Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

DOI: https://doi.org/10.46296/ig.v8i16.0282

REDES NEURONALES PARA LA PREVISIÓN DE LA DEMANDA EN

SISTEMAS DE DISTRIBUCIÓN PRIMARIA DE ENERGÍA

NEURAL NETWORKS FOR DEMAND FORECASTING IN PRIMARY POWER

DISTRIBUTION SYSTEMS

1

Diaz Abimael ; Pruna Jhon ; Proaño Xavier ; Quinatoa Carlos

1

Department of Electrical Engineering, University Technical of Cotopaxi. Latacunga, Ecuador.

Correo: carlos.quinatoa7864@utc.edu.ec.

Resumen

Se desarrolló un modelo predictivo para analizar los perfiles de carga en redes de distribución

eléctrica mediante redes neuronales artificiales. Se empleó una metodología mixta de análisis de

datos, que combina enfoques cualitativos y cuantitativos. La recopilación de datos se realizó a

intervalos de 15 minutos, generando 5856 registros de kilovatios (kW), kilovoltamperios (kVA) y

kilovoltamperios reactivos (kVAr). El modelo se dividió en un 70 % de datos para entrenamiento

(

4099,2 registros) y un 30 % para validación (1756,8 registros). La arquitectura del modelo consta

de tres capas: una capa de entrada, capas ocultas para análisis y una capa de salida. Se utilizó

un modelo secuencial con una capa de entrada LSTM de 200 neuronas y una capa de salida

densa con una sola neurona. El optimizador Adam y la medida de pérdidas mse se utilizaron

para el entrenamiento. La función de activación utilizada fue lineal. Finalmente, el modelo

demostró un error de pronóstico de ±3,57 % para el modelo 2 y de ±4,38 % para el modelo 3, lo

que lo hace eficaz para la planificación eléctrica.

Palabras clave: Redes artificiales, modelo, LSTM, optimizador, distribución.

Abstract

A predictive model was developed to analyze load profiles in electric power distribution networks

using artificial neural networks. A mixed data analysis methodology, combining qualitative and

quantitative approaches. Data collection was performed at 15-minute intervals, generating 5856

kilowatt (kW), kilovolt-ampere (kVA) and kilovolt ampere reactive (kVAr) records. The model was

divided into 70% data for training (4099.2 records) and 30% for validation (1756.8 records). The

model architecture consists of three layers: an input layer, hidden layers for analysis and an output

layer. A sequential model with a 200 - neuron LSTM input layer and a dense output layer with a

single neuron was used. The optimizer Adam and the loss measure mse were used for training.

The activation function used was linear. Finally, the model demonstrated a forecast error of ±

3

.57% for model 2 and ± 4.38% for model 3, making it effective for electrical planning.

Keywords: Artificial, Networks, Model, LSTM, Optimizer, Distribution.

Información del manuscrito:

Fecha de recepción: 15 de abril de 2025.

Fecha de aceptación: 27 de junio de 2025.

Fecha de publicación: 10 de julio de 2025.

37

Abimael-Diaz et al. (2025)

1

. Introduction

demand, it offers a valuable tool for

strategic planning of electric power

generation [5].

To begin with, it is mentioned that the

efficient management of electricity

demand peaks becomes crucial,

facing challenges in the network.

Strategies such as smart pricing and

anticipatory control, however, are

affected by consumer privacy [1].

The innovative strategy that detects

utility maximization without the need-

to-know specific details of the

consumer's utility function [2] And,

further that the study focuses on

improving the accuracy of day-ahead

electric forecasts through hybrid

approaches such as the combination

of Exponential Smoothing (EMA) and

Long Short-Term Memory networks

The rising cost of energy, driven by

the scarcity of fossil fuels, has

accelerated the search for more

sustainable sources. The integration

of neural networks in electricity

demand

forecasting

provides

accurate and efficient results,

enabling advanced and sustainable

energy planning [6]. The evaluation

of ANN against the Multiple

Regression model (MRA) in daily

forecasting becomes crucial for

Ceilán Electricity Board (CEB)

planning, looking for the optimal

method with minimum Root Mean

Square Error (RMSE),

Mean

(

LSTM) [3]. On the other hand, the

electricity industry is moving towards

competitive environment,

Absolute Percentage Error (MAPE)

and maximum R2 [7]. With more

reason to improve the prediction of

electricity demand using feedforward

neural networks. Parallel-hybrid

evolutionary algorithm Artificial Fish

Swarm Algorithm (AFSA) Particle

Swarm Optimization (PSO) Parallel-

a

abandoning its monopolistic model.

