B. da S. Macêdo, C. M. Saporetti,

“Electricity Energy Demand Prediction Using Computational Intelligence Techniques”,

Latin-American Journal of Computing (LAJC), vol. 11, no. 2, 2024.

Electricity Energy

Demand Prediction

Using Computational

Intelligence Techniques

ARTICLE HISTORY

Received 06 February 2024

Accepted 19 April 2024

Bruno da S. Macêdo

Federal University of Lavras

Lavras, Brazil

bruno.macedo2@estudante.uﬂa.br

ORCID: 0009-0009-4375-8464

Camila Martins Saporetti

Polytechnic Institute, Rio de Janeiro State University

Nova Friburgo, Brazil

camila.saporetti@iprj.uerj.br

ORCID: 0000-0002-8145-7074

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

Electricity Energy Demand Prediction Using

Computational Intelligence Techniques

Bruno da S. Macêdo

Systems and Automation Engineering Graduate Program

Federal University of Lavras

Lavras, Brazil

bruno.macedo2@estudante.ufla.br

ORCID: 0009-0009-4375-8464

Camila Martins Saporetti

Department of Computational Modeling

Polytechnic Institute, Rio de Janeiro State University

Nova Friburgo, Brazil

camila.saporetti@iprj.uerj.br

ORCID: 0000-0002-8145-7074

Abstract— Energy is an important pillar for the economic

development of a country. The demand for electricity is something

that continues to grow, one of the contributing factors is the

emergence of various technological equipment and the consequent

use by the population. There are several resources that can be

exploited to generate electricity, with hydroelectric power stations

being one of the most used resources. As electrical energy cannot be

stored, there is a need to estimate its consumption, looking for a way

to meet this energy demand. In this context, this study seeks to apply

machine learning techniques, using the Grey Wolf Optimization

(GWO) meta-heuristic to optimize regression models, to predict the

demand for electricity in Brazil, and it aims to estimate how much

energy should be produced. For the predictions, the period between

the years 2017 to 2022 was used, totaling around 2,190 samples. The

methodology involves pre-processing, crossvalidation, parameters

optimization and regression. The results show that Random Forest

performed well in the experiments carried out, presenting a

coefficient of determination (R

) of 0.8751, Root Mean Squared

Error (RMSE) of 0.0554 and Mean Absolute Error (MAE) of 0.0348

in the best model.

Keywords—Electric Energy, Machine Learning, Meta-

Heuristic, Grey Wolf Optimization

I. INTRODUCTION

Energy is a fundamental input in the current economy. The

economic and social development of countries is deeply

related to the growth and increase in the supply of electrical

energy [1]. It is estimated that global electrical energy

generation will increase by approximately 77% between 2006

and 2030 [2].

The demand for electrical energy is something that

continues to grow. One of the contributing factors is the

emergence of various technological equipment and the

consequent use by the population. Brazil has several resources

that can be explored to generate electrical energy, and one of

the most used resources is hydroelectric plants, which are

renewable energy sources [1].

The contribution of energy from hydroelectric plants is

around 63% in Brazil, being responsible for generating

approximately 70% of all energy used in the country. Despite

incentives for the use of other energy sources, it is estimated

that in the following years at least 50% of the energy

consumed will still come from hydroelectric plants [3].

Although Brazil has large sources of renewable energy, the

country has problems in terms of electricity supply, and issues

to be observed regarding energy-related investment [4]. There

is a segment of the Brazilian population that has difficulty

accessing electricity, with the majority of problems being in

the way energy is distributed.

As electrical energy cannot be stored, its production and

consumption must be accurately idealized in order to avoid

circumstances of energy insufficiency as well as

overproduction. Simultaneously, load and demand projections

serve as the foundation for several decisions made in the

energy markets, enabling the planning and operation in a way

that is secure, clear, effective, and meets industry demands.

Thus, one can observe the need to estimate energy

consumption, looking for some way to meet this energy

demand. In this context, computational tools can assist in the

prediction process and when it comes to the use of data,

machine learning techniques appear as an alternative. Given

the above, this study seeks to apply machine learning methods

to predict electricity demand in Brazil. Thus, it will be possible

to assist in decision-making for the distribution of electrical

energy.

