T.C.
ANKARA YILDIRIM BEYAZIT UNIVERSITY
BUSINESS SCHOOL
MANAGEMENT INFORMATION SYSTEMS
FORECASTING ELECTRICITY DEMAND IN TURKEY WITH
ARTIFICIAL NEURAL NETWORK
Alpaslan ADIYAMAN
Ufuk GÜLCEMAL
Uğur ALTUNTAŞ
Furkan KURT
Semih DÖNER
SENIOR PROJECT
Advisor
Prof.Dr. Ömür AKDEMİR
ANKARA (2018)
i
ABSTRACT
Energy is a field of strategic preoccupation that is vital to the development policies of
countries and the continuity of technology in the global world. Considering our country as a
base, the increasing need for energy increases rapidly due to increasing industrialization and
developing technologies, further increasing the importance of energy demand and energy
policies for the state. Especially electric energy has a vital proposition for the existence of the
developing industrialist and for the sustainability of the technology. In the world, the
sustainability of technology-supported industrialists is undoubtedly provided by electricity.
Undoubtedly, this situation is the same for our country. For the sustainability of technology-
supported industrialists in our country, future forecasts of electrical energy and interpretation
of these estimates are a vital issue. In this project, especially the estimation process is handled
for the future of Turkey's electricity demand. However, estimates of the data obtained after
the process was carried out according to the interpretation of Turkey's electricity generation
sources. In this context, demand and consumption estimation based on cause-effect
relationship with artificial neural networks was realized in the project. GDP, population,
imports, exports, total floor area of buildings and number of electricity subscribers data
between 1998-2016 were used as independent variables and an artificial neural network
model was established according to these data for forecast the net electricity demand in
Turkey. According to the analysis and comparisons it has been shown to perform more
successful result of artificial neural network estimates and Turkey between the years 2017-
2023 net electricity demand was estimated. Then these estimates will interpreted in economic,
strategic and political context.
ii
TABLE OF CONTENTS
ABSTRACT ............................................................................................................................... İ
LIST OF FIGURES ............................................................................................................... İİİ
LIST OF TABLES ................................................................................................................. İV
ABBREVIATIONS ................................................................................................................. V
1.INTRODUCTION ................................................................................................................. 1
2. ELECTRICITY MARKET IN TURKEY ......................................................................... 2
2.1 IMPORTANCE OF ELECTRICITY ...................................................................................................................... 2
2.2 DEVELOPMENT OF ELECTRICITY IN TURKEY ................................................................................................. 3
2.3 TURKEY ELECTRICITY GENERATION RESOURCES .......................................................................................... 4
3. LITERATURE RESEARCH .............................................................................................. 6
3.1 ENERGY DEMAND FORECASTING WITH ARTIFICIAL NEURAL NETWORKS ..................................................... 6
3.2 ENERGY DEMAND FORECASTING RESEARCHES IN TURKEY ........................................................................... 7
4. ARTIFICIAL NEURAL NETWORK ................................................................................ 9
4.1 THE DEFINITION OF ARTIFICIAL NEURAL NETWORK ...................................................................................... 9
4.2 STRUCTURE AND BASIC ELEMENTS OF ARTIFICIAL NEURAL NETWORKS ...................................................... 9
4.2.1 BIOLOGICAL NERVE CELL ........................................................................................................................ 10
4.2.2 ARTIFICIAL NEURAL CELL ........................................................................................................................ 10
4.2.2.1 INPUTS ................................................................................................................................................... 11
4.2.2.2 WEIGHTS ............................................................................................................................................... 11
4.2.2.3 SUMMING FUNCTION ............................................................................................................................. 11
4.2.2.4 ACTIVATION(TRANSFER) FUNCTION ...................................................................................................... 12
4.2.2.5 OUTPUTS ............................................................................................................................................... 14
4.3 ARCHITECTURE OF ARTIFICIAL NEURAL NETWORK .................................................................................... 14
4.3.1 FEEDBACK ARTIFICIAL NEURAL NETWORKS ............................................................................................ 14
4.3.2 FEEDFORWARD ARTIFICIAL NEURAL NETWORKS ..................................................................................... 15
4.3.2.1 FEEDFORWARD BACKPROPAGATION NETWORKS ................................................................................... 15
5. ELECTRICITY DEMAND FORECASTING MODEL ................................................ 16
iii
5.1 INTRODUCTION OF DATA TO BE USED IN FORECASTING MODEL .................................................................. 17
5.3 BUILDING ARTIFICIAL NEURAL NETWORK .................................................................................................. 19
5.4 TRAINING NETWORK ................................................................................................................................... 20
5.5 PERFORMANCE MEASUREMENT ................................................................................................................... 23
5.6 FORECASTING AND RESULTS ....................................................................................................................... 24
6. CONCLUSION ................................................................................................................... 26
REFERENCES ....................................................................................................................... 27
LIST OF FIGURES
Figure 1: Electricity Generation Resources ............................................................................... 3
Figure 2 : Biological Nerve Cell .............................................................................................. 10
Figure 3 : The Basic Flow for Designing Artificial Neural Network ...................................... 17
Figure 4 : View of Neural Network ......................................................................................... 20
Figure 5 : nntool Creating Network Screen ............................................................................. 20
Figure 6 : Training parameters screen ...................................................................................... 21
Figure 7 : Training performance Graph ................................................................................... 22
Figure 8 : Regression plot ........................................................................................................ 22
Figure 9 : Actual Demand and Model (net11) Output ............................................................. 24
Figure 10 : Simulate Tab of Neural Network Toolbox ............................................................ 25
Figure 11 : Artificial Neural Network Forecasting Graph ....................................................... 26
iv
LIST OF TABLES
Table 1 : Power Installed of Power Plants, Gross Generation and Net Consumption of
Electricity ................................................................................................................................... 3
Table 2 : Turkey Production and consumption of primary energy resources (2011) ............... 5
Table 3 : Summary of Energy Demand Forecasting Works by Artificial Neural Networks ..... 7
Table 4 : Summary of Energy Demand Forecasting in Our Country ......................................... 8
Table 5 : The Similarity Between Biological Nerve Cell and Artificial Neural Network ....... 11
Table 6 : Transfer Function and Graphs ................................................................................... 12
Table 7 : Explanation and Source of the Data of the Independent Variables .......................... 17
Table 8 : MAPE values of created models ............................................................................... 23
Table 9 : Forecasted Value of 2017 to 2023 ............................................................................ 25
v
ABBREVIATIONS
Abbreviations
ANN
GDP
GWH
KW
LR
MAD
MAE
MAPE
MATLAB
MENR
MFE
MPE
MSE
MW
NLR
SSE
TEDAŞ
TEİAŞ
TEP
Explanation
Artificial Neural Netwok
Gross Domestic Product
Gigawatt Hour
Kilowatt
Linear Regression
Mean Absolute Deviation
Mean Absolute Error
Mean Absolute Percentage Error
Matrix Laboratory
Ministry of Energy and Natural Resources
Mean Forecast Error
Mean Percentage Error
Mean Squarred Error
Megawatt
Nonlinear Regression
Sum of Squarred Error
Turkish Electricity Distribution
Corporation
Turkey Electricity Transmission
Corporation
Tons of Oil Equivalent
1
1. INTRODUCTION
"Scientifically, the definition of energy is, in short, defined as capacity or capability of doing
business. Energy, a universal concept with the power to do business in its most general sense,
is a necessary input to use in production processes. It also constitutes one of the basic
building blocks of economic and social development as a necessary service tool for raising
the welfare levels of societies." (Şahin, 1994).
