Exogenous Data for Load Forecasting: A Review
Ram
´
on Christen
1
, Luca Mazzola
1 a
, Alexander Denzler
1
and Edy Portmann
2 b
1
Information Technology, HSLU - Lucerne University of Applied Sciences and Arts, Suurstoffi 1, 6343 Rotkreuz, Switzerland
2
Human-IST Institute, University of Fribourg, Bd de P
´
erolles 90, 1700 Fribourg, Switzerland
Keywords:
Smart Grid, Energy Prediction, STLF, Feature Selection, Exogenous Data Analysis.
Abstract:
Electrical power load forecasting defines strategies for utilities, power producers and individuals that par-
ticipate in a smart grid. While it is well established in planning processes for production and utilities, the
importance of accurate forecasting increases for individuals. The ongoing deregulation of the electricity mar-
ket enables energy trading by individuals, requiring an accurate estimation of the production and consumption.
Research on forecast for aggregated demand shows that including features for the forecast from sources, called
exogenous, additional to the purely historical consumption data allows to obtain higher accuracy. In fact, their
usage demonstrated to be able to explain the large variability observed in the power demand, taking into ac-
count the individual influences. Anyway, the influence of exogenous data is hardly investigated for individual
forecasting, due to the minor prevalence of this analysis to date. This review shows the benefit of exogenous
data usage and the necessity of detailed research on the input features and their influence on detailed, individ-
ual level, forecasts of power demand. Eventually, this contribution is concluded by the presentation of open
issues and research directions for electric smart communities that the authors would like to address.
1 INTRODUCTION
Electrical power is hardly storable. In fact, the en-
ergy production depends directly on the demand. The
higher the demand, the more power must be produced.
That means, any change in the demand has an imme-
diate impact on the grid stability. Therefore, ensuring
a stable grid requires a regulation of the production.
On a producer level, power plants and grid operators
count on a serious planning in order to guarantee a
permanent power demand coverage. They expect to
know the future demand for maintaining short-term
grid balances and for planning grid extensions based
on long-term power demand.
In contrast, on consumer level, the ongoing dereg-
ulation of the energy market allows direct access to
the electricity market. Consumers with own power
production, so called prosumers, become able to share
their own productions with peers and trade energy on
local markets (Mazzola et al., 2020). This drives the
development of smart grids with the intention, to use
the power as close as possible to the source. Electrical
energy should be used more efficiently, close to pro-
duction and with less grid usage, thus incurring in less
a
https://orcid.org/0000-0002-6747-1021
b
https://orcid.org/0000-0001-6448-1139
infrastructure charges. However, a change to locally
produced sustainable energy increases the volatility
in the power grid and makes it hard to control with
respect to stability. Regarding this issue, it is again
essential for power plants to have an accurate esti-
mation of the power demand in advance, in order to
react on the volatile and fast changing demand. For
prosumers, on the other side, it is equally worthwhile
to estimate the expected power consumption and self-
production capabilities. In fact, this information al-
lows balancing two conflicting objectives: optimal
price control on network level with maximised self-
sufficiency for prosumer.
The end-user’s consumption itself is affected by
many external factors. Among others, weather condi-
tions and consumer’s behaviour and activities, define
the circumstances for the consumption. On the other
side, sustainable power production almost completely
depends on weather conditions, such as wind and sun
irradiation. Many studies evidence this dependency
of the power consumption and production on exter-
nal factors; the dependency on so called exogenous
variables. Therefore, load forecasting often include
data from exogenous variables in addition to histori-
cal load data. This additional data seems to provide
promising information for increasing forecast accu-
racy. Furthermore, the enhancement of everyday ob-
Christen, R., Mazzola, L., Denzler, A. and Portmann, E.
Exogenous Data for Load Forecasting: A Review.
DOI: 10.5220/0010213204890500
In Proceedings of the 12th International Joint Conference on Computational Intelligence (IJCCI 2020), pages 489-500
ISBN: 978-989-758-475-6
Copyright
c
2020 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
489
jects by electrification, the consequent growing diffu-
sion of consumer grade Internet of Things (IoT) de-
vices and the stream of information through social
networking services open up new possible sources for
exogenous data.
Load forecasting is an activity with differing chal-
lenges. Parameters such as the forecast time window
or the aggregation level can completely change the
focus and imply different challenges. That also ap-
plies for exogenous data as this information is used to
increase the forecast accuracy. Its value for contribut-
ing on better results strongly depends on the forecast
focus. Therefore, it is of paramount importance to
precisely know the level of support provided by each
variable for load forecasting. Although the influence
of several well studied exogenous variables on the
power consumption is known, the real increase for
forecast quality and certainty is broadly unknown.
This paper gives a review of previous work on
load forecasting with focus on the use of exogenous
variables for predicting power demand. The review
is based on literature searches run between March,
16th and April, 3rd 2020 on both IEEE Xplore and
Google Scholar libraries. While IEEE Xplore pro-
vides publications in computer science and engineer-
ing, Google Scholar has been considered for covering
a wide area of scientific publications. The literature
research comprises the terms: Short-Term Load Fore-
cast (STLF), power prediction, load forecasting, ex-
ogenous variables / features / data, feature selection
and social media. It considers more than 50 publi-
cations, with a focus on recent trends. In fact, about
85% appeared in 2010 or later (half of which in the
last 3 years). The analysis shows findings in: a) the
correlation of exogenous data to the power demand,
b) considered data in load forecasting and c) feature
selection approaches for accuracy improvements. In
addition, this paper gives an overview of the most
common applied forecasting methods and points out
the demand of an in-depth analysis on the use of ex-
ogenous data for increasing the prediction accuracy.
The remainder of the paper is organised as fol-
lows: Section 2 presents the use of exogenous data,
by disclosing the impact and advantage of consider-
ing it in load forecasting. Following, Section 3 re-
veals the broad variation of used exogenous data in
the analysed approaches. A grouping and mapping of
the various variables with respect to the application
incidence directs the focus of the research. Section 4
highlights, for various parameters, their information
value in power utilisation forecasting. Additionally,
Section 5 and Section 6 review common feature se-
lection practices. For covering all characteristics of
these variables in load estimation, this part also in-
clude a discussion of methodologies and issues ex-
isting. Based on this review, Section 7 discloses a
revealed knowledge gap about the true information
value of exogenous data in load forecasting, before
our conclusion (Section 8) terminate this contribution.
