Energy 144 (2018) 776e788

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Forecasting mid-long term electric energy consumption through

bagging ARIMA and exponential smoothing methods
Erick Meira de Oliveira a, b, *, Fernando Luiz Cyrino Oliveira a
a
Industrial Engineering Department, Pontifical Catholic University of Rio de Janeiro (PUC-Rio), Rua Marqu^ ~o Vicente, 225, Ga
es de Sa vea, 22451-900 Rio de
Janeiro, Rio de Janeiro, Brazil
b
Studies and Research Division, Brazilian Innovation Agency (Finep), Avenida República do Chile, 330, Torre Oeste - 17o andar, 20031-170 Rio de Janeiro,
Rio de Janeiro, Brazil

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

Article history: In the last decades, the world's energy consumption has increased rapidly due to fundamental changes in
Received 9 May 2017 the industry and economy. In such terms, accurate demand forecasts are imperative for decision makers
Received in revised form to develop an optimal strategy that includes not only risk reduction, but also the betterment of the
15 November 2017
economy and society as a whole. This paper expands the fields of application of combined Bootstrap
Accepted 11 December 2017
Available online 12 December 2017
aggregating (Bagging) and forecasting methods to the electric energy sector, a novelty in literature, in
order to obtain more accurate demand forecasts. A comparative out-of-sample analysis is conducted
using monthly electric energy consumption time series from different countries. The results show that
Keywords:
Electricity consumption
the proposed methodologies substantially improve the forecast accuracy of the demand for energy end-
Forecasting use services in both developed and developing countries. Findings and policy implications are further
Bagging discussed.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction Over the past decades, a large number of approaches have been
proposed to adequately estimate and forecast electric energy con-
Ensuring an adequate supply of energy is a pressing national sumption. In brief terms, the approaches can be divided into two
priority in almost every nation in the world. The issue is even more major categories on the basis of forecasting indicators: short-term
crucial when it comes to electric energy since, unlike other energy and mid-/long-term. Whilst the first is concerned with time
sources, it cannot be stored for large-scale consumption. From an frames of minutes to hours, the horizon for mid/long-term fore-
economic viewpoint, the supply must meet the demand for energy casting ranges from a few weeks to several years [1]. Forecasting
end-use services at any given time. In this connection, accurate electric energy demand accurately over the latter period is often
electric load forecasting is of the utmost importance for decision- regarded as a challenging task, due to the nonlinear, multidimen-
making processes in the electric sector, as the consequences of sional nature of this variable [2]. In addition, many unpredictable
overestimation or underestimation can be costly. For instance, factors exist in electricity demand modeling, such as structural
when delivered power is higher than the actual demand, the pro- breaks [3] and transitory effects from external variables [4].
vider not only wastes resources but may also bear expensive costs Nevertheless, mid/long-term load forecasting assumes a particular
due to strong spot market regulation in several countries. On the importance for electric power utility planning. Even though precise
other hand, underestimation naturally results in failures and short-term forecasting forms the basis of the electrical energy trade
shortages, which in turn translates into a loss of productive time and spot price calculation [5], several decisions are made on the
and quality and subjects the provider to sanctions and penalties. basis of mid/long-term energy demand forecasting, such as the
construction of new generation facilities, the purchase of existing
generating units, the development of transmission and distribution
systems, among others [1]. Kaboli et al. [6] add that mid/long-term
* Corresponding author. Industrial Engineering Department, Pontifical Catholic
^s de S~ vea,
forecasting is especially interesting for companies operating in a
University of Rio de Janeiro (PUC-Rio), Rua Marque ao Vicente, 225, Ga
22451-900 Rio de Janeiro, Rio de Janeiro, Brazil. deregulated environment, as it provides them precious insights on
E-mail addresses: emeira@finep.gov.br, [email protected] (E.M. de the market need of energy, the maintenance schedule of power
Oliveira), [email protected] (F.L. Cyrino Oliveira).

https://doi.org/10.1016/j.energy.2017.12.049
0360-5442/© 2017 Elsevier Ltd. All rights reserved.
 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788 777

plants, the fuel supplies and occasional needs of electrical energy sub-series is lower than the estimation error for the series as a
imports/exports. whole [28]. Recent examples concerning different industrial sectors
In light of the aforementioned, despite its drawbacks in terms of can be found in the works of Benaouda et al. [29], Theodosiou [28]
complexity and uncertainty, mid-/long-term energy demand and Qiu et al. [30].
planning remains a key issue in many parts of the world. In this Albeit still timid, another stream of literature that has emerged
connection, the pursuit of models that could enhance the predic- from the ideas of series decomposition and forecast combinations is
tion accuracy in forecasting long-term electric energy demand has the use of bootstrap aggregation (bagging) techniques to improve
remained a prominent issue in the fields of energy policy and the accuracy of forecasts for a given time series. The underlying idea
economic development. is to decompose the historical data on the variable of interest into
In this work we aim to generate two-year ahead (medium to key components, bootstrap one (or more) components to generate
long-term) forecasts for monthly electric energy demand in a random pool of similar time series, estimate and subsequently
different parts of the world, including both developing and devel- forecast each time series and average all forecasts into one single
oped countries. To do so, we use an approach that has never been output. Recent examples concerning different industrial sectors can
referenced in the electric energy sector. In brief terms, we combine be found in the works of Cordeiro and Neves [31], Bergmeir et al.
different decomposition and bootstrap aggregation (bagging) [32] and Dantas et al. [33].
techniques to simulate new series. Then, we forecast each gener- The literature review resulted in no evidence of the application
ated series using a wide variety of methods and pool the forecasts of a combined approach of Bootstrap aggregating (Bagging) stra-
into one single output. Finally, a comparative out-of-sample anal- tegies and time series forecasting methods to predict electric en-
ysis is conducted on the basis of several performance metrics. ergy consumption. It is our belief that this gap is worth pursuing, as
To the best of our knowledge, no work has explored the po- the results of such procedures were promising in other fields.
tential gains of using bagging techniques to enhance monthly
forecasts of total electric energy consumption. Furthermore, we 3. Methods
contribute to the literature by proposing a different variation of this
technique, which demonstrated satisfactory forecasting results in 3.1. Decomposition
terms of accuracy for several countries.
The paper unfolds as follows. Section 2 provides a brief overview As a first step, we decompose the time series under analysis into
of the most recent techniques used to estimate and forecast energy key components, in an attempt to categorize their patterns and
demand behavior. Section 3 reviews the methodologies and the behaviors. In this work, we use the Seasonal-Trend decomposition
forecasting approaches in details. Section 4 introduces the selected using Loess (STL decomposition) [34], to obtain the trend, seasonal
data and Section 5 summarizes the results of the quantitative and remainder components of the demand time series.
analysis. Section 6 discusses findings and policy implications. Broadly speaking, STL consists of a sequence of smoothing op-
Finally, Section 7 concludes the findings of the study and suggests erations that employ locally-weighted regression (Loess). In Loess, a
directions for future works. neighborhood is first defined for each data point and the points in
that neighborhood are subsequently weighted according to their
2. Literature review distances from the respective data point. A polynomial of degree
d is then fitted to these points - usually d ¼ 1 or d ¼ 2. Higher de-
There are roughly two major approaches to investigate the dy- grees do not improve much the fit. Indeed, Cleveland et al. [34]
namics of energy demand, whether in short, mid or long time argue that taking d ¼ 1 is reasonable if the underlying pattern in the
spans: (1) Traditional forecasting models such as time series and data has gentle curvature. The trend component is equal to the
regression-based, econometrics models and (2) Soft computing value of the polynomial at each data point.
techniques such as fuzzy logic, genetic algorithms, neural networks, STL uses Loess to divide the underlying time series into three
support vector regression models, among others [7]. additive components (trend, seasonal and remainder). In brief
In brief terms, the most popular methodologies for electricity terms, the steps performed during STL decomposition are: (i)
demand forecasting can be summarized in Table 1. We hasten to detrending; (ii) cycle-subseries smoothing, in which series are built
add that this classification is not exhaustive and does not include for each seasonal component, and smoothed separately; (iii) low-
combinatorial approaches (hybrid models). Indeed, there has been pass filtering of smoothed cycle-subseries, when the subseries are
so much work and discussion concerning applications of different put together again, and then smoothed; (iv) detrending of the
methodologies for electricity demand forecasting that a compre- seasonal series; (v) deseasonalizing the original series using the
hensive review of all techniques is virtually impossible in an orig- seasonal component calculated in the previous steps; and (vi)
inal research paper. The interested reader is referred to the smoothing the deseasonalized series to get the trend component.
compiling works of Suganthi and Samuel [7] (thorough overview) In this work, for each STL decomposition of the involved time
and Shao et al. [2] - this last one dedicated solely to decomposition series, we follow Bergmeir et al. [32] and use the R STL function
based approaches to electric load forecasting. A recent, systematic with its default parameters, i.e. d ¼ 1 in steps (iii) and (iv), and d ¼ 0
review over 113 different case studies is presented by Kuster et al. in step (ii).
[8], who identify three main approaches (in terms of popularity): The STL decomposition has some major advantages compared to
Time Series models, Regression-based formulations and Artificial other methods, such as: the possibility to handle any type of sea-
Neural Networks (ANNs). sonality (regardless of the frequency) and to change the seasonal
In order to cope with the difficulties presented by each tech- component over time; the possibility to control the smoothness of
nique in Table 1, different decomposition and combination tech- the trend-cycle; and its robustness to outliers when estimating the
niques have been proposed in recent years. The rationale behind trend-cycle and seasonal components [35]. Our bootstrap aggre-
such methods is that the isolation of key features of the data into gating (bagging) analysis is focused on the third component of the
distinct sub-series can enhance the forecasting performance of the STL decomposition, namely the remainder. In brief terms, we
models used for their estimation. Each sub-series can then be either: directly bootstrap the component by sampling data blocks
modelled according to its stylized facts. That way, the estimation of equal size until the desired series length is achieved; or estimate
error obtained from the further aggregation of the extrapolated specific models to eliminate possible autocorrelations, if any, in the
778 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788

