Proceedings of the

International Conference in Mathematics 2012

Chairman: Prof. Muhammad Syam

Editors: Dr. Philippe Poulin (Tome I), Dr. Simon Cowell (Tome II)

United Arab Emirates University

11-14 March 2012
 The Proceedings of the ICM 2012 is a collection of pre-prints corresponding to talks and
posters presented at the International Conference in Mathematics 2012 (Al Ain, United Arab
Emirates University, 11-14 March). By submitting their paper to a volume of the Proceedings the
Authors have accepted that it will be published on, and downloadable from, the ICM website for a
long period of time. Apart from that, the Authors retain the copyright of their text.
Contents

M. Montazeri, Y. Jafarzadeh, A. Tavasoli

Analytical solution of dual fully fuzzy linear system ................................................................................................................................ 1
Olaniyi S. Maliki, Adeola O. Maliki, Chukwuemeka O. Agwu
Application of Artificial Neural Network model in the forecasting of RGDP of Nigeria as a function of crude oil consumption and
other key economic variables ................................................................................................................................................................. 5
Heydari Aghile, Naseri Roghaye
Application of Daubechies wavelet for image compression ................................................................................................................... 11
Y. Jafarzadeh
Application of differential and integral equation in financial mathematics ............................................................................................ 19
Benhammouda Brahim
Automated exam generation with stratified dynamic databases .......................................................................................................... 23
David Allison, Andriy Didenko, Gary Miller
Can Microsoft Excel help to visualize eigenvalues and eigenvectors? .................................................................................................... 27
Mohammed Goma Tanko
Can Social Justice Methodology Enhance the Learning of Mathematics? .............................................................................................. 31
Anil Rajput, Pradeep Kumar Dwivedi, Archna Jain, Deepratna Mansulkar
Common fixed points of four mappings ................................................................................................................................................ 41
Ali Asghar Behroozpoor, Mohammad Mehdi Mazarei
A comparison between Runge-Kutta method and optimization method for nonlinear systems of ordinary differential equations ........ 45
Tamer Shahwan, Raed Said
A comparison of Bayesian methods and artificial neural networks for forecasting chaotic financial time series ................................... 49
Ahmad Kazemlou Sheikhi, Akbar Amiri Rick
The comparison of extended hyperbolic tangent and the extended (G'/G) expansion method in solution (2+1) dimension of
Schrdinger equation ............................................................................................................................................................................ 55
Boukemara Ibtissem, Djellit Ilhem, Ferchichi Mohamed Rda
Complex Dynamics in Piecewise non Linear maps ................................................................................................................................. 61
Fariba Pidani
Cutting plane methods for solving inexact problems ............................................................................................................................. 65
Ali Ismaiel Sorour EL-Bus
Effectiveness using of multiple graphical software in the development thinking skills of mathematics' students, Faculty of Edu-
cation .................................................................................................................................................................................................... 73
Mohammed Abdul Rahim, Syed Waseem Raja, M. V. Ramana Murthy, P. Muthu
Effect of slip velocity on fluid flow in a channel of varying cross-section ................................................................................................ 79
Zeana Z. Jamial, Eman A. Hussian
Equivalent between the criterion and the three open set's conditions in codisk-cyclicity ...................................................................... 97
Mahmoud Syam, Kamarulzaman Ibrahim, Amer Ibrahim Al-Omari
Estimating the population mean using stratified double ranked set sample .......................................................................................... 99
Madhvi Shakya, Manisha Jain
Finite element approach to study thermal disturbances of human body during wound healing process after plastic surgery .............. 105

iii
Muhammad Arshad, Akbar Azam
Fixed points solutions of sequence of locally contractive fuzzy mappings via iterative procedure ....................................................... 113
Aye Hande Erol, Nilay zelikkan, Asm Tokgz, Semih zel, Selim Zaim, mer F. Demirel
Forecasting Electricity Consumption Of Turkey Using Time Series Methods And Neural Networks ...................................................... 117
Kharfouchi Soumia
A generalized moments estimator for random coefficients auto-regressive model on a plane ............................................................ 129
Salah Derradji Lylia, Moumeni Abdelkader
Global solution for reaction-diffusion systems with non diagonal matrix ............................................................................................ 131
Samira Ghasemi, Alireza Moghali
Investigating the relationship between BPR and application of IT in PIDEC employees ....................................................................... 135
Mahmoud Syam, Ahmed Taleb
Investigating the use of some tools of Instructional Technology in teaching Mathematics; A case study ............................................. 141
Abdelaziz Amroune, Lemnaouar Zedam
Many valued logic and Zadeh's fuzzy sets: A Stone theorem generalization ........................................................................................ 145
Husnia Mohamed Eldanfour
Modified Newton's methods with fifth or sixth-order convergence .................................................................................................... 149
E. S. Ismail, N. Tahat
A new electronic signature based on hybrid-mode problems .............................................................................................................. 153
Husnia Mohamed Eldanfour
Newton's method with sixth-order convergence for solving systems of non-linear equations ............................................................. 157
Mohammad Tohidi, Jafar Emamipour, Amir-Masoud Eftekharimoghadam
A novel fuzzy similarity measure based on monotonic inclusion for image retrieval systems ............................................................. 163
,
Mohammad Mehdi Mazarei Ali Asghar Behroozpoor
Numerical solution of nonlinear partial differential equations using optimization method ................................................................ 167
M. A. Bokhari, H. Al-Attas
-1
On (1 t ) weighted orthogonal polynomials: Applications and computational aspects ................................................................... 171
Said Mohamed Said
On a nonlinear problem intervening in relativistic quantum mechanics perturbed by a factor of amortizement ................................. 175
Badis Abdelhafid
On minimal non-((periodic radicable abelian)-by-hypercenal) groups ................................................................................................. 181
Tahir H. Ismail1, Hind Y. Saleh, Barah M. Sulaiman
On some properties of convex galaxies and functions ......................................................................................................................... 183
Ismael Akray
Soc-+-Supplemented Modules ............................................................................................................................................................ 185
Faieza Samat, Fudziah Ismail, Norihan Md Arifin, Mohamed Suleiman
Solving second order initial value problems by an exponentially fitted explicit hybrid method .......................................................... 191
Seid Bahlali, Boulakhras Gherbal
Stochastic maximum principle for backward doubly control systems .................................................................................................. 195
N.Rajeswari, K.Indrani
Supply chain decision model for time dependent deteriorating items, variable holding cost and trade credit ..................................... 199

