Electrical Power and Energy Systems 77 (2016) 136144

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Daily peak electricity demand forecasting based on an adaptive hybrid

two-stage methodology
Abderrezak Laouafi a,, Mourad Mordjaoui a, Farida Laouafi a, Taqiy Eddine Boukelia b
a
Department of Electrical Engineering, University of 20 August 1955-Skikda, Skikda, Algeria
b
Department of Mechanical Engineering, University of Constantine 1, Constantine, Algeria

a r t i c l e i n f o a b s t r a c t

Article history: This paper describes daily peak load forecasting using an adaptive hybrid two-stage methodology.
Received 14 July 2015 Because the time series of electricity consumption is mainly influenced by seasonal effects, the double
Received in revised form 3 October 2015 seasonal HoltWinters exponential smoothing method is firstly used for next-day peak electricity
Accepted 17 November 2015
demand forecasting. In the second stage, the secondary forecasting model is applied taking into account
the benefits of Fuzzy c-means clustering; K-nearest neighbors algorithm; Wavelet packet decomposition;
and Adaptive Neuro-Fuzzy Inference System, for further improvement in forecasting accuracy. The whole
Keywords:
architecture of the proposed model will be presented and the results will be compared with neural net-
Daily peak load forecasting
Exponential smoothing
works and stand-alone adaptive neuro-fuzzy inference system based approaches by using a gathered
Data classification data from the Algerian power system. The results show that: (1) the proposed methodology is the best
Wavelet packet among all the considered schemes, (2) the FKW-ANFIS has satisfactory performance in both normal
Adaptive neuro-fuzzy inference system and special daily conditions.
2015 Elsevier Ltd. All rights reserved.

Introduction [1], short-term electricity demand forecasting has been extensively

Electric load forecasting is a very important tool for ensuring
economic and reliable operation in power systems. To achieve this
end, electric utilities use load forecasting models, to ensure that
the energy supplied meets the load of their customers plus the
energy lost in the system. For this specialized function, a trained
staff is needed to increase or decrease the produced power so as
to adjust the supplydemand balance at any time in the best con-
ditions of cost and safety. Accurate load forecasting models are
needed for a variety of time horizons, these could be listed as: very
short-term for loadfrequency control and economic dispatch
functions; short-term for the day-to day operation in power sys-
tem, medium-term for generators maintenance scheduling, and
long-term to build new lines and sub-stations or to upgrade the
existing systems. Very short-term (VSTLF), short-term (STLF),
medium-term (MTLF) and long-term load forecasts (LTLF) are ran-
ged respectively from few minutes to an hour, an hour to one
week, one week to one year, and one year to decades. Because of
its great importance for performing many operations in power sys-
tems, including generator unit commitment; hydro-thermal coor-
dination; load flow study; fuel allocation; and power interchange
Corresponding author.
E-mail address: laouafi_abderrezak@yahoo.fr (A. Laouafi).
studied in the literature [210]. Instead, a lower number of works
can be found in the literature about VSTLF [1113], MTLF [1417]
and LTLF [1719].
STLF forecasting covers the daily minimum and maximum elec-
tricity demand, total daily energy, and daily load curve as a series
of 24 hourly forecasted loads. Particularly, daily peak load forecast-
ing (DPLF) is very important task for decision making processes in
the national electricity supply system of a country, as the conse-
quences of over- or underestimation may increase the operating
cost. Overestimation of future electric load may cause the startup
of too many generating units and leads to a redundant reserve of
electric power. In contrast, underestimation of load causes failure
in providing enough electricity, which may lead into supply inter-
ruption and blackouts during periods of peak demand. If the entire
national electricity supply system were to shut down, it would take
days, possibly even weeks to restore [20].
In the last few decades, many techniques have been developed
to improve the accuracy of peak load forecasting. Reported models
in the literature of DPLF could be principally divided into two
groups: classical techniques and artificial intelligence based
approaches. Conventional techniques include time series models
[20,21], multivariate regression [22], and robust regression models
[23]. Artificial intelligence based approaches include neural
networks [1,24,25], fuzzy systems [26], and neuro-fuzzy systems