Table 1
Summary of previous studies on electricity demand forecasting (most frequent techniques and not including combinatorial approaches).

Models Energy Sampling Advantages Disadvantages Forecasting for

Type (most frequent)

Time Series Short, mid and Intuitive functional form; Pre-assumed form of the model; Monthly electric energy
long term adaptable; can represent several difficulty in trend extrapolation due consumption in eastern Saudi
types of time series; low to the presence of both a long-term Arabia [9]
computational complexity increasing trend and periodic Monthly electricity consumption of
waves with different frequencies the Asturian (Northern Spain)
and amplitudes in many countries domestic sector [10]
Daily electricity demand load in
Greece [11]
Energy consumption (and GHG
emission) for a pig iron
manufacturing organization in
India [12]
Annual Residential Electricity
Consumption in Brazil [13]
Regression- Mid and long Intuitive functional form; Pre-assumed form of the model; Monthly electricity demand in
based (very long) term straightforward interpretation of data independence assumption; England and Wales (using weather
results; low computational sensitive to outliers variables, gross domestic product,
complexity; Efficient for long and and population growth) [14]
very long prediction Monthly electricity demand in Italy
(using calendar and weather-
related variables) [15]
Monthly European and Italian
electricity demand (using seasonal
climate forecasts of temperature)
[16]
Monthly Australian electricity
demand (industrial, residential and
commercial sectors) using climate
variables [17]
Artificial Neural Short (very short) Provide good estimation in cases Results cannot be easily explained Monthly electricity consumption in
Networks and mid term where data is incomplete; Can as they are not mathematically Iran [18]
(ANNs) address complex nonlinear based; Computation is time Monthly energy demand in the
problems while demonstrating consuming; Extended data is United States (Industrial Sector)
robustness and fault tolerance; required; Model may never [19]
Efficient for short and very short converge in some cases - often get Annual gross electricity demand in
term (sub hourly) prediction stuck in local minimum, so Turkey [20]
choosing a proper dataset is Hourly load of the Polish power
paramount for obtaining quality system and half-hourly loads of the
forecasting results French, British and Victorian
(Australia) power systems [21]
Fuzzy Logic Short, mid and Mathematical-free model to map Lack of self-learning capability Monthly electricity consumption in
long term the input and output variables; Not (usually combined with ANNs to Iran [22]
necessary to have accurate input overcome this issue); Not capable of Hourly electricity load in Jordan
variables; Useful in practical generating a definite prediction [23]
situations where the intensity of a equation based on the input Annual electricity demand in
phenomenon is described by historical data; Proper criteria must Turkey [24]
linguistic variables (very low, low, be defined for accurate forecasting Annual electricity demand in Brazil
medium, high, very high); High (Require more fine tuning and [25]
accuracy achieved on total simulation before going
uncertainty scenarios (eg. economic operational)
crisis periods)
Support Vector Short, mid and Can address complex nonlinear Kernel function is crucial and Annual electricity demand in
Regressions long term problems with relative simplicity difficult to be determined; Real data Taiwan [26]
(SVR) (by creating an optimal separating are not perfectly separable; Sub-hourly (15 min) electricity load
hyperplane in a higher dimensional Computation on high dimension from four large office buildings in
feature space so that subsequent feature space can be very costly; eastern China [27]
observations can be classified into Depend a lot on the proper
separate subsets); Few parameters selection of the hyperparameters
need to be determined

remainder and then simulate different error time series by boot- of interest. Adapted versions of the original technique have been
strapping the residuals. These procedures are described in detail in developed for time series, since their data are typically autocorre-
the next subsections. lated. The Bootstrap aggregation (Bagging), in turn, is a supervised
machine learning technique, proposed by Breiman [37]. The aim is
to generate multiple versions of a predictor via Bootstrap and then
3.2. Bootstrapping the remainder
use the simulated results to get an aggregated predictor. Despite
the good results in the field of Machine Learning, there are just a
The bootstrap was first devised by Efron [36], growing out of an
few papers using Bagging to improve accuracy in time series
earlier work on the jackknife procedure. In its original form, the
forecasting [33]. In most of them, the authors have demonstrated,
technique consisted of re-sampling the underlying data, in order to
majorly with econometric applications, that the combination of
get an approximation of the sampling distribution of some statistic
 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788 779

techniques produces large improvements in forecast accuracy model to the data, we first decompose the original time series into
([38,39]). its trend, seasonal and remainder components, following the STL
One prerequisite for bootstrapping is the stationarity of the approach (Section 3.1). Then, provided that the last component is
series, which is often achieved for the remainder in a STL decom- already stationary, we estimate the best fit Autoregressive-Moving
position. In addition, care must be taken to ensure that every value Average - ARMA(p,q) - model for the remainder, using AIC with
from the original series can be placed anywhere in the boot- corrections (AICc) [44] or the most parsimonious formulation
strapped series. In this work, we deal with this issue for the which ensures that there are no autocorrelation issues in the re-
remainder in two different ways, which are briefly described in the siduals. Finally, we resample the centered residuals and use the
following paragraphs. ARMA(p,q) to obtain new series for the remainder.

