Electric Power Systems Research 148 (2017) 3547

Contents lists available at ScienceDirect

Electric Power Systems Research

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

Online electricity demand forecasting based on an effective forecast

combination methodology
Abderrezak Laoua a, , Mourad Mordjaoui a , Salim Haddad b , Taqiy Eddine Boukelia c ,
Abderahmane Ganouche a
a
Department of Electrical Engineering, University of 20 August 1955-Skikda, Skikda, Algeria
b
Department of Mechanical Engineering, University of 20 August 1955-Skikda, Skikda, Algeria
c
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 presents a new forecast combination methodology for generating very short-term electricity
Received 8 May 2016 demand predictions under both normal and anomalous load conditions. The main contribution of the
Received in revised form 17 February 2017 work is to propose an online load forecasting system that has the ability to achieve good forecasting
Accepted 19 March 2017
accuracy, avoid large forecasting errors and ensure low computation time. The real-time load data from
the French power system and the Australian dataset for the state of New South Wales are used as an
Keywords:
illustrative example to evaluate the performance of the proposed methodology. The results reect that
Online load forecasting
the developed approach has better forecasting performance than other methods considered in this study.
Forecast combination
Hampel lter
For example, the results from the public holidays in France showed an average mean absolute percentage
Anomalous load error (MAPE) of 0.863%, and the accuracy improvements over a simple average combination method, the
best individual method, and a weighted combination are 15.887%, 13.353%, and 3.034%, respectively. For
the case of the Australian load dataset, our forecasting system achieved an average MAPE of 0.860% and
the improvement in comparison to a benchmark algorithm from the literature is equal to 8.316%.
2017 Elsevier B.V. All rights reserved.

1. Introduction of renewable energy sources intermittence on grid stability and

reliability.
Electricity demand forecasting is the primary prerequisite for Over the past few years, researchers have studied this prob-
achieving the goal of sustainable energy management and eco- lem of improving VSTLF accuracy and various classical and articial
nomic and secure operation of modern power systems. The task intelligence techniques have been considered using different sim-
of knowing the electricity demand in advance is needed to sustain ulated datasets. Classical methods include BoxJenkins time series
supply/demand balance and to manage the process of electricity models [13,14], exponential smoothing [15,16], and Kalman l-
production, distribution and consumption on a variety of temporal tering [17,18]. Articial intelligence techniques include neural
scales: very short-term from few minutes to an hour, short- network [1921], fuzzy logic [13], adaptive neuro-fuzzy inference
term from an hour to one week [15], medium-term from one system [22,23], and support vector regression [24]. Hybrid methods
week to one year [69], and long-term from one year to decades can be also used to combine the advantages of several algorithms.
[1012]. In particular, the accurate very short-term load fore- A typical way to create a hybrid model is to use the empirical mode
casts (VSTLF) are needed for the real-time scheduling of electricity decomposition as a tool for decomposing the electric load time
generation as well as load-frequency control and economic dis- series into a limited number of intrinsic mode function compo-
patch functions. With the deregulation of electricity markets and nents, and to forecast the decomposed sequences with appropriate
the growing penetration of renewable energy sources into the computational intelligence models. However, the application of
energy matrix of todays power networks, such predictions are empirical mode decomposition requires a signicant execution
also of importance to market participants to mitigate the effects time and the forecaster may face difculty to handle the mode-
mixing problem, since the components of new time series may
signicantly differ from those used while training the forecasting
models [25]. In the other hand, wavelet decomposition is an ef-
Corresponding author. cient technique to unfold the inner features of the electric load
E-mail address: [email protected] (A. Laoua). time series. In this manner, many recent papers discussed the

http://dx.doi.org/10.1016/j.epsr.2017.03.016
0378-7796/ 2017 Elsevier B.V. All rights reserved.
36 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547

use of wavelet transform to decompose the data into constitu- task compared to the case of constructing a normal load forecasting
tive series that can be predicted more accurately than the original model.
signal [26]. There have been few research studies for developing accurate
However, it is not an easy task to perform the accurate pre- load forecasting techniques for special days. Some of them give
dictions of electricity consumption, since the electricity demand more attention to the quality of the load data and the rigor of
behavior is complex and the load time series is non-stationary in time series analysis. Some other papers aim to provide the appro-
mean and variance, containing long run trend, multiple cycles of priate output from a combination of forecasts. While some other
seasonal uctuations as well as random noise [27,28]. In fact, there works, like the study performed by Ghofrani et al. [37], focus on
is not any forecasting method which gives the best result for all the consideration of both the appropriate combination of different
times: some approaches may perform very well under certain cir- forecasts and the capture of the most relevant information of the
cumstances, and they fail to reect the variation laws and power time series as two key elements to perform the task of improv-
load information in others [29]. Hence, the methodology of fore- ing the forecast accuracy of special daily conditions. One of the
cast combination, also called ensemble forecasting, is a suitable existing methods in the literature that focus on the quality of the
tool which has been successfully applied in many disciplines for data to predict the atypical load is the triple seasonal exponen-
improving forecasting accuracy [30,31]. tial smoothing method developed by Arora and Taylor in [38]. The
New studies related to electric load forecasting research area methodological contribution of the authors work is to show how
show that the strategy of combining different models is quite useful the exponential smoothing method can be adapted to model the
to improve the forecasting performance. The simplest way to com- load for special days, when used in conjunction with an approach
bine forecasts is to assign equal weights to the individual forecasts. based on rules. The results of the study showed that the advan-
This method often work reasonably well when we are uncertain tage of this rule-based method is located on its ability to model
about which method is best. So practically, one can make important normal and abnormal load in a unied framework. However, this
performance improvements in electric load forecasting by sim- model is too complicated, requires a signicant amount of load
ply averaging the forecasts. In the study given by Taylor [15], a data, and a forecaster intervention is almost essential to restruc-
variety of forecasting methods have been evaluated using minute- ture the model. In another study, a research work to improve the
by-minute British electricity demand data, and the results showed short-term load forecasting for anomalous load conditions has been
that the simple average of the forecasts from the weather-based presented [39]. The method of this study uses articial neural net-
approach and the double seasonal Holt-Winters-Taylor (HWT) work (ANN) models to provide the forecast scaled load curve, and
method resulted in forecasts that outperformed all other methods fuzzy inference models to forecast maximum and minimum loads
beyond about an hour ahead. Another popular method for combin- of the special day. The 24 hourly forecast loads of the special day
ing forecasts is the unequal weighting based on the past forecast are predicted by combining the results of the ANN and the fuzzy
performance of models. In the approach proposed in [32], the short- inference method. The benet of this hybrid structure was to uti-
term load is obtained by combining the forecasts generated by the lize the advantages of both, i.e., the generalization capability of
seasonal autoregressive integrated moving average (SARIMA) and ANN and the ability of fuzzy inference for handling and formal-
the neural networks models, and the weights of the combination izing the experience and knowledge of the forecasters. The test
are determined using a variancecovariance approach. The results results showed that the hybrid forecasting method could provide
of this study showed an important improvement in forecasting a considerable improvement of the forecasting accuracy for the
accuracy compared to other methods. A combined short-term load special days. In Ref. [40] Ghayekhloo et al. presented a hybrid
forecasting model based on ensemble empirical mode decomposi- load forecasting framework with a data preprocessing algorithm to
tion (EEMD) and sub-section particle swarm optimization (SS-PSO) enhance the forecasting accuracy. The data preprocessing method
is proposed in [33]. Firstly, the load sequence is decomposed into a is a combination of the input selection, wavelet decomposition and
limited number of intrinsic mode function (IMF) components and a novel standardization to provide the most appropriate inputs for
one remainder by EEMD. Then, the decomposed sequences are fore- the Bayesian neural network algorithms. Genetic algorithm is used
casted with appropriate forecasting models. SS-PSO is proposed to optimize the weighting coefcients of different forecast com-
and used to optimize the linear combination weights. The numer- ponents and minimize the forecast error. The performance and
ical results indicate that the proposed method in this reference accuracy of this short-term load forecasting method have been
paper gets a higher forecasting accuracy rate when compared with evaluated using New England load data. The results from this study
other forecasting techniques. A number of recent research studies showed a signicant improvement in the forecast accuracy when
have suggested also the use of forecasts combination as a promising compared to other existing techniques.
tool for improving the electricity demand forecasting performance However, we would point out that most of the forecasting
[3436]. methods in the literature have been implemented off-line using
However, the combined models cited above have the drawback simulated datasets. In fact, the good use of any load forecasting
of ignoring the effect of large forecasting errors. Actually, most of algorithm in practice not only requires the consideration of rea-
the existing methods in the literature of electric load forecasting sonable accuracy, but requires also the consideration of many other
focus only to evaluate the performance in normal and typical con- issues, such as the online adaptation ability, the simplicity of use
ditions and they often ignore the effect of the anomalous load that and the computational efciency. This is particularly important for
occurs on special days (such as the case of public holidays). For the case where the lead time is as short as few minutes ahead and
example, if a combined model constructed by a weighted combi- the adopted method is a forecast combination. To our knowledge,
nation is implemented to forecast the load of a particular public there is a lack of works that discuss the use of combined methods for
holiday, a poor estimation could be obtained at the case where a a practical application of very short-term load forecasting. This lack
large forecasting error is obtained by the individual method that has motivated us to provide the present work to the development
had a higher weight. This poor estimation can be attributed to the of an appropriate online load forecasting method that can ensure
dissimilar load behaviors of holidays compared with those of nor- low computation time and achieve good forecasting performance
mal days. In addition, for a given special day, the pattern of the on both normal and special daily conditions.
load can vary from year to year, depending on the day and sea- The main objective of this paper is to present a new forecast
son of the year when the abnormal load is occurred. So making an combination method, namely HFCM (Hampel lter-based Forecast
accurate abnormal load forecasting model is a more complicated Combination Method), for performing the online predictions of the
 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547 37