iv
Forecasting Electricity Consumption Of Turkey Using Time Series
Methods And Neural Networks
Aye Hande Erol 1,*, Nilay zelikkan 2 , Asm Tokgz 3 , Semih zel 4 , Selim Zaim 5 and
mer F. Demirel 6
1
Department of Industrial Engineering, Marmara University, Institute For Graduate Studies In Pure And Applied Sciences ,
Gztepe, 34722, TURKEY , hande.erol@marmara.edu.tr,
2
Department of Industrial Engineering, Marmara University, Institute For Graduate Studies In Pure And Applied Sciences ,
Gztepe, 34722, TURKEY , nilay.ozcelikkan@gmail.com,
3
Department of Industrial Engineering, Marmara University, Institute For Graduate Studies In Pure And Applied Sciences ,
Gztepe, 34722, TURKEY , asimtokgoz@gmail.com,
4
Department of Industrial Engineering, Marmara University, Institute For Graduate Studies In Pure And Applied Sciences ,
Gztepe, 34722, TURKEY , semih.ozel@marmara.edu.tr4,
5
Department of Industrial Engineering, Marmara University, Institute For Graduate Studies In Pure And Applied Sciences ,
Gztepe, 34722, TURKEY , selimzaim@marmara.edu.tr5,
6 Department of Industrial Engineering, Fatih University, TURKEY , odemirel@fatih.edu.tr,
ICM 2012, 11-14 March, Al Ain

ABSTRACT form model. The former is a double-log linear demand model

The high growth rate of Turkish Economy in the last decade resulted
in substantial increase of energy consumption. Both commercial
and residential electricity consumption has increased significantly
and due to this Turkey is planning to make new investments to meet
this growing demand of electricity. The government is even consi-
dering building nuclear power plants which require high level of
initial investments. To forecast the future demand of electricity be-
comes a key issue for giving this type of vital decision. In this re-
search, we have forecasted the monthly electricity consumption of
Turkey by using time series methods and artificial neural networks.
The results of the methods are compared by using performance
measures based on the difference between real and forecasted
values. The best performing method has been proposed to make
predictions for the future electricity consumption. Keywords: Auto-
regressive Integrated Moving Average; Turkish electricity con-
sumption; Time series analysis, Holt-Winters Method, ARIMA,
ANN.
under which energy demand is assumed to be a direct linear func-
tion of energy price and real income. Kouris (1981), Drollas (1984)
and Stewart (1991) developed this model in their studies [23, 24,
25]. The second model is a disaggregated demand model that
separates energy demand into several number of demand equations
and treats it as an indirect function of energy price and real income.
Pindyck (1979) studied the second model [7].
There are many studies that concerns about the validity of the fixed
coefficients assumption that uses constant elasticities for the entire
sample period under study in the electricity demand equation
employed by these methods. This characteristic of the model is
questionable in view of the changes that could have taken place in
the economy over such a long period of time affecting the demand
for electricity [7].

Therefore, it is argued that if data is collected over a relatively long

1 INTRODUCTION
Modeling electric energy consumption is helpful in planning gen-
eration and distribution by power utilities [1]. Since the early
1970s, various estimation methods are used in various studies of
energy demand. In most of these studies, the purpose is to measure
the impact of economic activity and energy prices on energy de-
mand [7].
In most cases, energy demand studies have adopted two different
types of modeling; namely, reduced form model and structural
*To whom correspondence should be addressed.
time period to estimate an electricity demand function, the possi-
bility that the parameters in the regression may not be constant
should be considered [7].
Previous methods make use of time series data to estimate energy
demand but they do not analyze the data to set up its properties and
thus they completely assume the data to be stationary. Engle et. al,
(1989) [18] analyzed time series properties and estimating elastici-
tys based on co integration and Error Correction Method (ECM),
which enables full analysis of the properties of the relevant data
before actual estimation and commonly used is the Augmented
DickeyFuller (ADF) test. Subsequent improvements related to

117
Aye Hande Erol, et al.

this approach have been in the form of insertion of more specific 3 METHODOLOGY
energy-related variables in the model and the development of new
methods to identify co integrating relationships, using the Autore-
3.1 The Box-Jenkins (ARIMA) Methodology
gressive Distributed Lag Model (ARDL) and the Johansen Maxi-
mum Likelihood Model (JML) are especially popular [7].
Autoregressive integrated moving average modeling (ARIMA)
was greatly improved by G.E.P. Box and G.M. Jenkins. They have
contributed a great deal of effort to identify, fit and check the
model validity [8]. Box-Jenkins methodology is different from the
2 FORECASTING ENERGY CONSUMPTION other forecasting methods as it does not assume any regular pattern
in the history of the data series that will be forecasted. It heavily
There are many approaches include statistical models, neural net-
relies on its own past while describing the series and its an appro-
works, fuzzy logic and expert systems. A number of them focused
priate method for medium to long length time series data [4]. It
on mid-term and long-term electricity demand and incorporated
iteratively improves the model until residuals are small, stochasti-
climate factors [3].
cally distributed and show no regular pattern [7][8].
Ediger and Akar (2007) [6] improved the ARIMA and the seasonal
The ARIMA model is inherited from the ARMA(p,q) model which
ARIMA methods for forecasting the primary energy demand with
has the form as,
respect to types of fuel for the period 2005-2020 in 2007. nler
(2008) used the Particle Swarm Optimization (PSO) technique to
forecast the 2006-2025 period of energy demand based on the data Yt 0 1Yt 1 2Yt 2 ... pYt p t 1 t 1 2 t 2 ... q t q (1)