3.2.1. Moving blocks bootstrap

The Moving Blocks Bootstrap (MBB) approach was first devised 3.3. Forecasting
by Künsch [40]. In its original formulation, data blocks of equal size
are drawn from the series until the desired series length is ach- After having obtained all the bootstrapped time series using the
ieved. For a series of length n, with a block size of l, n  l þ 1 procedures described in Sections 3.1 and 3.2, we estimate and
(overlapping) possible blocks exist. However, care must be taken to subsequently forecast each generated series using different uni-
ensure that every value from the original series could possibly be variate methods. In the following subsections, we briefly explore
placed anywhere in the bootstrapped series. To meet that end, such methods.
Bergmeir et al. [32] propose drawing ½n=l þ 2 blocks from the
remainder series of a STL Decomposition, and discarding a random 3.3.1. Exponential smoothing methods and state space formulations
number of values, between zero and l  1, from the beginning of the Exponential smoothing consists of traditional procedures for
bootstrapped series. Then, to obtain a series with the same length continually revising a forecast in light of more recent information
as the (original) remainder series, they further discard as many about the estimated data. In brief terms, the methods assign
values as necessary to obtain the required length. This process exponentially decreasing weights as the observation gets older, i.e.,
ensures that the bootstrapped series do not necessarily begin or recent observations are given relatively more weight in forecasting
end on a block boundary. Finally, the trend and seasonality are than older ones.
combined with the bootstrapped remainder to get the final boot- There are several different approaches to exponential smooth-
strapped sample. ing. For an extensive list of the most commonly used exponential
In this work, we follow the same above-mentioned procedure. smoothing methods in the literature, the interested reader is
According to Bergmeir et al. [32], a MBB-based approach has the referred to the compiling works of Gardner Jr. ([45], [46]). In the
advantage that it makes no modeling assumptions other than sta- underlying paper, the total electric energy demand in different
tionarity, whereas other bootstrap methods assume that the fitted countries (and their bootstrapped versions) are estimated and
model on the underlying time series captures all serial correlation subsequently forecasted using three different exponential
in the data. Even though no assumptions have been formally made smoothing methods, briefly described in the following lines.
with regard to the predictive accuracies of the MBB, Bergmeir et al.
[32] showed that the approach, combined with exponential (i) The Holt-Winters additive model [47], appropriate for se-
smoothing methods, delivered good results when dealing with ries with a linear time trend and additive seasonal variation.
monthly time series from the M3-competition, the main dataset to Its component form can be written as follows:
evaluate and compare time series forecasting methods - see
Ref. [41] for further details. It is our belief that equally satisfactory b
y tþk ¼ a þ bk þ stþk (1)
results can be achieved in the electric energy sector.
where a and b are the permanent component (base signal) and the
3.2.2. Remainder Sieve Bootstrap linear trend parameters, respectively, and st are the additive sea-
The sieve bootstrap approach has its roots in the original sonal factors. These parameters, in turn, are defined by the
Bühlmann's [42] methodology, which was based on the idea of following recursive expressions:
fitting parametric models first and afterwards resampling from the
residuals. In brief terms, given a sample X1 ; …; Xn , from a stationary aðtÞ ¼ a ½yt  st ðt  mÞ þ ð1  aÞ ½aðt  1Þ þ bðt  1Þ
process, Bühlmann's [42] method involved: (i) selecting the order bðtÞ ¼ b ½aðtÞ  aðt  1Þ þ ð1  bÞ bðt  1Þ (2)
p ¼ p(n) of an autoregressive (AR) approximation using the Akaike st ðtÞ ¼ g ½yt  aðtÞ þ ð1  gÞ st ðt  mÞ
[43] Information Criterion (AIC); (ii) using the AR(p) model to filter
the residuals series, obtaining centered residuals and their empir- where 0 < a; b; g < 1 are the hyperparameters and m is the seasonal
ical cumulative distribution function; (iii) resampling the (sup- frequency (supposedly monthly). In these terms, forecasts are
posed) i.i.d. centered residuals; (iv) using the AR(p) for obtaining a computed by:
new series Xt* by recursion. Finally, given X1* ; …; XT* , the method
b
y Tþk ¼ aðTÞ þ bðTÞk þ sTþkm (3)
computes the estimation of the AR coefficients and then obtain
future bootstrap observations by recursion from the new series.
where the seasonal factors are used from the last s estimates.
In Cordeiro and Neves [31], a different approach is proposed: the
authors first suggest to fit an EXPOS model to the data and then to
(ii) The Holt-Winters multiplicative model [47], adequate to
proceed like the Bühlmann's [42] method over the residuals. The
series with a linear time trend and multiplicative seasonal
EXPOS model is named after the best fit model from a set of
variation. The smoothed series b
y t is given by:
exponential smoothing forecasting methods. The BOOT.EXPOS
procedure, as the method became known, demonstrated very good b
y tþk ¼ ða þ bkÞ stþk (4)
results in forecasting series with seasonal and trendy components.
In this paper, we propose a distinct version of the above- The three coefficients are now defined by the following
mentioned procedure. In lieu of fitting an exponential smoothing recursions:
780 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788

  which the component estimates are updated using the previous

yt
aðtÞ ¼ a þ ð1  aÞ ½aðt  1Þ þ bðt  1Þ period's estimates and the current prediction error - et . The mul-
st ðt  mÞ tiple functions are a notational device for writing the additive and
bðtÞ ¼ b ½aðtÞ  aðt  1Þ þ ð1  bÞ bðt  1Þ (5) multiplicative errors in compact form. With additive errors, we
  have k≡1, so that yt ¼ h ðxt1 ; qÞ þ et . In short, we may think of the
yt
rt ðtÞ ¼ g þ ð1  gÞ st ðt  mÞ updating smoothing equations as being weighted averages of a
aðtÞ term which depends on the current prediction error (and prior
Forecasts are then computed by: states), and one which depends on the prior states. The resulting
state space equations for an additive error, multiplicative trend and
b
y Tþk ¼ ðaðTÞ þ bðTÞkÞ sTþkm (6) multiplicative season model are:

b
y t ¼ at1 bt1 stm
et
at ¼ at1 bt1 þ a
(iii) State space based (exponential smoothing) formulations. stm
This approach consists of a set of different models obtained et (9)
bt ¼ bt1 þ ab
by considering variations in the combination of the error, ðstm at1 Þ
trend and seasonal components of a time series. According to st ¼ stm þ g et=ða
t1 bt1 Þ
the taxonomy proposed by Pegels [48] and extended by
Gardner Jr. [45], the possibilities for the trend and seasonal For more information regarding the alternatives on state space
components are depicted in Table 2. In addition, the error estimation methods, the interested reader is referred to [50].
term can also vary between additive or multiplicative. That
way, a total of 30 different formulations can be achieved. 3.3.2. Seasonal autoregressive integrated moving average models
(SARIMA)
Each model in a state space based formulation consists of two The SARIMA models, first proposed by Box and Jenkins [51],
sets of equations: (i) a measurement equation that describes the consist of an alternative approach to the traditional exponential
observed data; (ii) and some transition equations that describe how smoothing methods. In brief terms, SARIMA models are similar to
the unobserved components or states (level, trend, seasonal) exponential smoothing methods inasmuch as they are adaptive,
change over time. To express such formulation, let's consider one can model trends and seasonal patterns, and can be automated.
possible combination as an example: an additive error, multipli- They differ, however, in that they are based on autocorrelations
cative trend, multiplicative season model, or AMM, according to the (patterns in time) rather than a structural view of level, trend and
above-mentioned notation. First, we consider a p-dimensional state seasonality. It is argued that SARIMA formulations tend to succeed
vector xt ¼ ðat ; bt ; st ; st1 ; …; stm Þ0, with at and bt being the better than exponential smoothing methods for longer, more stable
contemporaneous estimates of the level and linear trend parame- data sets and not as well for noisier, more volatile data [52].
ters and s representing the included seasonal terms. We also let Non-seasonal ARIMA models are generally denoted by ARI-
b
y t ¼ at1 bt1 stm be the one-period ahead forecast of yt . Then, the MA(p,d,q) where parameters p, d, and q are non-negative integers, p
prediction error decomposition is being the order of the autoregressive model, d the degree of dif-
ferencing, and q the order of the moving-average model. Seasonal
yt ¼ b
y t þ et ¼ at1 bt1 stm þ et (7)
ARIMA models, in turn, are usually denoted by SAR-
Following Ord et al. [49], we may write a nonlinear dynamic IMA(p,d,q)x(P,D,Q)S and can be written as follows:
model representation of the exponential smoothing equations us-    
S V fðBÞ F B
VD Zt ¼ qðBÞ Q BS at
d S
ing a state space model with a common error term: (10)

yt ¼ h ðxt1 ; qÞ þ k ðxt1 ; qÞ et
(8) where:
xt ¼ f ðxt1 ; qÞ þ g ðxt1 ; qÞ et
 S refers to the number of periods in each season;
where h and k are known continuous scalar functions, f and g are
 the uppercase P,D,Q refer to the autoregressive, differencing, and
known continuous functions with continuous derivatives from
moving average terms for the seasonal part of the ARIMA model;
<p /<p and et  iid ð0; s2 Þ are the independent past realizations
 at is the error term;
of y and x.
 B is the backward shift operator (eg. Byt ¼ yt1 );
Conceptually, the yt equation represents how the various state
 fðBÞ and FðBS Þ are the non-seasonal and seasonal autore-
variable components ðat1 ; bt1 ; stm Þ are combined to express the
gressive polynomials, respectively;
series in terms of a smoothed forecast b y t ¼ hðxt1 ; qÞ and the pre-
 qðBÞ and QðBS Þ are the non-seasonal and seasonal moving-
diction error ðet Þ. The xt equations, in turn, outline the process by
average polynomials, respectively;
 Vd and VD S are the non-seasonal and seasonal differencing op-
Table 2 erators, respectively.
Possible variations for the trend and seasonal components under a state space based
approach.
3.4. Aggregation
Trend Component Seasonal Component