When using machine learning techniques, a very important

factor is to define attributes of the methods to maximize

performance. To overcome this situation, metaheuristics can

be applied to optimize the models, seeking the best parameters

to obtain estimates with the lowest error.

The aim of this study is to use energy load data made

available by ONS in order to predict demand based on

consumption from the previous seven days. To this end, we

propose the use of machine learning techniques commonly

used in other applications, such as MultiLayer Perceptron

(MLP), Random Forest (RF) and Support Vector Machine

(SVM), and the metaheuristic Grey Wolf Optimizer (GWO).

The results show that the computational methodology

developed performs well in prediction and can assist

specialists, providing direction for strategic management and

it will anticipate future needs, that will serve as a roadmap for

the development and execution of strategies. For businesses in

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

B. da S. Macêdo, C. M. Saporetti,

“Electricity Energy Demand Prediction Using Computational Intelligence Techniques”,

Latin-American Journal of Computing (LAJC), vol. 11, no. 2, 2024.

the energy sector, selecting the forecasting model and

approach for demand prediction is a crucial decision.

Companies operating in the energy sector can set their

strategic goals and have the opportunity to improve

performance based on this study.

The present study is divided into five sections: section 2

presents the research related to this study. Section 3 discusses

the study area, as well as the methodology used. In section 4,

there is a discussion of the results obtained and, finally, in

section 5, the conclusions of this study are presented.

II. RELATED

RESEARCHES

Forecasting energy demand in Brazil is a topic that has

been studied by many researchers. In these studies, analysis is

carried out and it is proposed tools in the context of Data

Mining to understand the problem and seek solutions to

predict the results.

Ruas et al. [5] carried out a study on predicting short-

term energy demand in the state of Paraná between 2004 and

2006. Artificial Intelligence methods were used to predict the

results, such as Recurrent Artificial Neural Networks (RNNs)

and Support Vector Machine (SVM). The SVM algorithm,

with 84 days of input, with sub-bands for the forecast, was the

one that obtained the best result.

Alves [6] conducted a study on short-term electrical load

forecasting, with historical data from periods of 24 and 48

hours forward, from a company in the electrical sector.

Multiple Linear Regression (MLR) and Multilayer Perceptron

(MLP) algorithms were used. The MLP was the one that

achieved the optimal results.

In the research by Drebes [7], the energy demand for a

given day was forecast for the Certel Cooperativa Operations

Center Company, responsible for the operation of distribution

systems, operation of substations and responsible for

controlling active demand. The algorithms used were the

MLP, Linear Regression (LR) and Random Forest (RF). The

LR algorithm was the one that presented really good results.

Schreiber et al. [8] made a prediction of the performance

of transformers at the State Electricity Distribution Company

in the city of Porto Alegre, Rio Grande do Sul. The MLR

algorithm was used. The best results showed an average

relative error of 0.050 of the real and estimated yield.

In Marcos and Júnior’s work [1], machine learning

techniques were used to predict electricity consumption in the

Northeast region of Brazil, between the years 2004 and 2019.

MLP and Convolutional Neural Networks (CNN) were those

that obtained the best outcomes.

Oliveira [9] used the GWO meta-heuristic to minimize the

objective function total cost of a shell and tube heat exchanger

project, which are used to heating and cooling in various

applications such as petroleum refineries, chemical

processing, among other applications.

In Pizzolato et al. [10], the GWO meta-heuristic was used

to obtain the optimal configuration of relay actuation and

optimize relay time, which allows faults to be identified, locate

and alert the operation of an electrical system so that circuit

breakers are open, isolating a given defect. Using GWO, it was

possible to coordinate the relays, maintaining the adjustments

to the protection system.

The papers found do not forecast energy demand for Brazil

as a whole, but rather for specific regions, in addition to not

using approaches to find the optimal model. The application

of machine learning algorithms is very promising and

employing meta-heuristics will help to find the best model,

making it possible to predict demand with less error.

III. METHODOLOGY

A. Database

The National Electric System Operator (ONS) has diverse

information about energy in Brazil. In this study, the variable

Energy Load (EL) was used, which indicates the population's

demand, that is, how much energy is used.