obtained in the period 1979-2005 [19]. Ediger and Tatldil (2002)

where Yt is the response (dependent) variable at time t, Yt-1, Yt-
used the cycle analysis method to forecast the primary energy
2,, Yt-p are response variables at time lags t-1, t-2,,t-p, s
demand based on the data from the 1950-1999 period [20]. Toksar
are autoregressive parameters to be estimated, s are the moving
(2009) estimated the electricity demand using Ant Colony Optimi-
average parameters to be estimated, s are random errors general-
zation (ACO) for the period 2007-2025 based on different Gross
ly assumed to be independent, identically distributed variables
Domestic Product (GDP), population and import growth rates [21].
sampled from a normal distribution with mean zero. Forecasts of
Hamzaebi (2007) forecasted the electricity energy consumption
the model rely on current and past values of the response Y and
for the period 2003-2020 by using Artificial Neural Networks
current and past values of the errors [8].
(ANN) from the period 1970-2004 [27]. Erdodu (2007) used co
integration analysis and ARIMA model to estimate 2005-2014
The ARIMA (p,d,q) model is the combination of ARMA (p,d) and
period electricity demand based on the data from the period 1923-
d times differencing applied to it. The notation of SARIMA model
2004 and ARIMA model for natural gas demand [7]. Albayrak
is represented as ARIMA(p,d,q) (P,D,Q)s where p is autoregres-
(2010) developed appropriate ARIMA models for forecasting of
sive order, q is moving average order and d is the times of diffe-
primary energy production and consumption levels and compare
rencing, P is the number of seasonal autoregressive (SAR) terms,
the results with previous studies [4].
D is the number of seasonal differences, Q is the number of sea-
sonal moving average (SMA) terms and s is the periodicity of the
The studies on energy demand forecasting in Turkey dates back to
season. Generally, the ARIMA(p,d,q)(P,D,Q)s equation is
1960s. The tradition of energy forecasting by using simple regres-
represented as follows [12],
sion techniques was initiated by the State Planning Organization
(SPO). Similar studies later have been continued by the Ministry of p ( B) P ( B s )(1 B s ) D (1 B) d Yt q ( B)Q ( B s ) t
Energy and Natural Resources of Turkey (MENR) and a number of
Where
academicians. These early forecasts consistently predicted much p ( B ) (1 1 B 2 B 2 ... p B p )
, the nonseason-
higher values than the consumptions that actually realized. Later,
al autoregressive operator of order p
starting from 1984, several econometric modeling techniques have P ( B L ) (1 1 B L 2 B 2 L ... P B PL ) , the
been employed. Among them, the Model for Analysis of Energy
seasonal autoregressive operator of order P
Demand (MAED), which was used in energy planning and policy q ( B) (1 1 B 2 B 2 ... q B q )
, the
making by MENR, has been the most commonly applied one [22].
nonseasonal moving average operator of order q
Q ( B L ) (1 1 B L 2 B 2 L ... Q B QL )
,
Energy consumption in Turkey also shows a tendency to increase
the seasonal moving average operator of order Q
rapidly in parallel with the developments throughout the world. ( B ) ( B ) , a constant term where is the mean
p P
L

The changes that occurred in socioeconomic balances especially

of stationary time series
after 1980 have a significant contribution to this tendency [22].
Phis(,) , thetas(,) and are unknown parameter that can be
calculated from the sample data. s are random shocks that are
assumed to be independent of each other. For abbreviation purpose
in this model, a backshift operator (B) is used to represent the time
series observation backward in time by k period and symbolized as
Bk, meaning that BkYt = Yt-k.