electricity demand. The proposed methodology consists of three annual cycle is characterized by an annual peak demand in win-
principal stages. Firstly, different models are adopted to forecast ter and a dip on summer. Fig. 2 presents the daily peak load in
the electricity consumption of the next-hour. The second part of France over four different weeks. This plot shows an intraweek
our methodology tries to detect any possible large forecasting error seasonal cycle including the weekdays with similar demand pat-
and replace it with a normalized one. This function is performed terns and the weekend, which is characterized by a decrease in the
by employing the Hampel lter in a data sequence formed by the consumption.
primary forecasts. However, the normal use of outlier detection Univariate classical methods such as BoxJenkins time series
methods is to detect spikes in load data and correct them [41]. models and exponential smoothing methods are frequently con-
Hence, we think that the present work is the rst in terms of using sidered to be sufcient for modeling and capturing the seasonal
a data pre-ltering technique as a tool for detecting over- or under- features of the data. However, it seems that the inclusion of the
estimation. At the third part of our methodology, the nal load intra-year cycle in a forecasting approach will not lead to any
estimation is obtained by considering the trimmed mean of the improvement in the accuracy of predictions as short as an hour
normalized forecasts. The basic idea in our work is to propose an ahead. In this regard, our choice fell on the consideration of the sea-
online load forecasting scheme that has the ability to ensure a low sonal autoregressive integrated moving average (SARIMA) and the
computation time, avoid large forecasting errors and achieve good double seasonal Holt-Winters-Taylor (HWT) exponential smooth-
forecasting accuracy on both normal and special daily conditions. ing as effective techniques for modeling intraday and intraweek
The real-time load data from the French power system are used as seasonal cycles in the electricity demand data.
an illustrative example to evaluate the online application and the The modeling process of HWT and SARIMA methods is based on
computation time of the proposed methodology, while the Aus- the sequential nature of the time series, where data are ordered
tralian data for the state of New South Wales are used for a purpose in time and future values are inuenced by past values. However,
of comparison with a benchmark algorithm from the literature. as a consequence of ignoring the special days during the modeling
The paper is organized as follows. Section 2 describes the char- process, the drawback of HWT and SARIMA approaches is found on
acteristics of the electric load data and presents our methodology its inability for given satisfactory results when subject to anoma-
for analyzing the electricity demand behavior. Section 3 presents lous load conditions such the case of the rapid change in load
the proposed forecasting method. Section 4 provides and explains demand during holidays. To incorporate this limitation, we pro-
forecasting results. Finally, Section 5 concludes the paper. pose a parallel processing approach for generating very short-term
load forecasts for both normal and special daily conditions. In this
approach, the electrical load time series is separated in autonomous
2. Electricity demand analysis and processing
points. A set of independent points means that the load at each
quarter-hour (or ve-minutes for the Australian data) of the day
To evaluate the performance of our approach, we used two elec-
is independent from the load at any other quarter-hour (or ve-
tricity load datasets from two different countries (Australia and
minutes). The main idea behind this approach is to partition the
France), with different time interval (5 min and 15 min). The France
electric load data into a dened number of parallel series and use
dataset is collected from the transmission system operator website
a separate model for each particular sub-series, where each sub-
[42]. It covers the period from April 07, 2013 to January 31, 2015,
series is related to a particular time of the day [44]. Our parallel
and is measured every 15 min. The Australian dataset is for the
approach considers two principal cases. For normal days, the paral-
state of New South Wales and is provided by the Australian Energy
lel load series are classied by calendar according to the type of day:
Market Operator [43]. The data is for two years, 2006 and 2007,
Saturdays, Sundays, and workdays. In previous works, this method
and is measured every 5 min. However, since we have opted to use
of arrangement showed satisfactory results for forecasting the very
the electricity demand information as the sole entry requirement
short-term load of normal days [16,23]. However, the particular
for developing an appropriate online load forecasting algorithm,
contribution of the present work is the consideration of the anoma-
it is therefore of rst importance to examine the main character-
lous load conditions. The consideration of such load in previous
istics of the load series and to understand the electricity demand
existing few works in the literature have been realized using com-
behavior. We present in Fig. 1 the power consumption in France
plex modeling process that requires a signicant amount of data,
for the period from January 01, 2014 to December 31, 2014. It is
and a forecaster intervention is almost essential to restructure the
clear from the plot that the electric load behavior is mainly inu-
model. In contract to those existing works, our special days model-
enced by daily and annually seasonal patterns where the electricity
ing algorithm is very simple and requires only few recent samples to
consumption typically begins with low values in early morning
perform predictions. Hence, each parallel series in our special days
and increases signicantly towards a daily peak load; whereas the
modeling process contains either the electricity demand informa-
tion from previous Sundays plus the load from the day to be forecast
(if it is a weekday), or the electricity demand information from
previous Saturdays or Sundays (if it is a weekend). The forecast-
ing methods designed to be used in a parallel implementation are
the back-propagation neural network (BPNN), the adaptive neuro-
fuzzy inference system (ANFIS), and a recently developed adaptive
Holts exponential smoothing (AHES) technique. For these paral-
lel models, the four step-ahead forecasts (or 12 for the Australian
data) can be obtained by performing one-step ahead forecast on
each separated parallel series related to the evaluation period (1 h).
According to this procedure, four (or 12) models with a single out-
put are implemented separately to provide the hour-ahead load
forecasts. For the methods based on the use of a sequential imple-
mentation, we made simply k step-ahead forecast from forecast
origin t; where k can up to 5 (or 13 for the Australian data), and
Fig. 1. Electricity demand data of the French power system: from January 01, 2014 t is a time related to the last observation of the electric load time
to December 31, 2014. series. The forecast horizon k is a term of up to 5 (or 13), since for
38 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547