The energy sector is a very important sector in terms of the development of the countries and,
accordingly, human life. Especially electric energy has an important and vital position in
developed and industrialized societies. In the global world, technology has developed and
with the developing technology, the muscle power has taken its place. The machines are
managed by brain power. In order to provide this management, electricity is the basis for the
use of machinery and machinery. So, where there is technology, there must be electricity.
With the development of technology, the consumption of electric energy has increased and
accordingly the production of electric energy has become important. The fact that the
consumption of electricity energy increases continuously from here has an important place on
the world agenda. This issue is one of the most important issues on the world agenda. So to
take place this issue on the agenda of Turkey is natural and inevitable.
In order to determine energy policies, it is very important to have information about the
energy resources of our country and make forecasts for the future by using this information.
In order to produce consistent policies for the future, it is necessary first to predict the future.
For this reason, it is very important for the future of our country to make successful
estimations in order to meet the demands of electricity, which is one of the sine qua non of
human life and technological infrastructure.
“Turkey is a transit point between Asia and Europe, as well as the required amounts of
consumption is a very important country in terms of energy policy because it is a transit
route. Especially in our country, electricity consumption in the year 2017 was 294.9 billion
kWh, increasing by 5.6% over the previous year. Electricity production increased by 7.7% to
295.5 billion kWh compared to the previous year.” (T.C. Enerji ve Tabii Kaynaklar
Bakanlığı, 2017).
It is very important to make reliable demand forecasts for the future in order to sustainably
meet the demand for electricity, which is increasing with increasing population, increasing
industrialization and urbanization. Forecasting methods; qualitative and quantitative
estimation methods. The strength of both forecasting methods is the observation values of the
relevant variables.
"The qualitative forecasting method is based on the opinion and experience of the individual
who is expert on the subject studied. Different people can be found in different estimates for
the same data. However, it is a positive aspect that it is easy to implement and does not
2
require much effort and time. They are also generally cheap and do not need statistical
abilities at advanced levels, which is an important influence.” (Çuhadar, 2006).
"Quantitative forecasting methods are methods based on mathematical models. Quantitative
forecasting methods can be used in situations where sufficient quantitative information is
available. The following three conditions must be fulfilled when estimating quantitative
forecasting methods: the existence of historical information, the fact that this information can
be expressed in a numerical form, and the idea that the variable's past display will continue in
the future. Historical observational values are used to determine the relationships that
contribute to the process and how these relationships shape the future. There are two basic
approaches to quantitative estimation: models based on cause-effect relationship and time
series analysis." (Hamzaçebi, 2011).
In this project, the artificial neural network prediction of quantitative forecasting methods
study of Turkey between the years 2017-2023 net electricity demand was realized. In addition
to the frequently used variables in the literature, artificial neural network (ANN) model was
established by determining different variables affecting net energy demand.
ANN provides very good results in forecasting studies and is able to model nonlinear
problems to avoid other estimation methods. For this reason, ANN estimation studies have
been done in the field of energy.
In this context, the importance of electrical energy and electrical energy in the second part of
the subject will be discussed in Turkey in general. In the third part, the literature research
section is included. In the literature research, studies conducted with ANN will be taken into
account first. In the fourth part, the ANN to be used in the study will be explained in detail
and in the fifth part the application part will be realized and findings will be obtained. In the
sixth part, the results will be presented and assessments will be made.
2. ELECTRICITY MARKET IN TURKEY
2.1 Importance of Electricity
“Electricity is the most versatile and easily controlled form of energy. At the point of use it is
practically loss-free and essentially non-polluting. At the point of generation it can be
produced clean with entirely renewable methods, such as wind, water and solar. Electricity is
weightless, easier to transport and distribute, and it represents the most efficient way of
consuming energy.” (International Electrotechnical Commission, 2018).
Today condition, there are several produce electricity such as hydraulic sources, thermal
sources, nuclear sources, wind power, solar energy and geotermal energy. International
Energy Agency members located in 29 countries of Turkey, in terms of electricity production
from fossil fuel use 9, it ranks 5th in the use of natural gas. It is ranked third in terms of
geothermal energy use after New Zealand and Italy. It has the 7th largest production capacity
3
in hydropower generation (OZAN, 2017). We can also understand the importance of
electricity by looking at the table below.
Figure 1: Electricity Generation Resources (TUIK, 2016)
According to data of Turkish Statistikal Institute (TUIK) , we supply the majority of energy
consumption from electricity.
2.2 Development of Electricity in Turkey
The first generation of electricity in our country, in the period of II.Abdulhamid in 1902. It
was realized with a small water turbine at 2 kW power installed in Tarsus. First The large
power plant was built by the Ottoman Electricity Generation Corporation in 1913 Istanbul
Thermal Power Plant has been established in Silahtarağa with a power of 15 MW
(YAVUZDEMİR, 2014).
After the government change in 1950, there was a change in the energy policy and it was
decided that distribution of energy production transmission would be carried out by privileged
companies (EMO Enerji Komisyonu, 2011). In these years, our installed power has reached
499,5 MW. Next time data from Turkey Electricity Transmission (TEIAS) we share with by
the shared data.
