Content uploaded by Salih Tutun
Author content
All content in this area was uploaded by Salih Tutun on Jun 08, 2018
Content may be subject to copyright.
A new forecasting framework for volatile behavior in net electricity
consumption: A case study in Turkey
*
Salih Tutun
a
,
b
,
*
,
1
,
2
, Chun-An Chou
b
, Erdal Canıyılmaz
c
a
Turkish Military Academy, Institute of Defense Science, Ankara, Turkey
b
State University of New York at Binghamton University, Department of Systems Science and Industrial Engineering, Binghamton, NY, USA
c
Erciyes University, Department of Industrial Engineering, Kayseri, Turkey
article info
Article history:
Received 29 April 2015
Received in revised form
10 August 2015
Accepted 17 October 2015
Available online xxx
Keywords:
Forecasting
Energy management
Regularization
Adaptive optimization
Time series analysis
abstract
Electricity is a significant form of energy that cannot be stored physically and is usually generated as
needed. In most research studies, the main aim is to ensure that sufficient electricity is generated to meet
future needs. In order to avoid waste or shortage, a good system needs to be designed to constantly
maintain the level of electricity needed. It is necessary to estimate independent factors because future
electricity volume is based not only on current net consumption but also on independent factors. In this
paper, a new framework is proposed to first estimate future independent factors using SARIMA (seasonal
auto-regressive iterative moving average) method and NARANN (nonlinear autoregressive artificial
neural network) method, both of which are called a ”forecasted scenario approach”. Subsequently, based
on these scenarios, a LADES (LASSO-based adaptive evolutionary simulated annealing) model and a
RADES (ridge-based adaptive evolutionary simulated annealing) model are applied to forecast the future
NEC (net electricity consumption). The proposed approaches are then validated with a case study in
Turkey. The experimental results show that our approach outperforms others when compared to pre-
vious approaches. Finally, the results show that the NEC can be modeled, and it can be used to predict the
future NEC.
©2015 Elsevier Ltd. All rights reserved.
1. Introduction
Electricity is one of the main forms of energy affecting the
development of modern life, and it does so in technical, social, and
economic ways [1]. With regards to development, electricity de-
mand planning is a vital part of energy policies in developed and
developing countries, allowing them to make cost efficient in-
vestments in capacity planning [2]. To obtain optimal planning,
policymakers have focused on the modeling and forecasting of
projections that are able to obtain quality and problem-free con-
ditions [3,4].
Due to limited primary energy sources, energy policies in many
countries depend on foreign countries to supply energy. With
optimal planning, governments need to purchase energy sources
from foreign countries. Moreover, electricity is a difficult energy
source for investment, and it is hard to measure physical flow
because electricity consumption has a volatile structure. The gov-
ernment needs a model that forecasts electricity demand for un-
stable situations [5]. Due to the fact that electricity cannot be
stored, and its production is costly, the net consumption of elec-
tricity energy has to be estimated by an optimal production model.
In the meantime, forecasting errors could cause either shortages or
excess capacity that are undesirable for energyplanning. In the case
of excess capacity, the remaining part is wasted. On the other hand,
an inadequate supply of electricity causes either increased cost or
cuts in electricity [6].
Governments use planning organizations to forecast energy
demands and consumption. In Turkey, energy forecasting studies
have been officially conducted by the SPO (State Planning Organi-
zation) and the MENR (Ministry of Energy and Natural Resources)
by means of the MAED (Model of Analysis of Energy Demand) [7].
The MAED is a simulation-based approach that has been used to
assess medium-long term energy demand by using historical data
for the last two decades [6]. However, this approach is not adequate
*
Fully documented templates are available in the elsarticle package on CTAN.
*Corresponding author. State University of New York at Binghamton University,
Department of Systems Science and Industrial Engineering, Binghamton, NY, USA.
E-mail addresses: stutun1@binghamton.edu (S. Tutun), cachou@binghamton.edu
(C.-A. Chou), erdalc@erciyes.edu.tr (E. Canıyılmaz).
1
Supported for Ph.D. Education from Turkish Military Academy.
2
Ph.D. Candidate in State University of New York at Binghamton University.
Contents lists available at ScienceDirect
Energy
journal homepage: www.elsevier.com/locate/energy
http://dx.doi.org/10.1016/j.energy.2015.10.064
0360-5442/©2015 Elsevier Ltd. All rights reserved.
Energy 93 (2015) 2406e2422
to represent future planning because of insufficient data gathered
during two decades [8] and significant errors in the data to be
analyzed [9]. For this reason, other improved forecasting tech-
niques could be used instead of the MAED model.
Many studies such as Toksarı[10,11], Erdo
gdu [12] and Ünler
[13] use future independent factors (scenarios), in which it is
assumed that independent factors increase at a constant growth
rate as days pass, although this is not practically feasible in dynamic
energy systems. At the same time, these studies have not consid-
ered the electricity consumption of Turkey in the medium term. In
order to compare energy models with the MAED, net electricity
consumption is estimated monthly. Therefore, we will give results
which are close to actual values. In this paper, the aim is to offer
new models for forecasting the NEC and new scenario approaches
that use reliable energy models. In order to obtain actual values of
future demand for scenarios, the SARIMA (seasonal auto-regressive
iterative moving average) method, and the NARANN (nonlinear
auto-regressive based on artificial neural network) method are
used in the forecasted scenario approach. Thereafter, the new
LASSO (Least Absolute Shrinkage and Selection Operator) base
adaptive evolutionary simulated annealing (LADES) and RADES
(ridge base adaptive evolutionary simulated annealing) energy
models with linear and quadratic behavior are constituted to
forecast the NEC to show the efficacy of the proposed approach.
Meanwhile, researchers such as Toksarı[10,11], Ünler [13],
€
Oztürk et al. [14] and Ceylan and €
Oztürk [15] have used some meta-
heuristic approaches to optimize the parameters of energy models
in the literature. However, they have not considered over-training
in their algorithms. In the proposed framework, the LASSO and
ridge regression are used to prevent over-training by adding reg-
ularization. ES (evolutionary strategy) and SA (simulated anneal-
ing) as meta-heuristic approaches are first studied to optimize
coefficients of energy models. The hybrid meta-heuristic approach
is used to get optimal coefficients for the proposed models because
they use the complex LASSO and ridge regression-based
formulation.
The rest of the paper is organized as follows. In the literature
review, applications are presented for the forecasting of energy
demand and consumption. In Section 2, the methods used in the
new proposed approaches are explained briefly. Section 3de-
scribes in detail how the proposed methodology is used to fore-
cast the NEC. The net electricity consumption is estimated
monthly and annually for the years 2010e2020. The proposed
framework is discussed for the forecasting of the NEC based on
new scenario approaches with sensitivity analysis. Finally, Section
4shows the improvement in forecasting and the contribution of
the paper.
1.1. Literature reviews
Since the early 1970s, several studies on energy demand have
been performed using various estimation methods. Many studies
have aimed to evaluate the impact of economic activity and energy
planing on energy demand [12]. In recent years, because predictive
models are of vital importance for policymakers, they have used
these models to forecast and model energy consumption and de-
mand (see Table 1). In order to accurately forecast future energy
demand and consumption, several studies have presented models
that use artificial intelligence, econometric and hybrid approaches.
