during rainy 
season almost all plants runs their full capacity (around 80MW). 
The main idea of this exercise is to develop a fairly accurate method of load 
forecasting by using Neural networks and prepare an Economic dispatch 
schedule at any given time, which is very useful for day to day power system 
operations. 
Neural network tool box functions & graphical user interface in MATHLAB 
version 6.5 is used to develop the neural network and to prepare the Machine 
dispatch schedule. 
IV 
Acknowledgement 
My heartfelt thanks also go to my course coordinator and lecturers of 
postgraduate study course in Electrical Engineering, University of Moratuwa, 
Sri Lanka, who gave me the theoretical knowledge and encouragement in 
bringing up this academic work in time with excellent corporation and 
guidance. I extend my very sincere thanks to my worthy supervisor Dr. Lanka 
Udawatta. 
My sincere gratitude is also extended to the support staff of the Department of 
Electrical Engineering for helping in various ways during the course of study. 
Finally I should thank many individuals, friends and colleagues, who have not 
been mentioned here personally, for making this educational product a success. 
May be I would not have been able to accomplish this task without their 
support. 
LIST OF FIGURES 
Figure Description Page 
Fig. l-l:Gross generation-2005 Fig. 1-2 rElectricity Consumption - 2005 • 2 
Fig. 1-3 Consumption by categoiy-2005 3 
Fig. 1-4 : Mahaweli Complex 3 
Fig. 1-5 :Laxapana Complex 4 
Fig. 2-1: Typical load curve 8 
Fig. 2-2 : Load curve with energy mix 9 
Fig. 2-3 : Biological neural network system 11 
Fig. 2-4 : The neuron model 14 
Fig. 2-5 : A simple neuron 15 
Fig. 2-6 : neural model 16 
Fig. 2-7 : A Simple network 18 
Fig. 2-8 : working principle of neural network 19 
Fig. 2-9 : A neural network for 24hr prediction 20 
Fig. 3-1 : Performance training curve 24 
Fig. 3-2 : Comparing the actual & predicted on 13/02/06 25 
Fig. 3-3 : Comparing of deferent averaging methods 26 
Fig. 3-4 :Temperature variation in year - 2005 27 
Fig. 3-5 : Relative Humidity variation in the year - 2005 28 
Fig. 3-6 : Electrical Energy variation in the year - 2005 28 
Fig. 3-7 : Variation of average temperature in year 2005 29 
Fig. 3-8 : Variation of average relative humidity in year 2005 29 
Fig. 3-9 : Comparison the architecture 12:6:6:48 30 
Fig. 3-10 : Comparison the architecture 240:96:96:48 31 
Fig. 3-11: Comparison the architecture 480:96:96:48 31 
Fig. 3-12 : Overall Comparison 32 
Fig. 3-13 : Error Comparison 33 
Fig. 3-14 : Comparing after few months 34 
Fig. 3-15 : Percentage error for 24hours 34 
Fig. 3-16 : Demand curve for a Week day 35 
Fig. 3-17 : Demand curve for a Week end 36 
Fig. 3-18: Demand curve for the temperature of 31 deg. 37 
Fig. 3-19 : Demand curve for the temperature of26deg. 37 
Fig. 3-20 : Demand curve for the Humidity of 90% 3 8 
Fig. 3-21 : Demand curve for the Humidity of 80% 39 
Fig. 3-22 : Demand curve for Month No. 8 40 
Fig. 3-23 : Demand curve for Month No.l 40 
Fig. 3-24 : Demand curve for a Week Day 41 
Fig. 3-25 : Demand curve for a Saturday ' 42 
Fig. 3-26 : Demand curve for one year ahead 43 
Fig. 4-1: Graphical Solution of EDC 50 
Fig. 4-2 : Availability form 51 
Fig. 4-3: A typical dispatch schedule 52 
Fig. 4-4 : Actual ECD 52 
Fig. 4-5 : Calculated ECD 53 
Fig. 4-6 : Actual ECD 53 
Fig. 4-7 : Calculated ECD 54 
Fig. 4-8 : Availability schedule 55 
Fig. 4-9 : Predicted load curve 56 
Fig. 4-10 : Comparison of Actual & Predicted load curve 56 
Fig. 4-11: Day minimum condition (Dispatch schedule) 57 
Fig. 4-12 : Day peak condition (Dispatch schedule) 57 
Fig. 4-13 : Night peak condition (Dispatch schedule) 58 
Fig. 5-1 : Long term neural network 61 
vi 
LIST OF TABLES 
Table Description 
Table 1-1 : Install capacity of Mahaweli Complex 
Table 1-2 : Installed capacity of Laxapana Complex 
Table 1-3 : Reservoir capacities of the system 
Table 1-4 : CEB Thermal plants 
Table 1-5 : IPP Thermal plants 
Table 3-1 : Comparison of different networks 
Table 3-2 : Performance of deferent architectures 
Page 
4 
5 
5 
6 
6 
23 
26 
ATTACHMENTS 
ANNEXTURE-1.1 : Grid substation capacities and peak loads for May-2006 
ANNEXTURE-1.2 : Transmission network 
ANNEXTURE-3.1: Percentage error for half an hour prediction for the day 
12/02/06. 
ANNEXTURE-5.1: MATLAB coding 
vii