Year Total Power Installed
(MW)
Gross Generation
(GWh)
Net Consumption
1975 4 186,6 15 622,8 13 491,7
1976 4 364,2 18 282,8 16 078,9
1977 4 727,2 20 564,6 17 968,8
1978 4 868,7 21 726,1 18 933,8
1979 5 118,7 22 521,9 19 633,1
1980 5 537,6 23 275,4 20 398,2
1981 6 638,6 24 672,8 22 030,0
1982 6 935,1 26 551,5 23 586,8
1983 8 461,6 27 346,8 24 465,1
Table 1 : Power Installed of Power Plants, Gross Generation and Net Consumption of Electricity
32,30%
16,30%15,90%
10,60%
7,10%
17,80% Natural Gas
Lignite
Coal
Electricity
Diesel
Others
4
Year Total Power Installed
(MW) Gross Generation
(GWh) Net Consumption
1984 9 121,6 30 613,5 27 635,2
1985 9 121,6 34 218,9 29 708,6
1986 10 115,2 39 694,8 32 209,7
1987 12 495,1 44 352,9 36 697,3
1988 14 520,6 48 048,8 39 721,5
1989 15 808,2 52 043,2 43 120,0
1990 16 317,6 57 543,0 46 820,0
1991 17 209,1 60 246,3 49 282,9
1992 18 716,1 67 342,2 53 984,7
1993 20 337,6 73 807,5 59 237,0
1994 20 859,8 78 321,7 61 400,9
1995 20 954,3 86 247,4 67 393,9
1996 21 249,4 94 861,7 74 156,6
1997 21 891,9 103 295,8 81 885,0
1998 23 354,0 111 022,4 87 704,6
1999 26 119,3 116 439,9 91 201,9
2000 27 264,1 124 921,6 98 295,7
2001 28 332,4 122 724,7 97 070,0
2002 31 845,8 129 399,5 102 948,0
2003 35 587,0 140 580,5 111 766,0
2004 36 824,0 150 698,3 121 141,9
2005 38 843,5 161 956,2 130 262,9
2006 40 564,8 176 299,8 143 070,5
2007 40 835,7 191 558,1 155 135,2
2008 41 817,2 198 418,0 161 947,6
2009 44 761,2 194 812,9 156 894,1
2010 49 524,1 211 207,7 172 050,6
2011 52 911,1 229 395,1 186 099,6
2012 57 059,4 239 496,8 194 923,4
2013 64 007,5 240 154,0 198 045,2
2014 69 519,8 251 962,8 207 375,1
2015 73 146,7 261 783,3 217 312,3
2016 78 497,4 274 407,7 231 203,7
Table 1(Continued): Power Installed of Power Plants, Gross Generation and Net Consumption of Electricity
2.3 Turkey Electricity Generation Resources
Electricity is obtained from various sources. Primary energy resources such as hydraulic,
thermal, nuclear, natural gas, wind and sun are utilized for the supply of this energy. These
resources have advantages and disadvantages relative to each other depending on the
geographical structure of the country. For this reason, countries have to use the resources that
are best suited to them. Countries with a lot of underground resources prefer thermal power
plants, while countries with rich streams prefer hydroelectric power plants. This can be done
in our country in the interconnection system, with seasonal differences and the supply and
demand relation in mind. In the developing world and in Turkey in the future it will not be
5
enough considering the current primary energy sources must be exploited to the optimum of
the available energy sources.
Source Type
Primary Energy Resources
Production
Consumption of Primary Energy
Resources
Energy
Production
(thousand tep)
Rate
(%)
Energy
Consumption
(thousand tep)
Rate
(%)
Coal 17.870 55,5 35.841 31,3
Natural Gas 652 2,0 36.909 32,2
Oil 2.555 7,9 30.449 26,6
Hydraulic 4.501 14,0 4.501 3,9
Biomass 3.555 11,0 3.573 3,1
Geothermal 1.463 4,5 1.463 1,3
Others 1.633 5,1 1.712 1,5
Total 32.229 100 114.480 100
Table 2 : Turkey Production and consumption of primary energy resources (2011)
Turkey is one of the country's demand for energy resources an intensely. As a matter of fact,
consumption of primary energy resources, which was 78.8 Mtep in 2000, reached 114.4 Mtep
in 2011 (EUAŞ, 2012). In the same period, the production of primary energy resources was
32.2 Mtep.The production of primary energy resources in Turkey in 1970 (14.5 Mtep) by the
year 2011 (32.2 Mtep) increased by 122% between the rate of the last 41 years in Table 1. In
the same period, consumption of primary energy sources increased by 508%. In this process,
the domestic rate of total energy consumption has decreased from 77% in 1970 to 28.1% in
2011 due to the rapidly rising energy demand. The local production is more than the increase
6
in the consumption of primary energy sources in Turkey's dependence on foreign sources of
energy increases with each passing year (YILMAZ, 2012).
As a result, Turkey's electric energy consumption in 2016, 231 billion kWh, while electricity
production was realized as 274 billion kWh.
3. LITERATURE RESEARCH
Reliable and accurate forecasting of energy policies in support of sustainable development is
the key to their work. For this reason, many demand prediction studies have been carried out
in the literature on energy and new prediction models have been derived. Many estimating
techniques have been used in these studies. Some of those are; Time Series, Gray Prediction,
Regression Model, Genetic Algorithm, Fuzzy Logic and Neural Networks. For example,
Lester and Himanshu (2008) using time series in Sri Lanka, Tonga and Lee (2011) using
genetic algorithms with gray models have predicted that China's demand for energy. On this
side of the working literature it will usually examin studies with Neural Networks that will be
used as an energy estimate of art and literature an important point in the estimation model
variables and used in the neural network model will be determined. After the analysis of these
studies has put, the energy demand forecasting studies put forward in our country which will
be reviewed.
3.1 Energy Demand Forecasting with Artificial Neural Networks
ANN gives very good results in hypothetical work. There are many studies with the ANN
technique which is widely used in estimation studies.
They used linear and non-linear statistical models along with ANN to examine the economic
factors (national income, population, gross domestic product (GDP)) and the impact of
electricity consumption in Taiwan. Models were created using data from January 1990 to
December 2002. When the results are compared, it is found that ANN gives better results than
linear models for electricity consumption prediction models (Pao, 2005).
They used ANN model to estimate South Korea's energy demand. The GDP for the years
1980-2007 calculated the electricity consumption of electricity between 2008 and 2025 using
population, import and export data. The obtained results are compared with the results
obtained with linear regression and exponential model. It has been found that ANN makes
better and healthier predictions than other methods (Roper & Geem, 2009).
The model of ANN was used to estimate total energy demand in Greece. Using data from
1992-2008, it estimated the total energy consumption for the years 2010, 2012 and 2015
(Ekonomou, 2010).