Accurate electricity price forecasting

is crucial to mitigate risks and

effectively

resources.

manage

energy

Subsequently the use of Radial Basis

Function Networks (RBFN), a more

efficient way to train and forecast

hybrid

evolutionary

algorithm

(APPHE), which combines AFSA

and PSO algorithms, is introduced,

showing more satisfactory and faster

performance compared to neural

network training [8].

electrical

loads,

outperforming

previous models based on Artificial

Neural Networks (ANN) [4]. Focusing

on economic factors and long-term

38

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

This is crucial given the depletion of

natural resources and the need to

develop alternative energies. In

forecasting the future distribution of

electricity demand in Taiwan as a

fundamental part of planning for the

efficient use of electricity [9]. It is

worth remembering that the accuracy

of the electric charge is essential in

meteorological conditions. It is

proposed to use a single hidden-

layer artificial neural network

optimized with techniques such as

Particle-swarm-optimization (PSO)

and Genetic algorithm optimization

In order to predict overall demand,

computational intelligence models

are evaluated and compared,

optimizing results through high-

performance

computing

infrastructure [12]. Then electricity

requires management, given the

unpredictability of the electrical load,

daily forecasting is crucial. RNA,

especially the multilayer perception

model, stands out as an effective tool

for this purpose, outperforming

traditional

methods

such

as

Autoregressive integrated moving

average (ARIMA) and Multiple linear

regression (MLR) in speed [13]. It is

necessary to emphasize that electric

power is essential, the importance of

short-term load forecasting in power

systems [8]. RNAs are presented as

a tool for analyzing and predicting

electricity consumption, considering

temporal factors, providing a solution

for three-phase electric power [14].

(GA) to improve the training

convergence and uniqueness of the

solution in strategic electrical

planning [10]. Incidentally, electricity

load forecasting, divided into short,

medium and long term, has

experienced an increase in short-

term forecasting due to the

deregulation of the electricity sector.

The application of methods, such as

Finally, the power grid has evolved

into a smart grid with technologies

such as advanced metering and

artificial intelligence [15]. Digital

twins, supported by statistical and

deep learning techniques, enable

efficient management and accurate

demand forecasting, improving grid

stacked

autoencoders

and

convolutional NRA, promise planning

and cost reduction [11]. In other

words, the energy sector often lacks

predictability, but the use of natural

sources and user demand can

improve in the short term.

39

Abimael-Diaz et al. (2025)

performance in a smart infrastructure

Finally, the study will also focus on

the assessment of the results

achieved and on the comparison with

[

16].

other approaches to analyse the

electrical charge distribution [17].

Table 1. Data base

2

. Materials and Methods

2.1

Previous Data

Data

856

Train

70% of

Validate

30% of 5856

5

The first step in the idea is gathering

freight data in November and

December of 2022. This data will be

processed to use only the specific

information needed for the creation

of the neural network, focusing on

power and time parameters. The

system in Table 1 collects data every

5

856

=

4099.2

= 1756.8

In general, an RNA architecture

composed of sets of neurons (layers)

divided into three groups is

proposed:

-

Input layer for receiving data.

Hidden layers for analysis.

The output layer to issue the

response.

15 minutes, generating a set of 5856

data including kilowatts (kW), kilovolt

amperes (kVA) and kilovolt amperes

reactive (kVAr).

In this research it is described in

Figure 1 below:

Fig. 1: Artificial neural network

40

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

close in time sequence can capture

long-term dependencies, thus

2.2

LSTM Networks

manifesting a type of extended

memory [18].

Long-term and short-term memory

LSTM) constitute a special category

(

within recurrent networks. The

distinguishing feature of recurrent

networks is their ability to retain

information by incorporating loops in

the network structure, essentially

allowing them to remember previous

states and use this information to

determine the next state. This

property makes them particularly

suitable for dealing with temporal

In Figure 2 a recurrent neural

network LSTM is composed of key

components:

the

memory

represented by the values of C, the

outputs represented by H, the inputs

by X, and the resulting outputs by Y.

Specific operations designed to

discard obsolete information, update

the

memory

with

relevant

information, and compute the new

hidden state of the memory are

performed [19].

sequences.

While

conventional

recurrent networks can model short-

term dependencies, then relations

Fig. 2: Schematic types of neural networks

is

accomplished.

To

control

2.3

How to Train a Neural

overfitting, the dropout technique is

applied after every layer. Lastly, a

conclusive probability is provided by

a dense layer with a single unit and

sigmoid activation.