RA (regression analysis), ARIMA (auto-regressive iterative moving
avarage), SARIMA (seasonal auto-regressive iterative moving avar-
age), cyclic patterns and grey theory have been presented as the
econometric approaches. For example, Ediger and Tatlıdil [9] pro-
posed a technique involving the analysis of cyclic patterns of annual
additional amounts relevant to energy consumption. Tunç et al. [16]
estimated the electricity consumption demand with RA.
Kavaklıo
glu [17] combined multivariate regression with SVD (sin-
gular value decomposition) so as to downsize the problem to es-
timate the electricity consumption. Afterwards, Ediger and Akar
[18] made estimates on the electricity energy demand by using
the ARIMA and SARIMA methods. For electricity consumption,
Chujai et al. [4] found a model to forecast by using the ARIMA
method. Moreover, in recent years, researchers have focused on
grey theory. The grey forecasting model was used by Lee and Tong
[2] to make an electricity consumption estimate. GPRM (grey pre-
diction with the rolling mechanism) was utilized by Akay and Atak
[19] for an electricity demand estimate. Thereafter, optimized grey
modeling was proposed by Hamzaçebi and Es [20] to forecast
electricity consumption. Discrete grey forecasting and the Markov
approach based on the quadratic programming model were used by
Nai-ming et al. [5] to forecast energy production and consumption.
The results was showed that these methods were not adequate to
capture nonlinear behavior of energy demand and consumption.
Furthermore, artificial intelligence approaches have been pre-
sented to propose some models capturing nonlinear characteristics
for modeling of energy demand and consumption. An ANN
Nomenclature
w(k,i) weight between k and i nodes
Y
1
(k)¼a
1
(k) output of cell K
εerror for cell j
C(j,i) cost for weights
r
learning factor
New w
1
(j,i) new weight between j and i
Y(t) predicted output at time t
dnumber of delays
funknown smooth function
f
i
εði¼1;2;3;…pÞset of weight parameters for lags
q
j
εðj¼1;2;…qÞset of weight parameters for random errors
e
t
random error at time t
xindependent variable
M
L
penetration parameter for the LADES model
M
R
penetration parameter for the RADES model
net(S
i
(k) net input value
F
R
final result for the RADES model
F
L
final result for the LADES model
b
0
and
b
i
scalar and P-vector respectively, namely coefficients of
models.
l
non-negative regularization parameter
Nthe number of observations
t
1
trend of gross production
t
2
trend of electricity energy imports
t
3
trend of transmitted electricity energy
t
4
trend of electricity energy exports
x
1
value of gross production
x
2
value of electricity energy imports
x
3
value of transmitted energy
x
4
value of electricity energy exports
OI output of import for forecasted scenarios
OE output of export for forecasted scenarios
OT output of transmitted energy for forecasted scenarios
OG output of gross production for forecasted scenarios
S. Tutun et al. / Energy 93 (2015) 2406e2422 2407
(artificial neural network) algorithm was used to estimate elec-
tricity demand and consumption by Kermanshahi and Iwamiya
[21],S
€
ozen et al. [22] and Kavaklıo
glu et al. [23]. In the meantime,
some researchers focused on comparing the ANN with other
methods in order to decide the best model for forecasting the
electricity consumption. They used linear and nonlinear methods to
show behavior of energy demand and consumption. The ANN
method and the RA method were compared as linear and nonlinear
models by Pao [24]. The ANN method and the ARIMA method were
compared by Hamzaçebi and Kutay [25]. The SVM (support vector
machine) method and the ANN method were used to estimate
demand and consumption. After the SVR (support vector regres-
sion) method was used by Kavaklıo
glu [3] to make an NEC estimate,
O
gcu et al. [26] compared the ANN with SVM methods in order to
estimate electricity consumption. The RA, ANN and LSSVM (least
squares support vector machines) were compared by Kaytez et al.
[27] to forecast electricity consumption. Azadeh et al. [28e30] used
methods such as ANN, GA (genetic algorithm), ANFIS (adaptive
neural fuzzy inference system), MCS (Monte Carlo simulation), PSO
(particle swarm optimization), and AIS (artificial immune system)
to compare forecasting results of electricity consumption. The re-
searchers used the artificial intelligence approach to show they
predicted better than econometric approach. However, they could
be improved for forecasting accuracy of energy models.
Finally, the researchers focused on hybrid approaches in an
attempt to improve energy models. The dimension reduction
approach (e.g. PCA (principal component analysis) and DEA (data
envelopment analysis)) and ANN were combined by Kheirkhah
et al. [31] to estimate electricity consumption. Furthermore, meta-
heuristic approaches were used to optimize the parameters of some
models using an artificial intelligence approach. Two different
nonlinear models that have quadratic and exponential behaviors
were developed using the GA method and the RA method by €
Oztürk
et al. [14] to estimate the energy demand. An ACOEDE (ant colony
optimization energy demand estimate) model is developed by
Toksarı[10]. Toksarı[11] established two different models to esti-
mate net electricity production and electricity demand by using
ACO (ant colony optimization). Ünler [13] developed an electricity
energy demand estimate model by using a PSO technique. Conse-
quently he made the energy estimate by using three different
scenarios and the results obtained were compared with the
ACOEDE model results of Toksarı[10]. Gürbüz et al. [32] used an
ABC (artificial bee colony) to optimize regression models for fore-
casting of electricity consumption. At the sametime, some re-
searchers focused on other artificial methods for hybrid methods.
Optimized regression and ANN using IPSO (improved particle
swarm optimization) were used by Ardakani and Ardehali [1] to
forecast electricity consumption. The ANN and PSO methods were
used to forecast electricity consumption by Jiang et al. [33]. Shi-wei
and Ke-jun [34] used a hybrid algorithm with GA and PSO methods
to forecast energy demand. As a result, when using hybrid methods,
energy models can be improved by the researchers. However, as the
researchers used these methods, they assumed that independent
factors are increased with a constant growth rate for scenarios. New
scenario approaches can be proposed for forecasting demand and
consumption. At the same time, because researchers used meta-
heuristic approaches, over-training needs to be prevented in en-
ergy models. The researchers need to propose better models in
order to obtain accurate results for energy demand and forecasting.
For example, they need to consider extensive data, over-training,
accurate scenarios for the future and hybrid optimization
algorithms.
2. Methodology
In this paper, the LADES and RADES energy models with
linear and quadratic function are developed for projections. The
adaptive evolutionary strategy is used to optimize the initial
coefficients of these models, while the adaptive simulated
annealing algorithm makes a local search to find the proper
Table 1
Summary of energy modeling and/or forecasting studies.