In order to estimate the transport energy demand in the study, it established ANN and
regression models using various variables such as GDP, population, oil prices, the number of
registered vehicles and the amount of passenger transport. All of them Estimate
7
transportation energy demand between 2008 and 2025 by using data from 1990-2007 (Geem,
2011).
Writers Variables
Used
Data Years Working
area
Forecast Years
Pao GDP,
Population,
Import,
Export
1990-2000
Electrical
Energy
Claim
Model Established
Roper,
Geem
GDP,
Population,
Import,
Export
1980-2007
Energy
Claim
2008-2025
Ekonomou GDP,
Electricity
Consumption
1992-2008 Energy
Claim
2010, 2012, 2015
Table 3 : Summary of Energy Demand Forecasting Works by Artificial Neural Networks
3.2 Energy Demand Forecasting Researches in Turkey
The field of energy plays a critical role in the future of our country. There are many studies in
our country as well as the work done in the field of energy in the world. The studies which
have been done on literature in our country are summarized and presented in table form.
They used artificial neural network technology to predict long-term electrical energy.
Moreover, They developed models using electric energy consumption values and population
information between 1970-2002 (Kutay & Hamzaçebi, 2007).
They estimated the total and sectoral electric energy demand with ANN in their work. By
using total and sectoral electrical energy data for the years 1970-2004 (Atak & Akay, 2007).
They have used the fuzzy logic method to estimate the electric energy of our country. Using
only 1980-2008 GDP data to estimate consumption of electricity demand between 2009 and
2014 (Barış & Küçükali, 2010).
8
They used this model to develop a new mathematical model to predict electricity energy. The
method gave very healthy results for the long-term estimation. Using the data from 1982 to
2007, They made electricity energy consumption estimation for 2025 (Filik, Gerek, &
Kurban, 2011).
In order to estimate the energy consumption of industry and housing in our country, ANN
used Linear Regression (LR), Nonlinear Regression (NLR) techniques. The results of these
estimations are compared with the results in the Ministry of Energy and Natural
Resources(MENR). These criterias show that the ANN technique yields are closer than the
other techniques (Bilgili, Şahin, Yaşar, & Şimşek, 2012).
To estimate the electricity energy of our country for 2008-2014; they used population, import,
export and employment variables, regression analysis and ANN technique. The estimation
results obtained are compared with the official results and have close results with the official
figures (Kankal, Akpınar, Kömürcü, & Özşahin, 2011).
Writers Variables Used Data Years Working area
Forecast
Years
Kutay,
Hamzaçebi
Electricity
Consumption
Values,
Population
1970-2002 Electric Energy
Prediction
2003-2010
Toksarı GDP, Import,
Export, Population
Information
1979-2005 Electric Energy
Prediction
2025
Filik & Friends GDP, Import,
Export, Population
Information
1982-2007 Electric Energy
Prediction
2025
Ünler GDP, Population
Imports, Exports
and Growth Rates
Data
1979-2005 Total Energy
Prediction
2006-2025
Table 4 : Summary of Energy Demand Forecasting in Our Country
9
4. ARTIFICIAL NEURAL NETWORK
4.1 The Definition of Artificial neural network
Artificial neural network does not have a commonly accepted definition. Different definitions
have been made by different people in the literature, as follows.
“An ANN network is a parallel distributed processor consisting of simple units that have a
natural tendency to store information. ANN resembles the human brain in two respects.
Information is obtained by the network through the learning process. The connecting forces
between neurons, known as synaptic weights, are used to store information.” (Haykin, 1994).
“An artificial neural network is an information-processing system that has certain
performance characteristics in common with biological neural networks. Artificial neural
networks have been developed as generalizations of mathematical models of human cognition
or neural biology.” (Fausett, 1994).
“ANN is a collection of many simple processor elements. These elements are interconnected
by "links" or "weights" that carry numerical representations that can be expressed in different
forms.” (Erler, Sağıroğlu, & Beşdok, 2003).
“Artificial neural networks are an information processing system that aims to acquire
abilities such as learning, generalizing, remembering by mimicking the work of nerves in the
brain.” (Sarı, 2016).
“Artificial neural networks are a information processing technology that is based on the
ability of the human brain to work and think. Artificial neural networks, in other words, are
computer programs that mimic biological neural networks.” (Ağyar, 2010).
“ANN is a computer system that automatically generates new data through learning that is
one of the human brain traits and can derive and discover capabilities without using any
support. With classical programming methods it is impossible to do such operations. ANN is a
scientific discipline dealing with adapter information processing developed to solve such
problems that are difficult to program.” (Öztemel, 2012).
4.2 Structure and Basic Elements of Artificial Neural Networks
Artificial neural networks are algorithms that perform the correlation between the input and
output effectively and reliably. Their parallel, multi-parametic charachters, and computing
speed make them powerfull computation tool, especially when the mathematical models are
complex. Information is proceed non-algorithmically in the ANN structure. Information is
obtained through learning, which means that the ANNs can be used to model complex
systems where mathematical descriptions or notations are neither available nor possible
(Rajpal, Shishodia, & Sekhon, 2006).
10
4.2.1 Biological Nerve Cell
In humans, the neural system consists of cells that called nerves. The nerves are the smallest
units of life's vital functions. The nerves, which are considered to have about 1010 in one
human being, are spread not only in the brain but also in the whole body on the central neural
system. The tasks of the nerves that make up the brain's communication system are to signal,
process and transmit electrochemical signals in neural networks (Şen, 2004).
Figure 2 : Biological Nerve Cell (Saraç, 2004)
The basic elements of the neural system are neurons. The neurons consist of dendrites, somas,
axons and synapses. The nerve cell carries the signals from the other nerve cell with the
extensions known as dendrites and generates an output signal by evaluating these signals
collected in the soma and these signals are sent to the other nerve cell through the axons.
Although there are multiple inputs from a nerve cell, while there is only one output. In the two
nerve cells, the connection between the axon, one of the output elements, and the input
dendrites of the other nerve cell, is called synaps and forms a synaptic linkage. Billions of
nerve cells in the body are connected to each other by synaptic connections to form the nerve
network (Sarı, 2016).
4.2.2 Artificial Neural Cell
Artificial neural networks are structurally and functionally similar because they are
constructed on the basis of biological nerve cells. Hence, artificial neural cell elements are
called processing elements that consist of 5 elements: inputs, weights, summing function,
activation(transfer) function, and output.