Network

By feeding the input data into a deep

neural network which consists of an

initial 100-unit LSTM layer and a

subsequent 50-unit layer anticipation

41

Abimael-Diaz et al. (2025)

It is crucial to put this network

through a training procedure in order

to get reliable predictions. Training is

developed as an iterative learning

process, where data instances,

specific segments of inputs, are

presented to the network along with

their known correct outputs. Each

instance is introduced individually,

adjusting the weights associated with

the input values at each layer

accordingly [20]. This process is

repeated several times with various

batches of data until the weights are

adjusted well enough to accurately

predict the correct label (or

probability) for any given input

sample.

to define an appropriate model that

satisfies the key aspects to ensure

an accurate assimilation of the

historical data relevant to this study.

In order to assess the created

model's

efficacy

in

properly

predicting demand patterns in

electrical load profiles, an evaluation

of its performance will be conducted

using several error metrics. Figure 3

shows the diagram model. The text

addresses the lack of research on

electrical load profiles on primary

feeders of a specific distribution

structure, highlighting the importance

of understanding and optimizing

performance in this area. The

application of deep learning models

and the use of available data to

develop more effective strategies for

electrical management of the

distribution network is proposed.

2.4

Procedure

The

approach

suggested

to

accomplish the goals centers on

applying artificial intelligence

techniques to create a prediction

model. This approach aims to

improve the forecasting capability of

electric load profiles in electric power

distribution

networks.

The

investigation of artificial neural

network models related to time series

databases is proposed with the

purpose of making predictions of

load profiles in electric power

distribution networks [21]. The aim is

42

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

Fig. 3: Neural network prediction model.

transform the data and build a

The need to examine data to

determine the optimal sampling

frequency and normalize it between

prediction model. The importance of

considering additional variables such

as climatic factors to improve model

accuracy is emphasized. Models

such as Vanilla and Stacked LSTM

with different configurations in layers

and neurons are proposed to identify

the optimal combination to maximize

performance. Subsequently, the

0

and 1 to facilitate its input to the

neural network is highlighted. In

addition, it is suggested to split the

data into training and test sets,

implement additional preprocessing

techniques such as anomaly

detection, and use matrices to

43

Abimael-Diaz et al. (2025)

flexibility of adjusting the number of

iterations for neural network training

is mentioned, highlighting the

computational cost associated with

this process [22]. It is suggested to

explore optimization strategies and

advanced training techniques, such

as transfer learning, to improve

computational efficiency. It is

trend is downward and does not

faithfully reproduce the test set.

Then, the second model is presented

in Figure 5, where it is possible to

notice that it fits more effectively to

the test set, and its trend is very

close.

The third model, which incorporates

a hidden layer and is visible in Figure

suggested

that

efficiency

be

6

, yields a similar result to the second

assessed in relation to variations in

the number of iterations and that the

best possible balance between

computing cost and accuracy be

found. Finally, the evaluation of

models using error metrics such as

MAE, MSE and R² is described, as

well as the possibility of exploring the

visualization of results using time

series plots to identify patterns and

areas of improvement in the

prediction of electrical load profiles.

LSTM model.

Looking at Figure 7, it is clear that in

the fourth model the neural network

is overtrained, which causes several

errors in its predictions. Due to this

reason, this model is excluded from

the analysis. It is essential to

emphasize the importance of

avoiding overfitting in the process of

learning

models,

as

it

can

significantly compromise their ability

to make accurate predictions in real-

world situations.

3. Results

In conclusion, all models can be

observed compared only with the

training data. This is reflected in

Figure 8.

After completing the training of the

LSTM models, we proceed to make

predictions for a period of 7 days and

evaluate their agreement with the

test set. Figure 4 is an example of the

kilowatt variable of the first LSTM

model. Although its performance is

acceptable, it is observed that the

44

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

3.1 Analysis of Results

Fig. 4: Prediction of the first LSTM model

Fig. 5: Prediction of the second LSTM model

Fig. 6: Prediction of the third LSTM model

45

Abimael-Diaz et al. (2025)

Fig. 7: Prediction of the fourth LSTM model

Fig. 8: LSTM model predictions

3

.2 Technical Validation of Results

Absolute Error (EAM) provides

insight into the accuracy of the model

without considering the direction of

the error, while the Mean Squared

Error (ECM) addresses this issue by

squaring the error, thus avoiding

biases in decision making.

After compiling all the models, as

shown in the previous section, it

could be noted that the second and

third models were the best fit to the

time series of load profiles. For this

reason, both models are directly

compared with the test set to analyze

their similarities. This process is

depicted in Figure 9.