Method used Type of method Author(s) Forecasted variable
Econometric approach RA Tunç et al. [16] Electricity consumption
ARIMA and SARIMA Ediger and Akar [18] Electricity demand
ARIMA Chujai et al. [4] Electricity consumption
Cyclic patterns Ediger and Tatlıdil [9] Energy consumption
RA and SVD Kavaklıo
glu [17] Electricity consumption
Grey theory approach Grey theory Lee and Tong [2] Electricity consumption
GPRM Akay and Atak [19] Electricity demand
Optimized Grey model Hamzaçebi and Es [20] Energy demand
Grey Theory, Markov approach Nai-ming et al. [5] Energy demand
Artificial inteligence approach ANN Kermanshahi and Iwamiya [21] Electricity demand
ANN Kavaklıo
glu [23] Electricity consumption
GA, ANN and Fuzzy Azadeh et al. [28] Electricity demand
ANFIS, GA and ANN Azadeh et al. [28] Electricity consumption
AIS, GA and PSO Azadeh et al. [30] Electricity consumption
ANN and SVM O
gcu et al. [26] Electricity consumption
SVR Kavaklıo
glu [3] Electricity consumption
ANN and RA Pao [24] Electricity consumption
ANN S€
ozen et al. [22] Energy consumption with
economic indicators
Hybrid approach GA €
Oztürk et al. [14] Energy demand
ACO Toksarı[10] Energy demand
PSO Ünler [13] Electricity demand
ACO Toksarı[11] Electricity demand
ANN, PCA and DEA Kheirkhah et al. [31] Electricity consumption
ANN and ARIMA Hamzaçebi and Kutay [25] Electricity consumption
ANN and PSO Jiang et al. [33] Electricity consumption
RA, ANN and LSSVM Kaytez et al. [27] Electricity consumption
RA and ANN Ardakani and Ardehali [1] Electricity consumption
ABC Gürbüz et al. [32] Energy consumption
GA, PSO Shi-wei and Ke-jun [34] Energy consumption
S. Tutun et al. / Energy 93 (2015) 2406e24222408
coefficients. Over-training is prevented by using the LASSO and
ridge regression methods in the models. In addition, we pro-
posed new approaches, namely forecasted scenarios, for fore-
casting the future values of independent factors such as imports,
exports, gross generation and transmitted energy. In the fore-
casted scenarios, the SARIMA and NARANN methods are imple-
mented respectively to obtain linear and nonlinear volatile
behavior for the future. Then, we define the best models for
forecasting by comparing the performance indicators. Finally, the
forecasting results are obtained by combining the proposed
approach and scenarios, as seen in Fig. 2.
2.1. A nonlinear auto-regressive model based on a neural network
ANNs (artificial neural networks) were originally developed to
mimic basic biological neural systems. The human brain has
interconnected simple processing elements (neurons or nodes) to
carry information. In ANNs, the information is carried by the
networks between input and output. In daily life, users draw
conclusions from the information obtained from samples, and
after that, they are able to make similar decisions in similar cases
and process incomplete information in uncertain cases. Each
neuron in ANNs takes an input signal from other neurons to
process an activation function for transforming output. ANNs can
make decisions by establishing relevant relationships between
events after gaining information with the help of data. After
training the network, it is possible to deal with incomplete in-
formation and give results even if there is incomplete informa-
tion on recently obtained examples. The information distributed
on the network has a distributed memory as numeric informa-
tion [35].
At the beginning in ANNs, weights are assigned randomly. For
the distribution of weights in the networks, each input value is
summed up by being multiplied by its own weight (summarization
function). In this way, the net input value that comes to the network
is calculated. Then the optimum weights can be reached as being
bound to the value.
Net ¼S
1
ðkÞ¼Xðw
1
ðk;iÞa
0
ðiÞ(1)
This shows that the net input value comes to the NET process
element. This value is acquired with the summarization function in
Eq. (1). Therefore, activation functions, which are the sigmoid
function in Eq. (2), detect the output that comes from net input
value.
Y
1
ðkÞ¼a
1
ðkÞ¼ 1
1þe
ð1s
1
ðkÞÞ
(2)
Therefore, the output of the node in Eq. (2) is the value that is
determined by the activation function.
ε
j
¼a
1
ðjÞy
r
ðjÞ(3)
The result is constituted on the output layer of the MLPANN
(multi-layer perceptron artificial neural network), which is
compared with the activation function. If there is a difference (error
signal in Eq. (3)) between the estimated value and actual value, the
weights in nodes are rearranged to reduce this difference. The
calculated outputs are compared with actual values and, if any, the
error is defined at the outcome. The error signal is used in changing
the weights in the output unit among the hidden layer elements.
The effects of each output unit on error are obtained by calculating
C(j,i) values in Eq. (4) to find optimal new weights [35].
C
1
ðj;iÞ¼X
p
r¼1
εðjÞa
1
ðjÞð1a
1
ðjÞÞ a
0
ðjÞ(4)
To achieve these procedures, the MLPANN can be used by first
assigning the value of zero to the C(j,i). At the same time,
r
(the
learning factor) is determined beforehand and new weight units
can be reached, as is seen in Eq. (5).
w
0
1
ðj;iÞ¼w
1
ðj;iÞrC
1
ðj;iÞfor j ¼1;…;ni¼1;…;n(5)
The procedure is repeated by deducting the error signals of
every system that has many hidden layers from the corrected
procedures of the previous layer. Finally, the procedure is continued
until the system finds the desired point by trial and error, which is
called the back propagation algorithm of the error [36]. The delta
rule is used in this system for the training process in Eq. (5). The
training algorithm dispersing the error back is an iterative gradient
algorithm that was developed for minimizing the square of the
errors between the outputs obtained from a forward distributed
network and acquired target outputs.
The NARANN (nonlinear auto-regressive based on neural
network) is mentioned as it displays the motivation for clear
research [37]. Many auto-regressive-approach based papers (e.g.
Valipour et al. [37], Jeong et al. [38], Ruiz-Aquilar et al. [39] and
Zhang et al. [40]) exist in the literature for non-stationary time
series, and the neural network based nonlinear auto-regressive
model is offered to improve the auto-regressive approach [41].In
Fig. 1. Defines the best model regarding the LASSO. Notes: Combinations are made with parameters such as sigma, number of movements, initial temperature, random walk, and
lambda, respectively.
S. Tutun et al. / Energy 93 (2015) 2406e2422 2409
this section, this method can be analyzed to forecast the next lag
value in time-series data. The neural network is predicted using a
time lag model in Eq. (6) [42].
YðtÞ¼fðyðt1Þ;yðt2Þ;…;yðtdÞÞÞ þεðtÞ(6)
The model assumes that error is IID (independent and identi-
cally distributed), and in order to find the optimal predictor with
minimum error of y(t), the model used is a conditional mean, as
seen in Eq. (7). We can find the output when past time lags as inputs
are given.
YðtÞ¼EðyðtÞjyðt1Þ;yðt2Þ;…;yðtdÞÞÞ
¼fðyðt1Þ;yðt2Þ;…;yðtdÞÞÞ As t dþ1 (7)
In our model, there are hidden layers with neurons, weights and
bias which are calculated using training algorithms such as Lev-
enbergeMarquardt, Bayesian Regularization, and Scaled Conjugate
Gradient.
2.2. Seasonal auto-regressive iterative moving average method
In the literature, the ARIMA (auto-regressive iterative moving
average) method has been used in many fields during the last three
decades [38]. It is a well-known method in the time-series
approach, which was proposed by Box and Jenkins [40]. This
method has three linear components: the AR (auto-regressive
term), the integration term (I), and the MA (moving average term).
After selecting a suitable factor, the model can forecast future
values by looking at the linear function of past observations and
random error. The Seasonal ARIMA model is also an extension of
the ARIMA model. If the time-series data contains seasonality, the
model is called SARIMA (p,d,q)(P,D,Q)S, in which pis the auto-
regressive order, qis the moving average order, dis the number
of differing operations, and P,D, and Qare seasonal orders of
p,dand q[39]. In this paper, the model
x
t
¼fa
t1
þfa
t1
þ…þfa
tp
þe
t
þqe
t1
qe
t2
…qe
tq
for scenarios is constituted using the SARIMA method with sea-
sonality [40]. Models can be defined by finding optimal orders for
this equation.