11
Biological Nerve Cell ANN
Neuron Processor Element
Dendrite Summing Function
Soma Transfer Function
Axon Articial Neuron Output
Synaps Weights
Table 5 : The Similarity Between Biological Nerve Cell and Artificial Neural Network. (Bulgan, 2013)
4.2.2.1 Inputs
Inputs (xi) in an artificial neuron contain information from the outside or from itself into an
artificial neural cell. This information is determined by the samples that will network to learn.
If entries (xi) do not consist of numbers, for example, entries of provincial and district names
used in meteorology need to be binarized before being given to the network. Another example
can be gender specified as 1 or 0.
4.2.2.2 Weights
One of the most important elements of artificial neural network is weights(wi) that named as
memories of artificial neural networks. Weights are the importance of information coming to
an artificial neuron and show the effect on the neuron. The weights carry the weight
coefficients of the input signals. In many network, success in the learning process is achieved
by changing the weight of the processing elements. The weights in each input pattern are
readjusted to try to minimize the output error. The fact that the values of weights are small or
large does not mean that it is important or not. The fact that the value of a weight is zero can
be the most important event for that network.
4.2.2.3 Summing Function
The net input to a cell is calculated by the summing function. Many different models can be
used as the summing function, but the most common one is the weighted sum model. In this
model, each incoming input value is multiplied by its own weight and the sum of these values
for all inputs is taken. The mathematical notation of weighted sum model as in the Formula 1.
(Formula 1)
12
X = Inputs
W = Weights
n = The total number of inputs coming into the cell.
4.2.2.4 Activation(Transfer) Function
The activation function is the function that determines the output of the cell by processing the
net input to the cell. It is also known as the function that limits the size of the output of the
artificial neural cell. It is also known as the function that limits the size of the output of the
artificial neural cell. Different functions are used in the activation function as well as in the
summing function.
The correct selection of the activation function has an influence on the performance of the
network. If the selected activation function is not linear, the slope parameter must be
specified. The slope parameter is an important factor that plays a role in reaching the
appropriate result at a satisfactory level (Erilli, Eğrioğlu, Yolcu, Aladağ, & Uslu, 2010).
The sigmoid function is generally used as the activation function in the most commonly
preferred Multi-Layer Perceptron model (Karasu, 2012). Transfer functions used in
MATLAB software are presented in Table 6.
Transfer Function Graph
Competitive
Table 6 : Transfer Function and Graphs
13
Transfer Function Graph
Hard-Limit
Symmetric Hard-Limit
Log-Simoid
Positive Linear
Linear
Radial Basis
Satlin
Table 6 (Continued) : Transfer Functions and Graphs
14
Transfer Function Graph
Satlins
Softmax
Tan-Sigmoid
Triangular Basis
Table 6 (Continued) : Transfer Function and Graphs
4.2.2.5 Outputs
The output is a value that determined by the activation function. Output in artificial neural
networks is solution of the problem. Output value can be used as input to another artificial
neural cell. Although an artificial neural network cell has more than one input, while there is
only one output.
4.3 Architecture of Artificial Neural Network
Artificial neural network architecture can be classified into two according to the directions of
the connections between neurons or the directions of the signals in the network. These are
feedforward artificial neural network and backforward artificial neural network
(ATASEVEN, 2007).
4.3.1 Feedback Artificial Neural Networks
In the feedback network, there are signals traveling in both directions by presenting a loop in
the network. Feedback networks are very powerful and can be very complex. Feedback
15
networks are dynamic; as long as they do not reach a balance point, their 'state' is constantly
changing. They remain at the balance point until the input changes and a new balance should
be found. Feedback architecture is also known as interactive or recurrent, although the latter
term is often used to refer to feedback in single-layer organizations (Stergiou & Siganos,
1996).
A network structure in which data flow is not only forward but also backward. In this network
structure, the network output can also be used as input at the same time. In feed forward
networks, neurons are found in layers, as in feedback networks. However, inter-layer
connections are bidirectional rather than one-way. A feedback network is a network structure
in which output and intermediate layer outputs are fed back to input units or previous
intermediate layers. Thus, inputs are transferred both forward and backward (KAYA,
OKTAY, & ENGİN, 2005).
4.3.2 Feedforward Artificial Neural Networks
Feed-forward neural networks divided into single layer and multi layer. Single-layered feed-
forward networks have only input and output layers, there are no hidden layers. Multi-layered
feed-forward networks have input layer, one or more hidden layer and output layer.
In the feed-forward network, artificial neurons are layered and the input neurons are
transmitted unilaterally from the input layer to the output layer. There is no connections
between neurons in the same layer. It is the network structure that the data stream is only
forward (Haykin, 1994) (Krose & Smagt, 1996).
There are no delays in feed-forward artificial neural networks, the process proceeds from the
inputs to outputs. The output values are approximated to the desired output value, and the
corresponding error signals are evaluated and the most suitable weight values are tried to be
found. In feed-forward neural networks, the connections between the processing elements do
not form a loop (Erler, Sağıroğlu, & Beşdok, 2003) (Yurtoğlu, 2005).
In the feed-forward artificial neural network models, the output of the network depends
entirely on the inputs to the network. There is no delay in such networks. The most common
learning algorithm is the "backpropagation" algorithm in the feed-forward networks.
4.3.2.1 Feedforward Backpropagation Networks
“Back propagation algorithms (Rummelhart and McClelland, 1986) are used in layered feed-
forward artificial neural networks. Neurons are arranged in layers which forward their
signals and the errors are propagated backwards.” (Onaran, 2010).
16
Back propagation network is a network type that introduces nonlinear solutions to complex,
unspecified problems. Most of the problems can be solved because of these solution methods.
In the learning process, "momentum" is used to shorten the learning process of the network.
Backpropagation: The output produced by the network for input to the network is compared
to the expected output of the network. The difference between them is regarded as an error.
The goal is to reduce this error. This error is distributed to the network weight values to
reduce the error in the next cycle. In order to minimize the total error, it is necessary to
distribute this error to the neurons that cause it. There are 2 situations to change the weights of
the network:
• Changing weights between hidden layer and output layer
• Changing weights between hidden layers or hidden layers and input layers
The amount of change in the weights of the connections between the neurons in the hidden
layer and the neurons in the output layer is calculated using the learning rate and momentum
coefficient and new weights are found. Similarly, the bias weights of the intermediate layer
threshold value unit are changed.
The learning rate determines the rate of change in connection weights and takes a value
between 0-1. The learning rate has an important effect on network performance. While
training of network in small value learning rate takes a long time, training time is shortened in
large value learning rates. Increasing the learning rate does not bring any improvement on the
total fault of the network. The addition of the momentum coefficient in the calculations was
found to be effective in network performance (Sönmez, 2015).