http://dx.doi.org/10.1016/j.ijepes.2015.11.046
0142-0615/ 2015 Elsevier Ltd. All rights reserved.
 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144
[27]. Although they are accurate in normal days, univariate con-
ventional techniques such as the seasonal autoregressive inte-
grated moving average (SARIMA) model cannot give satisfactory
results when subject to special daily conditions, because they
ignore important weather effects and cannot be updated for facing
the change in load demand during holidays. Even though multi-
variate classical methods such as multiple regression models
express the peak load as a function of exogenous inputs, the linear
modeling process faces difficulty to determine empirically the cor-
rect complex relationship that exists between the peak load and
the other explanatory inputs. In recent years, the advent of new
intelligent techniques such as neural networks and adaptive
neuro-fuzzy inference system (ANFIS) enabled to model automati-
cally complex nonlinear inputoutput relationships through learn-
ing process using a database of load and independent variables.
Current trends in research in the field of load forecasting are
focused on the use of hybrid and combined methods [2831].
The main objective of this paper is to present a novel model for
performing the DPLF. The proposed methodology can be viewed as
a two-stage forecasting model. Since the time series of peak elec-
tricity demand is mainly influenced by seasonal effects (daily and
weekly cycles), the double seasonal HoltWintersTaylor tech-
nique (HWT) is firstly used to perform a primary forecasted load.
In the second part, the primary forecasted peak load is verified
and improved by a secondary forecasting model; called FKW-
ANFIS, which uses the benefits of Fuzzy c-means (FCM) clustering,
k-Nearest neighbors (k-NN) algorithm, Wavelet packet decomposi-
tion (WPD), and Adaptive neuro-fuzzy inference system. Our idea
for this structure is to take the output of HWT model as an entry
to the secondary forecasting model. FCM and k-NN are used for
classifying data into somewhat similar patterns, which lead to
noise reduction and to an improvement on the quality of the
amount of data should be taken by the Wavelet packet-Adaptive
neuro-fuzzy inference system models; and therefore, to a higher
accuracy. To illustrate the suitability of the proposed methodology,
the approach is implemented to a peak load data from the Algerian
power system.
The rest of the paper is organized as follows. Section The Alge-
rian peak electricity demand time series presents the Algerian
peak electricity demand time series. Section Proposed daily peak
load forecasting methodology describes the proposed estimation
method. Section Results and discussion provides and explains
forecasting results. Finally, Section Conclusion concludes the
paper.
The Algerian peak electricity demand time series
In this study, we consider the power peak load data that consist
of thirty-five months of observations for electricity demand in
Algeria from 01 January 2012 to 30 November 2014. These data
were collected from OSE (Operateur du Systeme Electrique) web-
site [32]. The top subplot of Fig. 1 shows that the used data consist
of two time series. The first series records the morning peak elec-
tricity demand, while the second series consists of the evening
peak load observations. The figure illustrates that the daily peak
load usually occurs during the evening period. Therefore, we lim-
ited our interest in the present work to the evening peaks predic-
tion. The second subplot of Fig. 1 presents box-plots for the
evening electricity demand data. One can perceive an intraweek
seasonal cycle, where weekdays show similar patterns of demand,
but Fridays are somewhat different and the load on which is rela-
tively low compared to in workdays.
By considering both two series cited above, the data will con-
tain then daily seasonal cycle of duration two periods (morning
and evening peak), and weekly cycle of duration fourteen periods
137
(workdays and weekends). Hence, it seems natural to try to include
both seasonal cycles in a HWT forecasting model. This is the
approach taken by J.W. Taylor as part of their modeling of half-
hourly England and Wales electricity consumption for a lead time
up to a one day ahead [33]. This method compete well in normal
days, but it may faces some difficulty when subject to special daily
conditions such the case of the rapid change in load demand during
holidays. The whole architecture of the proposed methodology for
forecasting the peak load in both normal and special daily condi-
tions will be presented in the next section.
To evaluate the performance of the proposed strategy, the elec-
tricity peak load data for four different months: January 2014;
April 2014; July 2014; and November 2014, and for the public hol-
idays of 2014 were used as an illustrative example.
Proposed daily peak load forecasting methodology
The structure of the proposed DPLF method consists of two
principal stages. Through the use of morning and evening peak
electricity demand records, HWT method is applied as a first part
at the day J L
1 for ensuring a primary forecasted peak load for
the day J L . At its turn, the second part of our model consists of five
basic parts. First, the temperature observations are used as inputs
to the Fuzzy c-means clustering method for classifying the peak
load data into two classes (load of a hot or cold period). Then,
the morning peak loads; the evening peak electricity demand
information; and the primary forecast of HWT, are used as inputs
for k-NN to distinguish whether the day J L is workday or weekend
and so that to select closest samples that have the similar charac-
teristic of the day J L . The third part of the secondary forecasting
method decomposes the clustered morning and evening peak load
data; via wavelet packet method, into approximation parts associ-