The database has daily records of the energy load across

the country, where this information is separated by regions. As

the objective is to analyze the entire country, a sum of

information from all regions was carried out to obtain the

demand of the Brazilian population as a whole. The period

used for predictions is between the years 2017 and 2022,

around 8,764 samples.

B. Pre-Processing

There are four attributes available: id_subsistema,

nom_subsistema, din_instante and val_cargaenergiamwmed.

The id_subsistema attribute contains the initial letter of each

region of Brazil. For example, for the North region the

representation is N. The nom_subsistema attribute represents

the name of the regions of Brazil, being North, South,

Southeast and Northeast. Information for the Central-West

region is not available on the base. Furthermore, the

din_instante attribute indicates a respective date, in the format

(YYYY-MM-DD). Finally, the val_cargaenergiamwmed

attribute presents the load value in milliwatts (MW).

To predict energy demand, only the variables din_instante

and val_cargaenergiamwmed were considered. They were

renamed to DATE and EL, respectively. The DATE variable

represents a single date and the EL variable represents the sum

of energy loads between the North, South, Southeast/Central-

West and Northeast regions. After summing up the energy

load of the regions, the database had 2,191 samples.

Furthermore, a normalization of the EL variable was

performed, resulting in values from 0.10 to 0.90. The attributes

that refer to the energy load value of each region, as well as

those that identify a specific region, were excluded, as the

DATE and EL attributes, which contain the sum of the loads

between the regions, will be taken into consideration for the

analysis. A lag was also created in the database, creating 7

variables: EL1, EL2, EL3, EL4, EL5, EL6 and EL7. EL1 has

a charge from the second day of EL and as it contains one less

piece of information, this remains as NaN. EL2, from the third

day onwards, contains two NaN information and so on. These

samples containing NaN were excluded, 7 in total. After

exclusion, 2,184 samples were obtained. This way, it will be

possible to predict the energy load for the eighth day

considering the previous seven. With the creation of these

variables, a new base was created, having only the following

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

attributes: DATE, EL, EL1, EL2, EL3, EL4, EL5, EL6 and

EL7 and this data was used in this study.

McNemar's statistical test [11] was used to verify the

dependence between variables. When applying the test, some

factors must be considered, such as the variables are of the

same nature; are identical; have the same values; each variable

was entered only once in the sample. If the result is less than

(0.05) which is the significance level, the null hypothesis (HO)

is rejected, that is, there is an association between variables.

C. Cross Validation

Cross Validation (CV) is used to analyze how much a

method can generalize across a set of data. CV is widely used

in problems whose aim is to make predictions. Using this

approach, it is possible to divide the database into training and

test sets. The training set is applied to define the parameters

that will be used in the model and the test set is to evaluate the

model after training the method.

To evaluate the performance of the regression methods,

the K-Fold (KF) technique was used [12]. When using KF, k

subsets are divided from N samples, where K>1. After

separating the subsets, the k-1 subsets are used to train the

methods, and the rest of the sets is used to perform them. Thus,

at the end of the process, the validation error is calculated. This

procedure is repeated K times, using a different test set for

each iteration. In order for regression methods to be able to

predict future inputs, tests are repeated several times to best

train the models.

D. Methods

Random Forest (RF) [13] is a learning algorithm that

works as an ensemble. Builds k decision trees on a training

data set in k iterations. During each iteration, a set of samples

is randomly selected first. To construct a decision tree from

this subset, attributes are randomly chosen by the RF. In this

case, as the variable used is the EL and the lags, the decision

trees are created based on the randomly chosen lags. The

McNemar statistical test was applied where the null

hypothesis was not rejected, indicating that there is no

statistical evidence that there is an association between pairs

of variables in a way that allows the use of RF. Each decision

tree is constructed by considering independent random subsets

of features and samples. The prediction of a new sample is

made using an average or median of the individual tree

predictions.

SVR is the Support Vector Machine method for regression

[14]. The SVR can be linear or non-linear according to the

kernel functions employed. Given the data set of points {(x

), . . .(x

, y

)}, where

∈

ℜ

is a vector of features and

∈ℜ

is the vector of target values. It has the parameters >

0 and C > 0 and the SVR is formulated as the optimization

issue in (1).

subject to

ϕ(x

)+b− y

≤ε+ξ

−w

ϕ(x

)−b ≤ε+ξ

,ξ

' ≥0 ,i=1,...,l.