In order to make an appropriate

choice of the prediction model,

evaluations are carried out using

various types of errors. The Mean

46

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

Fig. 9: (a) Model 2 LSTM predictions; (b) Model 3 LSTM predictions

Likewise,

the

coefficient

of

errors of each model, it is highlighted

that the prediction of model #3 is

considerably higher, leading to the

choice of the second model with only

one input layer and one output layer.

determination

(R2) provides

information on the variability of the

model and the quality of the

prediction

fit.

The analysis of Table 2 reveals that,

although model #2 presents a lower

EAM compared to model #3, it is

crucial to keep in mind that negative

and positive errors can influence the

choice of model. Consequently, the

ECM suggests that model #3 exhibits

a lower value, indicating a more

accurate fit. In addition, the R2 metric

points out that both predictions are

close to zero, indicating similar

behavior between both models.

When examining the maximum

Table 2. Error analysis

Data

EAM

ECM

R2

Mod

el #2

318.4

5 Kw

347.0

9 Kw

0.3

1

Mod

el #3

343.7

8 Kw

437.7

5 Kw

-

0.0

9

Finally, to verify the probability

accuracy of the two models

investigated, the first 10 samples of

the test set are examined, analyzing

the variability in the predictions, the

percentage of hits for each model.

47

Abimael-Diaz et al. (2025)

Table 3 indicates that model #2

presents a forecast error of ± 3.57%,

while model #3 has an error of ±

3.3 Evaluation of impacts or

results

Once the model has been chosen, it

is trained using the entire data set to

make predictions of future samples.

4.38%. This approach allows a

deeper evaluation of the reliability

and consistency of each model in

terms of predictions, providing

valuable information about its

performance in specific situations of

the test set.

Following

the

same

criteria

established during the development

of this project, the prediction for the

next 7 days of January 2023 is

carried out.

Table 3. Errors in model analysis

In Figure 10, it can be seen that the

Model

Probability of

success

prediction

retains

the

same

Model #2

Model #3

± 3.57%

± 4.38 %

periodicity with similar peaks in

kilowatts, although it exhibits a more

linear trend. This behavior is

attributed to the limitation of the data

set, since its reduced size does not

allow it to capture more accurately

the periodicity of the time series.

Fig. 10: LSTM model predictions

4

. Conclusions

ability to forecast future samples for

electrical planning. This load profile

prediction study has evidenced the

usefulness of LSTM networks;

however, it is crucial to consider a

LSTM networks are capable of

generating predictive models very

close to the test values, ensuring the

48

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

data preprocessing that adapts them

to the needs of the model.

to

process

sequential

data,

understand complex patterns in the

data, and their ability to self-regulate

and adapt to changes in the

environment

Although this project has not

addressed the computational cost

associated with LSTM network

training and search in layer and

neuron selection, it is important to

note that high-performance servers

are required to perform these tasks.

By transforming and processing the

time series database, the data

needed to train deep learning

models, such as neural networks,

were obtained. This involved

adjusting the data to an appropriate

sampling frequency, scaling it

between 0 and 1, and subsequently

splitting it into training and test sets.

In addition, various statistical

methods, such as the Dickey- Fuller

test and the first difference

technique, were applied to evidence

the stationarity of the time series.

These results confirm the need for

the model to be complex in order to

forecast the data accurately. The

choice of using recurrent neural

networks to analyze electrical load

profiles on primary feeders of a

power distribution network appears

to be the most effective option to

achieve a predictive approach. This

is due to the ability of these networks

Acknowledgement:

We would like to express our most

sincere thanks to Xavier Proaño,

Carlos Quinatoa and José Luis

Camacho, who provided us with

valuable help and guidance during

the development of this research.

Their support, knowledge and

willingness were fundamental to

achieving the proposed objectives.

Thank you for sharing your

experience and guiding us through

each stage of the process.

References

[

1]

M. H. Sulaiman and Z.

Mustaffa, “Forecasting solar

power

evolutionary mating algorithm-

deep neural networks,”

Energy and AI, vol. 16, p.

generation

using

100371, May 2024, doi:

10.1016/J.EGYAI.2024.1003

71.

[

2]

W.

Krishnamurthy,

Nonparametric

Hoiles

and V.

“

demand

forecasting and detection of

energy aware consumers,”

IEEE Trans Smart Grid, vol. 6,

49

Abimael-Diaz et al. (2025)

no. 2, pp. 695–704, Dec.

Congress

Computer

on

Human-

Interaction,

2015,

doi:

10.1109/TSG.2014.2376291.

Optimization and Robotic

Applications, Proceedings,

[

3]

K. Theodorakos, O. M.

Agudelo, M. Espinoza, and B.