ated with low frequencies and detailed parts associated with high
frequencies. At the fourth part, the ANFIS is introduced to predict
the future patterns of each wavelet packet space. Finally, the eve-
ning peak load forecasting values of all the spaces are added up to
produce the final electricity demand forecasting result.
HoltWintersTaylor method
HoltWintersTaylor technique is an extension to the standard
HoltWinters exponential smoothing, which allows accommodat-
ing the intraday and intraweek seasonal cycles in the electricity
demand series. This is the approach taken by J.W. Taylor as part
of their modeling of half-hourly England and Wales electricity con-
sumption for a lead time up to a one day ahead [33]. However, the
present paper is the first in term of using HWT method for fore-
casting daily peak electricity demand. The formulation for HWT
method is given in the following expressions:
St ayt =Dt
s1 W t
s2 1
aSt
1 T t
1 1
T t cSt
St
1 1
cT t
1 2
Dt dyt =St W t
s2 1
dDt
s1 3
W t xyt =St Dt
s1 1
xW t
s2 4
b k
y t k St kT t Dt
s1k W t
s2k / yt
St
1 T t
1 Dt
s1 W t
s2
5
St and T t , are the smoothed level and trend; Dt and W t are respec-
tively the seasonal indices for the intraday s1 (s1 2) and intra-
week s2 (s2 14) seasonal cycles; a; c; d and x are the smoothing
parameters; yt is the actual value of the time series in period t;
and by t k is the k step-ahead forecast made from forecast origin t.
The term involving the parameter / in the forecast function expres-
sion Eq. (5) is a simple adjustment for first-order autocorrelation.
138 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144
Fig. 1. Peak electricity demand from 01 January 2012 to 30 November 2014.
In this paper, HWT method parameters are estimated by
minimizing the one-day ahead evening peak electricity demand
forecasting error in three principal stages. First we generated
2  103 vectors of parameters from a uniform random number gen-
erator between 0 and 1. For each of the vectors, we then evaluated
the absolute percentage error (APE) on the daily peak demand fore-
casting of 24 June 2013. The 200 vectors producing the lowest APE
are evaluated; in their turns, over the period from 24 June 2013 to
30 June 2013. Among the 200 vectors, the 10 vectors that have the
lowest mean absolute percentage error (MAPE) are evaluated at
last on the one-day ahead evening peak load forecasting over the
period from 01 June 2013 to 30 June 2013. The one producing
the lowest MAPE value was chosen as the final parameter
vector. The optimized values of HWT method parameters were:
a 0:189, c = 0.049, d 0:304; x 0:112, and / 0:660.
Fuzzy c-means clustering
Weather conditions have a major impact on electricity con-
sumption, especially for domestic and agricultural consumers.
The most important weather factor in short-term load forecasting
is the temperature. The changes considerably affect the load
requirement for heating purposes in winter and air conditioning
purposes in summer.
Fig. 2 shows morning (or evening) peak electricity demand in
Algeria; plotted against maximum (or minimum) temperature
from a weather station located in Algiers [34]. The figure shows
that the relationship existing between temperature and load con-
sumption is very complex and highly nonlinear. However, we can
perceive that during summer, an increase in temperature will
result in an increase in load. While in winter the decrease in tem-
perature will results in an increase of electric load. In fact, we are
not looking in the proposed methodology for modeling the rela-
tionship that exists between temperature and load, but we are just
trying to distinguish whether the climate for a given day is hot or
cold. To do that, the Fuzzy c-means clustering method is applied
for allowing the temperature data to be classified into two clusters;
related to hot or cold periods of the year.
Fuzzy c-means is a data clustering technique which allows one
point to belong to one or more clusters. The method was developed
by Dunn in 1973 [35] and improved by Bezdek in 1981 [36] and it
is frequently used in pattern recognition. The FCM algorithm
attempts to partition a finite collection of n elements
X x1 ; . . . ; xn into a collection of c fuzzy clusters with respect to
some given criterion. Given a finite set of temperature data, the
algorithm returns a list of two cluster centers C fc1 ; c2 g and a
partition matrix, W wi;j 2 0; 1; i 1; . . . ; n; j 1; 2 where each
element wi;j indicates the degree to which element xi belongs to
cluster cj . The FCM aims to minimize an objective function:
X
n X
c  2
Jm  
ij xi
c j
wm 6
i j
where, m is any real number greater than 1. The fuzzifier m deter-
mines the level of cluster fuzziness. wi;j is the degree of membership
of xi belonging to cluster j:
1
wi;j   2 7
Pc kxi
cj k m
1
k1 kxi
ck k
Once the temperature data classification is performed, it is easy
at that time to classify the peak load data by considering the appro-
priate loads; that have the same indexes of temperatures in the
original database. For a purpose of clarification, the two different
colors that appeared in Fig. 2 are related to the two clusters per-
formed with FCM clustering.
k-Nearest neighbors algorithm
k-Nearest neighbors algorithm is a very simple and powerful
method for classification and regression tasks based on the similar-
ity of the samples of a set. The technique was first described in the
1950s and later formalized in 1960s when increased computing
power became available.
There are three key elements to this approach: a set of stored
records; a distance metric to compute the distances between
records; and the value of k, the number of nearest neighbors to
retrieve. For an unlabeled object x, the algorithm computes the dis-
tance D (e.g. Chebychev distance) between x and the entire training
 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144 139