The momentum coefficient can be expressed as a factor that helps to restore the network
faster by adding some of the previous change to the next change cycle. Momentum coefficient
takes values between 0 and 1. When the momentum coefficient is added to the calculation,
there is a decline in the number of cycle and the total network error. When the momentum
coefficient is taken high, the total error in the network appears to be approaching zero with
too much slope. (Sönmez, 2015).
Bias is a special processing element with constant activation value. By using bias inputs in a
backpropagation network, it is attempted to bring the activation function to the original
balance by providing better learning (Mete, 2008).
5. ELECTRICITY DEMAND FORECASTING MODEL
The most widely used independent variables in the literature search are the gross domestic
product, imports, exports and population data. In addition to the four independent variables,
the total floor area of buildings in Turkey and number of electricity subscribers, has been
added to the model to improve the performance of the network and the originality of the
research. Forecasting is implemented in Matlab (R2017a version) software. While the
forecasting was applied, the flow in the Figure 3 was followed.
17
Figure 3 : The Basic Flow for Designing Artificial Neural Network (Al Shamisi, Assi, & Hejase, 2011)
5.1 Introduction of Data to be Used in Forecasting Model
A correlation analysis of the independent variables to be used in the model was made and it
was determined that the variables were appropriate for the model. Population, GDP, export,
import, total floor area of buildings datas have been obtained from Turkish Statistical Institute
and number of electricity subscribers data has been obtained by e-mail from TEDAŞ. Table 7
shows the sources and explanations of the data.
Independent Variable Explanation Source
X1 = Population
It was used as an independent variable in
the model because it is assumed that
there is a linear relationship between
population and electricity demand.
Turkish
Statistical
Institute
X2 = GDP
As GDP is a growth indicator, energy
demand is expected to be affected in
parallel with growth, and GDP is used as
an independent variable.
Turkish
Statistical
Institute
Table 7 : Explanation and Source of the Data of the Independent Variables
18
Independent Variable Explanation Source
X3 = Export
Import is one of the most important
variables that make up the trading
volume of an country and it is added as
an independent variable considering that
it will affect energy demand
Turkish
Statistical
Institute
X4 = Import
Considering that there is a linear
relationship between the export volume
and the output of a country, it is added as
an independent variable, considering that
it will also affect energy consumption
linearly.
Turkish
Statistical
Institute
X5 = Total Floor Area of
Buildings
It includes the total floor area of
structures such as house, apartment
house, commercial building, Industrial
building, medical and social building,
cultural building, religious building,
administrative buildings and other types
of building in square meters(m2). It is
thought that there is a linear connection
between building area and electricity
consumption and it is added as an
independent variable.
Turkish
Statistical
Institute
X6 = Number of Electricity
Subscribers
As number of electricity subscribers
increases the consumption will also
increase.
TEDAŞ
Table 7 (Continued) : Explanation and Source of the Data of the Independent Variables 5.2 Pre-process of Data
By normalizing the data in ANN, the non-linearity feature becomes meaningful. The method
chosen for the normalization of the data will directly affect the performance of the ANN. This
is because normalization allows the input data to be transferred from the active region of the
transfer function when transferred. The goal of data normalization is to protect the processor
elements from the negativeness of cumulative sums of data. In general, it is desirable to
normalize the range [0,1] or [-1,1]. (Sueri, 2016) This range may vary with different
normalization methods such as the D_min_max method used in this study. The D_min_max
method is used in this study because better performance is achieved with this method. In the
19
D_min_max method, the data are scaled from 0,1 to 0,9. The mathematical notation of
D_min_max method as in the Formula 2.
(Formula 2)
x’ = Normalized data
xi = Input value
xmin = The smallest number in the input set
xmax = The largest number in the input set
5.3 Building Artificial Neural Network
In this study, a feed-forward backpropagation artificial neural network was established in a
matlab environment. There are six neurons in the input layer and one neuron in the output
layer because there are six independent dependent variables and one dependent variable.
In the literature studies, only one hidden layer is preferred since only one hidden layer is
sufficient to obtain successful results (Zhang, Patuwo, & Hu, 1998). Increasing the number of
hidden layers has no effect on the performance of the network and has extended the training
process.
Studies in the literature have shown that the number of hidden neurons are not used more than
2n+1 of independent variables. “n” represents the number of independent variables.
Therefore, various models were created by trying with 3-13 hidden neurons.
In the builded artificial neural networks model, sigmoid activation function(tansig) is used as
hidden layer activation function. The sigmoid activation function is the most commonly used
activation function in applications because it is a continuous, nonlinear function that can be
derived. This function generates a value between 0 and 1 for each value of the net input
(Haykin, 1994).
Linear activation function(purelin) is used as output layer activation function. This function
which is used to solve linear problems, gives incoming net inputs directly as cell output.
The best result-giving method for the learning method was researched and Bayesian
regularization-based backpropagation algorithm(trainbr) was used. The goal of the trainbr is
minimizing a combination of squared errors and weights, and then determines the correct
20
combination by updating the weight and bias values according to Levenberg-Marquardt
optimization.
Figure 4 : View of Neural Network
5.4 Training Network
The initial training of the network begins with the supervised learning. In supervised
learrning, output values are taught for input values to the artificial neural network. The
network reinitiates weights until it minimizes the error between its output and the desired
output. Neural network toolbox(nntool) of Matlab R2017a software was used to train the
network.
Figure 5 : nntool Creating Network Screen
21
As shown in the figure above, input values and target output values are given to the network.
Trainbr has been used as training function. The network has been created in two (one hidden
layer and one output layer) layers and the tansig is used as the transfer function at the input
layer and purelin is used as the transfer function at the output layer.
Figure 6 : Training parameters screen
As shown in the figure above, the training parameters are set after the network is set up. In
order to extend the training process and increase the training performance, the max_fail
parameter changed from 0 to 1000 and the other parameters are left as default.
22
Figure 7 : Training performance Graph
Figure 8 : Regression plot
23
As shown in the table above, the regression value of the test set is 0.99 indicating that the
model is successful. In scientific studies, a regression value higher than 0.80 indicates that the
model is acceptable.