Fig. 2. Daily peak electricity demand in France for four different weeks.

an online implementation the electricity demand information of a exist between the load demand and the other exogenous factors. A
given actual time (denoted as t + 1) is not available at the forecasting fth individual model, named adaptive Holts exponential smooth-
moment. ing, is also included. We opted to use this fth method because it is
a recently developed approach that showed promising results for
handling the rapid variation of the electricity demand and ensuring
3. The proposed online load forecasting methodology
a low very short-term load forecasting error [16]. A simple com-
bination of the forecasts from the ve individual models is also
The structure of the proposed very short-term load forecasting
included to be a part of the primary forecasting models. To avoid
methodology consists of three principal stages. At the rst part, a set
large forecasting errors, the Hampel lter is introduced at the sec-
of basic, yet efcient, methods are used to perform primary electric
ond part of our approach to replace over- and under-estimations
load forecasts for an hour in advance. Since the time series of the
by normalized points. Then, the nal load estimation is obtained
electricity demand knows multiple variations, it is of rst impor-
by considering the trimmed mean of the normalized forecasts. The
tant to take into account several hypotheses and methods in order
general structure of our online load forecasting methodology is rep-
to treat these variations and capture the recent trend of the electric-
resented in Fig. 3, and the key elements and parameters of all the
ity demand. For this reason, two hypotheses have been considered
proposed methods are presented in Table 1. The blanks in Table 1
in our method. The rst is based on the use of a sequential approach
mean that the element or the parameter concerned is not included
and the second is based on the use of a parallel approach. Moreover,
for a specic forecasting model, while * designates the inclusion
a mixture between classical and computational intelligence meth-
of the element/parameter by the forecasting model.
ods has been arranged. Hence, two classical methods (HWT and
SARIMA) and two articial intelligence methods (BPNN and ANFIS)
have been used. The inclusion of these models is based on their rea- 3.1. Individual models
sonable accuracy, simplicity of use, and acceptable execution time.
The Holt-Winters exponential smoothing and the seasonal ARIMA 3.1.1. Seasonal autoregressive integrated moving average
approaches are, without doubt, the most popular classical meth- Exponential smoothing models and BoxJenkins (ARIMA) mod-
ods that have been used in the literature of load forecasting. This eling are amongst the most widely used time series models in the
popularity is attributed to the ability of these methods to handle eld of electric load forecasting. The popularity of these methods
the seasonal component of the electricity demand time series. In is attributed to their ability of handling the different components
the other hand, articial neural network and neuro-fuzzy systems that form the electricity demand time series. The modeling process
are among the computational intelligence methods that have great of exponential smoothing methods assumes that the time series
ability to model the non-linearity and the complex relationship that can be perfectly predictable using its past variations. It enables to

Fig. 3. General structure of the proposed online load forecasting method.

 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547 39

Table 1
The key elements and parameters for the different methods.

Methods Classical methods Computational intelligence methodsCombined methods

Elements SARIMA HWT AHES BPNN ANFIS SFCM HFCM- HFCM-TM

WAM

Mathematical model Eq. (1) Eqs. (2)(6) Eq. (7)(9) Not required Not required Required in Required in Required in
some parts some parts some parts
Over/Under estimation detection * *
Simple average * *
Weighted average *
Trimmed mean *
Sequential approach * * * * *
Parallel approach * * * * * *
Special days modeling * * * * * *

capture and forecast the level of the electricity demand data along Dt = (yt /St Wts2 ) + (1 )Dts1 (4)
with different types of trends and seasonal patterns. Hence, a
mathematical model is required to represents the load as the Wt = (yt /St Dts1 ) + (1 )Wts2 (5)
sum (additive model) or the product (multiplicative model) of a
number of components (just like the HWT model in Eqs. (2)(6)). yt (k) = (St + kT t )Dts1+k Wts2+k
BoxJenkins models are similar to exponential smoothing mod-
+ k (yt ((St1 + Tt1 )Dts1 Wts2 )) (6)
els in dealing with capturing patterns in the historical data and
extrapolating those patterns into the future. However, in contrast St and Tt , are the smoothed level and trend; Dt and Wt is the seasonal
to exponential smoothing where we suppose that the time series indices for the intraday s1 (s1 = 96 for the French data, s1 = 288 for
is stationary, autoregressive integrated moving average models the Australian data) and intraweek s2 (s2 = 672 for the French data,
are more benecial in cases where the time series show evi- s2 = 2016 for the Australian data) seasonal cycles, respectively; ,
dence of non-stationarity, where an initial differencing step (the , and are the smoothing parameters; yt is the actual value of
integrated part of the model) can be applied to reduce the non- the time series in period t; and yt (k) is the k step-ahead forecast
stationarity. Hence, the seasonal autoregressive integrated moving made from forecast origin t. The term involving the parameter  in
average method can be very useful in our load forecasting problem the forecast function expression Eq. (6) is a simple adjustment for
for modeling trend, cyclical, and irregular components of the elec- rst-order autocorrelation.
tricity demand time series. The general structure of this approach In this paper, HWT method parameters are estimated by tak-
use a combination of past observations and past errors as well ing into account the minimum forecasting error for three principal
as using a differencing scheme between an observation and the stages. For the French load series, we rst generated 2 103 vectors
corresponding observation from a previous period at the end of of parameters from a uniform random number generator between
translating the time series into smooth and a stationary process. A 0 and 1. For each of the vectors, we then evaluated the absolute
general expression of a SARIMA (p, d, q) (P, D, Q )s model can be percentage error (APE) on the daily peak demand forecasting for
given by: August 02, 2014. The 200 vectors producing the lowest APE are
   D   evaluated; in their turns, for predicting the load curve of August
p (B) P Bs (1 B)d 1 Bs yt = q (B) Q Bs t (1)
02, 2014 (96 load values). Among the 200 vectors, the 10 vec-
where yt and t are respectively the actual value and random error tors that have the lowest mean absolute percentage error (MAPE)
at time t. B is a backward shift operator satisfying: Bt yt = yt1 . P (Bs ), are evaluated at last for next-hour load prediction over the period
Q (Bs ) and (1 Bs )D are related to autoregressive, moving aver- from August 02, 2014 to August 08, 2014. The one producing
age and differencing for seasonal components. While p (B), q (B) the lowest MAPE value was chosen as the nal parameter vector.
and (1 B)d are related to autoregressive, moving average and dif- The optimized values of HWT method parameters were: = 0.987;
ference for the non-seasonal component. S is the seasonal order.  = 0.057; = 0.006; = 0.272; and  = 0.008. For the Australian load
The adopted SARIMA model for the France and the Australian load series, the parameters are estimated by a similar manner and the
data can be described as SARIMA (1, 1, 1) (2, 1, 1)96 and SARIMA optimized values were: = 0.996;  = 0.191; = 0.273; = 0.545;
(3, 1, 3) (2, 1, 1)288 , respectively. and  = 0.206.