The dual form of the optimization problem can be written

as (2).

subject to

(α −α')=0,

0 ≤α

,α

'≤C ,i=1,...,l .

where

K (x

, x

)=ϕ(x

ϕ(x

)

, and ϕ(.) is the kernel

function.

Solving (2) allows determining the parameters to build

the SVR approximation.

Then, SVR estimates are given by

Multilayer Perceptron (MLP) neural networks were

developed to solve problems that are non-linearly separable,

in which they cannot be separated by a hyperplane. The

structure of a MLP is composed of layers of neurons, where

the first layer is for data input, followed by one or more

hidden layers to process the information, which uses

activation functions, and a last output layer to return the

result. In the MLP training phase, the neuron weights are

updated to minimize errors using the Backpropagation

algorithm [15]. Fig. 1 illustrates an MLP neural network, with

a tgh activation function, with six and two neurons in the first

and second hidden layers, respectively.

Fig. 1. Architecture of a MLP

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

B. da S. Macêdo, C. M. Saporetti,

“Electricity Energy Demand Prediction Using Computational Intelligence Techniques”,

Latin-American Journal of Computing (LAJC), vol. 11, no. 2, 2024.

One of the more actual meta-heuristic swarm intelligence

techniques is the Grey Wolf Optimizer (GWO). Because of its

remarkable advantages over other swarm intelligence

techniques, namely, that it requires no derivation information

during the initial search and has very few parameters, it has

been extensively used for a wide range of optimization

problems. Additionally, it is straightforward, easy to apply,

adaptable, scalable, and has the unique ability to balance

exploration and exploitation in a way that promotes favorable

convergence during the search.

The GWO meta-heuristic is based on the social behavior

of grey wolves, seeking to simulate the social hierarchy of

wolves in a pack [16]. In the GWO, the inhabitants are

separated in alpha (⍺), beta (β), delta (δ) and omega (ω). The⍺

wolves that have more capability to survive environment are

the first denominated, β and δ that lead other wolves ω⍺

towards promising locations in the search space. During the

process, the wolves advance and modify their, β or δ⍺

positions as follows:

where t is the most recent epoch,

J =2a r

a, L=2r

X

is the prey position vector, X is the position vector of a

grey wolf, r



, r



are random vectors in [0,1] and the

parameter a decreases linearly from two to zero. The GWO

supposes that, β and δ are the assumed (optimal) prey⍺

position. The three best solutions obtained so far are

considerate, β and δ respectively. Other wolves are then⍺

indicated as ω and with the capacity to reposition themselves

in relation to, β and δ. The proposed mathematical model⍺ that

reports the location to be readjusted of the ω wolves:

M

=|L

X

−X| (5)



|



−



(6)



|



−



(7)

where

X

, X

present the position of, β and δ⍺

respectively,

L

, L

are random vectors and the



X is

the position of the current solution.

Equations (5), (6) and (7) calculate the distance between

the current solution and, β and δ. The final position of the ⍺

ǣ

where

X

, X

show the current location of the

wolves, β and δ, ⍺

J

, J

are randomly generated vectors

and t is the number of epochs.

As shown above, (5), (6) and (7) represent the step of the

wolf ω toward, β and δ respectively. Equations (8), (9), ⍺ (10)

and (11) are the final location of the ω wolves. There are two

vectors too, as can be seen:



J and



E. Metrics

The following metrics were used to evaluate performance:

Root Mean Square Error (RMSE), Mean Absolute Error

(MAE) and Coefficient of Determination (R

RMSE calculates the square root of the mean of the

difference between the true value and the estimated value for

the data set. The calculation of the difference is squared. The

higher the RMSE value in the calculation, the worse the model

will be [17]. In (12), y

the true value, ŷ

the estimated value

and n the data set number.

RMSE (12)

In MAE, the average difference between the true value and

the estimated value for the data set is calculated, but as there

may be negative values in the difference, the value in the

module is considered. The lower the value obtained in the

MAE calculation, the better the predicted results will be [17].

In (13), y

the true value, ŷ

the estimated value and n the data

set number.

measures the variance of a model data. The variability

value is between 0 and 1. The value obtained in the R

calculation indicates that the higher the value, the prediction is

closer to what is expected according to the original data [17].