2 IMPACT OF EXOGENOUS
DATA IN LOAD FORECASTING
This section answers the question about why exoge-
nous data is used in load forecasting. Predicting the
energy consumption is an old endeavor (Gross and
Galiana, 1987). Whether for long-term decisions in
respect to grid assets or short-term load balance es-
timations in modern smart grid approaches, the pre-
diction always needs to be as accurate as possible
with respect to its energy consumption. Balancing
the production and demand of power or trading en-
ergy on a deregulated market among others, require
a high agility of power producers and storage assets
to guarantee at the same time grid stability. Because
of this requirement, both may profit from an accu-
rate and reliable prediction. In an initial effort, ex-
perts purely estimated the power demand for the next
few hours, days or weeks based on the historical be-
haviour. This simple estimation features a high un-
certainty as it assesses the prospective behaviour on
history only, without considering any context to it.
Certainly, it suffices for rough decisions but not for
issues with a finer granularity or a higher complexity,
such as balancing significant and quick load changes.
The power demand as well as the production from
sustainable sources depend on higher-level circum-
stances. Power generation units such as Photo Voltaic
(PV), for instance, produce energy depending on the
intensity of solar irradiation. Similarly, the power
demand depends on the operation of electrical loads
which is driven by external factors. In fact, users de-
fine the consumption profile by turning on an off their
devices, under the needs determined by higher-level
circumstances. Accordingly, researchers try to use in-
formation of this external factors, so called exogenous
data, to improve the load forecasting quality. The cor-
relation between the climate and the power consump-
tion have already been discussed half a century ago
(Heinemann et al., 1966). And the results of many
dependency studies demonstrate a positive correlation
with the proposed forecasting approaches, when us-
ing exogenous data in load forecasting.
Using exogenous data as additional source for in-
put data in load forecasting allows for the extrac-
tion of context related features. This features provide
information from higher-level circumstances that di-
CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage
490
KWh
I
E1(t+x)
I
E1(t)
I
E3(t-x)
I
E4(t-x)
I
LD(t-x)
Historical
values
Instantanous
values
Forecast
values
Exogenous Data
Load Data
Input Data
Feature Extraction Feature Selection
Forecast
f
E1..4
, f
LD
f
E1
, f
E2
, f
LD
Feedback input for selection
Load
Forecast
Figure 1: Load forecast feature sources.
rectly influence the power time series. As depicted
in Figure 1 the information from exogenous data can
originate from recordings, current measurements or
forecasts. The extracted features from the exogenous
data can be used together with the information from
power load data for building a pool of features which
describe the target power time series. Alternatively,
exogenous data can also be used decoupled and there-
fore without load data as shown in (Kandil et al.,
2006) where they evaluated an approach for missing
historic load data. In a typical application, the ex-
tracted features pass a selection procedure that pick
a few features with high relevance which are subse-
quently forwarded to the input of the forecast algo-
rithm. For a refinement, the parametrisation of the
feature extraction and selection methods can also con-
sider the output of the load forecast in a feedback
loop.
Exogenous data can be seen as meta-data of the
power load. They provide information about higher-
level circumstances that affect the consumption as
well as the production. As in (L
´
opez et al., 2017) and
(Janicki, 2017), the literature shows that considering
this additional information can improve the load fore-
cast quality. However, this also shows the high com-
plexity of the dependency of power load on various
influencing variables. There is no single but also not
a fixed set of variables that completely describe the
consumption nor the production. The broad variation
of exogenous data and their impact on load forecast-
ing is discussed in the following sections.
3 EXOGENOUS DATA IN LOAD
FORECASTING:
CHARACTERIZATION
In this section, an exploration on which exogenous
data are used in load forecasting is provided. En-
ergy consumption forecast mainly extracts key infor-
mation from recordings of the target variable to pre-
dict the future demand. These recordings are scanned
for describing key values under the assumption that
time series similarly continue as in history. The ex-
tracted characteristics that accurately define history
time series such as frequencies, wavelet components
or patterns that follow a typical structure, serve as in-
put variables for the forecast as in (Chen et al., 2008;
Zheng et al., 2017; Silva et al., 2017; Jiang et al.,
2017). Some approaches, such as Rana and Koprin-
ska (Rana and Koprinska, 2012b; Rana and Koprin-
ska, 2013), try to fully include this source for bet-
ter forecast results. In this work, they claimed to
achieve better forecast results by means of a shift in-
variant transformation of the frequency components
in the exogenous data. However, despite the high in-
formative value of the history data, there is a possible
valuable improvement in forecasting accuracy of the
power load by extending the input variables with ex-
ogenous data.
The literature review revealed a wide range of ad-
ditionally included exogenous data. More than 50 dif-
ferent variables could be identified from 48 analysed
publications. By clustering them into typology-based
general categories, we were able to identify the fol-
lowing groups:
Weather data (humidity, precipitation, temperature,
wind-chill, etc.)
Calendar data (date, events, moving holidays, sum-
mer break, etc.)
Day information (before/after holiday, (non-) work-
ing day, weekday, etc.)
Socio-economic (economic trends, gdp, # of employ-
ments, etc.)
Demographic information (birth rate, dwelling
count, population, etc.)
Others (no. of sensors, occupants, devices, etc.)
Anyway, our analysis discovered a high variation
in the usage of these factors. Counting the occurrence
Exogenous Data for Load Forecasting: A Review
491
of all variables in the analysed forecast approaches
reveals large differences in the consideration received
by the higher-level categories. With a significant gap,
weather, calendar and day information are clearly the
default choices. 50% and more of the approaches in-
clude data from these groups. In contrast, only about
10% or less use socio-economic data, demographic
information or other data to enrich the input variables.
Table 1 shows the inclusion of exogenous data regard-
ing the different higher-level groups in a descending
order. It emphasises the large gap of favoured addi-
tional data sources to the minor ones.
Table 1: Inclusion of exo. data in load forecast.
Weather data 63%
Calendar data 55%
Day information 53%
Socio-economic data 8%
Demographic information 5%
Other data 5%
A breakdown of the variables in weather data
presents four main clusters: (i) temperature related
data with a usage proportion of approx. 37%. It ap-
pears to be the most relevant weather variable and in-
cludes any temperature data such as the air tempera-
ture, dry- and wet-bulb temperatures or the wind-chill
index. Followed by (ii) humidity and (iii) wind-speed
information that equally share about 20%. The sky
coverage (iv) has still a proportion of 12%. The re-
maining 30% comprises rarely used variables such as
precipitation or air pressure information.
A significant parameter in load forecast is the fore-
cast time window. Due to different problem repre-
sentations depending on the time window, the liter-
ature separate the forecast time windows predomi-
nantly in four time section: Very-Short-Term Load
Forecast (VSTLF), STLF, Medium-Term Load Fore-
cast (MTLF) and Long-Term Load Forecast (LTLF).
However, the time span definition for each of them
is differently designated in literature, as shown in
Table 2. This review integrates the VSTLF into
the group STLF because of the minimal differences
shown in terms of exogenous data usage between
those two categories.