In order to use this model, there are four steps: stationary check,
identification, diagnosis, and forecasting, as seen in Fig. 3. In the
first step, the time-series data are checked as to whether or not the
mean, variance, and auto-correlation function are stationary. If the
data display non-stationary behavior, regularization is made by
using differentiation for time lags until stationary. Thereafter, the
model parameters are calculated by comparing the estimated and
actual values. The model is then statistically checked for signifi-
cance. Finally, the forecasting is made using the SARIMA model
[38].
2.3. The proposed methodology for the forecasting of net electricity
consumption
The new energy models are offered using adaptive evolutionary
strategy and adaptive simulated annealing with LASSO and ridge
regression. At the same time, the new scenario approaches are
presented using the SARIMA and NARANN methods.
2.3.1. Hybrid approach
In the model, SA (simulated annealing) is a random search
technique and a trajectory found by using single-based optimi-
zation. The base of the idea was first presented by Metropolis in
1953. Then Kirkpatric et al. [43] offered a simulation search model
by using the annealing approach to find an optimal solution. This
Fig. 2. The flow chart of the new framework for forecasting net electricity consumption.
S. Tutun et al. / Energy 93 (2015) 2406e24222410
algorithm mimics the annealing process in materials physics as
metals freeze and cool into a crystalline state with minimum en-
ergy level by using bigger crystal sizes to decrease defects. The
efficiency of the algorithm for optimization depends on the control
of temperature and cooling schedule. Moreover, in order to move
to new solutions, the algorithm uses random walk, which de-
scribes the movement of the algorithm by searching randomly
from the current solution to a neighborhood solution in order to
explore the optimal feasible solution [43]. In addition, the tem-
perature is reheated when the new solution is not suitable for
movement, and the method is made adaptive to prevent prema-
ture convergence.
In ES (evolutionary strategy), new solutions as children are
compared with old solutions as parents. The ES is used to find a
good initial solution for the simulated annealing method because
the ES is a population-based algorithm, which can search out more
solutions for the global optimum in large search areas. At the same
time, these methods can cause over-fitting because a meta-
heuristic approach is used. In order to eliminate this situation,
regularization is added to the objective function. LASSO and ridge
regression are used for regularization of coefficients to find the
optimal regression model by optimizing parameters with a hybrid
based on the ES and SA algorithms.
The proposed algorithm is described in the following steps:
Step 1: Initialization with the ES algorithm that sets the bounds
of parameters. The initial values of the parameters are then
generated for the models. If the ES has better offspring as par-
ents, the standard deviation for movement to a new solution is
decreased for the adaptive model.
Step 2: Temporal state for the SA algorithm that makes a
random move to change the current system state by using the
optimal initial parameters for the ES.
Step 3: Acceptance checking that looks at the following equa-
tions to understand whether there is acceptance or rejection of
the temporal state. If there is rejection, the temperature is
reheated as the adaptive model in the SA.
>The temporal state is accepted if the energy of the new
solution >the energy of the old solution and p, which is a
random number, <P, which is the accepted rate with the new
solution as 0 p1.
>The temporal state is accepted if the energy of the new
solution the energy of the old solution.
>The temporal state is rejected, otherwise.
Step 4: Finding a solution with regularization that finds the
optimal solution by comparing all solutions. The algorithm with
regularization (the LASSO and ridge regression) is also checked
for over-training by comparing testing and training errors.
Step 5: Feature selection with the LASSO: The algorithm uses the
LASSO method to improve the subset of features for analysis. If
the algorithm uses the same features, go to Step 6; otherwise, go
to Step 2 with new features.
Step 6: Deciding the best scenario for future independent fac-
tors by comparing forecasted scenarios.
Step 7: Use as expert system. The new energy models predict
future electricity consumption through the forecasted scenario
approach for decision-making.
The proposed model is explained with all steps in Fig. 2. The best
model is finally found, which guides future planning.
2.3.2. Coupling evolutionary strategy with simulated annealing
In the proposed models, there can be different behavioral
models (e.g. quadratic, cubic, exponential and so on) that increase
the decision variables exponentially. In order to solve this problem,
the ES is used to find initial solutions for decision variables (coef-
ficient of the models) by giving initial ranges. Thereafter, by using
the SA based on a single solution, the algorithm searches the
neighborhood of the initial solution because a random walk is used
for the next solution. It moves to new solutions for decision vari-
ables by using a normal random number. This means that the al-
gorithm can get stuck unless it has a good initial solution. For
instance, as is seen in Fig. 4, it begins to find solutions from S
0
to S
3
.
Fig. 3. Flow chart of the SARIMA model.
S. Tutun et al. / Energy 93 (2015) 2406e2422 2411
After arriving at S
3
, the algorithm tends to accept this point as the
optimal solution for decision variables, but it is a local optimum.
The algorithm needs to search in a global way to find the opti-
mum solution. Thus the ES algorithm can find good (close to the
optimal solution) initial solutions that can be used in the SA algo-
rithm. When started with these solutions, the SA algorithm can find
the optimal solution by looking in the neighborhood of initial
solutions.
2.3.3. Formulation for new energy models
In the literature, researchers (e.g. Toksarı[10], Ünler [13], Toksarı
[11] and Ceylan and €
Oztürk [15]) use linear and quadratic regres-
sion to obtain new models for electricity consumption. However,
when they use a meta-heuristic approach for training, they need to
consider overtraining. In the models, the LASSO and ridge regres-
sion are used in order to prevent over-training. They are regression
methods that involve penalizing the absolute and squaresize of the
regression coefficients. In the formulation, forecasting models are
first decided using linear and quadratic regression, as seen in Eq. (8)
and Eq. (11), respectively.
F
1
¼b
5
þb
1
x
1
þb
2
x
2
þb
3
x
3
þb
4
x
4
(8)
Fp
L
¼M
L
ðb
5
þb
1
x
1
þb
2
x
2
þb
3
x
3
þb
4
x
4
Þ(9)
Fp
R
¼M
R
ðb
5
þb
1
x
1
þb
2
x
2
þb
3
x
3
þb
4
x
4
Þ(10)
F
2
¼b
15
þb
1
x
1
þb
2
x
2
þb
3
x
3
þb
4
x
4
þb
5
x
1
x
2
þb
6
x
1
b
3
þb
7
x
1
x
4
þb
8
x
2
x
3
þb
9
x
2
b
4
þb
10
x
3
x
4
þb
11
x
2
1
þb
12
x
2
2
þb
13
x
2
3
þb
14
x
2
4
(11)
The LADES and RADES energy models with linear and quadratic
behaviors are used as objective functions in order to optimize the
coefficients (decision variables) in Eqs. (12)e(15). LASSO as linear
and quadratic objective functions are used for the LADES energy
model, as is seen in Eqs. (12) and (13).
For a given value of
l
>0,
min
b
0
;b
1
;b
2
;b
3
;b
4
0
@
1
2NX
N
i¼1
ðY
i
ðF
1
ÞÞ
2
þlX
p
j¼1
b
j
1
A(12)
min
b
0
;b
1
;b
2
;b
3
;b
4
;…;b
15
0
@
1
2NX
N
i¼1
ðY
i
ðF
2
ÞÞ
2
þlX
p
j¼1
b
j
1
A(13)
As
l
increases, the number of nonzero components of
b
de-
creases [44].