None (N) Additive (A) Multiplicative (M) In this last step, we aggregate the forecasts obtained for each
None (N) N, N N, A N, M bootstrapped time series to generate the final output. To that end,
Additive (A) A, N A, A A, M we use two different methods: the simple mean (or equal weights
Additive Damped (Ad) Ad, N Ad, A Ad, M combination) and the simple median. Besides its simplicity, the
Multiplicative (M) M, N M, A M, M simple mean has proved to be a tough benchmark for forecasting
Multiplicative Damped (Md) Md, N Md, A Md, M
combinations, see for example Stock and Watson [53]. The median,
 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788 781

in turn, is less sensitive to outliers, reducing the effects of occa- measures to evaluate the accuracy of the forecasts, such as:
sional poor forecasts. Mean Absolute Percentage Error (MAPE):

Xh   !
Tþ  ybt  yt 
MAPE ¼  
4. Data and overall procedure  y  h  100% (11)
t¼Tþ1 t

Our empirical analysis is based upon monthly data of total Symmetric Mean Absolute Percentage Error (sMAPE):
electric energy consumption (GWh) in different developed and
8 2 3 9
developing countries, namely Canada, France, Italy and Japan > , >
>
< Tþ
Xh 6 b >
=
7
6 j yt  yt  j
(former case) and Brazil, Mexico and Turkey (for the latter). With
sMAPE ¼ 7 h  100% (12)
the exception of Brazil, the data were collected from the Interna- > 4 5 >
>
: t¼Tþ1 j ybt j þ jyt j 2 >
;
tional Energy Agency Monthly Electricity Statistics report, which
provides electricity production and trade data for all OECD Member
Countries [54]. For the Brazilian electric energy consumption, we Root Mean Squared Error (RMSE):
referred to the data provided by the major Brazilian electric utilities vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u Tþ h .
company, Eletrobras, available at the Brazilian Central Bank time u X
RMSE ¼ t
2
series database [55]. The time period of the analysis spans from July ð ybt  yt Þ h (13)
2006 (the first date available for OECD countries) to December t¼Tþ1

2016. We considered data from July 2006 to December 2014 as Theil Inequality Coefficient (TIC):
training set (in-sample) and the observations from January 2015 to
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
December 2016 as test set (out-of-sample). The original data can be Ph
Tþ .
2
found in the supplementary material (Appendix B). ð ybt  yt Þ h
Before proceeding to the methods described in Section 3, a t¼Tþ1
TIC ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (14)
transformed version of each original time series is generated by TþPh 2. TþPh 2
b
yt h þ yt h
means of a Box and Cox [56] transformation. The main rationale
t¼Tþ1 t¼Tþ1
behind this procedure is to stabilize the variance of a time series
and make highly skewed distributions less skewed. In addition, a In the above formulae, ybt is the predicted (forecasted) value
useful feature of the BoxeCox (BC) approach (along with other log whereas yt is the real (observed) value. h, in turn, is the number of
transformations) is that it constrains the forecasts to stay positive forecasting steps ahead.
on the original scale. Details on the BC procedure are not explained
here, however, it is worth noting that, for the transformation 5. Empirical findings
parameter ðlÞ, we follow Bergmeir et al. [32] and restrict it to lie in
the interval [0,1], then use the method of Guerrero [57] to choose its 5.1. Performance gains from bagging
value. Following the BC transformation, each series is decomposed
into three components (trend, seasonal and remainder), using the The empirical results (best highlighted in bold) for the devel-
STL Decomposition. The remainder is then either estimated by oped and developing countries are summarized in Tables 3 and 4,
means of an ARMA(p,q) process or directly bootstrapped using the respectively. A visualization of the forecasts generated by the best
MBB approach (see Subsections 3.2.1 and 3.2.2 for details). For each bagging approaches plotted against the actual values (for each
bootstrap approach, a total of 100 new series are generated. Finally, country) can be seen in Fig. 2. Figures illustrating the time series
the components are added together again, and the BC trans- generated using the procedures described in Sections 3.1 and 3.2
formation is inverted. The overall procedures are illustrated in the and their corresponding forecasts (obtained using the methods
flowchart of Fig. 1. addressed in Sections 3.3 and 3.4), along with other results, can be
In the second step of the analysis, several models are proposed found in the supplementary material (Appendix C).
to estimate and subsequently forecast the original and boot- Overall, the results indicate that the bootstrap aggregation
strapped versions of the total electric energy consumption time (bagging) approaches can substantially improve the forecast ac-
series. In this paper, we restrict our analysis to four different curacy of the demand for energy end-use services in different
methods: countries. With the exception of the Japanese case, where the best
forecast in terms of MAPE and sMAPE was achieved using a single
(i) an auto ARIMA approach. To that end, we use the auto.ar- ETS formulation, the STL and bagging procedures, combined with
ima() function in R, an algorithm which combines unit root traditional forecasting methods, led to considerably superior re-
tests, minimization of the AICc and Maximum Likelihood sults, in terms of accuracy. In several cases the gains were note-
Estimation (MLE) to obtain an ARIMA model; worthy when compared with single forecasts on the real data. For
(ii) a three parameter Holt-Winters additive model; the Italian electricity consumption, for instance, the sMAPE and the
(iii) a three parameter Holt-Winters multiplicative model; and RMSE obtained using a Remainder Sieve Bootstrap (RSB) ETS
(iv) an auto state space exponential smoothing approach. For this approach were almost 30% and 58% lower than the ones obtained in
case, we use the ets() function in R and let it decide which a single ETS forecast on the real data.
Error, Trend and Seasonal (ETS) combination best suits the It is worth noting that, for developed countries, the bagged
data. The selection is primarily based on the minimization of forecasts that used the RSB approach performed better, in terms of
the Akaike Information Criteria with corrections (AICc) [44]. MAPE and sMAPE, than the ones that used the MBB methodology.
In terms of RMSE and TIC, the only case where the MBB method-
Finally, the forecasts are combined using the aggregation pro- ology outperformed the RSB was for the French monthly electricity
cedures described in Section 3.4 and the predictive power is consumption. Even so, the difference between the error metrics
assessed by means of an out-of-sample experiment from January was not too significant. As for the developing countries, the MBB
2015 to December 2016. To that end, we make use of certain approach provided slightly better results on two of the three
782 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788

Fig. 1. Overall procedure - Flowchart.

Source: Authors' own.

involved countries (Mexico and Turkey). This is a substantial obtained using the simple median aggregation approach were
improvement over previous bagging methods, as the MBB considerably superior to the ones obtained by pooling the forecasts
approach proposed by Bergmeir et al. [32] has been regarded as a using equal weights (simple mean). Considering that the median is
benchmark for forecasting monthly data. less sensitive to outliers, this suggests that the outliers (whether in
Another interesting feature is the fact that the mean and the the original or the generated time series) exert a considerable effect
median of the forecasted values differed considerably in nearly on the overall results.
every occasion (the Brazilian case is the sole exception). The results There was no consensus concerning the superiority of the
 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788 783

Table 3
Forecast evaluation e developed countries (best results highlighted in bold).

Forecast Approach Statistic MAPE SMAPE RMSE TIC MAPE SMAPE RMSE TIC

(%) (%) (GWh) (%) (%) (GWh)

Canada France
MBB.Arima Mean 4.271 4.187 2275.225 0.024 4.056 4.067 1981.013 0.024
Median 3.881 3.958 1960.548 0.021 3.684 3.617 1442.878 0.019
RSB.Arima Mean 4.986 4.899 2629.646 0.027 5.116 5.132 2497.840 0.031
Median 5.004 5.004 2503.508 0.027 4.892 4.903 1784.197 0.023
Auto Arima Single 5.140 5.050 2718.240 0.028 5.946 6.014 2865.827 0.036

MBB.Add.H-W Mean 4.023 3.966 2146.523 0.022 4.157 4.239 2012.938 0.025
Median 3.433 3.424 1748.092 0.019 3.520 3.583 1618.959 0.021
RSB.Add.H-W Mean 3.911 3.817 2251.833 0.023 5.248 5.388 2413.163 0.030
Median 3.174 3.176 1589.605 0.017 4.320 4.416 2019.551 0.027
Add H-W Single 4.206 4.078 2474.736 0.026 5.634 5.792 2463.391 0.031

MBB.Mult.H-W Mean 4.011 3.948 2149.314 0.022 3.807 3.880 1944.163 0.024
Median 3.294 3.347 1743.229 0.019 3.722 3.793 1582.883 0.021
RSB.Mult.H-W Mean 3.884 3.793 2239.347 0.023 4.834 4.959 2351.984 0.029
Median 3.296 3.253 1575.038 0.017 4.558 4.664 1722.974 0.023
Mult H-W Single 4.141 4.023 2427.717 0.025 5.033 5.158 2319.074 0.029