118

The term represents the d times nonseasonal difference while the
term represents the D times seasonal difference.
The Box-Jenkins method is iteratively follows the identification,
estimation and model checking steps until a satisfactory model
found and after that the forecasting is made. First, in order to iden-
tify the model correctly, the series data must be stationary, that is,
the series must vary about a fixed level, and inconstant variance,
trend and seasonality must be removed from the series data. In
order to convert non stationary series to stationary series some
modifications are applied to the series data. The inconstant va-
riance can be removed from the data by applying methods such as
the natural logarithm, square root or dividing the data to its moving
standard deviation. Seasonality is corrected by taking the average
of each seasonal/cyclic variables and subtracting it from the cor-
responding seasonal data. The trend in the non stationary data is
often removed by differencing. The differencing can be done
iteratively until data become stationary, but over differencing may
introduce dependence while none exists.
After obtaining a stationary series data, the form of model must be
identified. This is accomplished by matching the autocorrelation
and partial autocorrelation functions of the series data with a theo-
retical pattern. Usually, if sample autocorrelations die out mildly
and sample partial autocorrelations cut off, the model will require
autoregressive terms. In case the sample autocorrelations cut off
and the sample partial autocorrelations die out, the model will need
moving average terms. If both sample autocorrelations and sample
partial autocorrelations die out, autoregressive and moving average
terms are needed. The number of significant terms in sample auto-
correlations and sample partial autocorrelations determines the
orders of the MA and AR parts [8]. The model identification phase
ends up with determining the values of p, q, and d of ARIMA
(p,d,q). If series data contains seasonality then seasonal values P, Q
and D must be identified additionally. When the model is
SARIMA, the seasonal components are determined according to
spikes where the autocorrelation functions cut the confidence
limits [6][8].
After selecting a rough model, the parameters (phis and thetas)
for that model must be estimated. The parameters of the ARIMA
models are estimated by minimizing the sum of squares of the
fitting models. The significance of the parameters is tested with t-
statistics and p-value. As the t-statistic is itself not informative, it is
used for the calculation of p-value for hypothesis testing [8]. The
null hypothesis (H0) states that all parameters of the model are
zero against the alternative hypothesis (Ha) that some parameters
are different than zero. If the p-value is less than , representing
the level of confidence as (1- ) %, then null hypothesis is rejected,
meaning that there are some parameters different than zero.
Before forecasting with model, the adequacy of the model must be
checked by examining the information about the residuals, which
should be random and normally distributed. This can be done very
easily by examining the histogram, or the normal probability plot
or a time sequence plot of residuals. The residual autocorrelations
should be small and within the confidence intervals at the ACF and
Forecasting Monthly Electricity Consumption in Turkey
PACF plots. The overall adequacy of the model is provided by a
chi-square (2) test based on the Ljung-Box Q statistics. If the p-
value regarding the Q statistic is very small (say, p-value <
=0.05), the built model is considered as inadequate. In another
word, the built model can be considered as adequate, if the signi-
ficance level of Ljung-Box Q statistics is greater than [4][8].
3.2 Holt-Winter's Method
Time series data is usually plotted as series on the vertical axis
against the time usually on horizontal axis. The aim is to get a
visual perception of the series data components such as trend,
seasonality which can help an analyst to select the appropriate
model with the potential to produce the best forecasts. Time series
forecasting usually assumes that a time series is composed of a
pattern and some random error. It is tried to separate the pattern
from the random error by understanding the trend, its long-term
fluctuations, and its seasonality, caused by seasonal use and de-
mand, exist in the pattern. Time series often exhibit seasonal beha-
viors and this behavior has a tendency to repeat itself every L
periods [9].
Holt-Winters method is also known as triple exponential smooth-
ing and consists of three components, namely, level, trend and
seasonality that change by time. If the fluctuation at a certain time
of each seasonal term is nearly constant then the seasonality is
considered as additive. On the other hand, if this seasonal fluctua-
tion is proportional to a certain factor then, the seasonality is natu-
rally considered as multiplicative. For this reason, there are two
versions of Holt-Winters method which take into account the
seasonality.
3.2.1.Holt-Winter's Multiplicative Method
The following equations are (taken from Russell Cheng web site
[10]);
Yt
Lt (1 )( Lt 1 bt 1 ) (3)
St s
bt ( Lt Lt 1 ) (1 )bt 1 (4)
Yt
St (1 ) St s
Lt (5)
Ft m ( Lt bt m ) S t s m
(6)
where s is the number of periods in one cycle of seasons e.g. num-
ber of months or quarters in a year. Lt and bt are respectively (ex-
ponentially smoothed) estimates of the level and linear trend of the
series at time t, whilst Ft+m is the m-step-ahead forecast starting
from time period t. For initialization, one complete cycle of data,
i.e. s values, is necessary. The parameters , , should be be-
tween the interval (0, 1), and can be selected by minimizing mean
absolute error (MAE) and square root of mean squared error values
(RMSE).
Then set
119
Aye Hande Erol, et al.

1 artificial neural networks go to the fore by predicting the informa-

Ls (Y1 Y2 ... Ys ) (7) tion by learning from experience. As learning methods are very
s vital in artificial intelligence, many studies are still being con-
ducted to produce best learning rules. Some of these rules are
For initialization of the starting trend, s + k time periods are Hebbs rule, Hobfields rule, Deltas rule and Kohonens rule
needed. If the series is long enough then it is a good choice to use among which the Hebbs rule is the oldest and most used one.
two complete cycles for this purpose.
In a neural network system model, it is supposed that the input is
Yk processed in neurons and it is transformed between neurons by
Sk where k = 1,2,...,s (8) signals. Each unit has its own value and these values are multiplied
Ls with signals. The sum of these values is called as net input. The
output is calculated with activation function. The input of the
Initial seasonal indices can be computed as activation function is our net input. Here is an example of a neural
network, [13] as seen at Figure 1.
1 Ys 1 Y1 Ys 2 Y2 Y Yk
bs ... s k
(9)
k s s s

3.2.1. Holt-Winter's Additive Method

The equations are (taken from Russell Cheng web site [10]);

Lt (Yt St s ) (1 )( Lt 1 bt 1 ) (10)

bt ( Lt Lt 1 ) (1 )bt 1 (11)
St (Yt Lt ) (1 ) St s
(12)
Ft m Lt bt m St s m
(13)
where s is the number of periods in one cycle. The parameters , , Fig. 1. An example of neural network
should be between the interval (0, 1), and can again be selected
by minimizing error values.
As you see in the Figure 1, the neural network consists of inputs, weights,
The initial values of Ls and bs can be computed as in the multip- summation function, activation function and outputs.
licative case. The initial seasonal indices can be taken as

Sk Yk Ls where k = 1,2,...,s (14)

3.3 Artificial Neural Network

Since the emergence of artificial intelligence concept, the most

controversial issue is whether machines are intelligent or not.
There are still many ongoing arguments on this subject. Even
though a lot of people still believe that machines cannot be intelli-
gent, many ongoing studies on automation systems are being con-
ducted to gain ability to make decisions by using artificial intelli-
gence. Today, control mechanisms have features that almost dont
require any human intervention which, eventually, ends up with
Fig. 2. Multilayer perceptron [17]
more efficient and faster production. Consequently, the ultimate
goal is to make machines more intelligent by programming.
As learning mechanism, mostly the backward chaining is used in
There are many approaches within the concept of artificial intelli- multilayer perception. After backward chaining is invented, neural
gence such as genetic algorithms, fuzzy logic, expert systems, networks are more popular as a research area [16]. This algorithm
artificial neural networks and intelligent agents which are pro- helps decreasing the errors starting with outputs and ending with
grammed to solve different problems effectively. These approach- inputs. That is why it is called backward chaining.
es/methods can be adapted to solve each technical problem accord-
ing to the characteristics of the problem. Among these approaches,

120
 Forecasting Monthly Electricity Consumption in Turkey

In literature, neural network was used especially for energy con- After the application of natural logarithm transformation to the
suming forecasting. As an example of time series study, the inter- data (see Figure 5), the variance became constant but, trend and
ested person can investigate the study of [15]. Zhang (2003) [14] seasonality can be seen.
used ARIMA and artificial neural network together in his study.