Table 3 to 5 provide an overview of the distribu-
tion of three forecast key values and the inclusion of
exogenous data accordingly. Table 3 divides the fore-
cast time window in STLF, MTLF and LTLF. The
forecast resolution and the aggregation level are com-
pared in table 4 and table 5 respectively. In all ta-
bles, the first column represents the proportion of the
variables of all evaluated approaches. The columns
after the double line separation represent the distribu-
tion of included exogenous information for each vari-
able. The review focus on the use of exogenous data
in power demand forecasting. Approaches for other
variable forecasts or without exogenous data are out
of this scope. Hence, a change of the focus may yield
different results in statistics.
In Table 3 it is obvious that researcher pay more
attention to STLF when considering exogenous data
as supporting input data. Mid- and long-term fore-
cast have only a proportion of 10% and 5% respec-
tively. However, decoupled from the forecast time
window, historic load data always seems to be a key
input source for load forecast. For all time window,
they have a prominent proportion of approx. 30% of
the considered input data. In contrast, only long-term
forecasts use socio-economic and demographic infor-
mation seriously as additional input variables. Mid-
and short-term forecast time window rather consider
weather and day information whereas calendar data
seems to provide valuable information for all time
horizons.
According to the mainly provided forecast reso-
lution in literature (see Table 4), it is separated in
hourly, daily or yearly forecasts. The forecast resolu-
tion represents the time span of a single forecast value
comprising either an instant demand or a peak value.
Comparing Tables 3 and 4, the proportion of the fore-
cast resolution shows a similar distribution as of the
forecast time windows. This high correlation implies
that forecasts with short time windows usually pro-
vide higher resolutions. The longer the forecast time
window, the lower the forecast resolution. By consid-
ering the inclusion of exogenous data, it shows that
only the proportion of other exogenous data differs
from the relation to forecast window. This difference
appears due to the evaluation of the study from Chen
et al. that analyses the forecast behaviour by increas-
ing the resolution (Chen and Cook, 2012).
The third key variable of the analysed studies in
load forecasting, the node aggregation level presented
in Table 5, points out a strong focus on utility level.
More than 80% of all studies relate to forecasts on
this level. That means, they consider the forecast of
the power demand of numerous end users up to a sum-
marised node on a utility level (i.e. of a grid branch,
accommodation or production unit). The table also
shows, that forecasts on utility and local levels include
load data with a considerably higher proportion than
individual forecasts. However, socio-economic and
demographic data are only used on utility level. On
the contrary, only the individual level includes other
exogenous data in the inputs parameter list.
As evident from statistics, weather variables are
the best scrutinised and most considered exogenous
CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage
492
Table 2: Definition of forecasting time horizons.
(Hong, 2010)
(Ma and Ma, 2017)
(Matija
ˇ
s et al., 2011)
(Mirowski et al., 2014)
(Raza and Khosravi, 2015)
(Mustapha et al., 2016)
(Zor et al., 2017)
(Hammad et al., 2020)
VSTLF 1d <30’ 1h - - - <1h <1h
STLF 2W 30’ - 6h <30d 1h - 1W 1h - 1W 1h - 1W 1h - 2W 1h - 1W
MTLF 3Y 6h -1d <1Y 1W - 1Y 1M - 1Y 1W - 1Y 2W - 3Y 1W - 1Y
LTLF 30d 1d - 1W >1Y >1Y 1Y - 10Y >1Y >3Y >1Y
Table 3: Forecast Time Window and Exogenous Data [%].
%
Load
Weather
Calendar
Day
Socio-economic
Demographic
Others
STLF 85 29 25 21 22 1 0 2
MTLF 10 27 27 27 18 0 0 0
LTLF 5 29 0 14 0 29 29 0
Table 4: Forecast Resolution and Exogenous Data [%].
%
Load
Weather
Calendar
Day
Socio-economic
Demographic
Others
Hourly 80 33 17 22 22 0 0 6
Daily 15 29 26 21 21 1 0 1
Yearly 5 20 10 20 10 20 20 0
data. Heinemann et al. studied the relationship be-
tween weather and power load data already about 55
years ago (Heinemann et al., 1966). The study de-
scribes a method to extract weather sensitive compo-
nents from the total daily peak load. However, it only
concerns the load during summer time. In the follow-
ing years, several studies investigated the dependency
of the power demand on the weather behaviour. And
all agree on the existence of a notable correlation be-
tween this two domains. Furthermore, all emphasise
possible improvements in load forecasts by including
weather data such as in (Rahman and Hazim, 1993;
Hern
´
andez et al., 2012; Sahay and Tripathi, 2014;
Janicki, 2017).
As a result, numerous approaches tried to increase
the forecast accuracy by extending the input vari-
ables with information from various weather variables
(Rahman and Hazim, 1993; Mirasgedis et al., 2006;
Howe, 2010; Chu et al., 2011). Recently, Silva et al.
even defined weather variables and mainly tempera-
ture, humidity and wind speed as the most significant
exogenous influences in STLF (Silva et al., 2019b).
In contrast to utility or local level, the forecast on
individuals takes the direct environment more into ac-
count. So in (Chen and Cook, 2012), where Chen et
al. propose an STLF approach on individual level that
additionally uses the activity in the building. They
Table 5: Forecast Aggregation Level and Exogenous Data
[%].
%
Load
Weather
Calendar
Day
Socio-economic
Demographic
Others
Indiv. 13 15 23 23 23 0 0 15
Local 5 33 33 17 17 0 0 0
Utility 82 31 25 19 19 4 2 0
spread several motion sensors and monitored a whole
residential area. The recorded motion data allowed
to derive a pattern that provides extended information
to the consumption, which they used to enrich the in-
put variables for the power forecast. Similarly, Wang
et al. proposed an approach that considers the occu-
pancy information to forecast the energy usage of ed-
ucational buildings (Wang et al., 2018). Furthermore,
Tascikaraoglu and Sanandaji looked for relational pat-
terns among correlations of time series of surround-
ing houses (Tascikaraoglu and Sanandaji, 2016). The
proposed approach showed a considerable improve-
ment for short term forecasts against various bench-
mark models using real and high-quality data.
Other than in STLF, forecasts with long time
windows use exogenous data related to higher-level
consumption influences. This information does not
comprise changes having a direct influence on single
end user’s behaviour but represent large-scale events.
In fact, load forecasts ahead for several years are
more affected by long term changes such as eco-
nomic, demographic or climatic movements. Hence,
LTLF approaches as in (Chui et al., 2009; Khatoon
et al., 2014) include information about the popula-
tion, Gross Domestic Product (GDP) or geographi-
cal related changes. Huang et al. showed in their
study (Huang et al., 2016a) that the rapidly increase
of the power demand in a city in northeast China
matches with the population growth and the develop-
ment of the local society. The End-use model, a de-
tailed modelling approach for LTLF, breaks down the
energy consumption to single consumers. This model
is applied to estimate the energy consumption for long
time windows. It is based on extensive information
Exogenous Data for Load Forecasting: A Review
493
about the end user that ranges till to the device level.