3.1.2. Holt-Winters-Taylor method 3.1.3. Back-propagation neural network

Exponential smoothing method has established itself as one Due to their inherent capability for solving any complex rela-
of the robust and accurate strategies for forecasting time series tionships between inputs and outputs through a learning process,
of different characteristics. In particle, the Holt-Winters exponen- the use of neural networks models is nowadays an alternative
tial smoothing is a very successful method for dealing with time method for dealing with the electric load forecasting problem
series which contain both trend and seasonal variations. However, [46,47]. In this regard, a univariate back-propagation neural net-
the standard Hold-Winters is suitable for accommodating only one work model is used in this work. The conguration of the proposed
seasonal pattern on the data. In this work we apply an extended neural network is very similar to the one used in [48]. The proposed
version of the standard Holt-Winters method, originally developed BPNN model contains one hidden layer with four sigmoidal neurons
by Taylor [45], for accommodating both the intraday and intraweek followed by an output layer of linear neuron. The load information
seasonal cycles in the electricity demand series. The formulation for from a previous parallel series is used as input, and the load data
the double seasonal Holt-Winters-Taylor method is given in the for the quarter-hour (or the ve-minutes) to be forecast is adopted
following expressions: as the output of the network. Such univariate modeling procedure
St = (yt /Dts1 Wts2 ) + (1 )(St1 + Tt1 ) (2) is frequently considered to be sufcient for very short-term load
forecasting because the weather variables tend to change in a
Tt = (St St1 ) + (1 )(Tt1 ) (3) smooth manner over short time scales, and the effects of exogenous
40 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547

factors can be captured in the electricity demand time series itself. forecasting accuracy. This method is based on assigning an arith-
During the building process of our neural network model, the load metic mean to the individual forecasts for a time m and often
data of one year are divided into three subsets: 70% for training, generates reliable forecasts for normal days. Its mathematical rep-
15% for validation and 15% for test. The LevenbergMarquardt algo- resentation is given below:
rithm is considered to train the network and the mean square error
1
n
(MSE) criteria is used as the performance function criteria. The c
Ym = Ym,i (10)
number of epochs for training the network is limited to a maximum n
i=1
of 20 epochs.
c is the combined forecast; Y
where Ym m,i is the individual forecast
3.1.4. Adaptive neuro-fuzzy inference system obtained from the ith model; n is equal to 5, the total number of
The adaptive neuro-fuzzy inference system is one of the suc- proposed individual models.
cessful schemes that combine the merits of human reasoning from
fuzzy logic and the adaptive learning ability of articial neural 3.2. Data cleaning lter to avoid large forecasting errors
networks. Such hybrid system is a fuzzy TakagiSugeno model
where the parameters and the optimal topology are obtained by Sometimes, even the more advanced combination method may
applying a learning process of an adaptive articial neural net- fails to improve the forecasting accuracy, especially for the anoma-
work [49]. It can be viewed as a multi-layer feed-forward network lous loads where the accurate prediction is almost considered as a
in which each node (neuron) performs a particular function on difcult task. For such cases, we can nd frequently that the fore-
incoming signals. In this architecture, the rst layer executes the cast of one of the individual models is signicantly outside the range
fuzzication process and gives the degree of fuzzy membership of of the majority of methods. To avoid large forecasting errors, the
the input. The second computes the ring strengths of the fuzzy Hampel lter is introduced at the second part of our approach to
rules. The third layer normalizes the ring strengths from the pre- replace over- and under-estimations by normalized points, which
vious layer. The fourth layer gives the consequent part of the fuzzy can help to improve the forecasting accuracy when combining the
rules. Finally, the fth layer presents a summation of all incoming normalized forecasts.
signals from the fourth layer [23]. The proposed neuro-fuzzy infer- The Hampel lter is a moving window nonlinear data clean-
ence structure uses generalized bell-shaped membership function, ing lter that searches outliers or abnormal local data in a time
two fuzzy rules, and 20 epochs for identifying the fuzzy inference sequence and replaces them with a more reasonable value when it
system (FIS) parameters and the optimal topology of the ANFIS. The nds them [50]. For a data sequence (Y1 , Y2 , . . ., Yn ), the lter forms
input/output pairs are similar to that of BPNN model. a collection of samples from which it computes the median and the
median absolute deviation (MAD) and performs the outlier test.
3.1.5. Adaptive Holts exponential smoothing The median of a data sequence is obtained by rst rank-ordering
AHES is an adaptive load forecasting approach which has been it from smallest to largest, i.e. Y1 Y2 Y3 Yn1 Yn and then
introduced in [16] to produce one step-ahead prediction. The taking either the middle value (if it is odd) or the average of the
method is adaptive at the manner that their smoothing parame- middle two values (if it is even) [51]. The Hampel lter denes out-
ters are updated every time when a new data point is recorded. liers as those data points Yj whose absolute difference from the
It is found that the adaptive scheme of this methodology has the median of the short series in the moving window is greater than a
advantage to perceive the variation that could be effected in the pre-determined threshold:
electric load time series, and so that the ability of providing a slight  
error when forecasting the electricity consumption of a particular Yj Y0  > T S0 (11)
time. The formulation for the Holts exponential smoothing could
be set up as: where Y0 represents the median value of the rank-ordered samples
of the data sequence of length n, T is a scalar threshold, and S0 is
St = yt + (1 )(St1 + Tt1 ) (7) the median absolute deviation, as dened by:
 