5.5 Performance Measurement
Evaluating the performance of the regression is a fundamental element of artificial learning. A
model takes a training set as input and creates a regression model that can predict an unseen
case. The model can be evaluated using the evaluation criterion. Evaluating is important to
select the most acceptable model from a particular set of models and to understand the quality
of the model. There are several criteria for evaluating models such as RMSE (Root Mean
Square Error), MFE (Mean Forecast Error), MAE (Mean Absolute Error), MAD (Mean
Absolute Deviation), MAPE (Mean Absolute Percentage Error), MPE (Mean Percentage
Error), MSE (Mean Squared Error), and SSE (Sum of Squared Error) (Koç, 2017).
In this study SSE was used as the error performance function of the network and the MAPE
criterion, which is frequently used in the literature, was used to evaluate the performance of
the created models. The mathematical notation of the MAPE criterion as in the Table 8.
MAPE = ( Formula 3)
Model Name Number of Hidden
Neurons
MAPE Result (%)
net1 3 3,444529023
net2 4 2,405237055
net3 5 2,156614641
net4 6 4,933429941
net5 7 4,130396932
net6 8 3,298814559
net7 9 11,10834092
net8 10 2,492953096
net9 11 2,414800761
net10 12 3,931451323
net11 13 1,935154164
Table 8 : MAPE values of created models
24
The above table shows the MAPE values of the created models and the net11 model gives the
lowest MAPE value. so the forecasting process will continue with net11 model.
Figure 9 : Actual Demand and Model (net11) Output
The graph in Figure 9 shows the actual value of the electiricity demand and output of the
artificial neural network model which is model net11.
5.6 Forecasting and Results
Future values of the independent variables should be known in order to estimate the future net
electricity demand. Therefore, the future values of the independent variables are forecasted by
the forecasting extension of Excel. Forecasting was made from 2017 to 2023 and 7 years of
forecasting was applied for independent variables. The future values of the independent
variables are normalized after being forecasted and added to the MATLAB software as a
sample matrix. Afterwards, this matrix is given as input to the model ‘net11’ network via the
simulate tab of nntool toolbox.
0,00
50000,00
100000,00
150000,00
200000,00
250000,00
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Actual Demand (Gwh) Model Output (Gwh)
25
Figure 10 : Simulate Tab of Neural Network Toolbox
The artificial neural network gave the forecasting results as output value. Since these values
are normalized, denormalization has been applied to reach the actual values. Forecasted
values for 2017 to 2023 are as in Table 9.
Year Forecasted Value (Gwh)
2017 236194,433
2018 243398,298
2019 250348,191
2020 257093,159
2021 263663,545
2022 270076,907
2023 276342,314
Table 9 : Forecasted Value of 2017 to 2023
26
Figure 11 : Artificial Neural Network Forecasting Graph
6. CONCLUSION
In this study, Turkey's annual electricity demand was forecasted by using artificial neural
network forecasting models that is widely used in energy sector. Unlike the studies in the
literature, the total floor area of buildings in Turkey and the number of electricity subscribers
are used in the models. The MAPE value is used to measure the performance of the models to
be used for forecasting.
In this study, electricity demand was forecasted for the years 2017-2023 in Turkey. The
results are very important in the decision-making process for future investment plans of all
institutions related to energy production, especially the Ministry of Energy and Natural
Resources.
As can be understood from the results of the research, the electricity demand is increasing
from year to year in Turkey. Turkey needs to increase its total installed capacity to meet this
demand. Using local resources to be not dependent on the outside in energy production is very
important in this matter. The mission of the ministry of energy and natural resources is to
reduce foreign dependency by using local resources. In line with this mission, it is expected to
increase the total installed capacity by using local resources in order to meet the growing
electricity demand in Turkey. Our country constitutes 16% of the economic potential of
Europe in terms of hydraulic resources. It is beneficial for our country to turn to hydraulic
0,00
50000,00
100000,00
150000,00
200000,00
250000,00
300000,00
19
98
19
99
20
00
20
01
20
02
20
03
20
04
20
05
20
06
20
07
20
08
20
09
20
10
20
11
20
12
20
13
20
14
20
15
20
16
20
17
20
18
20
19
20
20
20
21
20
22
20
23
Actual Demand (Gwh) Model Output (Gwh)
27
resources to evaluate this potential, to reduce foreign dependency, to turn to renewable energy
and to meet electricity energy demand which is increasing year by year. Reduction of foreign
dependency by the use of local resources can result in cheaper electricity production and
distribution. The electricity generated at a cheaper price will also increase the growth rate of
the industry.
In future studies, it is possible to develop models in which more accurate forecasting results
can be obtained by increasing independent variables or adding new independent variables.
Bitcoin (cryptocurrency) production, which has increased in recent years, has a significant
effect on electricity consumption. The datas of this production can be shown as one of the
independent variables that can be used in future forecasting models.
REFERENCES
Ağyar, Z. (2010). Yapay Sinir Ağlarının Kullanım Alanları ve Bir Uygulama. Mühendis ve
Makina, 22.
Al Shamisi, M. H., Assi, A. H., & Hejase, H. A. (2011). Using MATLAB to Develop
Artificial Neural Network Models for Predicting Global Solar Radiation in AI Ain
City - UAE. A. H. Assi içinde, Engineering Education and Research Using MATLAB
(s. 226). InTech.
Atak, & Akay. (2007). Grey Prediction with Rolling Mechanism for Electricity Demand
Forecasting of Turkey .
ATASEVEN, B. (2007). Satış Öngörü Modellemesi Tekniği Olarak Yapay Sinir Ağlarının
Kullanımı: Petkim'de Uygulanması.
Barış, & Küçükali. (2010). Turkey's Short-term Gross Annual Electricity Demand Forecast by
Fuzzy Logic Approach.
Bilgili, M., Şahin, B., Yaşar, A., & Şimşek, E. (2012). Electric Energy Demands of Turkey in
Residential and Industrial Sectors. Renewable and Sustainable Energy Reviews.
Bulgan, A. (2013). Pnömatik Devrelerde Optimum Tasarım ve Yapay Sinir Ağları ile
Titreşim Analizi. Erciyes Üniversitesi Fen Bilimleri Enstitüsü.
Çuhadar, M. (2006). TURİZM SEKTÖRÜNDE TALEP TAHMİNİ İÇİN YAPAY SİNİR
AĞLARI KULLANIMI VE DİĞER YÖNTEMLERLE KARŞILAŞTIRMALI
ANALİZİ. Isparta: Süleyman Demirel Üniversitesi Sosyal Bilimler Enstitüsü.
Demuth, H., & Beale, M. (1992). Neural Network Toolbox For Use with MATLAB. Natich,
MA: The MathWorks, Inc.
Ekonomou. (2010). Greek Long-term Energy Consumption Prediction Using Artificial Neural
Networks.