Ridge regression as linear and quadratic objective functions is
used for the RADES energy model, as seen in Eqs. (14) and (15).
min
b
0
;b
1
;b
2
;b
3
;b
4
0
@
1
2NX
N
i¼1
ðY
i
ðF
1
ÞÞ
2
þlX
p
j¼1
b
j
2
1
A(14)
min
b
0
;b
1
;b
2
;b
3
;b
4
;…;b
15
0
@
1
2NX
N
i¼1
ðY
i
ðF
2
ÞÞ
2
þlX
p
j¼1
b
j
2
1
A(15)
2.3.4. Evaluation criteria of forecast performance
Measuring the accuracy of the method is achieved by finding the
difference between the actual value and the estimated value in
keeping with the rareness of error values. Mean absolute percent-
age error (MAPE) in Eq. (16) is a measure ofaccuracy for building up
fitted time series values. It commonly asserts accuracy as a per-
centage. The results can be obtained more clearly with a
percentage.
MAPE ¼P
n
i¼1
Y
i
F
i
F
i
n100 (16)
RMSE (root mean square error) in Eq. (17) performs sample
standard deviation of differences between estimated (F) values and
actual values (Y). It is the square root of variance, which is called
standard deviation.
RMSE ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
n
i¼1
ðY
i
F
i
Þ
2
n
s(17)
MSE (mean square error) in Eq. (18) is a variant of estimator. This
approach is used as objective function in the proposed models. We
evaluated MSE with LASSO and ridge regularization.
MSE ¼P
n
i¼1
ðY
i
F
i
Þ
2
n(18)
MAE (mean absolute error) in Eq. (19) is a batch that measures
how close estimations are to possible results.
Fig. 4. Coupling the ES and the SA to explain how to prevent getting stuck in a local optimum.
S. Tutun et al. / Energy 93 (2015) 2406e24222412
MAE ¼P
n
i¼1
Y
i
F
i
n(19)
SSE (sum square error) in Eq. (20) measures any contradiction
between estimated and actual values.
SSE ¼X
n
i¼1
ðY
i
F
i
Þ
2
(20)
The results can be compared to find the best model structure by
using these performance indicators.
3. Forecasting of Turkey's net electricity consumption
Net electricity consumption in Turkey is forecasted by using pre-
processed data. The necessary data, including each variable and
covering a period of 35 years, are divided into two sets: 336 and 84
monthly observations data as training and testing, obtained from
the TEIAS (Turkish electricity transmission company). As inde-
pendent factors, the transmitted energy, gross generation, imports
and exports, which have high efficiency, are chosen according to
previous studies (e.g. Hamzaçebi [7], Toksarı[11],S
€
ozen [45] and
S€
ozen [46]) conducted on the NEC of Turkey. At the same time, our
analysis determined that the NEC is influenced by these indepen-
dent factors.
In order to make a better analysis of Turkey's situation, it is
necessary to review indicators such as imports, gross generation,
exports and transmitted energy. Imports and exports are financial
transactions of international trade. For energy, exports mean
shipping goods and services for energy out of the port of a country;
imports mean receiving goods and services for energy from a
foreign country. The energy imports and exports of Turkey are
strong indicators of manufacturing activity. Electricity energy,
which can be easily transmitted to homes, is related to consump-
tion. Optimal transmitted energy improves living standards
because it can prevent electricity cuts and increase the delivery of
electricity for consumption. At the same time, gross generation is
important as it allows the country to provide electrical energy on
time. We tried to understand how the relationship between easy
energy and price affects net electricity consumption by using these
independent factors.
In this paper, the projections for independent factors are
determined by forecasted scenarios using the SARIMA and
NARANN methods. The data from January 1990 to December
2005 formed the training set, and those from January 2006 to
December 2010 formed the testing set of independent factors. At
the same time, forecasting errors are calculated from 2001 to
Fig. 5. Forecasted values of the NARANN of gross production for testing data. (Actual and estimated values are almost the same as is seen for R
2
value).
S. Tutun et al. / Energy 93 (2015) 2406e2422 2413
2010 one-by-one to show how the scenarios work better than
previous approaches. The best estimate value is reached by
calculating the MAPE, RMSE and R
2
values. The results are
compared by looking at performance indicators to achieve
optimal future planning. This approach is better than other ap-
proaches in the literature for preparing future independent fac-
tors and projections.
3.1. Forecasted scenarios
The constant growing values used in the literature for future
independent factors do not reflect actual future values of inde-
pendent factors because they always assume an augmentation of
estimated factors such as linear behavior. For this reason, forecast-
based scenarios are offered to forecast future independent factors.
The SARIMA and NARANN methods are used to forecast the value of
each future independent factor.
3.1.1. The results of nonlinear auto-regressive model based on
neural network scenario
The nonlinear approach (NARANN) in Figs. 5e8is used in
forecasting to decide the best scenario structure. This algorithm is
shown in Eqs. (21)e(24) with high Rvalues for the forecasted
scenario of independent factors. The method is constituted by
deciding the number of hidden neurons, the number of delays, the
percentage of training validation, and testing data as shown in
Table 3. The results are evaluated by MSE (mean squared error) and
R
2
.
OI ¼0:88target þ6:8 (21)
OE ¼0:78target þ11 (22)
OG ¼0:99target þ52 (23)
OT ¼0:98target þ110 (24)
In order to compare the results of the scenarios, as seen in
Figs. 5e8, the Rvalues are calculated for the testing data of inde-
pendent factors by using the NARANN method. After using Eqs.
(21)e(24), forecasting values for the independent factors are found
for the future years. Target is the lag value for the NARANN method.
When using past values as targets in equations, scenarios are
forecasted for independent factors. However, we realize that these
equations work well for the short term (e.g. two or threeyears). For
long-term forecasting, we need to improve on this approach. For
this reason, the SARIMA method is proposed to compare the results.
Fig. 6. Forecasted values of the NARANN of import for testing data. Note: actual and estimated values are in keeping with high R
2
value.
S. Tutun et al. / Energy 93 (2015) 2406e24222414
3.1.2. The results of an auto-regressive iterative moving average
scenario
Predictive models are found for independent factors by using
the SARIMA method in Fig. 9. The future values of independent
factors are forecasted by using these models. As seen in Table 2, the
best parameters are decided by using the SARIMA method for in-
dependent factors when comparing the testing data for the years
2006 through 2010. The future independent factors are then esti-
mated for forecasting with the proposed energy models.
Therefore, the best model in the SARIMA method can use these
forecasted values instead of growth rate because when looking at
Figs.11e12, we can see that the forecasted scenarios approach has a
low SSE for ten years. In order to show how forecasted scenarios
work, forecasted values are obtained by SARIMA as the best sce-
nario approach for the years 2006 through 2010, as seen in
Figs. 9e10. After showing the results for 2006 through 2010,
detailed results are demonstrated to prove scenarios will work for
the future years. Therefore, the best model found in the study is
used to forecast future net electricity consumption. The NEC of
Turkey from 2011 through 2020 is forecasted to constitute new
capacity plans, as seen in Fig. 16.