Tt = (St St1 ) + (1 )(Tt1 ) (8) S0 = median Yj Y0  (12)
yt (k) = St + kT t (9) If the Hampel lter declares a data point to be an outlier, its value
St and Tt are the smoothed level and trend; and  are the smooth- is then replaced by the median of its neighbors. For the Hampel lter
ing parameters; yt is the actual value of the time series in period t; used in our load forecasting model, the time sequence is dened
and yt (k) is the k step-ahead forecast made from forecast origin t. by the forecasts from the individual models and from the simple
For the purpose of predicting the electricity consumption at a average-based forecast combination method.
particular time m; we create at a primary step two parallel load
series: a rst series for a previous time m 1, and a second series 3.3. Final electric load estimation
for the time m. In the second stage, we look to nd a pair of smooth-
ing parameters that give the smallest forecasting error when going Once the Hampel lter is performed, the nal load forecasting is
from the penultimate to the last data point of the rst electricity obtained by taking into account the trimmed mean of the normal-
demand time series. Since the exponential smoothing methods are ized forecasts (the mean of the normalized forecasts, excluding the
mostly inuenced by newest data points, and especially the abil- highest and lowest value). For the purpose of comparison, we pro-
ity of the Holts method to identify the tendency in non-seasonal pose also a second variation of HFCM. In this method, the weighted
time series, it seems that the fact of searching an optimal model average of the normalized forecasts is employed to combine fore-
assuring the minimum error at the time m 1 may signicantly casts from primary models. This combined forecast can be written
help to provide a small error for predicting the load of the specied as follows:
time m.
F (f1 (Y ) , f2 (Y ) , . . ., fn (Y )) = w1 f1 (Y ) + w2 f2 (Y ) + + wn fn (Y ) (13)

3.1.6. Simple average-based forecast combination method where fi is related to a primary forecasting method transforming
In this method, a simple average of the primary forecasts from the real electricity consumption Y of a previous quarter-hour (or
the individual models is performed with the aim of improving the ve-minutes) to its prediction Y . In the present work, the inverse
 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547 41

Fig. 4. HFCM-TM results for the French load data: from August 09, 2014 to August 12, 2014.

error weights are used [1], so that a model with a smaller error 4. Results and discussion
obtains more weight:
In this section, we describe an evaluation of the Hampel
lter-based forecast combination method for hour-ahead load fore-
e1 casting on both normal and special daily conditions. We will also
wi = n i , i = 1, 2, . . ., n (14) compare the results with a number of benchmark algorithms and
e1
i=1 i forecast combination methods. In addition, a discussion about the
computation time of our provided tool for realizing the online pre-
diction of the electricity demand will be presented.
where ei designates the absolute percentage error realized by the
ith prediction method. By this approach, the forecasting technique
with the lowest value of the absolute percentage error receives a 4.1. Evaluation of forecasts for the French load data
higher weight in composing the combined prediction. For the rest
of this paper, the method based on the weighted arithmetic mean Figs. 47 show the results of HFCM-TM for next-hour elec-
algorithm will be dened by HFCM-WAM, while the proposed tricity demand prediction in France along the week from 09th to
method based on the trimmed mean will be noted by HFCM-TM. 15th August 2014, and the week from 10th to 16th January 2015,

Fig. 5. HFCM-TM results for the French load data: from August 13, 2014 to August 15, 2014.
42 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547

Fig. 6. HFCM-TM results for the French load data: from January 10, 2015 to January 13, 2015.

Fig. 7. HFCM-TM results for the French load data: from January 14, 2015 to January 16, 2015.

respectively. It is obvious from this gure that satisfactory results where Y m is the predicted load for the time m, Ym is the actual load
were obtained by using the Hampel lter-based forecast combi- at the time m, n is the total number of forecasting cases.
nation approach, since the shape of predicted load curves is very All detailed forecasting performances are listed in Table 2. These
similar to that of actual load curves. However, we need to compare results clearly demonstrate the effectiveness of the proposed fore-
the performance of our HFCM-TM against the performance all other casting strategy and show that HFCM-TM outperformed all other
considered methods. For the evaluation of the methods considered methods. It can be seen from the relative performance that the
in this study, we used APE, and MAPE criteria: average MAPE value of HFCM-TM in normal days is equal to
0.478%, which is 5.720% and 22.023% more accurate than the simple
|Y m Ym | average-based forecast combination method and the best individ-
APE = 100 (15)
Ym ual forecasting approach (AHES), respectively. The results show
also that the method based on a weighted mean (HFCM-WAM) has
1  |Ym Ym |
n been slightly outperformed by HFCM-TM. Among the individual
MAPE = 100 (16) forecasting methods, the recently developed AHES has showed a
n Ym
m=1 very good forecasting accuracy. The obtained performance of this
 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547 43

Table 2
The relative error (MAPE) for forecasting the electricity consumption in France over two weeks.

Time SARIMA HWT BPNN ANFIS AHES SFCM HFCM-WAM HFCM-TM

09/08/2014 1.006 0.566 0.979 0.843 0.721 0.583 0.556 0.570

10/08/2014 1.337 0.758 1.080 0.906 0.739 0.661 0.641 0.642
11/08/2014 0.666 0.569 1.420 0.840 0.583 0.582 0.565 0.552
12/08/2014 0.533 0.488 0.685 0.740 0.393 0.406 0.400 0.404
13/08/2014 0.430 0.634 0.703 0.788 0.479 0.404 0.399 0.376
14/08/2014 0.437 1.511 0.877 0.777 0.334 0.499 0.416 0.421
15/08/2014 1.088 1.710 0.982 0.737 0.743 0.649 0.740 0.642
10/01/2015 1.027 0.483 1.095 0.513 0.703 0.472 0.460 0.438
11/01/2015 1.204 0.777 1.058 0.649 1.040 0.690 0.663 0.648
12/01/2015 1.004 0.608 0.524 0.446 0.685 0.439 0.437 0.430
13/01/2015 0.806 0.758 0.519 0.490 0.617 0.470 0.431 0.430
14/01/2015 0.684 0.728 0.560 0.500 0.589 0.493 0.486 0.486
15/01/2015 0.574 0.728 0.454 0.521 0.641 0.444 0.425 0.415
16/01/2015 0.452 0.829 0.520 0.530 0.445 0.447 0.393 0.407
Normal days 0.781 0.726 0.806 0.657 0.613 0.507 0.483 0.478
All days 0.803 0.796 0.818 0.663 0.622 0.517 0.501 0.490

Table 3
The relative error (MAPE) for the public holidays of 2014 in France.

Time AHES SFCM HFCM-WAM HFCM-TM

01/01/2014 1.433 1.517 1.382 1.400

21/04/2014 0.603 0.992 0.722 0.747
01/05/2014 1.128 1.111 0.901 0.837
08/05/2014 1.115 0.900 0.789 0.783
29/05/2014 0.946 1.190 0.917 0.898
09/06/2014 1.218 0.883 0.797 0.730
14/07/2014 0.649 0.803 0.628 0.636
15/08/2014 0.743 0.644 0.740 0.642
01/11/2014 1.132 0.812 0.835 0.799
11/11/2014 0.572 0.991 0.670 0.639
25/12/2014 1.563 1.586 1.547 1.525
All days 0.996 1.026 0.890 0.863