MBB.ETS Mean 3.855 3.787 2141.358 0.022 2.781 2.784 1663.936 0.020
Median 3.385 3.389 1681.856 0.018 2.098 2.100 815.629 0.011
RSB.ETS Mean 4.234 4.152 2306.102 0.024 2.994 3.004 1785.948 0.022
Median 3.879 3.956 1794.504 0.019 1.955 1.954 867.615 0.011
Auto ETS Single 4.040 3.944 2268.954 0.023 2.489 2.479 1534.811 0.019

Italy Japan
MBB.Arima Mean 2.533 2.595 1024.809 0.020 3.494 3.570 3507.538 0.022
Median 2.527 2.559 625.712 0.012 3.426 3.486 2957.626 0.018
RSB.Arima Mean 2.562 2.619 1243.208 0.024 3.575 3.645 3516.507 0.022
Median 1.455 1.456 377.912 0.007 3.585 3.609 2922.618 0.018
Auto Arima Single 3.314 3.407 1221.386 0.024 3.229 3.288 3267.591 0.020

MBB.Add.H-W Mean 2.502 2.556 917.959 0.018 3.494 3.570 3507.538 0.022
Median 2.409 2.439 607.370 0.012 3.426 3.486 2957.626 0.018
RSB.Add.H-W Mean 2.126 2.163 848.341 0.016 3.575 3.645 3516.507 0.022
Median 1.583 1.583 402.300 0.008 3.585 3.609 2922.618 0.018
Add H-W Single 1.904 1.943 803.252 0.015 3.229 3.288 3267.591 0.020

MBB.Mult.H-W Mean 2.458 2.512 928.555 0.018 3.909 3.994 3638.494 0.023
Median 2.317 2.344 635.191 0.012 3.997 4.058 3053.883 0.019
RSB.Mult.H-W Mean 2.012 2.049 856.750 0.016 5.624 5.805 5008.852 0.031
Median 1.419 1.409 371.320 0.007 6.128 6.322 5071.051 0.032
Mult H-W Single 1.829 1.868 818.579 0.016 3.735 3.770 3345.802 0.021

MBB.ETS Mean 1.745 1.773 755.870 0.015 3.526 3.588 3255.967 0.020
Median 1.609 1.596 402.184 0.008 3.594 3.660 2818.551 0.017
RSB.ETS Mean 1.855 1.860 748.582 0.014 3.711 3.781 3490.502 0.022
Median 1.305 1.296 327.862 0.006 3.044 3.057 2635.629 0.016
Auto ETS Single 1.806 1.838 768.508 0.015 2.274 2.233 2687.012 0.016

forecasting methods proposed in Section 3.3. The (auto) ARIMA This is mainly due to the fact that independent variables also
approach seems to perform better for the Brazilian and Mexican need to be forecasted (first) so their predicted outcomes can
cases, whereas the exponential smoothing methods adapted well then be used to forecast the dependent variable. It should be
for the monthly consumption in developed countries. highlighted that several steps ahead forecasts are not to be
confounded with long run forecasts (such as annual, quin-
5.2. Comparison with other methods quennial or even larger horizons), which usually deals with a
small number of steps ahead (up to 12 steps, in the vast majority
In what follows, we compare our results with forecasts from of cases);
other univariate methods established in the literature. Care was - Second, the selection of the most appropriate independent
taken to choose models that deal with different stylized facts in variables is linked to a huge variety of criteria, being unfeasible
electricity demand time series, such as nonlinearities, stochastic to address each case (country) in full particularity.
components (trend, seasonality, residuals), heteroscedasticity,
among others. In addition, it should be noted that the choice for Particularly, we employ the following univariate methods for
other univariate methodologies in lieu of regression-based forecasting comparison:
methods is due to two main reasons:
- a feedforward Artificial Neural Network (ANN) model [58,59], to
- First, regression-based methods fail to perform well when address complex nonlinear behavior;
forecasting several steps ahead (in our case, 24 months ahead).
784 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788

Table 4
Forecast evaluation e developing countries (best results highlighted in bold).

Forecast Approach Statistic MAPE SMAPE RMSE TIC MAPE SMAPE RMSE TIC

(%) (%) (GWh) (%) (%) (GWh)

Brazil Mexico
MBB.Arima Mean 4.724 4.603 1933.379 0.025 3.531 3.585 1091.476 0.023
Median 4.627 4.522 1784.111 0.023 3.041 3.046 680.274 0.014
RSB.Arima Mean 4.368 4.264 1805.707 0.023 3.503 3.559 1100.538 0.023
Median 4.359 4.266 1675.724 0.021 3.280 3.276 750.486 0.016
Auto Arima Single 4.677 4.550 1943.011 0.025 3.092 3.122 968.572 0.020

MBB.Add.H-W Mean 6.789 6.536 2803.963 0.035 4.554 4.674 1375.443 0.029
Median 6.700 6.483 2574.868 0.032 4.241 4.333 1083.106 0.023
RSB.Add.H-W Mean 6.447 6.222 2658.016 0.034 4.947 5.058 1446.868 0.030
Median 6.250 6.061 2383.104 0.030 4.257 4.350 1106.579 0.023
Add H-W Single 7.170 6.884 2961.887 0.037 5.128 5.298 1558.574 0.033

MBB.Mult.H-W Mean 6.647 6.405 2745.442 0.035 4.608 4.728 1364.842 0.029
Median 6.588 6.378 2495.533 0.031 4.156 4.162 995.764 0.021
RSB.Mult.H-W Mean 6.381 6.162 2629.106 0.033 4.566 4.657 1345.790 0.028
Median 6.122 5.940 2335.423 0.030 3.832 3.908 952.716 0.020
Mult H-W Single 7.180 6.891 2973.818 0.037 4.779 4.911 1428.301 0.030

MBB.ETS Mean 6.471 6.242 2661.628 0.034 6.192 6.441 1780.398 0.038
Median 6.570 6.361 2502.552 0.032 6.086 6.278 1442.600 0.031
RSB.ETS Mean 6.411 6.188 2649.286 0.033 6.341 6.610 1853.388 0.040
Median 6.195 6.009 2366.104 0.030 6.046 6.234 1463.423 0.031
Auto ETS Single 7.214 6.927 2965.903 0.037 6.921 7.228 1953.420 0.042

Turkey
MBB.Arima Mean 2.644 2.632 712.075 0.016
Median 2.151 2.138 490.369 0.012
RSB.Arima Mean 2.744 2.729 724.887 0.017
Median 2.507 2.511 556.709 0.013
Auto Arima Single 2.277 2.279 681.329 0.016

MBB.Add.H-W Mean 3.079 3.038 755.223 0.017

Median 2.756 2.718 595.679 0.014
RSB.Add.H-W Mean 3.326 3.275 807.492 0.018
Median 2.993 2.949 642.197 0.015
Add H-W Single 2.623 2.594 707.484 0.016

MBB.Mult.H-W Mean 2.740 2.701 707.057 0.016

Median 2.035 2.015 452.767 0.011
RSB.Mult.H-W Mean 2.995 2.938 795.979 0.018
Median 2.392 2.364 495.507 0.012
Mult H-W Single 2.421 2.383 698.838 0.016

MBB.ETS Mean 2.512 2.527 724.285 0.017

Median 2.489 2.459 547.975 0.013
RSB.ETS Mean 2.686 2.728 830.497 0.019
Median 2.224 2.224 460.625 0.011
Auto ETS Single 2.913 2.984 932.234 0.022

- a feedforward Artificial Neural Network model with prior Box- seeks to filter the noise and forecast the signal of an underlying
Cox transformation (BC-ANN), in an attempt to ensure that re- time series using multiple steps (Embedding, Singular Value
siduals will be roughly homoscedastic [56]; Decomposition, Grouping and Diagonal Averaging). In this work,
- a univariate Support Vector Regression (SVR), an advanced we employ both the Recurrent SSA (RSSA) and the Vector SSA
machine learning algorithm, able to learn from training data and (VSSA) variations [63].
form complex non-linear decision boundaries [60]. To select the
best subset of variables for prediction (in our case, the lagged The results obtained using the above-mentioned methods are
values of the electricity demand with the most predictive po- depicted in Tables 5 and 6, for the developed and developing
wer), the Correlation-based feature selection (CFS) algorithm countries, respectively. To conserve space, references and expla-
[61] was used in each country's training set; nations on the alternative methods are provided in the supple-
- the Theta method [62], a technique equivalent to a simple mentary material (Appendix A).
exponential smoothing with drift (with a particular restriction The results outlined in the above Tables endorse the superiority
for this last component). The technique has performed partic- of the proposed bagging methodologies in terms of more accurate
ularly well in the M3-competition [41] for monthly series and forecasts. The Japanese case remains the only exception, but now
for microeconomic data; results are not conclusive in terms of the best forecasting technique
- two variations of the univariate Singular Spectrum Analysis for the 2015e2016 period. The Single ETS approach performed
(SSA) technique - a decomposition-reconstruction method that better in terms of MAPE and sMAPE for the Japanese electricity
 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788 785