The approach is discussed by Ghods and Kalantar in
(Ghods and Kalantar, 2008).
4 VALUE OF EXOGENOUS DATA
FOR DIFFERENT LOAD
FORECAST PARAMETERS
This passage narrates on the correlation between ex-
ogenous data types for different forecast parameters
(e.g. forecast horizon, location, climate). Higher-
level conditions have different influence on the energy
consumption. Derived from statistics in section 3, it
is obvious that the usable information contained in a
specific exogenous data type mainly depends on the
forecast parameters. Fast or slowly changing values,
for instance, have more or less impact on the accu-
racy depending on the forecast time window. While
economic or climate aspects have a significant influ-
ence in long-term forecasts, they provide barely any
useful information for short-term forecasts. In fact,
they appear to be basically stationary for the fore-
casting time period considered. Therefore, their con-
tribution in increasing the accuracy in STLF is very
small. In the study (Gul et al., 2011) Gul et al. inves-
tigate the relationship between electrical power de-
mand and the slowly evolving economic and demo-
graphic variables. On a country level case study for
Pakistan, they discovered a high correlation between
the selected variables and the power demand. Gen-
erally, it can be shown that long-term forecasts use
averaged data or slow changing variables with low
sampling rates. In contrast, fast changing data with
high sampling rates such as sky coverage are primary
meaningful for short-term forecasts.
In (Wang et al., 2018), a study from Wang et al.
on the energy prediction of two educational build-
ings, they show a variation of the most influencial
factors from one semester to another. The study dis-
cuss the variable importance in a case study of an
hourly energy prediction. As a consequence of the
observed variation in the influence, they conclude that
the energy usage of these educational buildings follow
rather a semester than an annual basis pattern.
Equally to the effect on the forecast time win-
dow, exogenous data also differently influences the
forecast accuracy at various aggregation levels. The
lower the forecast level (e.g: room or apartment), the
higher the observable impact generated by single en-
ergy consuming appliances. Information about the
energy consumption for charging a battery of an e-
vehicle, for instance, highly affect the consumption
behaviour of a single end user. On the other hand, the
aggregated load on a city or country level (i.e. on a
utility level) is not considerable affected by the con-
sumption of a single battery charge. The large num-
ber of end users cancels out significant changes in a
single load behaviour, due to the so called averaging
effect. This effect of the aggregation level is analysed
by Mirowski et al. in (Mirowski et al., 2014) and Ger-
wig et al. in (Gerwig, 2015) and seems to have a rel-
evant impact in terms of the prediction accuracy.
The geographic location of the considered energy
consumption is one of the variables that define the
higher-level conditions. The specific climate, eco-
nomic or lifestyle among others depend on the lo-
cation and influence the energy consumption signif-
icantly. The effect of the geographic location, how it
defines the information value of exogenous variables,
becomes evident by observing the weather variables.
Howe features in his thesis (Howe, 2010) the depen-
dency of the region and the influence of the temper-
ature on the energy consumption. So, in Philadel-
phia for instance, the temperature swings are limited
thanks to the proximity to the ocean although the city
experiences large temperature ranges than Chicago,
from bitter cold winter to hot summer days. This has a
certain impact on the energy consumption. In fact, the
temperature influence varies among the regions due to
a different use of cooling and heating devices. This
effect is also discussed by De Felice et al. (De Felice
et al., 2013) and Silva et al. in (Silva et al., 2019b) for
the climate in Italy and South America respectively.
Furthermore, in their study De Felice et al. discov-
ered that from all included weather variables only the
temperature demonstrated an evident influence on the
daily load variation. On the other hand, they also con-
clude that in some cases the use of weather informa-
tion does not show an evident benefit for daily load
forecasting. A similar statement was made by Kandil
et al. in (Kandil et al., 2006). For Hydro-Quebec they
investigated the effect of various weather variables on
the power load for the province of Quebec, in Canada.
Thereby, they discovered that only temperature has a
serious influence. Other weather variables like sky
condition (cloud cover) and wind velocity showed no
relation to the load. A brief overview of the influ-
ence of the weather factors on the electrical demand
is given in (Janicki, 2017).
Using weather variables in load forecasting
mainly means the inclusion of some weather vari-
able forecasts in the input variable list. Douglas et
al. investigated in (Douglas et al., 1998) the ef-
fects of the uncertainty of the variable forecasts on
STLF. The performed analysis presents differing im-
pacts of the temperature forecast errors in the vari-
CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage
494
ous annual seasons. Also L
´
opez et al. discovered a
non-linear dependency of the load on temperature in
(L
´
opez Garc
´
ıa et al., 2013). According to this study,
the non-linearity makes raw temperature data insuffi-
cient for using in load forecasting. The data need to be
contextualised for a meaningful use. However, they
also proposed a STLF approach for Balearic Islands
that considers the solar radiation, cloudiness and wind
velocity without temperature and claimed, that the use
of all variables in combination outperforms all other
variants (L
´
opez et al., 2017).
As exemplary shown on weather variables, the
various exogenous data provide information with a
different value. It highly depends on the target of
the forecast - the location, time window, aggrega-
tion level and so on. Therefore, it is not possible to
uniquely order the variables according to a specific
relevance or prevalence rank. A proper selection of
the including exogenous data and thereof the applying
features is indispensable. The following section give
an overview about the various feature selection meth-
ods and discusses the need for identifying the most
appropriate forecast methodology.
5 FEATURE SELECTION
METHODS
In this part it is concisely explored the role of fea-
ture selection and its relation to exogenous variables
for different settings of load forecasting. Using ex-
ogenous data in load forecasting makes a careful fea-
ture selection fundamental. The data should provide
a considerable added value to the basic information
from historic load data for reaching more accurate
forecasting results. The power demand is a high com-
plex system that depends on many external factors;
and some provide information with a higher value
than others. Typically, it is straight forward to assume,
that more information would lead to a higher forecast
accuracy. However, the forecast accuracy does not
strictly monotonous increase with a growth in feature
number. Too many and redundant features may even
drag down the forecasting performance, either by in-
troducing inessential information or by accumulating
the effects of the noise present in any real measure-
ment. This behaviour is discussed by Cheng et al.
in (Cheng et al., 2017). As a consequence, it is cru-
cial to include only features producing a considerable,
non-negligible increase in the forecast performance.
However, those selected features do not necessarily
show always a high correlation with the target data.