higher performance on the validation period. Therefore, if an over

Fig. 8. The exponential smoothing parameters of AHES over a week in summer and or under-estimation of the load is obtained by this model, we will
a week in winter. get then an erroneous forecast if the weighted combination method
is not able to detect and to replace this over or under-estimation.
method is mainly attributed to its adaptive character where the To incorporate this limitation it will be more appropriate to assign
smoothing parameters, as shown in Fig. 8, are updated every time equal weights to each individual model, or better; to exclude the
when a new observation is recorded. A satisfactory performance highest and lowest data value from the normalized forecasts and
of a MAPE value equal to 0.743% has been also achieved by this use then the arithmetic mean of the four remaining values. Thus,
method for the public holiday of August 15th 2014. For this spe- this is the case with our HFCM-TM.
cic day, a further improvement in forecasting accuracy has been In order to clarify the key idea of our HFCM-TM, we present
obtained using HFCM-TM. in Fig. 9 and in Table 4 the forecasting results for December 26,
However, we cannot be denitively sure about the superiority
of our combined method for forecasting atypical loads if we base
only on the results of just one comparing day. So an extended eval-
uation period is extremely necessary to make a clear idea about the
performance of each method during the atypical load conditions.
For that reason, a thorough assessment of AHES, SFCM, HFCM-WAM
and HFCM-TM is performed off-line on the length of the public hol-
idays in France for the year of 2014 and the results of the MAPEs
are listed in Table 3. It can be seen from the results that the average
MAPE of HFCM-TM is less than all the average MAPEs of the other
methods. Compared to the case of the normal days where HFCM-
TM showed only 5.720% higher accuracy over SFCM, we can notice
for the case of anomalous days that a very signicant improvement
is achieved by HFCM-TM; with an average MAPE value lower by
15.887%. By comparing the results obtained by HFCM-TM with the
results obtained by HFCM-WAM, we can clearly observe that this
last method has a lower performance. We think that the main rea-
son for this is attributed to the weighted algorithm of HFCM-WAM.
In this method of combination, the forecasting model that has a
lower forecasting error in a training period gets a higher weight in
composing the combined forecast. However, a model with higher Fig. 9. Results of HFCM-TM and HFCM-WAM for the forecasting process of
performance in a training period does not mean a model with December 26th 2014 at 00:00.
44 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547

Table 4 occurred for more than 85% and 65% of the evaluation period of
Forecasting results for December 26, 2014 at 00:00 AM.
normal and special days, respectively. However, the results from
Results Actual load (MW) Predicted load (MW) APE (%) these gures show that HFCM-TM is more effective since it avoided
Method the most signicant number of large forecasting errors.
SARIMA 60363 1.285
HWT 60005 0.685 4.2. Evaluation of forecasts for the Australian load data
BPNN 59439 0.265
ANFIS 59597 59560 0.062
AHES 61300 2.857 In this subsection, we carry out a further evaluation of our
SFCM 60133 0.895 method through performing a simulation of the Australian load
HFCM-WAM 60027 0.721 data for the state of New South Wales. The main reason for this
HFCM-TM 59919 0.540
application is to compare our model with the one proposed in
[26]. The model proposed in the cited reference, called Advanced
2014 at 00:00. We can perceive three important points from these Wavelet Neural Network (AWNN), combines an advanced wavelet
results: (1) an over-estimation of the load has been occurred by the algorithm for load decomposition, mutual information for feature
best individual forecasting method, (2) HFCM-TM is outperformed selection and neural networks as prediction algorithm. AWNN was
by BPNN and ANFIS models, (3) Although that ANFIS model has tested on the Australian load dataset of 2007 and achieved a MAPE
received the second lower weight in composing the combined equal to 0.938% for a forecast horizon of an hour-ahead. For a fair
forecast of HFCM-WAM, the obtained results indicated that this comparison, we used the same evaluation period (2007) to build
individual approach has achieved the best forecasting performance hour-ahead load forecasting with the prediction algorithm pro-
among all the considered methods. This is consistent with what posed in the current study. Fig. 12 presents the predicted load with
we have reported above about the non-existence of a forecasting HFCM-TM as a function of an actual load and plots also the actual
method which gives the best results for all times. However, the and predicted load values for a typical working day (December
advantage of our proposed HFCM-TM is located on its ability to 31, 2007). This gure indicates the high correlation between the
reduce the uncertainty and provide electricity demand predictions predicted and actual load values. Table 5 compares the predictive
which are very close to the actual values of the load. This can be accuracy results of HFCM-TM with those of AWNN. The results in
clearly observed by Figs. 10 and 11, where we compared the APE this table show that HFCM-TM achieved overall prediction errors as
distribution for AHES, SFCM, HFCM-WAM and HFCM-TM on nor- follows: MAPE = 1.048% for the public holidays and MAPE = 0.860%
mal and anomalous load conditions, respectively. Although that for all the days of 2007. Furthermore, HFCM-TM showed better
the simple averaging used in SFCM is among the most commonly forecasting results and the MAPE improvement by using HFCM-TM
used approaches to produce the combined forecast, we can perceive compared to AWNN is equal to 8.316%.
that this method of combination has not allowed to any improve-
ment in the case of the anomalous load conditions. In contrast, 4.3. Presentation of the proposed online load forecasting tool
we can observe that HFCM-WAM and HFCM-TM give almost the
best results and show excellent forecasting performances, since an In addition to an effective model that ensures a very high
absolute percentage error lower than one percent has been accuracy, a friendly forecasting tool takes also an important role

Fig. 10. APE distribution of AHES, SFCM, HFCM-WAM, and HFCM-TM for normal loads.

Fig. 11. APE distribution of AHES, SFCM, HFCM-WAM, and HFCM-TM for anomalous loads.
 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547 45

Fig. 12. Predicted load by HFCM-TM plotted against actual load from the Australian dataset.

Table 5 led us; through an extensive study, to the development of the

The relative error (MAPE) for the load data of the New South Wales.
Hampel lter-based forecast combination method. For the online
Day type All the public holidays of 2007 All the days of 2007 implementation of this method, many considerations, algorithms
Forecasting method and functions have been taken into account for developing suit-
AWNN 0.938 able software using MATLAB R2014b environment. As shown in
HFCM-TM 1.048 0.860 Fig. 13, our method is currently accompanied with a forecasting
tool and working very well for the online electricity demand predic-
tion of the French power system. This forecasting tool can perform
at the real implementation of any electric load forecasting tech- an hour ahead forecasting from any moment selected from the
nique. As a natural extension of the work in [16], the AHES is user. This is important, because in practice, the power compa-
the method that has been implemented for performing the online nies need to perform the load predictions in an iterative manner,
electricity demand prediction in France over the summer week not once every an hour. However, for an evaluation purpose, the
and winter week of the evaluation period. We opted to use the same quarter-hours of a given day are considered to form the
French power data for the simple reason that we found the RTE rst step in an hour ahead load forecasting over all the evalua-
(transmission system operator of France) website very useful and tion period of our study. Hence, the rst steps in the 1-h ahead
accurate in providing the electric load data in real-time. How- forecasting are 00:15, 01:15, 02:15, . . ., 23:15. Thus, in this manner
ever, the aim of achieving forecasting error not higher than one we get only one forecasted value for each specic moment of the
percent under both normal and anomalous load conditions has day.

Fig. 13. Developed tool for online electricity demand prediction.

46 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547

Table 6
The computation time of all considered methods in four different forecasting cases.