Fig. 2. Electricity demand by country (best forecasts in red, actual values in blue). (For interpretation of the references to colour in this figure legend, the reader is referred to the
Web version of this article.)
Source: Authors' own.

demand, whilst the Theta forecasts were slightly more accurate in For the Brazilian case, for instance, the Industrial Sector accounted
terms of RMSE and TIC. for almost 43% of the total electric energy consumption between
the years of 2006 and 2014 (training period). Another point worth
noting is the important role that the Gross Domestic Product (GDP)
6. Discussion plays in electric energy consumption behavior ([24, 64]). By quickly
glancing the GDP data in Brazil, one may notice substantial falls in
The performance gains demonstrated by the STL þ bagging its common trend, reflecting the recent political and economic
approaches are remarkable as accurate forecasts are decisive for turmoil in the country. Along with the energy rationing and the
assertive profit/cost management and investment decisions, as well lower industrial output, this might have been an important factor
as for the definition of sectoral policies in a local or national scale. for the substantial decline in the demand for energy end-use ser-
For the energy sector, particularly, precise mid/long-term demand vices in Brazil in the last years.
forecasting is of the utmost importance for several decision-making Notwithstanding the above, as previously mentioned, formula-
processes, such as the construction of new generation facilities, the tions that consider external influences on the variable of interest
purchase of existing generating units, the development of trans- usually yield satisfactory results when simulating historical data
mission and distribution systems, among others. In a more general but fail to perform well in forecasting several steps ahead (more
sense, accurate forecast results are also paramount to reach than 12 steps, as in our case). On these grounds, the combination of
agreements between different stakeholders (generators, trans- bagging methodologies and univariate forecast methods emerges
mitters, distributors, traders, consumers, investors, government as a promising alternative to predict mid-/long-term behavior for a
and national regulation institutes). broad variety of time series in different economic sectors.
It should be noted that a considerable amount of the variation in
the monthly electric energy consumption is due to external factors,
which cannot be captured by univariate forecasting methods. Some 7. Conclusions and future directions
remarkable examples are the influences of the electric energy
generation and, particularly, Industrial Output in several countries. In light of the strategic importance of having credible forecasts
786 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788

Table 5
Comparison with other methods e developed countries (best results highlighted in bold).

Forecast Approach MAPE SMAPE RMSE TIC Forecast Approach MAPE SMAPE RMSE TIC

(%) (%) (GWh) (%) (%) (GWh)

Canada France

RSB.Add.H-W - Median 3.174 3.176 1589.605 0.017 RSB.ETS - Median 1.955 1.954 867.615 0.011
RSB.Mult.H-W - Median 3.296 3.253 1575.038 0.017 MBB.ETS - Median 2.098 2.100 815.629 0.011
ANN (1, 1, 2) [12] 4.137 4.078 2334.346 0.024 ANN (2, 1, 2) [12] 3.406 3.440 1957.902 0.024
BC-ANN (1, 1, 2) [12] 4.138 4.085 2307.694 0.024 BC-ANN (1, 1, 2) [12] 3.303 3.370 2002.234 0.025
SVR (6, 20, 27, 84, 96) 8.509 8.550 4977.226 0.052 SVR (8, 96) 8.199 8.074 3881.359 0.048
Thetha 4.137 4.022 2385.858 0.025 Thetha 2.846 2.861 1666.439 0.021
RSSA (35, 33) 4.746 4.697 2661.387 0.028 RSSA (35, 7) 4.413 4.422 1980.674 0.024
VSSA (27, 26) 5.857 5.672 3245.611 0.033 VSSA (34, 11) 4.289 4.322 1897.870 0.023

Italy Japan

RSB.ETS - Median 1.305 1.296 327.862 0.006 RSB.ETS - Median 3.044 3.057 2635.629 0.016
RSB.Mult.H-W - Median 1.419 1.409 371.320 0.007 Single ETS 2.274 2.233 2687.012 0.016
ANN (12, 1, 6) [12] 3.863 3.960 1405.311 0.027 ANN (7, 1, 4) [12] 4.318 4.310 4637.648 0.028
BC-ANN (12, 1, 6) [12] 3.393 3.520 1498.755 0.029 BC-ANN (6, 1, 4) [12] 5.804 5.856 5782.212 0.036
SVR (34, 99) 3.516 3.590 1613.364 0.031 SVR (2, 100) 7.368 7.243 6883.180 0.042
Thetha 2.065 2.107 839.800 0.016 Thetha 2.392 2.370 2615.827 0.016
RSSA (29, 21) 3.078 3.053 1107.605 0.021 RSSA (21, 7) 5.276 5.443 5254.590 0.033
VSSA (30, 16) 2.791 2.866 1127.547 0.022 VSSA (23, 11) 4.424 4.591 4869.373 0.030

Notes: For each country, the first and second rows refer to the two best (most accurate) forecasting methods from Table 3 (developed countries) or 4 (developing countries).
For ANN formulations, the selected model is given in the form ANN (p,P,k) [m], where p is the number of lagged inputs (autoregressive terms), P is the number of autoregressive
terms for the seasonal part of the time series, k is the number of nodes in the hidden layer and m is the seasonal frequency. For SVRs, numbers in parentheses are the lag
variables selected by the CFS algorithm for the training set. For the SSA models, the parameters refer to the window length (L) and the number of eigenvalues/eigentriples (r),
in that order. The selection was made on the basis of the lowest Root Mean Squared Error (RMSE) for the calibration period (24 months before the out-of-sample period), i.e.
the L and r parameters are the same from the model which demonstrated the lowest RMSE when forecasting for the period January 2013eDecember 2014.

Table 6
Comparison with other methods e developing countries (best results highlighted in bold).

Forecast Approach MAPE SMAPE RMSE TIC Forecast Approach MAPE SMAPE RMSE TIC

(%) (%) (GWh) (%) (%) (GWh)

Brazil Mexico

RSB.ARIMA - Mean 4.368 4.264 1805.707 0.023 MBB.ARIMA - Median 3.041 3.046 680.274 0.014
RSB.ARIMA - Median 4.359 4.266 1675.724 0.021 Single ARIMA 3.092 3.122 968.572 0.020
ANN (1, 1, 2) [12] 5.531 5.360 2259.505 0.029 ANN (1, 1, 2) [12] 6.762 7.019 2042.153 0.043
BC-ANN (1, 1, 2) [12] 5.415 5.251 2212.039 0.028 BC-ANN (2, 1, 2) [12] 5.810 5.998 1750.736 0.037
SVR (2, 3, 4, 100) 5.069 4.910 2160.605 0.027 SVR (2, 100) 15.040 16.749 4538.936 0.102
Thetha 5.153 5.012 2086.524 0.026 Thetha 6.545 6.820 1860.834 0.040
RSSA (26, 13) 6.318 6.072 2751.797 0.035 RSSA (35, 7) 3.557 3.567 1080.091 0.022
VSSA (20, 11) 7.033 6.722 3059.962 0.038 VSSA (24, 10) 3.303 3.341 1042.185 0.022

Turkey

MBB.Mult.H-W - Median 2.035 2.015 452.767 0.011

MBB.ARIMA - Median 2.151 2.138 490.369 0.012
ANN (1, 1, 2) [12] 3.604 3.490 1320.487 0.030
BC-ANN (2, 1, 2) [12] 3.387 3.463 1028.813 0.024
SVR (2, 100) 5.723 6.022 1887.737 0.045
Thetha 2.914 2.975 898.173 0.021
RSSA (25, 7) 3.559 3.503 884.791 0.020
VSSA (39, 24) 6.174 5.941 1457.826 0.033

Notes: See Table 5.