In the case of sudden changes in the contextual situa-
tion, even features without a high correlation lead to
a more robust forecaster, as described by Drezga and
Rahman in their study (Drezga and Rahman, 1998).
In a common forecast pipeline, valuable features
are defined in an extraction and selection process
that precedes the definition or training of the fore-
cast model. The forecast output, on the other hand,
may have a direct feedback on the preceding fea-
ture definition process as illustrated in Figure 1. The
state-of-the-art feature selection approaches usually
apply a multi-step selection procedure. Additional
steps break down the feature set to a few crucial vari-
ables. Commonly used distance related metrics for
defining feature’s information value comprise corre-
lation functions, Mutual Information (MI) or Fisher
Information (FI). This distance metrics quantify the
feature’s similarity to the load data and are used as
measure for the influence on it (Rana and Koprinska,
2012a; Cai et al., 2018; Hu et al., 2015; Huang et al.,
2016a). However, the relation between two variables
can also show a dispersed non-linear characteristics
such as it is often shown for the temperature and load.
For this case, Silva et al. claim more accurate rela-
tion indices when using models that estimate the ex-
istence condition and not only the linear dependency
as the Pearson Linear Correlation method (Silva et al.,
2019a). The approaches for the selection refinement
are very broad and comprise methods of Machine
Learning (ML) (Niu et al., 2010; Rana and Koprinska,
2012a), wrappers (Hu et al., 2015), minimum Redun-
dancy Maximum Relevance (mRMR) (Huang et al.,
2016a), Permutation Importance (PI) (Huang et al.,
2016b) or Random Forest (RF) (Cheng et al., 2017).
Finally, a qualitative feature selection requires
clean data sets, affected by a minimised noise com-
ponent. As all raw data, exogenous data usually con-
tain outliers or spikes that do not correlate with target
data and distort the valuable information. For a rea-
sonable benefit from the additional information, the
anomalies need to be rejected. The advantage of such
a preprocessed filtering is shown in several studies as
in (Guan et al., 2013; Mustapha et al., 2016; Saleh
et al., 2016). The same holds, if only certain parts
of exogenous data provide valuable information. In
such a case, filters need to cancel out the remaining
parts before the model uses the data for training and
forecast.
6 LOAD FORECASTING
METHODOLOGIES
In this section, on top of the basic approaches reported
in literature, the selection importance of some spe-
cific features is detailed. A broad review of various
Exogenous Data for Load Forecasting: A Review
495
STLF methodologies is presented in Srivastava (Sri-
vastava et al., 2016). This study, in accordance with
other literature analysed, separates the models mainly
in two groups: statistical and machine learning ap-
proaches. Statistical models usually explicitly de-
scribe a mathematical relationship between multiple
variables. Therefore their applicability is limited in
case of large number of variables as well as for highly
non-linear, complex dependencies. Based on this lim-
itation, machine learning approaches gained more at-
tention in recent years. According to comprehensive
reviews of multiple applied methods and models by
Gerwig et al. (Gerwig, 2015) and Hammad et al.
(Hammad et al., 2020), STLF approaches favourably
use machine learning models. For STLF on individual
level, the results in (Marinescu et al., 2013) indicate
equally good performances provided by the analyzed
Artificial Intelligence (AI) and Autoregressive (AR)
methods. Similarly, (Gerwig, 2015) state compara-
ble results for Artificial Neural Network (ANN), AR
and hybrid methods of both for individual households
up to 1000 end-users. In contrast, Linear Regression
(LR) shows comparable results only for individual
users while Support Vector Regression (SVR) works
well for more than 32 households. Additionally, this
study also remarks a possible accuracy improvement
by combining clustering methods with ANN or auto-
regressive methods.
Independently from the method chosen, each load
forecast is based on two main components: a model
and a selection of input variables presenting a consid-
erable influence on the target dimension. Thereby, the
input elements may include recordings, instant values
or other forecast measurements. That holds for ex-
ogenous data but also for load data, as illustrated in
Figure 1. For weather input variables, models mostly
use forecast variables. These seem to provide a higher
information value than historical data or instant val-
ues. In fact, the use of historic weather data requires
that load forecasts also comprise weather models in
order to map the information from historical data to
forecast load behaviours. The other way round, when
using forecasts of the weather variables, customised
models do the mapping of the information from his-
torical data to future scenarios separately (Zhu et al.,
2018). Yet, in this case, the uncertainty of the weather
variable forecast directly enters into the load forecast
model. This yields to a forecast, that implicitly com-
prises the uncertainty of the input variable.
Recursive forecast approaches show a similar ef-
fect. Models such as recursive Kalman filters or
Bayesian estimations use the previously calculated
data point of the same time series as base for the cal-
culation of the subsequent point, as reported in (Dou-
glas et al., 1998). Due to the recursive calculation,
the forecast errors of previous data points directly af-
fect the calculation (i.e. the forecast of the next data
point). Because of this effect - the accumulation of
the prediction errors - all recursive models can have a
significant drift in the forecast results.
The effect of forecast errors in weather variables
on load forecasting models is discussed in (Douglas
et al., 1998; Fay and Ringwood, 2010). Douglas et al.
note that a sizeable portion of the load forecast error
is due to a lack of accuracy in the weather forecast.
In order to minimise this effect, Taylor and Buizza
present an approach in (Taylor and Buizza, 2003) that
uses an ensemble prediction system. It estimates the
midday power demand from the density function of
an ensemble of 51 weather-related demand scenarios.
Another approach propose Fay and Ringwood with a
model fusion technique (Fay and Ringwood, 2010).
They claim, that fused forecasts are often more accu-
rate than any individual model forecasts. The model
in the applied case study comprises four independent
sub-models that feed a fusion algorithm. In a for-
mer study they found already, that decomposing load
data into parallel series is advantageous due to the
degree of independence of parallel series. Equally,
Tascikaraoglu points out the positive effect of decom-
posed forecasting in (Tascikaraoglu and Sanandaji,
2016) and Boroojeni et al. proposed a similar ap-
proach in (Boroojeni et al., 2017). In the latter, they
separate the forecast in multiple models by means
of an extended Seasonal Auto-Regressive Integrated
Moving Average Model (SARIMA). This allows to
map multiple seasonality cycles to the power demand
forecasting.
The various load forecasting methodologies pro-
cess the input data in different ways. This has, as
shown in literature, a notable impact on the forecast
quality. Therefore, it is meaningful to define the fore-
cast methodology depending on the input variable list.
7 FURTHER RESEARCH
This part presents some gaps we noted and our fu-
ture research directions in this domain. Knowledge
about the future power demand is valuable for plan-
ning and controlling the power distribution system.