In (14), y

represents the true value, ŷ

the value to be predicted

and ȳ the average value for y.

IV. RESULTS

AND DISCUSSIONS

The computational methodology was implemented using

the programming language Python, which is a language used

to perform data analysis and which has several libraries with

optimized functions [18]. All experiments were run on a

computer with the following specifications: Intel(R)

Core(TM) i5-1135G7, 8 GB RAM, and Windows 10

operating system. The Scikit-Learn and PyGMO libraries

were used. Scikit-Learn is a library that allows you to work

with machine learning, it has a set of resources, such as

algorithms to perform data analysis, metrics for prediction,



M=|



(t)−



X (t)|

(3)



(t+1)



(t)−



(4)

X

=X

−J

M

(8)



−



(9)



−



(10)

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

among others [19]. PyGMO is a library used to work with

optimization problems, it has several optimization algorithms

to be used in conjunction with machine learning algorithms for

a better performance [20]. As for the library, 30 independent

iterations were carried out to evaluate the methodology. A KF

with a value of K equal to 5 was used.

Table 1 presents the description of the models used in the

GWO meta-heuristic, indicating the method, the parameters

for the model to perform well, the description of these

parameters and their configuration, indicating the values used

during the executions. In MLP, the activation function settings

to be used in GWO were represented as follows: 0: Identity, 1:

Logistic, 2: Tanh and 3: ReLU, the configuration being [0, 3].

In this configuration, for example, it means that the values will

be generated randomly in the range from 0 to 3. The máximum

number of GWO generations was 30.

TABLE I. DESCRIPTION OF THE MODELS

Methods

Parameters Description Settings

MLP

hidden_laye

r_sizes

Number of hidden

layers and Number of

neurons in each layer

[1, 4] and

[1, 50]

activation Activation Function

0: Identity; 1:

Logistic; 2:

Tanh; 3: ReLU

RF n_estimators

Number of trees in the

forest

[100, 200]

max_depth

Maximum depth of the

tree

[2, 10]

SVM C

Adjusts the penalty for

regression errors

[20, 200]

gamma

Defines how far the

influence of a single

training example

extends

[0.001, 0.1]

epsilon

Sets a limit on

insensitivity to errors in

predictions

[0.001, 0.1]

Table 2 presents the results obtained using the average and

standard deviation, the R

metrics of the iterations, RMSE and

MAE. Using the GWO meta-heuristic, it was possible to find

the ideal parameters to maximize the performance of the

regression models. The best model is highlighted in bold.

The model that achieved the best performance was RF,

with a R

of 0.8704, MAE of 0.0352 and RMSE of 0.0564,

thus indicating that it was the method that obtained the best

predictions in relation to the EL variable. The SVM also

showed good results, with a R

of 0.8618, MAE of 0.0353 and

RMSE of 0.0583. The closer the value of R

is to 1 and the

lower the values of MAE and RMSE, the better the model

performance.

TABLE II. MODELS PERFORMANCE

Methods

RMSE MAE

MLP 0.7982 ± 0.0288 0.0703 ± 0.0047 0.0496 ± 0.0042

RF 0.8704 ± 0.0016 0.0564 ± 0.0003 0.0352 ± 0.0003

SVM 0.8618 ± 0.0005 0.0583 ± 0.0001 0.0353 ± 0.0003

Table 3 shows the optimal MLP, RF and SVM parameters

to maximize performance. The best result obtained by RF has

a number of trees equal to 130, a maximum tree depth equal

to 10, R

of 0.8751, RMSE of 0.0554 and MAE of 0.0348. The

results show that using ensemble methods generates good

results, as the method combines several models.

TABLE III. BEST MODELS

Fig. 2 presents boxplots for the R

, RMSE and MAE

metrics, showing the performance of the RF, SVM and MLP

regression models in predicting the load during the 30

iterations. It can be seen that RF was the model that presented

the best results, with the lowest values for RMSE and MAE

and the highest for R

. The MLP, considering the network

topology adopted, presented some variations in the prediction

of the load variable, resulting in higher values for MAE. For

the MLP, some MAE values were not concentrated like the

other predictions, thus indicating that for some data the

predictions were not correct. Such analysis is appropriate

considering the search range for the parameters used and the

meta-heuristic employed.