Method
SARIMA HWT BPNN ANFIS AHES SFCM HFCM-WAM HFCM-TM
Results

Evaluation period Sat, 09 Aug 2014, 12:45-13:30

Computation time (s) 4.38 5.16 6.08 4.29 4.71 7.96 11.85 8.39
Evaluation period Thu, 25 Dec 2014, 23:15-00:00
Computation time (s) 6.57 7.39 9.20 5.28 7.66 11.24 15.83 11.61
Evaluation period Thu, 27 Oct 2016, 21:15-22:00
Computation time (s) 6.77 7.28 7.84 5.97 8.51 11.39 16.61 12.35
Evaluation period Tue, 01 Nov 2016, 21:30-22:15
Computation time (s) 6.27 6.93 8.35 5.84 8.25 10.84 13.24 11.06

However, there are three principal tasks that can be performed combined method can be easily implemented at any power system
by using our provided forecasting tool. Obviously, the main task of for performing the online predictions of the electricity demand.
our tool is to enable the online prediction of electricity demand.
All what the user is required to do for achieving this function is to
select the forecasting method and the type of day in which he would 5. Conclusion
like to perform the predictions and the software complete then the
task by passing through four stages: (1) getting the current time in In this paper, a new technique was developed to perform the
France, (2) connecting to RTE website and downloading the avail- online predictions of the electricity demand for a lead time of an
able load data for the forecasting day, (3) updating the database by hour ahead. Such very-short term forecasts are important for the
adding up the downloaded data to the main le the include histor- real-time scheduling of electricity generation as well as the load-
ical load data, (4) performing the required algorithms and showing frequency control and the economic dispatch functions. However,
the load forecasting results for the next-hour. In addition to the since the load demand behavior is complex and mainly affected
online implementation, the proposed tool enables also to perform by several factors, there is no forecasting method which gives the
the off-line predictions. This is important if the user would like to best result for all times. It is therefore necessary to integrate differ-
see the previous forecasting performance by taking into account ent models for reducing the uncertainty of the predictions. In this
different error measurement criteria. The third task that can be regard, the proposed method is categorized in combined forecast-
achieved by our provided software is the ability to update the his- ing methods. To deal with the seasonal features and the non-linear
torical load data in case where the user neglected the use of the character of the electric load time series, two classical methods and
tool for several days. In that case, an intervention from the user is two computational intelligence approaches, have been considered.
required to select a desired day from Update The Historical Data A recently developed adaptive exponential smoothing method and
panel in order include the missed data to the main historical data a simple combination of the forecasts from the ve individual mod-
le. els were also used as candidate models. To improve the forecasting
We present in Table 6 the computation time of our software on a accuracy when combining forecasts from the primary models, a
laptop of 4 GB of RAM, i3 380 M processor, a hard drive of 5200-rpm, data cleaning lter and the trimmed mean were employed as two
and operating in Windows 10 platform. The table shows the com- key elements at this function of improving the performance and
putation time of each considered method in four different cases: avoiding the large forecasting errors. The obtained results showed
the rst case (09 August 2014) is related to an off-line execution that the proposed forecasting scheme has satisfactory performance
for performing hour ahead load prediction during a weekend; the in both normal and special daily conditions, with an average mean
second case (25 December 2014) is related to an off-line implemen- absolute percentage error not higher than one percent in both two
tation for performing hour ahead load prediction during a special cases. Moreover, our method is also accompanied with a user-
day; the third case (27 October 2016) is related to an online execu- friendly graphical interface and the execution time of this software
tion for performing hour ahead load prediction during a weekday; on a laptop of 4 GB of RAM; i3 380 M processor; and a hard drive of
and the fourth case (01 November 2016) is related to an online 5200-rpm, is lower than 15 s. Therefore, the proposed methodology
execution for performing hour ahead load prediction during a pub- can be easily implemented at any power system for performing the
lic holiday. One can clearly perceive from these results that the online predictions of the electricity demand.
individual models require less computation time compared to the
combined methods. However, the difference in computation time
is very acceptable if we take into account that the combined meth- References
ods require the simultaneous use of many algorithms and methods
to perform predictions. Moreover, the execution time of our HFCM- [1] J.X. Che, J.Z. Wang, Short-term load forecasting using a kernel-based support
TM method is lower than 15 s. This time includes the time to update vector regression combination model, Appl. Energy 132 (1) (2014) 602609.
[2] S. Li, P. Wang, L. Goel, Short-term load forecasting by wavelet transform and
the historical load data and to write the forecasting results on an evolutionary extreme learning machine, Electr. Power Syst. Res. 122 (2015)
XLS le. If we compare the execution time of HFCM-TM for the pub- 96103.
lic holiday of 2014 with that of the public holiday of 2016, we can [3] A. Laoua, M. Mordjaoui, F. Laoua, T.E. Boukelia, Daily peak electricity
demand forecasting based on an adaptive hybrid two-stage methodology, Int.
clearly perceive that both cases require practically the same exe- J. Electr. Power Energy Syst. 77 (2016) 136144.
cution time of about 11 s. This is can be easily justied if we take [4] G. Dudek, Pattern-based local linear regression models for short-term load
into account that we have limited the data to be used by each indi- forecasting, Electr. Power Syst. Res. 130 (2016) 139147.
[5] L. Zjavka, V. Snsel, Short-term power load forecasting with ordinary
vidual method to a predened amount of samples. For example,
differential equation substitutions of polynomial networks, Electr. Power
even if the main historical load data is consisted by twenty years Syst. Res. 137 (2016) 113123.
of electricity demand observations, the HWT method is supposed [6] E. Gonzalez-Romera, M.A. Jaramillo-Moran, D. Carmona-Fernandez, Monthly
to use only the last year of observations as a sufcient amount of electric energy demand forecasting based on trend extraction, IEEE Trans.
Power Syst. 21 (4) (2006) 19461953.
data for handling the seasonality of the time series and performing [7] N. Abu-Shikhah, F. Elkarmi, Medium-term electric load forecasting using
the predictions. Hence, with an execution time lower than 15 s, our singular value decomposition, Energy 36 (7) (2011) 42594271.
 A. Laoua et al. / Electric Power Systems Research 148 (2017) 3547 47