Fig. 2. Fuzzy c-means clustering: results for 30 November 2014.

Fig. 3. k-Nearest neighbors clustering: results for 30 November 2014.

objects y to determine its k-nearest-neighbor list on the training
set of n elements:
Di maxi jx
yi j; i 1; . . . ; n
Once the distance between the query-instance and all the train-
ing samples is calculated, the distances are then sorted to deter-
mine nearest-neighbor list based on the k-th minimum distance.
In the k-NN method proposed in this paper, each individual is
defined by the morning and the evening peak load values corre-
sponding to a particular day J i . Thus, the k-NN classifier tries to find
pairs which are similar to the actual morning peak and the primary
forecasted evening peak load of the day J L . Consequently, the algo-
rithm allows the identification of the load patterns, and by the way
to generally distinguish whether the day J L is a weekday or a week-
end. An example of the k-Nearest neighbors algorithm results is
presented in Fig. 3. The point to be classified, shown with X; is
related to the peaks electricity demand information for 30 Novem-
ber 2014. This figure shows that data classifiers not only allow to
reduce the dimension of the problem, but also to improve the
quality of the amount of data should be taken by the Wavelet
8 packet-Adaptive neuro-fuzzy inference system models.
Wavelet packet decomposition
The wavelet transform (WT) is powerful and flexible tool that
has been widely used in electric load forecasting for decomposing
power load series into constitutive series that can be predicted
more accurately than the original signal [3739]. Wavelet trans-
form is able to reveal breakdown points, discontinuities in higher
derivatives and self-similarity aspects of data [37]. WT can be
divided in two categories: continuous wavelet transform (CWT),
and discrete wavelet transform (DWT).
In continuous wavelet transform, each wavelet is created by
scaling and translating operations in a mother wavelet. The mother
140 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144

Fig. 4. Three level wavelet packet tree.

Fig. 5. WPD results for forecasting the peak load of 30 November 2014.

wavelet is an oscillate function with finite energy and zero average. and positions usually based on powers of two [40]. In multiresolu-
The CWT of a continuous time signal f t is defined as: tion analysis, the signal is passed through high-pass and low-pass
Z 1 filters. The high-pass filter gives the detailed resolution of the
CWT a; b f twa;b tdt 9 signal, while low-pass filter gives the approximate information of

1 the signal. An approximation (A) holds the general trend of a given
where w is the mother wavelet, a 2 R is scaling parameter, b 2 R is signal, whereas a detail (D) depicts high-frequency components of
translating parameter, and: it.
  By setting a ac0 ; b db0 ac0 ; with a and b 2 Z, the DWT of the
1 t
b sampled signal f t of length T is given by:
wa;b t p w 10
a a X
T
1

However, the CWT is difficult to put into practice and the infor-
DWT c; d f twc;d t 11
t0
mation that has been picked up may overlap and result in redun-
dancy. It is therefore evident that a discretization of the where:
transform should be considered if it is desired to obtain a non-  
1 t
db0 ac0
redundant transformation. Therefore, instead of doing that, the wc;d t pc w 12
a0 ac0
mother wavelet can be scaled and translated using certain scales
 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144 141

Table 1
The key elements, parameters, and inputs for the different models.