3.2. The results of the proposed energy models
In the proposed energy models, when increasing independent
factors, the decision variables are increased exponentially for the
quadratic model, and there are square and absolute values in
objective functions in Eqs. (12)e(15) of the proposed energy
models. This means that this is a non-polynomial hard problem in
which coefficients must be optimized with meta-heuristic ap-
proaches. After using the proposed methods to optimize the co-
efficients of the models in Eqs. (8)e(11), the models are found to
forecast future values in Eqs. (25)e(28).
The best structure for energy models is constituted through
training and testing sets for the proposed approach. As parameters
of the proposed algorithms for optimization, the Boltzmann con-
stant, initial temperature, number of movements, standard devia-
tion for random walk, and tuning parameter (
l
) are given
respectively in order to find the best model structure for the LASSO
method.
The best parameters of the hybrid approach are defined for the
LADES and RADES energy models in order to avoid over-training.
With an increased tuning parameter, the MAPE for training
increased while the MAPE for testing decreased. The best model
framework with the LASSO based formulation is decided for pa-
rameters as seen in Table 4. Because
l
is more than 150, the testing
error does not change sharply in Fig. 1.
Parameters in Table 5 are decided for the best model struc-
ture with ridge regression based formulation because as long as
l
is increased, the testing errors do not change sharply in Fig. 13.
This shows that overfitting is prevented for the best energy
model.
Fig. 7. Forecasted values of the NARANN of transmitted energy for testing data.
S. Tutun et al. / Energy 93 (2015) 2406e2422 2415
Fð1Þ
L
¼150:1763 þð0:5440Þx
1
þð0:0012Þx
2
þð0:2946Þx
3
þð0:0027Þx
4
þð0:1241Þð150:1763 þð0:5440Þx
1
þð0:0012Þx
2
þð0:2946Þx
3
þð0:0027Þx
4
Þ
(25)
Fð1Þ
R
¼150:119 þð0:4380Þx
1
þð0:0458Þx
2
þð0:4269Þx
3
þð0:1102Þx
4
þð0:0097Þð150:119 þð0:4380Þx
1
þð0:0458Þx
2
þð0:4269Þx
3
þð0:1102Þx
4
Þ
(26)
Fð2Þ
L
¼146:81 þð0:0756Þx
1
þð0:033Þx
2
þð0:0124Þx
3
þð0:1103Þx
4
þð0:0223Þx
1
x
2
þð0:0352Þx
1
x
3
þð0:0671Þx
1
x
4
þð0:0037Þx
2
x
3
þð0:0185Þx
2
x
4
þð0:0328Þx
3
x
4
þð0:0319Þx
2
1
þð0:024Þx
2
2
þð0:0236Þx
2
3
þð0:1272Þx
2
4
(27)
Fð2Þ
R
¼146:83 þð0:0962Þx
1
þð0:0682Þx
2
þð0:0050Þx
3
þð0:106Þx
4
þð0:0245Þx
1
x
2
þð0:0508Þx
1
w
3
þð0:1151Þx
1
x
4
þð0:01842Þx
2
x
3
þð0:0246Þx
2
w
4
þð0:0461Þx
3
x
4
þð0:0324Þx
2
1
þð0:04240Þx
2
2
þð0:0566Þx
2
3
þð0:1288Þx
2
4
(28)
Scenarios from the SARIMA method are used for future values of
independent factors from 2010 through 2020 to show accuracy of
the proposed framework. Forecasting, which is found according to
the best model, is done for future months and years by using the
estimated values of independent factors. Hence, future demand is
predicted monthly and annually by using these approaches.
The main aim of a modeling study is to produce a model which
can present the nature of the problem. Almost all relationships in
the real world are nonlinear, and the nature of the model should
capture non-linearity. We have added penetration, which is the
mean value of errors for past years, to the proposed models. This
penetration allows the linear models to catch nonlinear behavior
for the future. At the same time, a quadratic model is used to
capture the behavior of the net electricity consumption. For
Fig. 8. Forecasted values of the NARANN of export for testing data.
S. Tutun et al. / Energy 93 (2015) 2406e24222416
penetration, we calculated the mean value of difference, which is
the difference between the actual and the forecasted values for
each year. Thereafter, we calculatedthe percentage of difference for
the models. Finally, we added this percentage according to the
models for years 2011 through 2020.
As a result, better forecasting is possible for future years when
compared with the literature because in the new framework
monthly data are used to train the best model, and forecasted
scenarios are used to predict future independent factors instead of
the constant mean growth rate [11]. After finding the values of the
future independent factors, the proposed energy models are
constituted as mentioned in the methodology. Using the new en-
ergy models with scenarios, the projection is presented for future
planning, as seen in Fig. 16. By using the new models in Eqs.
(25)e(28), as seen in Figs. 14e15, the results show that the models
work well to forecast the net electricity consumption. The LADES
energy model (the best model) can be used in planning energy
needs for both the medium term and the long term.
3.3. Sensitivity analysis
The model in this study more accurately forecasts energy con-
sumption than do the models in other studies conducted to
determine the NEC. Better results are obtained than in the studies
carried out by Hamzaçebi et al. [25], MENR [6] and Erdo
gdu [12] for
estimated electricity consumption. When the conducted studies are
analyzed in Table 6, more realistic values are obtained for the years
2008e2009. A very small error value occurred between the
Table 2
Comparison of testing data for independent factors with the SARIMA with (p,d,q)(P,D,Q).
Imports Exports
Years (301) (101) Actual values Mape errors (101) (101) Actual values Mape errors
2006 805.89 573.20 0.29 1722.89 2235.70 0.23
2007 925.96 864.33 0.07 2312.32 2422.22 0.05
2008 998.05 789.40 0.21 1546.99 1122.20 0.38
2009 923.02 811.95 0.12 1376.76 1545.85 0.11
2010 1031.93 1143.83 0.11 1966.41 1917.59 0.03
Transmitted energy Gross generation
Years (101) (332) Actual values Mape errors (211) (131) Actual values Mape errors
2006 139242.90 143015.90 0.03 171777.90 176299.80 0.03
2007 2312.32 2422.22 0.05 194634.70 191558.10 0.02
2008 178289.30 172635.20 0.03 205248.50 198417.90 0.03
2009 177545.90 172187.70 0.03 188294.90 194812.90 0.03
2010 188852.70 184334.90 0.02 211304.50 211207.80 0.0005
Table 3
Parameters of NARANN models for independent factors. (Note: Dataset are divided into training, validation and testing subsets).
Parameters Gross production (x
1
) Imports (x
2
) Transmitted energy (x
3
) Export (x
4
)
Hidden Layers 10 5 10 10
Lags 7 3 4 4
Data-sets 75-15-15 75-15-15 75-15-15 75-15-15
Fig. 9. Forecasted values obtained by SARIMA for five years.
S. Tutun et al. / Energy 93 (2015) 2406e2422 2417
Fig. 10. MAPE errors between actual and forecasted values for five testing years.
Fig. 11. Forecasted values obtained by net electricity consumption for 10 years using the best model with SARIMA.
Fig. 12. The MAPE errors of net electricity consumption for 10 years.
Table 4
The best model structure for the LADES.
Training level Testing level LADES parameters
SSE 19,006 SSE 4,106,300 Sigma 0.95
MSE 56.57 MSE 48,885 Number of movements 1500
RMSE 0.41 MRSE 24.12 Initial Temperature 120
MAE 6.21 MAE 189.53 Random Walk 0.01
MAPE 0.18 MAPE 1.60 Lambda 150
Table 5
The best model structure for the RADES.