The increasing electrification of consumers devices
and the deregulation of the energy market enhance the
demand for accurate forecasts. Especially for end-
user and prosumer, it is beneficial for participating
on deregulated energy markets in the near future. In
parallel, the growing data recordings from innumer-
able consumers devices opens access to new exoge-
CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage
496
nous data. The diffusion of social media platforms,
digital social networking services and the electrical
enhancement of everyday objects can constitute new
sources of exogenous data for energy behaviour pre-
diction. To list only a few, posts on social media plat-
forms, status information from in-house and mobility
smart appliances or wearable device data embed in-
formation relevant for factor estimation of electricity
consumption. Anyway, in literature these new sources
represent a minimal fraction of the exogenous vari-
ables considered for power demand forecasting, as
can be observed from the category Other data in Ta-
ble 1 (see Section 3). For instance, details about a
planned home party from social media channels con-
tain direct information of the end-user behaviour and
may contribute in describing the upcoming power de-
mand, but are currently rarely considered for this pur-
pose. Another peculiarity is their end-user level gran-
ularity. This aggregation scale is better suited for
more precisely describing the energy behaviour of in-
dividuals.
The literature shows a clear potential for increas-
ing the load forecast accuracy by using exogenous
data. Several studies highlight the correlation be-
tween these data series and the power demand and
show the consequent advantage by their usage. Yet for
reasonable accuracy improvements, the interaction of
the consumption behaviour and the exogenous vari-
ables as well as their influence need to be fully under-
stood.
Despite an intense research on forecast method-
ologies that consider exogenous data, the value pro-
vided by exogenous data is not sufficiently explored.
Indeed, studies compare several exogenous variables
with the consumption behaviour and analyse the fore-
cast improvements when considering this additional
information; but they mostly fail to provide an analy-
sis of the real influence. To date, it seems to be widely
unexplored, which information (i.e. which part of
the exogenous variables) correlate with anomalies in
the consumption or defines the higher-level contextual
conditions such as the humidity in combination with
the ambient temperature that influence the perception
of the temperature. This would especially be of in-
terest for demand forecasts on an individual level, as
this level generally presents a high volatile consump-
tion behaviour.
In that issue, we discovered a certain research de-
mand in the analysis of the information value of ex-
ogenous and historical data on the individual fore-
cast level. We see a demand in research on the im-
pact of the describing variables and expect to have
a weighting for the variables; possibly based on a
fuzzification of the information value. This will play
a twofold role: on one side, structurally considering
and account for the uncertainty, while, on the other
side, allowing the use of a more fine grained rea-
soning, such as for categorical dimensions and their
mapping from/to continuous values, using member-
ship degrees. A detailed knowledge about the influ-
ence of exogenous data on the power load as well as
the information value from historical data on a future
load can be a key value for further research on fore-
cast methods.
For a better use of the inclusion of exogenous data
in load forecasting, the influence of these informa-
tion needs to be analysed in more detail. Thereby, the
choice of a suitable metric for the information value
builds the base for a proper analysis of their influ-
ence. Secondly, an information value quantification
for each candidate exogenous variable is required, by
examining its point-wise correlation with respect to
the power demand time series. Finally, weighting fac-
tors are necessary to control the exogenous variables
effect on the power consumption forecast. To achieve
this, it is crucial to identify the factors averaged quan-
tified influence on the power behaviour. In this way,
it should be possible to extract and apply only the key
information from additional variables to power fore-
casts. It is expected that selective inclusion of ex-
ogenous variables in forecasting achieves higher ac-
curacy, by removing redundant and irrelevant noise.
This will be of paramount importance for improve-
ments on energy management at the local community
level, where individual power influx estimation plays
a major role.
8 CONCLUSION
Exogenous data provides additional information us-
able for increasing the load forecast accuracy. This
review shows the significantly additional value that
exogenous data can contribute to better predictive per-
formances. Additionally, it demonstrates the need
for a detailed analysis of the extracted features. De-
pending on several factors such as the aggregation
level, area or the forecast time horizon, the variables
show different influence on the load. Additional vari-
ables from exogenous data show a noteworthy infor-
mation value for increasing forecast accuracy. Espe-
cially weather variables, but also others, show a high
correlation with power demand. This emerges from
several studies ranging from short-term to LTLF and
considering high aggregation levels. In contrast, there
is still a significant lack of attention about the use of
exogenous data in load forecast for residential and in-
dividual level. In fact, forecast on individual level is
Exogenous Data for Load Forecasting: A Review
497
generally affected by the high volatility issue. Nev-
ertheless, this last category is increasingly becoming
relevant due to the energy market deregulation and
the possibility for end-users to actively participate in
smart grids. As a consequence a detailed research on
the feature information value from historic or exoge-
nous data show an increased interest.
REFERENCES
Boroojeni, K. G., Amini, M. H., Bahrami, S., Iyengar, S. S.,
Sarwat, A. I., and Karabasoglu, O. (2017). A novel
multi-time-scale modeling for electric power demand
forecasting: From short-term to medium-term hori-
zon. Electric Power Systems Research, 142:58–73.
ZSCC: 0000109.
Cai, S., Liu, L., Sun, H., and Yan, J. (2018). Fisher In-
formation Based Meteorological Factors Introduction
and Features Selection for Short-Term Load Forecast-
ing. Entropy, 20(3):184. ZSCC: 0000001.
Chen, C. and Cook, D. J. (2012). Behavior-Based Home
Energy Prediction. In 2012 Eighth International Con-
ference on Intelligent Environments, pages 57–63.
ZSCC: 0000038.
Chen, Y., Luh, P. B., and Rourke, S. J. (2008). Short-term
load forecasting: Similar day-based wavelet neural
networks. In 2008 7th World Congress on Intelligent
Control and Automation, pages 3353–3358. ZSCC:
0000030.
Cheng, Y., Xu, C., Mashima, D., Thing, V. L. L., and Wu,
Y. (2017). PowerLSTM: Power Demand Forecasting
Using Long Short-Term Memory Neural Network. In
Cong, G., Peng, W.-C., Zhang, W. E., Li, C., and
Sun, A., editors, Advanced Data Mining and Applica-
tions, Lecture Notes in Computer Science, pages 727–
740, Cham. Springer International Publishing. ZSCC:
NoCitationData[s0].
Chu, W.-C., Chen, Y.-P., Xu, Z.-W., and Lee, W.-J. (2011).
Multiregion Short-Term Load Forecasting in Consid-
eration of HI and Load/Weather Diversity. IEEE
Transactions on Industry Applications, 47(1):232–
237. ZSCC: 0000000.
Chui, F., Elkamel, A., Surit, R., Croiset, E., and Douglas, P.
(2009). Long-term electricity demand forecasting for
power system planning using economic, demographic
and climatic variables. European J. of Industrial En-
gineering, 3(3):277. ZSCC: 0000026.