Method Primary forecast Secondary forecast

Parameters P-BPNN P-NARX P-ANFIS S-BPNN S-NARX S-ANFIS FKW-BPNN FKW-NARX
Number of inputs/outputs 3/1 3/1 3/1 3/1 3/1 3/1 1/1 1/1
Minimum temperature
Maximum temperature
Morning peak load
Evening peak load
Number of hidden layers 1 1 1 1 1 1
Number of nodes in hidden layer Not fixed Not fixed Not fixed Not fixed Not fixed Not fixed
Training epochs 200 200 200 200 200 200 200 200
Number of membership functions 2 2
Wavelet packet decomposition
Fuzzy c-means clustering
K-nearest neighbors algorithm

Wavelet packet analysis is an extension of wavelet analysis that
offers better frequency resolution and richer signal analysis. The
process of WPD by filters is similar to the DWT. The difference
between wavelet packet decomposition and wavelet decomposi-
tion is that only low-frequency components are divided into two
parts continuously in multiresolution analysis, while both the
detail and approximation coefficients are decomposed in WPD.
This attribute makes the WPD very suitable in our case studies
for peak electricity demand forecasting, since relevant information
may have located in higher frequency components of the load data.
For a better understanding of wavelet packet decomposition
process, a third-order WPD is shown in Fig. 4. In wavelet packet
decomposition, the original signal S is firstly decomposed into
detail (D1 ) and approximation (A1 ) coefficients. In the second level,
not only the approximation part is decomposed into new detail
(DA2 ) and approximation (AA2 ), but detail coefficient vector (D1 )
is also decomposed into two parts (DD2 ; AD2 ) using the same
approach as in approximation vector splitting. This process is
repeated in a similar manner at the third level of the wavelet
packet tree. However, there is not a unique manner to reconstruct
the original signal S from its wavelet coefficients. In this paper, the
following nodes from the three level wavelet packet tree are
selected: fAAA3 ; DAA3 ; ADA3 ; DDA3 ; AD2 ; DD2 g. The used mother
wavelet function is the Daubechies of order 2 (abbreviated as
Db2) wavelet function. An example for wavelet packet decomposi-
tion results is shown in Fig. 5, where we have presented the
obtained wavelet sub-series from the forecasting process of 30
November 2014.
Adaptive neuro-fuzzy inference system
The adaptive neuro-fuzzy inference system is one of the most
successful schemes which combine the benefits of artificial neural
network and fuzzy inference system (FIS) paradigms into a single
capsule [41]. An ANFIS works by applying neural learning rules
to identify and tune the parameters and structure of a Fuzzy Tak-
agiSugeno Inference System. The network is a multi-layer feed-
forward that consists of a collection of artificial neurons in which
each node (neuron) performs a particular function on incoming
signals. The first layer executes the fuzzification process. The

Fig. 6. General structure of the proposed DPLF methods.

142 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144

Fig. 7. Results of FKW-ANFIS for four different months: January, April, July, and November 2014.

Table 2
The relative error (MAPE) of all proposed methods.

Method Primary forecasts Secondary forecasts

MAPE (%) HWT P-BPNN P-NARX P-ANFIS S-BPNN S-NARX S-ANFIS FKW-BPNN FKW-NARX FKW-ANFIS
January 1.872 2.021 2.535 1.985 1.889 2.187 1.719 2.000 1.553 1.717
April 1.465 1.693 1.998 1.985 1.273 1.934 1.301 1.382 1.353 1.434
July 2.848 3.143 2.828 2.845 2.360 2.630 2.571 2.157 2.908 2.238
November 1.329 2.034 2.486 1.742 1.655 2.164 1.610 1.335 1.704 1.176
All 1.879 2.223 2.462 2.048 1.794 2.229 1.800 1.719 1.879 1.641

Generalized bell-shaped membership function, two fuzzy rules,

and two hundred epochs for identifying the fuzzy inference system
parameters and the optimal topology of the ANFIS.

Final peak load forecasting result

The final peak load forecasting result is produced by means of

adding up all the forecasted components from each wavelet space.

Results and discussion

In this section, we evaluate the forecasting accuracy for HWT

Fig. 8. MAPE results of HWT and FKW-ANFIS for each day of the week.
second layer performs the fuzzy AND of the antecedent part of
fuzzy rules. The third layer normalizes the membership functions.
The fourth layer executes the consequent part of the fuzzy rules,
and finally the fifth layer calculates the output of fuzzy system
by summing up the outputs of the fourth layer [13].
At this part of the proposed forecasting methodology, the ANFIS
is introduced to predict the future patterns of each wavelet space.
However, since there are six nodes from the wavelet packet tree,
we achieve eventually six separate ANFIS models. For each model,
the input is associated, to the approximation or detail part of the
morning peak load signal. Whereas the output is associated to
the approximation or detail part of the evening peak load series.
The proposed adaptive neuro-fuzzy inference structure uses
and FKW-ANFIS models. We also compare their performance
against the primary forecasts of the well-known feed-forward
back-propagation neural network (P-BPNN); nonlinear autoregres-
sive neural network with external input (P-NARX); and adaptive
neuro-fuzzy inference system (P-ANFIS) methods, and against the
secondary forecasts of S-BPNN; S-NARX; S-ANFIS; FKW-BPNN;
and FKW-NARX models. We opted to implement the nonlinear
autoregressive neural network with external input in our study
because we were attracted by the structure of this network: the
output depends not only on previous values of inputs, but also
on previous values of the output signal. We also opted to imple-
ment the stand-alone adaptive neuro-fuzzy inference system and
the feed-forward back-propagation neural network, because in
our previous works, these models have showed excellent perfor-
mance for forecasting the quart-hourly electricity demand of the
metropolitan France [13,42]. The evening load data from previous
day (e.g. day J L
1 ) and from the same day type of the two weeks
above the day to be forecast (e.g. days J L
7 and J L
14 ) are adopted
as inputs for the primary computational intelligence models
considered in this paper. Minimum temperature, maximum
 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144 143