28
EMO Enerji Komisyonu. (2011). Türkiyenin Enelektirik Enerjisi Sektörünün Tarihsel
Gelişimi.
Erilli, N. A., Eğrioğlu, E., Yolcu, U., Aladağ, Ç. H., & Uslu, V. R. (2010). Türkiye'de
Enflasyonun İleri ve Geri Beslemeli Yapay Sinir Ağlarının Melez Yaklaşımı ile
Öngörüsü. Doğuş Üniversitesi Dergisi, 42-55.
Erler, M., Sağıroğlu, Ş., & Beşdok, E. (2003). Mühendislikte Yapay Zeka Uygulamaları - 1,
Yapay Sinir Ağları. Kayseri: Ufuk Kitap Kırtasiye.
EUAŞ, E. (2012). YENİLENEBİLİR ENERJİ KAYNAKLARI VE ENERJİ JEOPOLİTİĞİ.
Fausett, L. (1994). Fundamentals of Neural Networks. New Jersey: Prentice-Hall.
Filik, Ü. B., Gerek, Ö. N., & Kurban, M. (2011). A Novel Modeling Approach for Hourly
Forecasting of Long-term Electric Energy Demand. Energy Conversion and
Management.
Geem. (2011). Energy Demand Estimation Using Artificial Neural Network.
Hamzaçebi, C. (2011). Yapay Sinir Ağları Tahmin Amaçlı Kullanımı Matlab ve
Neurosolutions Uygulamalı. Bursa: Ekin Yayınevi.
Haykin, S. (1994). Neural Network, A Comprehensive Foundation. New York: Macmillan
College Publishing Company.
International Electrotechnical Commission. (2018). The Strategic Importance of
Electrification.
Kankal, M., Akpınar, A., Kömürcü, M. İ., & Özşahin, T. Ş. (2011). Modeling and Forecasting
of Turkey’s Energy Consumption Using Socio-economic and Demographic Variables.
Applied Energy.
Karasu, F. (2012). Petrokimya Sektöründe Talep Tahmininde Yapay Sinir Ağlarının
Kullanılması " Petkim A.Ş. Örneği". Dokuz Eylül Üniversitesi Sosyal Bilimler
Enstitüsü.
KAYA, İ., OKTAY, S., & ENGİN, O. (2005). Kalite Kontrol Problemlerinin Çözümünde
Yapay Sinir Ağlarının Kullanımı. Erciyes Üniversitesi Fen Bilimleri Enstitüsü
Dergisi, 2(21), 92-107.
Koç, İ. (2017). Türkiye'de İç Hatlarda Havayolu Yolcu Taleplerinin Yapay Sinir Ağları
Kullanarak Tahmini. İstanbul Üniversitesi Fen Bilimleri Enstitüsü.
Krose, B., & Smagt, P. v. (1996). An Introduction to Neural Networks. Netherlands: The
University of Amsterdam.
Kutay, & Hamzaçebi. (2007). Forecasting of Turkey's Net Electricity Energy Consumption on
Sectoral Bases .
29
Mete, T. (2008). Kesikli Bir Biyoreaktörde Yapay Sinir Ağlarıın Kullanımı. Ankara: Ankara
Üniversitesi Fen Bilimleri Enstitüsü.
Onaran, B. D. (2010). Forecasting Excahnge Rates Using Artificial Neural Networks. İstanbul
Bilgi Üniversitesi Sosyal Bilimler Üniversitesi.
OZAN, N. (2017). Türkiye’de Yenilenebilir Kaynaklardan Elektrik Enerjisi Üretimi.
Öztemel, E. (2012). Yapay Sinir Ağları. İstanbul: Papatya Yayıncılık.
Pao. (2005). Comparing Linear and Nonlinear Forecasts for Taiwan's Electricity
Consumption.
Rajpal, P., Shishodia, K., & Sekhon, G. (2006). An Artificial Neural Network for Modeling
Reliability, Avalibility and Maintainability of a Repariable System. Reliability
Engineering and System Safety, 91, 809-819.
Roper, & Geem. (2009). Energy Demand Estimation of South Korea Using Artificial Neural
Network .
Saraç, T. (2004). Yapay Sinir Ağları. Ankara: Gazi Üniversitesi Endüstri Mühendisliği
Bölümü.
Sarı, M. (2016). Yapay Sinir Ağları ve Bir Otomotiv Firmasında Satış Talep Tahmini
Uygulaması. Sakarya Üniversitesi Fen Bilimleri Enstitüsü.
Sönmez, İ. (2015). Seydişehir Bölgesinin Orta Vadedeki Elektrik Enerjisi Talebinin Yapay
Zeka ile Tahmini. Konya: Selçuk Üniversitesi Fen Bilimleri Enstitüsü.
Stergiou, C., & Siganos, D. (1996). Neural Networks.
https://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/cs11/report.html adresinden
alındı
Sueri, M. (2016). Kanalizasyon Hattı Maliyetlerinin Yapay Sinir Ağları ile Tahmini. Gazi
Üniversitesi Fen Bilimleri Enstitüsü.
Şahin, V. (1994). ENERJİ SEKTÖRÜNDE GELECEĞE BAKIŞ " Arz, Talep ve Politikalar ".
İstanbul: TÜSİAD.
Şen, Z. (2004). Yapay Sinir Ağları İlkeleri. İstanbul: Su Vakfı.
T.C. Enerji ve Tabii Kaynaklar Bakanlığı. (2017). T.C. Enerji ve Tabii Kaynaklar Bakanlığı.
T.C. Enerji ve Tabii Kaynaklar Bakanlığı: enerji.gov.tr/tr-TR/Sayfalar/Elektrik
adresinden alındı
TUIK. (2016). Gross Electrical Energy Generation.
YAVUZDEMİR, M. (2014). Türkiye'nin Kısa Dönem Yıllık Brüt Elektirik Enerjisi Talep
Tahmini.
30
YILMAZ, M. (2012). Türkiye’nin Enerji Potansiyeli ve Yenilenebilir Enerji Kaynaklarının
Elektrik Enerjisi Üretimi Açısından Önemi.
Yurtoğlu, H. (2005). Yapay Sinir Ağları Metodolojisi ile Öngörü Modellemesi: Bazı Makro
Ekonomik Değişkenler için Türkiye Örneği. Ankara: DPT.
Zhang, G., Patuwo, B. E., & Hu, M. Y. (1998). Forecasting with Artifical Neural Networks:
The state of the art. International Journal of Forecasting, 42-44.