Training level Testing level RADES parameters
SSE 434,080 SSE 6,926,200 Sigma 0.95
MSE 1291.9 MSE 82,455 Number of movements 150
RMSE 1.96 MRSE 31.33 Initial Temperature 100
MAE 30.89 MAE 244.49 Random Walk 0.01
MAPE 1.03 MAPE 1.96 Lambda 1100
Fig. 13. Determine the best model regarding with ridge regression. Defines the best model regarding the ridge. Notes: Combinations are made with parameters such as sigma,
number of movements, initial temperature, random walk, and lambda, respectively.
Fig. 14. Scattering and distribution graphics of training level and testing level for the best LADES energy model.
S. Tutun et al. / Energy 93 (2015) 2406e2422 2419
estimated values and actual values because while the actual values
are 161.95 and 156.89 TWh for 2008e2009 in the data, the model
has predicted values of 159.5958 and 157.3347 TWh. Percentage
errors are estimated at 1.475% and 0.2826% MAPE error ratio, which
is lower than in previous studies, as shown in Table 6.
In order to demonstrate how the framework functions in energy
planning, we can check our forecasted results with actual values.
We know monthly data for independent and dependent variables
until December 2010. We also know annual values of net electricity
consumption for 2011 and 2012. The scenarios are forecasted by
Fig. 15. Scattering and distribution graphics of training and testing level, respectively, for the best RADES energy model.
Fig. 16. Monthly forecasting of the NEC with two scenarios between 2011 and 2020.
Table 6
Comparison with MAPE errors in the literature for forecasting of net electricity consumption. Note: the bold values show that NEC is forecasted better in the literature.
Years Actual value Forecasted values for the LADES (TWh) with mean absolute percent error (MAPE)
(TWh) This study MAPE MENR MAPE Hamzacebi MAPE Erdogdu MAPE Kavaklioglu MAPE
2008 161.95 159.643574 (1.475) 168.60 (4.1) 173.59 (7.2) 146.37 (9.6) 165.94 2.46
2009 156.89 157.168505 (0.283) 184.40 (13.86) 189.47 (16.99) 145.14 (10.4) 175.04 11.57
S. Tutun et al. / Energy 93 (2015) 2406e24222420
using the SARIMA based approach (the best scenario approach) for
these years. At the same time, the proposed models as LADES and
RADES are used to find future values. As seen in Table 7, actual net
electricity consumption is 186.100 TWh for 2011. When using sce-
narios and proposed methods, forecasted net electricity con-
sumption is found as 186.323 TWh in 2011. The NEC in 2011 and
2012 is forecasted with 0.96% and 0.09%, respectively, which is less
than one percent MAPE error, by using the LADES model. The
RADES model also forecasted future NEC with 0.62% and 5.46%
MAPE errors, respectively. This proves that our framework can
forecast the future net electricity consumption efficiently, as seen in
Fig. 16.
4. Conclusion
In modern life, forecasting is extremely important in the effec-
tive application of energy policies. Governments need to know how
much electricity must be generated to meet the energy demand
and consumption. In Turkey, the NEC (net electricity consumption)
for projections is officially obtained from the MAED simulation
technique in MENR with high forecasting errors. Forecasts need to
guide the MENR in developing the best energy policy.
The primary conclusion of this paper is that electricity con-
sumption of Turkey is modeled as the new LADES and RADES en-
ergy models with linear and quadratic behavior. New energy
models are used in such forms that future forecasting is possible.
We also present the significance of alternative forecasting methods.
Scenarios in the literature, which assume that independent factors
increase at a constant growth rate over time, are improved so as to
forecast the future values of independent factors by using the
SARIMA method and the NARANN method in the forecasted sce-
nario approach.
In the light of the results and discussion presented so far in this
study, the NEC is estimated to show how the framework works for
the future by using proposed scenarios and the best energy model.
The proposed best model forecasts Turkeys electricity consumption
with 1.59% MAPE error ratio on average for 34 years, while the
MENR forecasts more than 10% error ratio for some years. This
means that this framework can be used by the Turkish government
and related organizations to forecast future values in order to
ensure good future planning. These models can be used in different
countries as well. A new planning strategy can be developed with
this study by looking at the future values. Policymakers can use this
framework both to plan new investments and to determine
appropriate export and import amounts. Moreover, the new energy
models can be defined by using different evaluation criteria of er-
rors (e.g. SSE, MAE, MAPE and so on.) as objective function to
improve models. New energy models with hybrid techniques can
be developed to conduct better studies.
To conclude, in Turkey and other nations, inadequate forecasting
of energy demand has often led to power shortages and outages.
This hinders the development of the economy and leads to irrita-
tion and inconvenience for the average citizen. By forecasting
actual energy demand, the model proposed in this study would
help avoid these power outages, thus allowing Turkey to develop
more rapidly and to improves the quality of life for its citizen using
electrical power.
Acknowledgments
The authors wish to thank the Turkish Electricity Transmission
Company, and the Turkish Minister of Energy and Natural Re-
sources for their help in providing data.
References
[1] Ardakani FJ, Ardehali MM. Long-term electrical energy consumption fore-
casting for developing and developed economies based on different optimized
models and historical data types. Energy 2014;65:452e61.
[2] Lee YS, Tong LI. Forecasting energy consumption using a grey model improved
by incorporating genetic programming. Energy Convers Manag 2011;52(1):
147e52.
[3] Kavaklıo
glu K. Modeling and prediction of Turkey's electricity con-
sumption using support vector regression. Appl Energy 2011;88(1):
368e75.
[4] Chujai P, Kerdprasop N, Kerdprasop K. Time series analysis of household
electric consumption with ARIMA and ARMA models. In: Proc. IMECS Conf.,
Hong Kong; 2013.
[5] Nai-ming X, Chao-qing Y, Ying-jie Y. Forecasting Chinas energy demand and
self-sufficiency rate by grey forecasting model and Markov model. Int J Electr
Power Energy Syst 2015;66:1e8.
[6] World Energy Council-Turkish National Committee (WEC-TNC). Energy re-
ports, Ankara, Turkey, http://www.dektmk.org.tr/upresimler/Enerji_Raporu_
20106.pdf; [accessed 08.01.15] [in Turkish].
[7] Hamzaçebi C. Forecasting of Turkey's net electricity energy consumption on
sectoral bases. Energy Policy 2007;35(3):2009e16.
[8] CE. Altas¸M,€
Ozkan HF. Energy statistics, the Turkish ministry of energy and
natural resources of Turkey, http://www.dektmk.org.tr/; [accessed 08.01.15].
[9] Ediger VS¸ , Tatlıdil H. Forecasting the primary energy demand in Turkey and
analysis of cyclic patterns. Energy Convers Manag 2002;43(4):473e87.
[10] ToksarıMD. Ant colony optimization approach to estimate energy demand of
Turkey. Energy Policy 2007;35(8):3984e90.
[11] ToksarıMD. Estimating the net electricity energy generation and demand
using the ant colony optimization approach: case of Turkey. Energy Policy
2009;37(3):1181e7.
[12] Erdo
gdu E. Electricity demand analysis using cointegration and ARIMA
modelling: a case study of Turkey. Energy policy 2007;35(2):1129e46.