Table 3
Obtained results when HWT and FKW-ANFIS are used to forecast the peak load in 2014 public holidays.

Results Real peak (MW) HWT FKW-ANFIS

Day Forecasted peak (MW) Error (MW) APE (%) MAPE (%) Forecasted peak (MW) Error (MW) APE (%) MAPE (%)
01/01/2014 8237 8709 472 5.730 3.283 7813
424 5.147 2.796
14/01/2014 8294 8127
167 2.013 8156
138 1.664
19/03/2014 7795 7668
127 1.629 7741
054 0.693
01/05/2014 7400 7471 071 0.959 7475 075 1.013
29/07/2014 9093 9570 477 5.246 9316 223 2.452
30/07/2014 9258 9198
060 0.648 9470 212 2.290
05/10/2014 7292 7205
087 1.193 6910
382 5.239
06/10/2014 8031 7061
970 12.078 7505
526 6.550
03/11/2014 7602 7598
004 0.053 7611 009 0.118

temperature, and morning peak load of the day J L are used as
inputs for S-BPNN, S-NARX, and S-ANFIS models. The architecture
of FKW-BPNN and FKW-NARX is very similar to that of
FKW-ANFIS, where only the morning peak of the day on which
we would like to predict the load; is used as input to the neural
networks. The key elements and parameters of these approaches
are indicated on Table 1, and schematized in Fig. 6. The blanks in
Table 1 mean that the input or the parameter concerned is not
included for a specific forecasting model, while designates
the inclusion of the input/parameter by the forecasting model.
For the evaluation of the methods considered in this study, we
used APE, and MAPE criteria:
jb
y i
yi j
APE :100
yi
1X n
jby i
yi j
MAPE :100
n i1 yi
where b y i is the forecasted peak demand of the day i; yi is the real
load at the day i. n is the total number of forecasting cases.
Fig. 7 shows the results of FKW-ANFIS method for evening peak
demand prediction along four different months: January 2014,
April 2014, July 2014, and November 2014. This figure illustrates
that the proposed FKW-ANFIS model has successfully predict the
daily peak demand, since the shape of the predicted peak curves
follows almost the shape of the actual peak curves. Hence, detailed
forecasting results; as indicated on Table 2; clearly demonstrate
the effectiveness of the proposed forecasting strategy and show
that FKW-ANFIS outperformed all others methods. The results
from MAPE illustrate that HWT method is the more accurate
among the primary forecasting models, while the best technique
among the secondary forecasting models is the proposed FKW-
ANFIS. They also show that there is a notable improvement in fore-
cast accuracy when using ANFIS secondary forecasting model
rather than the double seasonal HWT technique, and that there is
a further noticeable improvement achieved by using the FKW-
ANFIS model. From these results, we can notice that average MAPE
of FKW-ANFIS has 08.838%, and 12.678%, higher accuracy over
S-ANFIS, and HWT, respectively. The most important improvement
was in summer, with a MAPE value lower respectively by 12.958%,
and 21.418%. This is a very important point, since we know that the
electricity consumption reaches its highest levels in summer as a
result of huge utilization of air conditioning appliances. In the rest
of this section, we are interested in further comparison between
HWT and FKW-ANFIS models.
We present in Fig. 8 the MAPE results of the four months dis-
cussed above; consolidated according to the type of day. This figure
shows that HWT gives the minimum MAPE values for Fridays and
Saturdays, and FKW-ANFIS gives the minimum MAPE values for
workdays. For Mondays, the MAPE value obtained by FKW-ANFIS
is somewhat around the half of the obtained value by HWT. We
present in Table 3; the forecasting results of HWT and FKW-
ANFIS for the public holidays of 2014. The Table shows that MAPE
value of HWT and FKW-ANFIS is respectively 3.283% and 2.796%,
meaning that FKW-ANFIS model has a 14.834% higher accuracy
over HWT. Therefore, the proposed FKW-ANFIS model in this paper
is very suitable tool for forecasting the daily peak electricity
demand in both normal and special daily conditions.
Conclusion
This paper has addressed the problem of daily peak load fore-