4 ENERGY CONSUMPTION FORECAST IN

TURKEY: 1970-2010

4.1 The ARIMA Modeling of the Energy Consump-

tion
The statistical software package of IBM SPSS Statistics 19 and
Minitab 16 are employed to develop and use the ARIMA models.
This section contains how the four stages, namely identification,
estimation, diagnostic checking and forecasting, of ARIMA model
applied to the electric consumption data.

4.1.1 Identification
Fig. 5. Natural Logarithm Transformation of the Electric Consumption
As a first step to model identification, monthly electric consump- Data.
tion time series data is plotted for the years 1970 to 2010. When
the time series plot is examined Figure 3, the main features of the
data are: (1) the overall level rises almost linearly until 1985 and
after that there are exponential fluctuations. (2) the variance is not
constant meaning that transformation is necessary (3) there is also
seasonal pattern of monthly length 12 which can be more easily
seen when the yearly data superimposed on the same graph as seen
in the Figure 4.

Fig. 6. Natural Logarithm and Differencing (1) Transformation of the

Electric Consumption Data.

After natural logarithm and differencing (1) transformation, time

series plot seems as stationary but, from the ACF (Figure 7) and
PACF (Figure 8) plots it can be still seen the signs of non statio-
nary and autocorrelations. Nonseasonal and seasonal lags show no
known pattern yet.

Fig. 3. Time Series Plot of Electric Consumption between 1970 and 2010.

Time Series Plot of 2007, 2008, 2009, 2010

18000
2007
2008
17000 2009
2010

16000
Consumption

15000

14000

13000

12000
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Month

Fig. 4. Seasonal Pattern of the Electric Consumption Data. Fig. 7. ACF of Natural Logarithm and Differencing (1) Transformation of
the Electric Consumption Data.

121
Aye Hande Erol, et al.

Fig. 9. Natural Logarithm, Differencing (1) and Seasonal Differencing (1,

period 12) Transformation of the Electric Consumption Data.

Fig. 8. PACF of Natural Logarithm and Differencing (1) Transformation

of the Electric Consumption Data.

Taking the additional nonseasonal differences, it still is not mea-

ningful about ACF and PACF plots that seem stationary. So it is
applied the seasonal differencing to the data and get the Figure 9,
Figure 10 and Figure 11. ACF and PACF now seem more statio-
nary and have only a few significant coefficients at the seasonal
and nonseasonal levels either they cut off or die down fairly quick-
ly. Therefore, the final transformed series data can be considered
as stationary and ARIMA modeling procedures can be imple-
mented now. Both ACF and PACF have some significant spikes at
nonseasonal, seasonal and near-seasonal lags. Even though addi-
tional seasonal differencing may help reduce these extra spikes, it
is not common to use seasonal second differencing, nor do more Fig. 10. ACF of Natural Logarithm, Differencing (1) and Seasonal Diffe-
rencing (1,period 12) Transformation of the Electric Consumption Data.
than two total differencing (nonseasonal and seasonal). That is
why; it is decided to ignore these spikes at near-seasonal and near-
half-seasonal lags [1]. These significant lags also may be the ref-
lections of the strong seasonality and often disappear at the resi-
duals of ACF (Figure 10) and PACF (Figure 11) whose model
dont introduce the coefficients of those lags. It is also known that
excessive differencing may also introduce artificial patterns that
result in forecast errors with large variance [11].

The ACF in Figure 10 cuts off to zero suddenly after lag 1 at the
nonseasonal level while the PACF in Figure 11 almost dies down
meaning MA (1), nonseasonal moving average. Both ACF and
PACF dies down at the seasonal or near seasonal lags that suggests
to add SAR (1) and SMA(1) to the model. Thus the final model is
the combination of them and stated as ARIMA (0,1,1)(1,1,1)12.
The PACF in Figure 11 also reminds of the AR(1) pattern, that will
result in ARIMA (1,1,1)(1,1,1)12 model as a whole but, the for-
mer model is simple and has slightly better residual errors than the
later. Also principle of parsimony advises us to prefer simple
Fig. 11. PACF of Natural Logarithm, Differencing (1) and Seasonal
models over complex ones. Differencing (1,period 12) Transformation of the Electric Consumption
Data.

122
 Forecasting Monthly Electricity Consumption in Turkey

4.1.2 Estimation Table 2: SARIMA Model Parameters

Estimate SE t Sig.
The parameters of the model ARIMA (0,1,1)(1,1,1)12 are estimated
Consumption Natural Difference 1
by using the IBM SPSS 19 package. The model fit statistics are Log MA(1) 0.540 0.039 13.943 0.000
Stationary R-squared 0.411, R-squared 0.997, RMSE 241.743, AR, Seasonal (1) 0.270 0.066 4.081 0.000
MAPE 2.719, MAE 146.693, MaxAPE 13.382, MaxAE 1331.077 Seasonal 1
Difference
Ljung-Box Q(18) Statistics 39.715 DF 15 Sig. 0.001.
MA, Seasonal(1) 0.816 0.044 18.665 0.000