De Felice, M., Alessandri, A., and Ruti, P. M. (2013).
Electricity demand forecasting over Italy: Potential
benefits using numerical weather prediction models.
Electric Power Systems Research, 104:71–79. ZSCC:
0000053.
Douglas, A., Breipohl, A., Lee, F., and Adapa, R. (1998).
The impacts of temperature forecast uncertainty on
Bayesian load forecasting. IEEE Transactions on
Power Systems, 13(4):1507–1513. ZSCC: 0000152.
Drezga, I. and Rahman, S. (1998). Input variable se-
lection for ANN-based short-term load forecasting.
IEEE Transactions on Power Systems, 13(4):1238–
1244. ZSCC: 0000223.
Fay, D. and Ringwood, J. V. (2010). On the Influence of
Weather Forecast Errors in Short-Term Load Forecast-
ing Models. IEEE Transactions on Power Systems,
25(3):1751–1758. ZSCC: 0000063.
Gerwig, C. (2015). Short Term Load Forecasting for
Residential Buildings—An Extensive Literature Re-
view. In Neves-Silva, R., Jain, L. C., and Howlett,
R. J., editors, Intelligent Decision Technologies, Smart
Innovation, Systems and Technologies, pages 181–
193, Cham. Springer International Publishing. ZSCC:
NoCitationData[s0].
Ghods, L. and Kalantar, M. (2008). Methods for long-term
electric load demand forecasting; a comprehensive in-
vestigation. In 2008 IEEE International Conference
on Industrial Technology, pages 1–4. ZSCC: 0000110
ISSN: null.
Gross, G. and Galiana, F. (1987). Short-term load fore-
casting. Proceedings of the IEEE, 75(12):1558–1573.
ZSCC: 0000921.
Guan, C., Luh, P. B., Michel, L. D., Wang, Y., and Fried-
land, P. B. (2013). Very Short-Term Load Forecasting:
Wavelet Neural Networks With Data Pre-Filtering.
IEEE Transactions on Power Systems, 28(1):30–41.
ZSCC: NoCitationData[s0].
Gul, M., Qazi, S. A., and Qureshi, W. A. (2011). Incorporat-
ing economic and demographic variablesfor forecast-
ing electricity consumption in Pakistan. In 2011 2nd
International Conference on Electric Power and En-
ergy Conversion Systems (EPECS), pages 1–5. ZSCC:
0000015 ISSN: null.
Hammad, M. A., Jereb, B., Rosi, B., and Dragan, D. (2020).
Methods and Models for Electric Load Forecasting:
A Comprehensive Review. Logistics & Sustainable
Transport, 11(1):51–76. ZSCC: 0000000.
Heinemann, G. T., Nordmian, D. A., and Plant, E. C.
(1966). The Relationship Between Summer Weather
and Summer Loads - A Regression Analysis. IEEE
Transactions on Power Apparatus and Systems, PAS-
85(11):1144–1154. ZSCC: 0000127.
Hern
´
andez, L., Baladr
´
on, C., Aguiar, J. M., Calavia, L.,
Carro, B., S
´
anchez-Esguevillas, A., Cook, D. J., Chi-
narro, D., and G
´
omez, J. (2012). A Study of the
Relationship between Weather Variables and Elec-
tric Power Demand inside a Smart Grid/Smart World
Framework. Sensors, 12(9):11571–11591. ZSCC:
NoCitationData[s0].
Hong, T. (2010). Short Term Electric Load Forecasting.
PhD thesis, North Carolina State University.
Howe, K. J. (2010). AN ANALYSIS OF WEATHER FORE-
CASTS IN THE CONTEXT OF ELECTRICITY USE.
PhD thesis, Pennsylvania State University. ZSCC:
0000000.
Hu, Z., Bao, Y., Xiong, T., and Chiong, R. (2015). Hybrid
filter–wrapper feature selection for short-term load
forecasting. Engineering Applications of Artificial In-
telligence, 40:17–27. ZSCC: 0000106.
Huang, N., Hu, Z., Cai, G., and Yang, D. (2016a).
Short Term Electrical Load Forecasting Using Mu-
CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage
498
tual Information Based Feature Selection with
Generalized Minimum-Redundancy and Maximum-
Relevance Criteria. Entropy, 18(9):330. ZSCC:
0000009.
Huang, N., Lu, G., and Xu, D. (2016b). A Permuta-
tion Importance-Based Feature Selection Method for
Short-Term Electricity Load Forecasting Using Ran-
dom Forest. Energies, 9(10):767. ZSCC: 0000027.
Janicki, M. (2017). Methods of weather variables introduc-
tion into short-term electric load forecasting models
- a review. Przeglad Elektrotechniczny, 1(4):72–75.
ZSCC: 0000001.
Jiang, P., Liu, F., and Song, Y. (2017). A hybrid fore-
casting model based on date-framework strategy and
improved feature selection technology for short-term
load forecasting. Energy, 119:694–709. ZSCC:
0000065.
Kandil, N., Wamkeue, R., Saad, M., and Georges, S. (2006).
An Efficient Approach for Shorterm Load Forecasting
using Artificial Neural Networks. In 2006 IEEE In-
ternational Symposium on Industrial Electronics, vol-
ume 3, pages 1928–1932. ZSCC: 0000203 ISSN:
2163-5145.
Khatoon, S., Ibraheem, Singh, A. K., and Priti (2014). Ef-
fects of various factors on electric load forecasting:
An overview. In 2014 6th IEEE Power India Interna-
tional Conference (PIICON), pages 1–5. ZSCC: NoC-
itationData[s0] ISSN: null.
L
´
opez, M., Valero, S., Senabre, C., and Gabald
´
on, A.
(2017). Analysis of the influence of meteorological
variables on real-time Short-Term Load Forecasting
in Balearic Islands. In 2017 11th IEEE International
Conference on Compatibility, Power Electronics and
Power Engineering (CPE-POWERENG), pages 10–
15. ZSCC: 0000004 ISSN: 2166-9546.
L
´
opez Garc
´
ıa, M., Valero, S., Senabre, C., and Ga-
bald
´
on Mar
´
ın, A. (2013). Short-Term Predictability
of Load Series: Characterization of Load Data Bases.
IEEE Transactions on Power Systems, 28(3):2466–
2474. ZSCC: 0000011.
Ma, J. and Ma, X. (2017). State-of-the-art forecasting al-
gorithms for microgrids. In 2017 23rd International
Conference on Automation and Computing (ICAC),
pages 1–6. ZSCC: 0000006 ISSN: null.
Marinescu, A., Harris, C., Dusparic, I., Clarke, S., and
Cahill, V. (2013). Residential electrical demand fore-
casting in very small scale: An evaluation of forecast-
ing methods. In 2013 2nd International Workshop on
Software Engineering Challenges for the Smart Grid
(SE4SG), pages 25–32. ZSCC: 0000053 ISSN: null.