casting using an adaptive hybrid two-stage structure, which has
obtained better forecast results in comparison with double sea-
sonal HoltWinters, neural networks, and stand-alone adaptive
13
neuro-fuzzy inference system based approaches. The obtained
results and forecasting performance showed that: (1) the average
14 MAPE value of HWT primary forecaster and FKW-ANFIS secondary
forecaster in normal days is 1.879% and 1.641%, respectively, and
(2) the average MAPE value of HWT and FKW-ANFIS in public hol-
idays is 3.283% is 2.796%, respectively. Therefore, the proposed
two-stage methodology in this paper is an efficient tool for improv-
ing the peak load forecasting accuracy in both normal and special
days.
References
[1] Abdel-Aal RE. Modeling and forecasting electric daily peak loads using
abductive networks. Int J Electr Power Energy Syst 2006;28(2):13341.
[2] Hagan MT, Behr SM. The Time Series Approach to Short Term Load Forecasting.
IEEE Trans Power Syst 1987;2(3).
[3] Papalexopoulos AD, Hesterberg TC. A regression-based approach to short-term
system load forecasting. IEEE Trans Power Syst 1990;5(4).
[4] Dillon TS, Sestito S, Leung S. Short term load forecasting using an adaptive
neural network. Int J Electr Power Energy Syst 1991;13(4):18692.
[5] Yao SJ, Song YH, Zhang LZ, Cheng XY. Wavelet transform and neural networks
for short-term electrical load forecasting. Energy Convers Management
2000;41(18):197588.
[6] Mandal P, Senjyu T, Urasaki N, Funabashi T. A neural network based several-
hour-ahead electric load forecasting using similar days approach. Int J Electr
Power Energy Syst 2006;28(6):36773.
[7] Taylor JW. Short-Term Load Forecasting With Exponentially Weighted
Methods. IEEE Trans Power Syst 2012;27:45864.
[8] Li P, Li Y, Xiong Q, Chai Y, Zhang Y. Application of a hybrid quantized Elman
neural network in short-term load forecasting. Int J Electr Power Energy Syst
2014;55:74959.
[9] Barakati SM, Gharaveisi AA, Rafiei SMR. Short-term load forecasting using
mixed lazy learning method. Turkish J Electr Eng Comput Sci 2015;23(1):
20111.
[10] Zhai MY. A new method for short-term load forecasting based on fractal
interpretation and wavelet analysis. Int J Electr Power Energy Syst 2015;69:
2415.
[11] Charytoniuk W, Chen MS. Very Short-Term Load Forecasting Using Artificial
Neural Networks. IEEE Trans Power Syst 2000;15:2638.
[12] Laouafi A, Mordjaoui M, Dib D. Very short-term electricity demand forecasting
using adaptive exponential smoothing methods. In: 15th International
conference on sciences and techniques of automatic control & computer
engineering STA2014, Hammamet, Tunisia; 2123 December 2014. p. 553
7.
[13] Laouafi A, Mordjaoui M, Dib D. One-hour ahead electric load forecasting using
neuro-fuzzy system in a parallel approach. Computational intelligence
144 A. Laouafi et al. / Electrical Power and Energy Systems 77 (2016) 136144
applications in modeling and control. Studies in computational intelligence,
vol. 575; 2015. p. 95121.
[14] Gonzalez-Romera E, Jaramillo-Moran MA, Carmona-Fernandez D. Monthly
electric energy demand forecasting based on trend extraction. IEEE Trans
Power Syst 2006;21(4):194653.
[15] Jaramillo-Morn MA, Gonzalez-Romera E, Carmona-Fernndez D. Monthly
electric demand forecasting with neural filters. Int J Electr Power Energy
Syst 2013;49:253263, 201.
[16] Lee WJ, Hong J. A hybrid dynamic and fuzzy time series model for mid-term
power load forecasting. Int J Electr Power Energy Syst 2015;64:105762.
[17] Laouafi A, Mordjaoui M, Laouafi F. An evaluation of conventional and
computational intelligence methods for medium and long-term load
forecasting in Algeria. In: 3rd International conference on control,
engineering & information technology CEIT2015, Tlemcen, Algeria; 2527
May 2015.
[18] Zhang Z, Ye S. Long term load forecasting and recommendations for China
based on support vector regression. In: 2011 International conference on
information management, innovation management and industrial engineering
(ICIII); 2627 November 2011. p. 597602.