[8] Z. Hu, Y. Bao, R. Chiong, T. Xiong, Mid-term interval load forecasting using [28] G. Dudek, Pattern similarity-based methods for short-term load forecasting.
multi-output support vector regression with a memetic algorithm for feature Part 2: Models, Appl. Soft Comput. 36 (2015) 422441.
selection, Energy 84 (2015) 419431. [29] C.E. Borges, Y.K. Penya, I. Fernandez, Evaluating combined load forecasting in
[9] O. Ahmia, N. Farah, Multi-model approach for electrical load forecasting, in: large power systems and smart grids, IEEE Trans. Ind. Inf. 9 (3) (2013)
2015 SAI Intelligent Systems Conference (IntelliSys), London, 1011 15701577.
November 2015, 2015, pp. 8792. [30] L.M. de Menezes, D.W. Bunn, J.W. Taylor, Review of guidelines for the use of
[10] H.M. Al-Hamadi, S.A. Soliman, Long-term/mid-term electric load forecasting combined forecasts, Eur. J. Oper. Res. 120 (1) (2000) 190204.
based on short-term correlation and annual growth, Electr. Power Syst. Res. [31] R. Clemen, Combining forecasts: a review and annotated bibliography, Int. J.
74 (3) (2005) 353361. Forecast. 5 (1989) 559583.
[11] U. Perwez, A. Sohail, S.F. Hassan, U. Zia, The long-term forecast of Pakistans [32] J. Wang, J. Wang, Y. Li, S. Zhu, J. Zhao, Techniques of applying wavelet
electricity supply and demand: an application of long range energy de-noising into a combined model for short-term load forecasting, Int. J.
alternatives planning, Energy 93 (2) (2015) 24232435. Electr. Power Energy Syst. 62 (2014) 816824.
[12] A. Laoua, M. Mordjaoui, F. Laoua, An evaluation of conventional and [33] Z. Liu, W. Sun, J. Zeng, A new short-term load forecasting method of power
computational intelligence methods for medium and long-term load system based on EEMD and SS-PSO, Neural Comput. Appl. 24 (3) (2014)
forecasting in Algeria, in: 3rd International Conference on Control, 973983.
Engineering & Information Technology CEIT2015, Tlemcen, Algeria, 2527 [34] L. Xiao, J. Wang, R. Hou, J. Wu, A combined model based on data pre-analysis
May, 2015, 2015. and weight coefcients optimization for electrical load forecasting, Energy 82
[13] K. Liu, S. Subbarayan, R.R. Shoults, M.T. Manry, C. Kwan, F.L. Lewis, J. (2015) 524549.
Naccarino, Comparison of very short-term load forecasting techniques, IEEE [35] J. Nowotarski, B. Liu, R. Weron, T. Hong, Improving short term load forecast
Trans. Power Syst. 11 (2) (1996). accuracy via combining sister forecasts, Energy 98 (2016) 4049.
[14] A. Kotillov, Very short-term load forecasting using exponential smoothing [36] L. Xiao, W. Shao, T. Liang, C. Wang, A combined model based on multiple
and ARIMA models, J. Inf. Control Manag. Syst. 9 (2) (2011). seasonal patterns and modied rey algorithm for electrical load
[15] J.W. Taylor, An evaluation of methods for very short-term load forecasting forecasting, Appl. Energy 167 (2016) 135153.
using minute-by-minute British data, Int. J. Forecast. 24 (2008) 645658. [37] M. Ghofrani, M. Ghayekhloo, A. Arabali, A. Ghayekhloo, Hybrid short-term
[16] A. Laoua, M. Mordjaoui, D. Dib, Very short-term electricity demand load forecasting with a new input selection framework, Energy 81 (2015)
forecasting using adaptive exponential smoothing methods, in: 15th 777786.
International Conference on Sciences and Techniques of Automatic Control & [38] S. Arora, J.W. Taylor, Short-term forecasting of anomalous load using
Computer Engineering STA2014, Hammamet, Tunisia, 2123 December rule-based triple seasonal methods, IEEE Trans. Power Syst. 28 (3) (2013)
2014, 2014, pp. 553557. 32353242.
[17] D.J. Trudnowski, W.L. McReynolds, J.M. Johnson, Real-time very short-term [39] K.H. Kim, H.S. Youn, Y.C. Kang, Short-term load forecasting for special days in
load prediction for power-system automatic generation control, IEEE Trans. anomalous load conditions using neural networks and fuzzy inference
Control Syst. Technol. 9 (2) (2001). method, IEEE Trans. Power Syst. 15 (2) (2000) 559565.
[18] C. Guan, P.B. Luh, L.D. Michel, Z. Chi, Hybrid Kalman lters for very short-term [40] M. Ghayekhloo, M.B. Menhaj, M. Ghofrani, A hybrid short-term load
load forecasting and prediction interval estimation, IEEE Trans. Power Syst. 28 forecasting with a new data preprocessing framework, Electr. Power Syst. Res.
(4) (2013) 38063817. 119 (2015) 138148.
[19] W. Charytoniuk, M.S. Chen, Very short-term load forecasting using articial [41] C. Guan, P.B. Luh, L.D. Michel, Y. Wang, P.B. Friedland, Very short-term load
neural networks, IEEE Trans. Power Syst. 15 (1) (2000) 263268. forecasting: wavelet neural networks with data pre-ltering, IEEE Trans.
[20] P. Qingle, M. Zhang, Very short-term load forecasting based on neural network Power Syst. 28 (1) (2013) 3041.
and rough set, in: 2010 International Conference on Intelligent Computation [42] RTE, Transmission System Operator of France, Website: www.rte-france.com.
Technology and Automation, vol. 3, 1112 May 2010, 2010, pp. 11321135. [43] AEMO, Australian Energy Market Operator, Website: www.aemo.com.au.
[21] D. Chen, M. York, Neural network based very short term load prediction, IEEE [44] D. Fay, J.V. Ringwood, M. Condon, M. Kelly, 24-h electrical load data sequential
PES Transm. Distrib. Conf. Expos. 2124 (2008) 19. or partitioned time series? Neurocomputing 55 (34) (2003) 469498.
[22] L.C.M. de Andrade, I.N. da Silva, Very short-term load forecasting using a [45] J.W. Taylor, Short-term electricity demand forecasting using double seasonal
hybrid neuro-fuzzy approach, in: 2010 Eleventh Brazilian Symposium on exponential smoothing, J. Oper. Res. Soc. 54 (2003) 799805.
Neural Networks (SBRN), 2328 October 2010, 2010, pp. 115120. [46] D. Niu, H. Shi, D.D. Wu, Short-term load forecasting using Bayesian neural
[23] A. Laoua, M. Mordjaoui, D. Dib, One-hour ahead electric load forecasting networks learned by Hybrid Monte Carlo algorithm, Appl. Soft Comput. 12 (6)
using neuro-fuzzy system in a parallel approach, in: Computational (2012) 18221827.
Intelligence Applications in Modeling and Control, Studies in Computational [47] X. Lu, J. Wang, Y. Cai, J. Zhao, Distributed HS-ARTMAP and its forecasting
Intelligence, vol. 575, 2015, pp. 95121. model for electricity load, Appl. Soft Comput. 32 (2015) 1322.
[24] A. Setiawan, I. Koprinska, V.G. Agelidis, Very short-term electricity load [48] A. Laoua, M. Mordjaoui, D. Dib, One-hour ahead electric load and wind-solar
demand forecasting using support vector regression, in: Proceedings of power generation forecasting using articial neural network, in: 2015 6th
International Joint Conference on Neural Networks, Atlanta, Georgia, USA, International Renewable Energy Congress (IREC), Sousse, Tunisia, 2426
2009. March 2015, 2015, pp. 16.
[25] Y. Wang, L. Wu, On practical challenges of decomposition-based hybrid [49] J.S.R. Jang, ANFIS: adaptive network based fuzzy inference system, IEEE Trans.
forecasting algorithms for wind speed and solar irradiation, Energy 112 Syst. Man Cybern. 23 (3) (1993) 665685.
(2016) 208220. [50] F.R. Hampel, The inuence curve and its role in robust estimation, J. Am. Stat.
[26] M. Rana, I. Koprinska, Forecasting electricity load with advanced wavelet Assoc. 69 (346) (1974) 383393.
neural networks, Neurocomputing 182 (2016) 118132. [51] R.K. Pearson, Outliers in process modeling and identication, IEEE Trans.
[27] G. Dudek, Pattern similarity-based methods for short-term load forecasting. Control Syst. Technol. 10 (1) (2002).
Part 1: Principles, Appl. Soft Comput. 37 (2015) 277287.