The formulation of the ARIMA(0,1,1)(1,1,1)12 model corresponds

to the 0 ( B)1 ( B12 )(1 B12 )1 (1 B)1Yt 1 ( B)1 ( B12 ) t
equation which can be expanded as in Appendix 1 and represented
from to the 1 ( B12 )(1 B12 )(1 B)Yt 1 ( B)1 ( B12 ) t
Yt t 1 t 13 1 t 1 11 t 13 (Yt 1 Yt 12 Yt 13 ) 1 (Yt 13 Yt 12 Yt 24 Yt 25 )
In this work, it is used the natural logarithmic transformation for
As it can be seen at Table 2 which deals with the final estimate of
parameters for ARIMA (0, 1, 1)(1,1,1)12, it is observed that the p-
values of MA (1), SAR (1) and SMA (1) parameters are less than
0.05 and are therefore significant.
ACF for Residuals of Ln(Consumption) for ARIMA(0,1,1)(1,1,1)12
(with 5% significance limits for the autocorrelations)

the electric consumption data before application of the SARIMA, 1.0

so the above formulation is based on this data. 0.8

Yt 1 t 13 1 t 1 11 t 13 (Yt 1 Yt 12 Yt 13 ) 1 (Yt 13 Yt 12 Yt 24 Yt 25 )
0.6
0.4

Autocorrelation
0.2
And 0.0

t Y Yt -0.2
-0.4

Zt eYt
-0.6
The real estimates (Zs) will be
-0.8
-1.0
The estimated ARIMA Model Parameters and their corresponding 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
values are 0.540 for the MA(1) ,1, 0.270 for the SAR(1), 1, and Lag

0.816 for the SMA(1), 1, while addition of 1 set of nonseasonal

and seasonal differencing is applied. Fig. 12. ACF of Residuals

4.1.3 Diagnostic checking PACF for Residuals of Ln(Consumption) for ARIMA(0,1,1)(1,1,1)12

1.0
From plotting and looking the residual errors, they look random, 0.8
normally distributed, no outliers and their mean is approximately 0.6
Partial Autocorrelation

zero. Although there are several slight amounts of autocorrelations 0.4

at some lags, the overall appearance of the plots is good (see Fig- 0.2
0.0
ure 12, Figure 13 and Figure 14 ).
-0.2
-0.4
The Ljung-Box statistic is a test of the relationship between the -0.6
residuals and a large value shows that the residuals are related. -0.8

When the observations are completely random (i.e. form a white -1.0

noise process), it is usually expected that there would be no signif- 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

Lag
icant autocorrelations among the observed values of the series and
there will be no need to fit a model. The Ljung-Box Q Statistics
null hypothesis states that the series is white noise, meaning uncor- Fig. 13. PACF of Residuals
related. For our model, the test statistics rejects the no-
autocorrelation hypothesis at a high level of significance (see Residual Plots of Ln(Consumption) for ARIMA(0,1,1)(1,1,1)12
Normal Probability Plot Versus Fits
Table 1, p-value=0.001). This means that the residuals are not 99.9
0.10
99
white noise, and the proposed model is fully adequate for this 90 0.05
Residual
Percent

series with significant autocorrelations. 50 0.00

10
-0.05

1 -0.10
0.1
-0.10 -0.05 0.00 0.05 0.10 6 7 8 9 10
Table 1. Model Statistics Residual Fitted Value

Histogram Versus Order

Model Fit statistics Ljung-Box Q 60 0.10

Stationary R- 0.05
RMSE 45
Frequency

MAPE MAE Statistics DF Sig.

Residual

R-squared squared 0.00

30
0.411 0.997 241.743 2.719 146.693 39.715 15 0.001 -0.05
15
-0.10
0
-0.12 -0.09 -0.06 -0.03 0.00 0.03 0.06 0.09 1 50 100 150 200 250 300 350 400 450
Residual Observation Order

Fig. 14. Plots of Residuals

123
Aye Hande Erol, et al.

PACF of Residuals
4.1.4 Forecasting
1.0
0.8
After the diagnostic checking of the model, the forecasting is made
0.6
for the next 12 months by using the constructed model and pre-

Partial Autocorrelation
0.4
sented at the Results and Discussions part. 0.2

4.2 Holt-Winter's Method 0.0

-0.2
-0.4
The optimum value of parameters ( , and ) of Holt-Winters -0.6
method can be found by using the maximum likelihood estimation -0.8
and minimization of the sum of the squared errors. In this study, -1.0
the parameters of both the additive and multiplicative Holt-
1 5 10 15 20 25 30 35 40 45 50 55 60 65
Winters Method are estimated by using Stata/SE 11.0 software Lag
package.
Fig. 17. PACF Plots of Residuals

4.2.1 Additive Method

4.2.2 Multiplicative Method

The estimated optimal weights of model parameters are 0.5804 for

alpha, 0.0020 for beta and 0.2909 for gamma with accuracy meas- The estimated optimal weights of model parameters are 0.5039 for
ures of MAE 166.3 and RMSE 283.23378. The residuals plots of alpha, 0. 0063 for beta and 0.2343 for gamma with accuracy meas-
the data seem to satisfy the randomness (see Figure 15, Figure 16 ures of the MAE 162.26 and RMSE 273.4088. The residuals plots
and Figure 17). of the data seem to satisfy the randomness (see Figure 18, Figure
19 and Figure 20).