Matija
ˇ
s, M., Cerjan, M., and Krajcar, S. (2011). Features
affecting the load forecasting error on country level.
In Proceedings of the 2011 3rd International Youth
Conference on Energetics (IYCE), pages 1–7. ZSCC:
0000008 ISSN: null.
Mazzola, L., Denzler, A., and Christen, R. (2020). To-
wards a Peer-to-Peer Energy Market: an Overview.
arXiv:2003.07940 [physics]. ZSCC: NoCitation-
Data[s0] arXiv: 2003.07940.
Mirasgedis, S., Sarafidis, Y., Georgopoulou, E., Lalas, D. P.,
Moschovits, M., Karagiannis, F., and Papakonstanti-
nou, D. (2006). Models for mid-term electricity de-
mand forecasting incorporating weather influences.
Energy, 31(2):208–227. ZSCC: 0000258.
Mirowski, P., Chen, S., Ho, T. K., and Yu, C.-N. (2014). De-
mand forecasting in smart grids. Bell Labs Technical
Journal, 18(4):135–158. ZSCC: 0000085.
Mustapha, M., Mustafa, M. W., Khalid, S. N., Abubakar,
I., and Abdilahi, A. M. (2016). Correlation and
Wavelet-based Short-Term Load Forecasting using
Anfis. Indian Journal of Science and Technology,
9(46). ZSCC: 0000006.
Niu, D., Wang, Y., and Wu, D. D. (2010). Power load
forecasting using support vector machine and ant
colony optimization. Expert Systems with Applica-
tions, 37(3):2531–2539. ZSCC: 0000305.
Rahman, S. and Hazim, O. (1993). A generalized
knowledge-based short-term load-forecasting tech-
nique. IEEE Transactions on Power Systems,
8(2):508–514. ZSCC: 0000325.
Rana, M. and Koprinska, I. (2012a). Electricity load fore-
casting using non-decimated wavelet prediction meth-
ods with two-stage feature selection. In The 2012
International Joint Conference on Neural Networks
(IJCNN), pages 1–8. ZSCC: 0000005 ISSN: 2161-
4407.
Rana, M. and Koprinska, I. (2012b). Shift Invariance and
Border Distortion in Wavelet-Based Electricity Load
Forecasting. In 21st International Conference on
Pattern Recognition (ICPR 2012), page 4. ZSCC:
0000000.
Rana, M. and Koprinska, I. (2013). Wavelet Neural Net-
works for Electricity Load Forecasting – Dealing with
Border Distortion and Shift Invariance. In Mlade-
nov, V., Koprinkova-Hristova, P., Palm, G., Villa, A.
E. P., Appollini, B., and Kasabov, N., editors, Artifi-
cial Neural Networks and Machine Learning ICANN
2013, Lecture Notes in Computer Science, pages 571–
578, Berlin, Heidelberg. Springer. ZSCC: NoCitation-
Data[s0].
Raza, M. Q. and Khosravi, A. (2015). A review on artifi-
cial intelligence based load demand forecasting tech-
niques for smart grid and buildings. Renewable and
Sustainable Energy Reviews, 50:1352–1372. ZSCC:
0000334.
Sahay, K. B. and Tripathi, M. (2014). Day ahead hourly
load forecast of PJM electricity market and iso new
england market by using artificial neural network. In
ISGT 2014, pages 1–5. ZSCC: 0000029 ISSN: null.
Saleh, A. I., Rabie, A. H., and Abo-Al-Ez, K. M. (2016). A
data mining based load forecasting strategy for smart
electrical grids. Advanced Engineering Informatics,
30(3):422–448. ZSCC: 0000034.
Silva, L. N., Abaide, A. R., Figueir
´
o, I. C., Martinuzzi,
D., and Rigodanzo, J. (2017). Development of an
ANN model to multi-region short-term load forecast-
ing based on power demand patterns recognition. In
2017 IEEE PES Innovative Smart Grid Technologies
Conference - Latin America (ISGT Latin America),
pages 1–6. ZSCC: 0000001.
Exogenous Data for Load Forecasting: A Review
499
Silva, L. N., Abaide, A. R., Negri, V. G., Capeletti, M.,
and Lopes, L. F. (2019a). Impact Evaluation of Fea-
ture Selection to Short-Term Load Forecasting Mod-
els considering Weather Inputs and Load History. In
2019 54th International Universities Power Engineer-
ing Conference (UPEC), pages 1–6. ZSCC: 0000000
ISSN: null.
Silva, L. N., Abaide, A. R., Negri, V. G., Capeletti, M.,
Lopes, L. F., and Cardoso, G. (2019b). Diagnostic and
Input Selection Tool applied on Weather Variables for
Studies of Short-Term Load Forecasting. In 2019 8th
International Conference on Modern Power Systems
(MPS), pages 1–6. ZSCC: 0000000 ISSN: null.
Srivastava, A. K., Pandey, A. S., and Singh, D. (2016).
Short-term load forecasting methods: A review. In
2016 International Conference on Emerging Trends
in Electrical Electronics Sustainable Energy Systems
(ICETEESES), pages 130–138. ZSCC: NoCitation-
Data[s0].
Tascikaraoglu, A. and Sanandaji, B. M. (2016). Short-
term residential electric load forecasting: A compres-
sive spatio-temporal approach. Energy and Buildings,
111:380–392. ZSCC: 0000052.
Taylor, J. W. and Buizza, R. (2003). Using weather en-
semble predictions in electricity demand forecasting.
International Journal of Forecasting, 19(1):57–70.
ZSCC: 0000290.
Wang, Z., Wang, Y., Zeng, R., Srinivasan, R. S., and
Ahrentzen, S. (2018). Random Forest based hourly
building energy prediction. Energy and Buildings,
171:11–25. ZSCC: 0000059.
Zheng, H., Yuan, J., and Chen, L. (2017). Short-Term
Load Forecasting Using EMD-LSTM Neural Net-
works with a Xgboost Algorithm for Feature Impor-
tance Evaluation. Energies, 10(8):1168. ZSCC:
0000123.
Zhu, G., Chow, T.-T., and Tse, N. (2018). Short-term
load forecasting coupled with weather profile gener-
ation methodology. Building Services Engineering
Research and Technology, 39(3):310–327. ZSCC:
0000018.
Zor, K., Timur, O., and Teke, A. (2017). A state-of-the-art
review of artificial intelligence techniques for short-
term electric load forecasting. In 2017 6th Interna-
tional Youth Conference on Energy (IYCE), pages 1–7.
ZSCC: 0000026 ISSN: null.
CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage
500