[19] Kandil MS, El-Debeiky SM, Hasanien NE. Long-term load forecasting for fast
developing utility using a knowledge-based expert system. IEEE Trans Power
Syst 2002;17(2):4916.
[20] Chikobvu D, Sigauke C. Regression-SARIMA modelling of daily peak electricity
demand in South Africa. J Energy Southern Africa 2012;23(3):2330.
[21] Amjady N. Short-term hourly load forecasting using time-series modeling with
peak load estimation capability. IEEE Trans Power Syst 2001;16(4):798805.
[22] Sigauke C, Chikobvu D. Daily peak electricity load forecasting in South Africa
using a multivariate non-parametric regression approach. ORiON: J ORSSA
2010;26(2):97111.
[23] Jin L, Lai YJ, Long TX. Peak load forecasting based on robust regression model.
In: 2004 International conference on probabilistic methods applied to power
systems; 2004. p. 1238.
[24] Amin-Naseri MR, Soroush AR. Combined use of unsupervised and supervised
learning for daily peak load forecasting. Energy Convers Management 2008;49
(6):13028.
[25] Iizaka T, Matsui T, Fukuyama Y. A novel daily peak load forecasting method
using analyzable structured neural network, 2002: Asia Pacific. In: IEEE/PES
transmission and distribution conference and exhibition 1; October 2002. p.
3949.
[26] Kiartzis SJ, Bakirtzis AG, Theocharis JB, Tsagas G. A fuzzy expert system for
peak load forecasting. Application to the Greek power system. In: MELECON
2000: 10th mediterranean electrotechnical conference 3; 2931 May 2000. p.
1097100.
[27] Son SY, Lee SH, Chung K, Lim JS. Feature selection for daily peak load
forecasting using a neuro-fuzzy system. Multimedia Tools Appl; March 2014.
[28] Saini LM, Soni MK. Artificial neural network-based peak load forecasting using
conjugate gradient methods. IEEE Trans Power Syst 2002;17(3):90712.
[29] Fan S, Mao C, Chen L. Peak load forecasting using the self-organizing map. In:
Advances in neural networks. Lecture notes in computer science, vol. 3498;
2005. p. 6407.
[30] Khan GM, Khan S, Ullah F. Short-term daily peak load forecasting using fast
learning neural network. In: 11th International conference on intelligent
systems design and applications (ISDA), Cordoba; 2224 November 2011.
[31] Moazzami M, Khodabakhshian A, Hooshmand R. A new hybrid day-ahead peak
load forecasting method for Irans National Grid. Appl Energy 2013;101:
489501.
[32] Oprateur du systme electrique, subsidiary of Sonelgaz Group. <http://www.
ose.dz/courbes.php>.
[33] Taylor JW. Short-term electricity demand forecasting using double seasonal
exponential smoothing. J Oper Res Soc 2003;54:799805.
[34] http://www.wunderground.com/history/airport/DAAG.
[35] Dunn JC. A fuzzy relative of the ISODATA process and its use in detecting
compact well-separated clusters. J Cybernet 1973;3:3257.
[36] Bezdek JC. Pattern recognition with fuzzy objective function algoritms. New
York: Plenum Press; 1981.
[37] Mordjaoui M, Boudjema B, Bouabaz M. A hybrid wavelet transform and ANFIS
model for short term electric load prediction. In: 2012 2nd International
conference on advances in computational tools for engineering applications
(ACTEA), Beirut, Liban; 1215 December 2012. p. 2925.
[38] Guan C, Luh PB, Michel LD, Wang Y, Friedland PB. Very short-term load
forecasting: wavelet neural networks with data pre-filtering. IEEE Trans Power
Syst 2013;28(1):3041.
[39] Sudheer G, Suseelatha A. Short term load forecasting using wavelet transform
combined with HoltWinters and weighted nearest neighbor models. Int J
Electr Power Energy Syst 2015;64:3406.
[40] Catalao JPS, Pousinho HMI, Mendes VMF. Hybrid intelligent approach for
short-term wind power forecasting in Portugal. IET Renew Power Gener
2011;5(3):2517.
[41] Jang JSR. ANFIS: adaptive network based fuzzy inference system. IEEE Trans
Syst Man Cybernet 1993;23(3):66585.
[42] Laouafi A, Mordjaoui M, Dib D. One-hour ahead electric load and wind-solar
power generation forecasting using artificial neural network. In: 2015 6th
International renewable energy congress (IREC), Sousse, Tunisia; 2426 March
2015. p. 16.