Residual Plots for Natural Log(Consumption) Residual Plots for Natural Log (Consumption)
Normal Probability Plot Versus Fits Normal Probability Plot Versus Fits
99.9 99.9
0.1 0.10
99 99

90 90 0.05
Residual

Residual
Percent

Percent

0.0 0.00
50 50

-0.05
10 10
-0.1 -0.10
1 1
0.1 0.1
-0.1 0.0 0.1 6 7 8 9 10 -0.10 -0.05 0.00 0.05 0.10 6 7 8 9 10
Residual Fitted Value Residual Fitted Value

Histogram Versus Order Histogram Versus Order

60 0.1
60 0.10

45 0.05
Frequency

45
Frequency
Residual

0.0 Residual
0.00
30
30
-0.05
15 15
-0.1 -0.10
0 0
-0.12 -0.08 -0.04 0.00 0.04 0.08 1 50 100 150 200 250 300 350 400 450 -0.12 -0.08 -0.04 0.00 0.04 0.08 1 50 100 150 200 250 300 350 400 450
Residual Observation Order Residual Observation Order

Fig. 15. Plots of Residuals

Fig. 18. Plots of Residuals

ACF of Residuals ACF of Residuals

1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
Autocorrelation

Autocorrelation

0.2 0.2
0.0 0.0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1.0 -1.0

1 5 10 15 20 25 30 35 40 45 50 55 60 65 1 5 10 15 20 25 30 35 40 45 50 55 60 65
Lag Lag

Fig. 16. ACF Plots of Residuals

Fig. 19. ACF Plots of Residuals

124
 Forecasting Monthly Electricity Consumption in Turkey

PACF of Residuals

1.0
0.8
0.6
Partial Autocorrelation

0.4
0.2
Fig. 21. ANN Model

0.0
-0.2 Backpropogation with momentum learning method is selected while creat-
-0.4
ing the model. The parameters used in this study are: momentum rate 0.8,
learning rate 0.5 and error rate 0.0001. The software run stops when the
-0.6
error rate decreased to the value that was set for training. According to the
-0.8 defined parameters, Neuroph software was executed and the training
-1.0 process was performed. The training error rates were demonstrated on the
Table 4 below. In the next step, the model was verified with the test data
1 5 10 15 20 25 30 35 40 45 50 55 60 65
and the learning level was measured as in Table 4.
Lag

Table 3. Training and Test Measures

MAE RMSE
Fig. 20. PACF Plots of Residuals Training 272.44 389.59
Test 1136.36 1329.27

4.3 ANN Model During the training phase, the trend was usually easy part for ANN
model to capture but, the major challenge was to overlap the fitted
values with the actual values to meet the seasonality pattern needs.
In this work, an electricity consumption estimation model is devel-
oped by using artificial neural network. In order to estimate the When the ANN model is trained normally, the fitted values
consumption, multilayer perception (MLP) structure is used. The showed a very similar behavior to simple linear regression and a
MLP training process is based on supervised learning algorithm. seasonal pattern cannot be seen. However, when the model is
To generate ANN network, Java based, open source Neuroph trained up to a significant level in order to minimize the errors, the
software is used (http://neuroph.sourceforge.net/index.html, Janu- ANN model produced fitted values that match the seasonal pat-
ary 2012). Neuroph is simple and functional software with Java terns somehow. But this overtraining did not end up with good
libraries for developing ANN models. Training and test data are predictions as it almost removed the trend from the ANN weights.
separately given to the software with the other necessary selected This can be easily seen at Table 3 by comparing the training meas-
model parameters. With training data set, the ANN network learns ures with the test measures. There is great deal of difference be-
the structure of the data, and then with test data set, the quality of
tween training and test measures. Even though the training seems
the learning of the network is tested.
significant, the ANN failed to produce good estimates for testing
and it can be seen at the Figure 22.
The consumption data from 1970 to 2009 are considered as train-
ing data and 2010 consumption data as test data. Years and months
of data are selected as input while the consumption is considered
as output. As it is selected the sigmoid function for activation
function, normalization of the data fields between (0,1) is neces-
sary. The normalization of the input output data is maintained by
formulation. For normalization of consumption data, DataMin
value is set as 350 and DataMax value is set as 20.000. The years
are normalized between 1970 and 2025 while the months are nor-
malized dividing each month to 12. The model is constructed with;
an input layer with two nodes, two hidden layers with six nodes
each and an output layer with one node. Additionally a bias node is
included in each layer except the output layer. The model structure
created with Neuroph software can be seen in the Figure 21 below.
Fig. 22. ANN Test Data Comparison