of the demand for electric energy in almost every country, the RSB procedure for simulation, forecast and aggregation delivered
continuous search for more accurate methods is key to competitive the best results, especially in terms of MAPE and sMAPE, in most
advantage among different stakeholders. This is especially true cases. This is a substantial contribution to the literature, as the MBB
when it comes to methods which are robust to uncommon events, approach proposed by Bergmeir et al. [32] has been regarded as a
such as economic distress and rationing. benchmark for forecasting monthly data.
This paper applies for the first time, as long as the authors are A promising aspect in estimating future monthly electric energy
aware, a combination of decomposition and Bootstrap aggregating consumption is that much of its variation is dependent on the
(Bagging) techniques in order to enhance univariate forecasts for common patterns which are natural to demand time series - such
monthly electric energy demand across different countries. as trends and seasonality. On these grounds, even though models
Furthermore, a new variation of an existing bagging approach is can sometimes benefit from the inclusion of exogenous economic
proposed, here name after Remainder Sieve Bootstrap (RSB). The variables, the combination of decomposition and bagging meth-
obtained results attest that the proposed methodologies can odologies with univariate methods seems to perform better when
improve forecast accuracy. Moreover, the approaches that used the predicting energy consumption behavior several steps ahead, as the
 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788 787

forecasting of external variables is often regarded as a complex, consumption in Asturias (northern Spain). Energy 1999;24(3):183e98.
https://doi.org/10.1016/S0360-5442(98)00099-1.
sensitive task. The empirical results confirmed this belief, as the
[11] Pappas SS, Ekonomou L, Karamousantas DC, Chatzarakis GE, Katsikas SK,
values obtained for the MAPE and sMAPE were fairly low Liatsis P. Electricity demand loads modeling using autoregressive moving
(approximately 1.3% for the Italian case, for instance) considering average (ARMA) models. Energy 2008;33(9):1353e60. https://doi.org/
the large number of forecast steps proposed (24 months ahead). 10.1016/j.energy.2008.05.008.
[12] Sen P, Roy M, Pal P. Application of ARIMA for forecasting energy consumption
In general, it can be stated that the results from the proposed and GHG emission: a case study of an Indian pig iron manufacturing orga-
approaches were appropriate and provided consistently accurate nization. Energy 2016;116:1031e8. https://doi.org/10.1016/
forecasts for the mid-/long-term electric energy consumption. This j.energy.2016.10.068.
[13] Maçaira PM, Souza RC, Oliveira FLC. Modelling and forecasting the residential
give us confidence to conclude that the bagging methodologies, electricity consumption in Brazil with Pegels exponential smoothing tech-
particularly the MBB and the proposed RSB procedure, are an niques. Procedia Comp Sci 2015;55:328e35. https://doi.org/10.1016/
important forecast tool for the energy industry, despite its gap in j.procs.2015.07.057.
[14] Hor C-L, Watson SJ, Majithia S. Analyzing the impact of weather variables on
related literature. monthly electricity demand. IEEE Trans Power Syst 2005;20(4):2078e85.
Finally, it is worth noting that other methodologies may be https://doi.org/10.1109/TPWRS.2005.857397.
proposed to investigate total electric energy consumption behavior, [15] Apadula F, Bassini A, Elli A, Scapin S. Relationships between meteorological
variables and monthly electricity demand. Appl Energy 2012;98:346e56.
and that the selection of the most appropriate models is linked to a https://doi.org/10.1016/j.apenergy.2012.03.053.
huge variety of criteria. Further studies might benefit, for instance, [16] De Felice M, Alessandri A, Catalano F. Seasonal climate forecasts for medium-
from a top-down disaggregation approach before proceeding to the term electricity demand forecasting. Appl Energy 2015;137:435e44. https://
doi.org/10.1016/j.apenergy.2014.10.030.
methods here developed. For the electric sector, particularly, this
[17] Vu DH, Muttaqi KM, Agalgaonkar AP. A variance inflation factor and backward
means applying the STL þ bagging methods for each subsystem of elimination based robust regression model for forecasting monthly electricity
the total consumption (Industrial, Commercial, Residential and demand using climatic variables. Appl Energy 2015;140:385e94. https://
Other Sectors classes), forecasting the simulated series and then doi.org/10.1016/j.apenergy.2014.12.011.
[18] Azadeh A, Ghaderi SF, Sohrabkhani S. A simulated-based neural network al-
pooling together the results into one single output. Such sector- gorithm for forecasting electrical energy consumption in Iran. Energy Pol
specific studies not only would provide a more in-depth under- 2008;36(7):2637e44. https://doi.org/10.1016/j.enpol.2008.02.035.
standing of the demand for electric energy across different coun- [19] Kialashaki A, Reisel JR. Development and validation of artificial neural
network models of the energy demand in the industrial sector of the United
tries, but could also possibly further enhance the accuracy of States. Energy 2014;76:749e60. https://doi.org/10.1016/
forecasts for total electric energy consumption. The study of other j.energy.2014.08.072.
decomposition, bagging and forecasting methods is also a logical [20] Günay ME. Forecasting annual gross electricity demand by artificial neural
networks using predicted values of socio-economic indicators and climatic
extension of this work. conditions: case of Turkey. Energy Pol 2016;90:92e101. https://doi.org/
10.1016/j.enpol.2015.12.019.
Acknowledgements [21] Dudek G. Neural networks for pattern-based short-term load forecasting: a
comparative study. Neurocomputing 2016;205:64e74. https://doi.org/
10.1016/j.neucom.2016.04.021.
Funding: This work was supported by CAPES; CNPq [304843/ [22] Azadeh A, Saberi M, Seraj O. An integrated fuzzy regression algorithm for
2016-4, 443595/2014-3] and FAPERJ [E-26/202.806/2015]. energy consumption estimation with non-stationary data: a case study of
Iran. Energy 2010;35(6):2351e66. https://doi.org/10.1016/
j.energy.2009.12.023.
Supplementary data [23] Mamlook R, Badran O, Abdulhadi E. A fuzzy inference model for short-term
load forecasting. Energy Pol 2009;37(4):1239e48. https://doi.org/10.1016/
j.enpol.2008.10.051.
Supplementary data related to this article can be found at [24] Kucukali S, Baris K. Turkey's short-term gross annual electricity demand
https://doi.org/10.1016/j.energy.2017.12.049. forecast by fuzzy logic approach. Energy Pol 2010;38(5):2438e45. https://
doi.org/10.1016/j.enpol.2009.12.037.
[25] Torrini FC, Souza RC, Oliveira FLC, Pessanha JFM. Long term electricity con-
References sumption forecast in Brazil: a fuzzy logic approach. Socio Econ Plan Sci
2016;54:18e27. https://doi.org/10.1016/j.seps.2015.12.002.
[1] Al-Hamadi HM, Soliman SA. Long-term/mid-term electric load forecasting [26] Hong WC. Electric load forecasting by support vector model. Appl Math
based on short-term correlation and annual growth. Elec Power Syst Res Modelling 2009;33(5):2444e54. https://doi.org/10.1016/j.apm.2008.07.010.
2005;74(3):353e61. https://doi.org/10.1016/j.epsr.2004.10.015. [27] Chen Y, Xu P, Chu Y, Li W, Wu Y, Ni L, et al. Short-term electrical load fore-
[2] Shao Z, Chao F, Yang S-L, Zhou K-L. A review of the decomposition method- casting using the support vector regression (SVR) model to calculate the de-
ology for extracting and identifying the fluctuation characteristics in elec- mand response baseline for office buildings. Appl Energy 2017;195:659e70.
tricity demand forecasting. Renew Sustain Energy Rev 2017;75:123e36. https://doi.org/10.1016/j.apenergy.2017.03.034.
https://doi.org/10.1016/j.rser.2016.10.056. [28] Theodosiou M. Forecasting monthly and quarterly time series using STL
[3] Dogan E. Are shocks to electricity consumption transitory or permanent? Sub- decomposition. Int J Forecast 2011;27(4):1178e95. https://doi.org/10.1016/
national evidence from Turkey. Utilities Pol 2016;41:77e84. https://doi.org/ j.ijforecast.2010.11.002.
10.1016/j.jup.2016.06.007. [29] Benaouda D, Murtagh F, Starck J-L, Renaud O. Wavelet-based nonlinear
[4] Apadula F, Bassini A, Elli A, Scapin S. Relationships between meteorological multiscale decomposition model for electricity load forecasting. Neuro-
variables and monthly electricity demand. Appl Energy 2012;98:346e56. computing 2006;70(1e3):139e54. https://doi.org/10.1016/
https://doi.org/10.1016/j.apenergy.2012.03.053. j.neucom.2006.04.005.
[5] Castelli M, Vanneschi L, De Felice M. Forecasting short-term electricity con- [30] Qiu X, Ren Y, Suganthan PN, Amaratunga GAJ. Empirical mode decomposition
sumption using a semantics-based genetic programming framework: the based ensemble deep learning for load demand time series forecasting. Appl
South Italy case. Energy Econ 2015;47:37e41. https://doi.org/10.1016/ Soft Comput 2017;54:246e55. https://doi.org/10.1016/j.asoc.2017.01.015.
j.eneco.2014.10.009. [31] Cordeiro C, Neves M. Forecasting time series with BOOT.EXPOS procedure. Rev
[6] Kaboli SHA, Fallahpour A, Selvaraj J, Rahim NA. Long-term electrical energy Stat J 2009;7:135e49. Available at: https://www.ine.pt/revstat/tables.html
consumption formulating and forecasting via optimized gene expression (accessed 13 February 2017).
programming. Energy 2017;126:144e64. https://doi.org/10.1016/ [32] Bergmeir C, Hyndman RJ, Benítez JM. Bagging exponential smoothing
j.energy.2017.03.009. methods using STL decomposition and BoxeCox transformation. Int J Forecast
[7] Suganthi L, Samuel AA. Energy models for demand forecasting d a review. 2016;32:303e12. https://doi.org/10.1016/j.ijforecast.2015.07.002.
Renew Sustain Energy Rev 2012;16(2):1223e40. https://doi.org/10.1016/ [33] Dantas TM, Oliveira FLC, Repolho HMV. Air transportation demand forecast
j.rser.2011.08.014. through Bagging Holt winters methods. J Air Transport Manag 2017;59:
[8] Kuster C, Rezgui Y, Mourshed M. Electrical load forecasting models: a critical 116e23. https://doi.org/10.1016/j.jairtraman.2016.12.006.
systematic review. Sustain Cities Soc 2017;35:257e70. https://doi.org/ [34] Cleveland RB, Cleveland WS, McRae J, Terpenning ISTL. A seasonal-trend
10.1016/j.scs.2017.08.009. decomposition procedure based on loess. J Offic Stat 1990;6:3e73. Available
[9] Abdel-Aal RE, Al-Garni AZ. Forecasting monthly electric energy consumption at: http://www.jos.nu/Articles/abstract.asp?article¼613 (accessed 11
in eastern Saudi Arabia using univariate time-series analysis. Energy February 2017).
1997;22(11):1059e69. https://doi.org/10.1016/S0360-5442(97)00032-7. [35] Hyndman RJ, Athanasopoulos G. Forecasting: principles and practice. first ed.
[10] Chaveza SG, Bernata JX, Coallab HL. Forecasting of energy production and Perth: University of Western Australia; 2013.
788 E.M. de Oliveira, F.L. Cyrino Oliveira / Energy 144 (2018) 776e788