[13] Ünler A. Improvement of energy demand forecasts using swarm intelligence:
the case of Turkey with projections to 2025. Energy Policy 2008;36(6):
1937e44.
[14] €
Oztürk HK, Ceylan H, Canyurt OE, Hepbas
¸lıA. Electricity estimation using
genetic algorithm approach: a case study of Turkey. Energy 2005;30(7):
1003e12.
[15] Ceylan H, €
Oztürk HK. Estimating energy demand of Turkey based on economic
indicators using genetic algorithm approach. Energy Convers Manag
2004;45(15):2525e37.
[16] Tunç M, ÇamdalıÜ, Parmaksızo
glu C. Comparison of Turkey's electrical energy
consumption and production with some European countries and optimization
of future electrical power supply investments in turkey. Energy Policy
2006;34(1):50e9.
[17] Kavaklıo
glu K. Robust electricity consumption modeling of Turkey using
singular value decomposition. Int J Electr Power Energy Syst 2014;54:268e76.
[18] Ediger VS¸ , Akar S. ARIMA forecasting of primary energy demand by fuel in
Turkey. Energy Policy 2007;35(3):1701e8.
[19] Akay D, Atak M. Grey prediction with rolling mechanism for electricity de-
mand forecasting of Turkey. Energy 2007;32(9):1670e5.
[20] Hamzacebi C, Es HA. Forecasting the annual electricity consumption of Turkey
using an optimized grey model. Energy 2014;70:165e71.
[21] Kermanshahi B, Iwamiya H. Up to year 2020 load forecasting using neural
nets. Int J Electr Power Energy Syst 2002;24(9):789e97.
[22] S€
ozen A. Future projection of the energy dependency of Turkey using artificial
neural network. Energy Policy 2009;37(11):4827e33.
[23] Kavaklıo
glu K, Ceylan H, €
Oztürk HK, Canyurt OE. Modeling and prediction of
Turkey's electricity consumption using artificial neural networks. Energy
Convers Manag 2009;50(11):2719e27.
[24] Pao HT. Comparing linear and nonlinear forecasts for Taiwan's electricity
consumption. Energy 2006;31(12):2129e41.
[25] Hamzaçebi C, Kutay F. Forecasting of electricity energy consumption in Turkey
with artificial neural network until 2010 year. J Fac Eng Archit Gazi Univ
2004;19(3):227e33 [in Turkish].
Table 7
Sensitive analysis for two models (the LADES and RADES results).
Years Actual values for the NEC The LADES model for NEC The RADES model for the NEC MAPE for the LADES model MAPE for the RADES model
2011 186099.55 184323.47 184951.72 0.96 0.62
2012 194923.35 195094.78 206180.79 0.09 5.46
S. Tutun et al. / Energy 93 (2015) 2406e2422 2421
[26] O
gcu G, Demirel OF, Zaim S. Forecasting electricity consumption with neural
networks and support vector regression. Procedia Social Behav Sci 2012;58:
1576e85.
[27] Kaytez F, Taplamacioglu MC, Cam E, Hardalac F. Forecasting electricity con-
sumption: a comparison of regression analysis, neural networks and least
squares support vector machines. Int J Electr Power Energy Syst 2015;67:
431e8.
[28] Azadeh A, Saberi M, Ghaderi S, Gitiforouz A, Ebrahimipour V. Improved esti-
mation of electricity demand function by integration of fuzzy system and data
mining approach. Energy Convers Manag 2008;49(8):2165e77.
[29] Azadeh A, Saberi M, Gitiforouz A, Saberi Z. A hybrid simulation-adaptive
network based fuzzy inference system for improvement of electricity con-
sumption estimation. Expert Syst Appl 2009;36(8):11108e17.
[30] Azadeh A, Taghipour M, Asadzadeh S, Abdollahi M. Artificial immune simu-
lation for improved forecasting of electricity consumption with random var-
iations. Int J Electr Power Energy Syst 2014;55:205e24.
[31] Kheirkhah A, Azadeh A, Saberi M, Azaron A, Shakouri H. Improved estimation
of electricity demand function by using of artificial neural network, principal
component analysis and data envelopment analysis. Comput Ind Eng
2013;64(1):425e41.
[32] Gürbüz F, €
Oztürk C, Pardalos P. Prediction of electricity energy consumption
of Turkey via artificial bee colony: a case study. Energy Syst 2013;4(3):
289e300.
[33] Jiang X, Ling H, Yan J, Li B, Li Z. Forecasting electrical energy consumption of
equipment maintenance using neural network and particle swarm optimi-
zation. Math Probl Eng 2013;2013.
[34] Shi-wei Y, Ke-jun Z. A hybrid procedure for energy demand forecasting in
China. Energy 2012;37(1):396e404.
[35] €
Oztemel E. Yapay sinir a
gları. 2nd ed. Papatya Press; 2006.
[36] Donahoe JW, Palmer DC. The interpretation of complex human behavior:
some reactions to parallel distributed processing, edited by JL McClelland, De
Rumelhart, and the PDP research group 1. J Exp Anal Behav 1989;51(3):
399e416.
[37] Valipour M, Banihabib ME, Behbahani SMR. Comparison of the ARMA, ARIMA,
and the auto-regressive artificial neural network models in forecasting the
monthly inflow of Dez dam reservoir. J Hydrol 2013;476:433e41.
[38] Jeong K, Koo C, Hong T. An estimation model for determining the annual
energy cost budget in educational facilities using SARIMA (seasonal auto
regressive integrated moving average) and ANN (artificial neural network).
Energy 2014;71:71e9.
[39] Ruiz-Aguilar J, Turias I, Jim
enez-Come M. Hybrid approaches based on SAR-
IMA and artificial neural networks for inspection time series forecasting.
Transp Res Part E Logist Transp Rev 2014;67:1e13.
[40] Zhang X, Pang Y, Cui M, Stallones L, Xiang H. Forecasting mortality of road
traffic injuries in china using SARIMA model. Ann Epidemiol 2015;25(2):
101e6.
[41] Sheremetov L, Cosultchi A, Martínez-Mu~
noz J, Gonzalez-S
anchez A, Jim
enez-
Aquino M. Data-driven forecasting of naturally fractured reservoirs based on
nonlinear auto-regressive neural networks with exogenous input. J Petrol Sci
Eng 2014;123:106e19.
[42] Pucheta JA, Rivero CMR, Herrera MR, Salas CA, Patino HD, Kuchen BR. A feed-
forward neural networks-based nonlinear auto-regressive model for fore-
casting time series. Comput Sist 2011;14(4):423e35.
[43] Kirkpatrick S. Optimization by simulated annealing: quantitative studies.
J Stat Phys 1984;34(5e6):975e86.
[44] The Mathworks, The LASSO and ridge regression, http://www.mathworks.
com/help/stats/lasso.html; [accessed 07.28.15].
[45] S€
ozen A, Arcaklıo
glu E, €
Ozkaymak M. Turkey's net energy consumption. Appl
Energy 2005;81(2):209e21.
[46] S€
ozen A, Arcaklıo
glu E. Prediction of net energy consumption based on eco-
nomic indicators (GNP and GDP) in Turkey. Energy Policy 2007;35(10):
4981e92.
S. Tutun et al. / Energy 93 (2015) 2406e24222422