125
Aye Hande Erol, et al.

5 RESULTS AND DISCUSSIONS Mucuk, M., Uysal, D., (2009). Turkeys Energy Demand, Current Research Journal of
Social Sciences 1(3), 123-128
Mirasgedis, S., Sarafidis, Y., Georgopoulou, E., Lalas, D.P., Moschovits, M., Kar a-
The real consumption values of 2011 and the forecasted values giannis, F., Papakonstantinou, D., (2006). Models for mid -term electricity demand
forecasting incorporating weather influences, Energy 31, 208227.
produced by each method are given in Table 4. Their mean abso-
Albayrak, A.S., (2010). ARIMA forecasting of primary energy production and con-
lute mean error values and root mean squared error values are also sumption in Turkey: 19232006, Enerji, piyasa ve dzenleme, vol.1, no.1, 24-50.
given. In forecasting 2011 monthly consumption values Winters Box, G.E.P. and Jenkins, G.M., (1976). Time series analysis: forecasting and control,
additive method performs the best; both MAE and RMSE values Holden Day, San Francisco.
are lower than other methods. SARIMA is the second best per- Ediger, V.., Akar, S.B., (2007). ARIMA Forecasting of Primary Energy Demand by
former. ANNs performance is quite poorer than others. This is Fuel in Turkey, Energy Policy, 35, 1701-1708.
also obvious in figure 23 the ANN forecasts are constantly lower Erdodu, E. (2007). Electricity Demand Analysis Using Cointegration and ARIMA
than the real values. Modeling: A Case Study of Turkey, Energy Policy, Volume 35, 1129-1146.
Hanke, J.E. (2005). Business Forecasting, Pearson Education International, Eight
Edition.
We recommend to use Winters additive method to forecast Prajakta, S., Kalekar, B. (2004). Time series Forecasting using Holt-Winters Exponen-
tial Smoothing, Technology Issue: 04329008 (online)
monthly electricity consumption of Turkey. Our results have
http://www.it.iitb.ac.in/~praj/acads/seminar/04329008_ExponentialSmoothing.pdf
shown that this method is really successful for predicting the real (Access date: January 02, 2012).
consumption values. The forecasted values can be used confidently Russell Cheng's Home Page (online)
for making short term energy planning such as purchasing and http://www.personal.soton.ac.uk/rchc/Teaching/MATH6011/ Forecasting-
selling electricity decisions. Chap3.htm (Access date: January 02, 2012).
Pankratz, A., (1983). Forecasting with Univariate Box-Jenkins Models: Consepts and
Cases, Wiley, New York.
Table 4. The Comparison of 2011 Forecasts Permanasari, A.E., Dayang, R.A.R. and Dominic, P.D.D., (2009). Prediction of
Zoonosis Incidence in Human using Seasonal Auto Regressive Integrated Moving
Winters Winters Average (SARIMA), International Journal of Computer Science and Information
Real ANN SARIMA
Multip. Add. Security, Vol. 5, No. 1.
16926.19 14704.33 16785.68 16774.95 16483.81 Fausett, L.. (1994). Fundamentals of Neural Networks: Architectures, Algorithms and
15392.17 14299.54 15224.35 15315.34 15052.93 Applications, Prentice Hall.
16596.39 13993.00 16057.08 16174.70 15995.09 Zhang, G., Patuwo, B. E., Hu, M. Y., (1998). Forecasting with artificial neural net-
15279.88 13776.85 14851.74 15109.59 15034.08 works: the state of the art, International Journal of Forecasting, 14, 35-62.
15213.85 13684.49 15066.03 15427.42 15429.74
Ghiassi, M., Saidane, H. , Zimbra, D. K. (2005) A dynamic artificial neural network
model for forecasting time series events, International Journal of Forecasting 21,
15636.83 13928.15 15383.88 15873.9 15870.10
341362.
18414.47 15545.35 17226.36 17832.08 17940.77
Haykin, S., (1994). Neural Networks, Maxwell-McMillan, Eight Edition, Ontario-
17971.84 16478.72 17535.16 18233.59 18565.25 Canada, 1-45.
16577.64 14364.38 15495.92 16206.44 16026.07 ztemel, E., (2003). Yapay Sinir Alar, Papatya Yaynclk, Istanbul.
16378.25 13945.84 15583.17 16404.40 15910.24 Engle, R.F., Granger, C.W.J., Hallman, J.J., (1989). Merging short-and long-run
16586.5 13936.01 16257.19 17192.66 15848.78 forecasts: an application of seasonal cointegration to monthly electricity sales fo-
18355.08 13998.89 18081.95 19160.00 17746.06 recasting. Journal of Econometrics 40-1, 4562.
MAE 2222.80 481.72 326.94 459.28 nler, A., (2008). Improvement of energy demand forecasts using swarm intelligence:
The case of Turkey with projections to 2025. Energ. Policy 36, 1937-1944.
RMSE 2375.28 591.29 397.80 487.23
Ediger, V.. and Tatldil, H., (2002). Forecasting the primary energy demand in
Turkey and analysis of cyclic patterns. Energ. Convers. Management, vol. 43,
473-487.
Toksar, M.D., (2009). Estimating the net electricity energy generation and demand
using the ant colony optimization approach: Case of Turkey. Energy Policy 37,
1181-1187.
Mucuk, M. and Uysal, D., (2009). Turkeys Energy Demand, Current Research
Journal of Social Sciences 1(3), 123-128
Kouris, G., (1981). Elasticities - Science or fiction?. Energy Economics, 3:2, 66-126.
Drollas, L.P., (1984). The demand for gasoline: Further evidence. Energy Economics,
6:1, 71-82.
Stewart, J., (1991). Econometrics, Philip Allan: Hemel Hempstead.
Pindyck, R.S., (1979). The Structure of World Energy Demand, Mass.: M.I.T. Press:
Cambridge.
Hamzaebi, . (2007). Forecasting of Turkeys Net Electricity Energy Consumption
on Sectoral Bases, Energy Policy, 35, 2009-2016.

APPENDIX 1
Fig. 23. The Comparison Plot of 2011 Predictions with 2011 Real Values The general SARIMA model is;

REFERENCES p ( B) P ( B s )(1 B s ) D (1 B) d Yt q ( B)Q ( B s ) t

Abdel-Aal , R.E. and Al-Garni, A.Z., (1997), Forecasting monthly electric energy
consumption in eastern Saudi Arabia using univariate time-series analysis, Energy The formulation of the ARIMA(0,1,1)(1,1,1)12 model expressed as
Volume 22, Issue 11, 1059-1069 =>
0 ( B)1 (B12 )(1 B12 )1 (1 B )1Yt 1 (B )1 (B 12 ) t

126
 Forecasting Monthly Electricity Consumption in Turkey

=> 1 ( B12 )(1 B12 )(1 B )Yt 1 (B )1 (B12 ) t

=>
(1 1B12 )(1 B12 )(1 B )Yt (1 1B )(1 1B12 ) t
=>
(1 1B12 )(1 B12 )(Yt Yt 1 ) t (1 1 B )(1 1 B12 )

=>
(1 1B12 )(Yt Yt 1 Yt 12 Yt 13 ) ( t 1 t 1 )(1 1B12 )

=>
Yt Yt 1 Yt 12 Yt 13 1Yt 12 1Yt 13 1Yt 24 1Yt 25 t 1 t 13 1 t 1 1 1 t 13

=>
Yt (Yt 1 Yt 12 Yt 13 ) 1(Yt 13 Yt 12 Yt 24 Yt 25) t 1 t 13 1 t 1 1 1 t 13

=>
Yt t 1 t 13 1 t 1 11 t 13 (Yt 1 Yt 12 Yt 13) 1(Yt 13 Yt 12 Yt 24 Yt 25)

127