[36] Efron B. Bootstrap methods: another look at the jackknife. Ann Stat 1979;7(1): [50] Hyndman RJ, Koehler AB, Ord JK, Grose S. Forecasting with exponential
1e26. https://doi.org/10.1214/aos/1176344552. smoothing: the state space approach. first ed. Berlin: Springer-Verlag; 2008.
[37] Breiman L. Bagging predictors. Mach Learn 1996;24(2):123e40. https:// [51] Box GEP, Jenkins GM. Time series analysis: forecasting and control. first ed.
doi.org/10.1007/BF00058655. San Francisco: Holden Day; 1970.
[38] Hillebrand E, Medeiros MC. The benefits of bagging for forecast models of [52] Makridakis S, Andersen A, Carbone R, Fildes R, Hibon M, Lewandowski R, et al.
realized volatility. Econometric Rev 2010;29(5e6):571e93. https://doi.org/ The forecasting accuracy of major time series methods. first ed. New York:
10.1080/07474938.2010.481554. John Wiley & Sons; 1984.
[39] Zontul M, Aydin F, Dogan G, Sener S, Kaynar O. Wind speed forecasting using [53] Stock JH, Watson MW. Combination forecasts of output growth in a seven-
reptree and bagging methods in Kirklareli-Turkey. J Theor Appl Info Tech country data set. J Forecast 2004;23(6):405e30. https://doi.org/10.1002/
2013;56(1):17e29. Available at: www.jatit.org/volumes/Vol56No1/ for.928.
3Vol56No1.pdf (accessed 20 March 2017). [54] IEA (International Energy Agency) [dataset]. Monthly electricity statistics.
[40] Künsch HR. The jackknife and the bootstrap for general stationary observa- 2017 Feb 1 [cited 2017 Feb 4]. February 1, 2017. Available from: http://www.
tions. Ann Stat 1989;17(3):1217e41. https://doi.org/10.1214/aos/ iea.org/statistics/monthlystatistics/monthlyelectricitystatistics/.
1176347265. [55] Eletrobras (Centrais Ele tricas Brasileiras S.A.) [dataset]. Consumo de energia
[41] Makridakis S, Hibon M. The M3-Competition: results, conclusions and im- eletrica por classe de consumidor e por regia ~o. 2017 Feb 1 [cited 2017 Feb 4].
plications. Int J Forecast 2000;16(4):451e76. https://doi.org/10.1016/S0169- February 1, 2017. Available from: https://www3.bcb.gov.br/sgspub/.
2070(00)00057-1. [56] Box GEP, Cox DR. An analysis of transformations (with discussion). J Royal Stat
[42] Bühlmann P. Sieve bootstrap for time series. Bernoulli 1997;3(2):123e48. Soc: Series B 1964;26:211e52. http://www.jstor.org/stable/2984418.
https://doi.org/10.2307/3318584. [57] Guerrero V. Time-series analysis supported by power transformations.
[43] Akaike H. A new look at the statistical model identification. IEEE Trans J Forecast 1993;12(1):37e48. https://doi.org/10.1002/for.3980120104.
Automat Contr 1974;19(6):716e23. https://doi.org/10.1109/ [58] Rumelhart DE, Hinton GE, Williams RJ. Learning internal representations by
TAC.1974.1100705. error propagation (No. ICS-8506). San Diego: University of California; 1985.
[44] Sugiura N. Further analysis of the data by Akaike's information criterion and [59] Auer P, Burgsteiner H, Maass W. A learning rule for very simple universal
the finite corrections. Commun Stat Theor Meth 1978;7(1):13e26. https:// approximations consisting of a single layer of perceptrons. Neural Network
doi.org/10.1080/03610927808827599. 2008;21(5):786e95. https://doi.org/10.1016/j.neunet.2007.12.036.
[45] E.S. Gardner Jr. Exponential smoothing: The state of the art. J Forecast [60] Smola AJ, Scho € lkopf B. A tutorial on support vector regression. Stat Comput
198;4(1):1e28. doi: 10.1002/for.3980040103. 2004;14(3):199e222. https://doi.org/10.1023/B: STCO.0000035301.49549.88.
[46] Gardner Jr ES. Exponential smoothing: the state of the artdPart II. Int J [61] Hall M. Correlation-based feature selection for machine learning. Hamilton
Forecast 2006;22(4):637e66. https://doi.org/10.1016/j.ijforecast.2006.03.005. (NZ): Waikato University; 1999.
[47] Winters PR. Forecasting sales by exponentially weighted moving averages. [62] Assimakopoulos V, Nikolopoulos K. The theta model: a decomposition
Manag Sci 1960;6:324e42. https://doi.org/10.1287/mnsc.6.3.324. approach to forecasting. Int J Forecast 2000;16(4):521e30. https://doi.org/
[48] Pegels CC. Exponential forecasting: some new variations. Manag Sci 1969;12: 10.1016/S0169-2070(00)00066-2a.
311e5. Available at: http://www.jstor.org/stable/2628137 (accessed 10 [63] Golyandina N, Nekrutkin V, Zhigljavsky A. Analysis of time series structure:
February 2017). SSA and related techniques. Boca Raton: Chapman & Hall/CRC; 2001.
[49] Ord JK, Koehler AB, Snyder RD. Estimation and prediction for a class of dy- [64] Burke PJ, Csereklyei Z. Understanding the energy-GDP elasticity: a sectoral
namic nonlinear statistical models. J Am Stat Assoc 1997;92(440):1621e9. approach. Energy Econ 2016;58:199e210. https://doi.org/10.1016/
https://doi.org/10.1080/01621459.1997.10473684. j.eneco.2016.07.004.