Methods Parameters R

RMSE MAE

MLP

hidden_layer_si

zes = (47, 49,

48, 44),

activation =

ReLU

0.8244 0.0656 0.0450

n_estimators:

130,

max_depth: 10

0.8751 0.0554 0.0348

SVM

C = 100,

gamma = 0.1,

epsilon =

0.0229

0.8626 0.0580 0.0353

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

B. da S. Macêdo, C. M. Saporetti,

“Electricity Energy Demand Prediction Using Computational Intelligence Techniques”,

Latin-American Journal of Computing (LAJC), vol. 11, no. 2, 2024.

Fig. 2. Boxplot for R

, RMSE and MAE metrics

Fig. 3 shows a comparison of the true energy load already

normalized with that predicted by the best MLP, RF and SVM

models in the analyzed period. It can be observed that the error

between the predicted and the true is less for the RF comparing

to other models. The time series formed by these values have

similar behavior. This shows that the RF model has the

capacity to assist in the energy load prediction process, which

helps to verify population demand.

Fig. 3. Energy Load True x Energy Load Predict – MLP, RF and Svm

respectively

Fig. 4 illustrates the parameter distribution of the MLP, RF

and SVM models. One can see that for MLP the ReLU was

the activation func chosen in all runs and 4 hidden layers in

the most runs. For RF, the max depth was 10 in all runs and

No. estimators around 130 in 8 runs followed by 200 in 5 runs.

In the SVM case, C equal 100 in all runs, ℽ equal 0.1 and ε

values were well distributed between 0.001 and 0.0275.

Fig. 4.

MLP, RF and SVM parameters distribution

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

10.5281/zenodo.12192271

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

V. CONCLUSION

In this paper, the performance of three regression methods

(MLP, RF, SVM) for forecasting electricity demand in Brazil,

it was analyzed for the algorithms to have good results, the

GWO meta-heuristic was used to improve their performance.

The RF and SVM methods showed the best results. However,

the RF was the one that presented the best result, with R

0.8751, RMSE = 0.0554 and MAE = 0.0348 and, which shows

that using ensemble methods generates good results, by

combining a set of models. For the MLP, some predictions

were not correct, due to the fact that some metric values were

not concentrated. This analysis was carried out considering the

search space for the parameters adopted and the meta-heuristic

used.

Future research includes applying deep learning

techniques, such as Long Short-Term Memory, to evaluate

whether methods considered more robust will perform better

in predicting energy demand. In addition, analyzing the

insertion of other variables that affect daily energy

consumption and that can assist in prediction.

Furthermore, the proposed approach proved to be

effective, as it used cross-validation techniques to enable the

model ability to generalize data from the tests carried out.

Moreover, by using the GWO meta-heuristic, it was possible

to search for the best parameters to maximize the performance

of the regression models, as well as by the use of an ensemble

algorithm that combines multiple models.

EFERENCES

[1] I. P. Marcos and A. P. P. Júnior,” Forecast of Electricity Consumption

in the Northeast Region of Brazil”. Journal of Engineering and Applied

Research. v. 6, n. 3, p. 21-30, 2021.

[2] L. C. Morais, Study on the panorama of electrical energy in Brazil and

future trends. Dissertation - Electrical Engineering. UNESP. 2015.

[3] B., Stearns, F. Rangel, F. Firmino F, Rangel and J. Oliveira, Predicting

performance of enem candidates through socioeconomic data. In:

Proceedings of the XXXVI SBC

Scientific Initiation Paper Competition. SBC, 2017.

[4] Energy research company, Dea Technical Note 22/16, Projection of

electricity demand for the next 10 years (2016 – 2026). 2016.

[5] G. I. S. Ruas, T. A. C. Bragatto, M. V. Lamar, A. R. Aoki and S.

M. Rocco,” Forecasting electrical energy demand using artificial neural

networks and support vector regression”. VI National Artificial

Intelligence Meeting. p. 1262-1271, 2007.

[6] M. F. Alves, Forecasting electrical load demand by stepwise variable

selection and artificial neural networks. Dissertation Electrical

Engineering. UNESP. 2013.

[7] F. Drebes, Electricity demand forecast using artificial intelligence. 2020.

Monograph (Graduation in Electrical Engineering) – University of Vale

do Taquari - Univates, Lajeado, 03 Dec. 2020.

[8] J. F. Schreiber, I. E. M. Kühne, L. A. Destefani, A. T. Z. R. Sausen, M.

Campos and P. S. Sausen, “Intelligent Networks: Data Mining as a

Support Tool for the Analysis of Large Volumes of Data in Underground

Energy Substations”. In: Brazilian Automatic Congress-CBA. 2020.

[9] C. T. Oliveira, Optimization of a shell and tube heat exchanger using the

grey wolf algorithm. Dissertation - Mechanical Engineering. Unisinos.

2015.2018.

[10] G. P. Pizzolato, E. M. dos Santos, A. R. Fagundes, J. O. dosSantos and

H. Hasselein, Optimizing the Operating Time of Overcurrent Relays

Using the Grey Wolf Algorithm. Brazilian Symposium on Electrical

Systems, 1(1). 2020.

[11] P. A. Lachenbrunch, McNemar test, Wiley StatsRef: StatisticsReference

Online. 2014.

[12] R., Kohavi, “A study of cross-validation and bootstrap foraccuracy

estimation and model selection”, Ijcai. Vol. 14. No. 2. 1995.

[13] L. Breiman, Random forests. Machine learning, v. 45, p. 5-32,2001.

[14] A. J. Smola, Learning with Kernels, PhD Thesis. TechnicalUniversity of

Berlin, 1998.

[15] S. Haykin, Neural networks: a comprehensive foundation. Prentice Hall

PTR, 1998.

[16] S. Mirjalili, S. M. Mirjalili and A. Lewis, “Grey wolf

optimizer”.Advances in engineering software, v. 69, p. 46-61, 2014.

[17] B., Stearns, F. Rangel, F. Firmino F, Rangel and J. Oliveira,Predicting

performance of enem candidates through socioeconomic data. In:

Proceedings of the XXXVI SBC

Scientific Initiation Paper Competition. SBC, 2017.

[18] A. Boschetti and L. Massaron, Python data science essentials. Packt

Publishing Ltd, 2016.

[19] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B.Thirion, O.

Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J.

Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrotand and

É. Duchesnay, “Scikit-learn: Machine learning in Python”. the Journal

of machine Learning research, v. 12, p. 2825-2830, 2011.

[20] F. Biscani and D, Izzo, “A parallel global multiobjectiveframework for

optimization: pagmo”. Journal of Open Source Software, v. 5, n. 53, p.

2338, 2020.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2024

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XI, Issue 2, July 2024

AUTHORS

Mestrando em Engenharia de Sistemas e Automação na Universidade

Federal de Lavras. Possui graduação em Engenharia da Computação

pela Universidade do Estado de Minas Gerais e curso técnico

proﬁssionalizante em Informática para Internet pelo Centro Federal

de Educação Tecnológica de Minas Gerais (2018). Tem desenvolvidos

trabalhos na área de Inteligência Computacional.

Doutora em Modelagem Computacional pela Universidade Federal de

Juiz de Fora. Mestre em Modelagem Computacional pela Universidade

Federal de Juiz de Fora. Graduada em Bacharelado em Ciências

Exatas pelo Instituto de Ciências Exatas da Universidade Federal

de Juiz de Fora, em Engenharia Computacional pela Faculdade de

Engenharia da Universidade Federal de Juiz de Fora e em Ciência

da Computação pelo Instituto de Ciências Exatas da Universidade

Federal de Juiz de Fora. Atualmente é professora adjunta do Instituto

Politécnico da Universidade do Estado do Rio de Janeiro e membro

permanente do Programa de Pós-Graduação em Modelagem

Computacional do Instituto Politécnico. Tem experiência na área de

Inteligência Computacional, atuando principalmente nos seguintes

temas: aprendizado de máquina e metaheurísticas.

Bruno da S. Macêdo

Camila Martins Saporetti

B. da S. Macêdo, C. M. Saporetti,

“Electricity Energy Demand Prediction Using Computational Intelligence Techniques”,

Latin-American Journal of Computing (LAJC), vol. 11, no. 2, 2024.