De Moor, “Decomposition-

Residuals Neural Networks:

Hybrid System Identification

Applied to Electricity Demand

Forecasting,” IEEE Open

Access Journal of Power and

Energy, vol. 9, pp. 241–253,

Institute of Electrical and

Electronics Engineers Inc.,

Dec.

2021.

doi:

1

4

0.1109/HORA52670.2021.9

61186.

[7]

[8]

[9]

S. Karunathilake and H.

Nagahamulla, “Artificial neural

networks for daily electricity

demand prediction of Sri

Lanka,” in Conference: 2017

2

1

2

022,

doi:

0.1109/OAJPE.2022.31455

0.

Seventeenth

International

Conference on Advances in

ICT for Emerging Regions

[

4]

H. Alghamdi et al., “Bayesian

neural networks for solar

power forecasts in advanced

(ICTer), 2017, pp. 1–6. doi:

1

3

0.1109/ICTER.2017.825782

.

thermoelectric

systems,”

Case Studies in Thermal

Engineering, vol. 61, p.

L. Camacho, S. Marrero, C.

Quinatoa, and C. Pacheco,

“Emulation of a PEM Fuel Cell

Stack from its Generic and

1

0

04940, Sep. 2024, doi:

0.1016/J.CSITE.2024.10494

.

Polynomial

Simulink,”

TRANSACTIONS

Model

using

WSEAS

ON

[

5]

P. Chévez and I. Martini,

Applying neural networks for

short and long-term hourly

electricity consumption

“

CIRCUITS AND SYSTEMS,

vol. 23, pp. 92–103, Mar.

forecasting in universities: A

simultaneous approach for

energy management,” Journal

of Building Engineering, vol.

2

1

024, doi:

0.37394/23201.2024.23.9.

S. Li, J. Wang, H. Zhang, and

Y. Liang, “Enhancing hourly

electricity forecasting using

fuzzy cognitive maps with

sample entropy,” Energy, vol.

9

1

2

7, p. 110612, Nov. 2024, doi:

0.1016/J.JOBE.2024.11061

.

[

6]

Y. E. Unutmaz, A. Demirci, S.

M. Tercan, and R. Yumurtaci,

298, p. 131429, Jul. 2024, doi:

10.1016/J.ENERGY.2024.13

1429.

“Electrical Energy Demand

Forecasting Using Artificial

Neural Network,” in HORA

[10] A. Jarndal and S. Hamdan,

“Forecasting of peak

2021

-

3rd International

50

Revista Científica ‘‘INGENIAR”: Ingeniería, Tecnología e Investigación. Vol. 8 Núm. (16) 2025. ISSN: 2737-6249

Neural networks for demand forecasting in primary power distribution systems.

electricity

ANNGA

demand

and ANN-PSO

using

Proceedings - 2017 6th IIAI

International Congress on

approaches,” in Conference:

017 7th International

Conference on Modeling,

Advanced

Applied

2

Informatics, IIAI-AAI 2017,

Institute of Electrical and

Electronics Engineers Inc.,

Nov. 2017, pp. 835–840. doi:

10.1109/IIAI-AAI.2017.89.

Simulation,

Optimization

Sharjah,

and

Applied

(ICMSAO),

United

Arab

Emirates: IEEE, 2017, pp. 1–

[

14] E. Aljarrah, “AI-based model

for Prediction of Power

consumption in smart grid-

smart way towards smart city

using blockchain technology,”

5

1

8

.

doi:

0.1109/ICMSAO.2017.7934

42.

[

11] S. Hosein and P. Hosein,

Load forecasting using deep

“

Intelligent

Systems

with

neural networks,” in 2017

IEEE Power & Energy Society

Applications, vol. 24, p.

200440, Dec. 2024, doi:

10.1016/J.ISWA.2024.20044

0.

Innovative

Technologies

Smart Grid

Conference

(

ISGT), Washington, DC,

[

15] M. Perera, J. De Hoog, K.

Bandara, D. Senanayake, and

S. Halgamuge, “Day-ahead

USA: IEEE, 2017, pp. 1–5.

doi:

1

0.1109/ISGT.2017.8085971.

regional

solar

power

[

12] R. Porteiro and S.

forecasting with hierarchical

temporal convolutional neural

networks using historical

Nesmachnow, “Forecasting

hourly electricity demand of

Uruguay for the next day using

artificial neural networks,” in

power

generation

and

weather data,” Appl Energy,

2020 IEEE PES Transmission

vol. 361, p. 122971, May

and Distribution Conference

and Exhibition - Latin America,

T and D LA 2020, Montevideo,

Uruguay: Institute of Electrical

and Electronics Engineers

Inc., Sep. 2020, pp. 1–6. doi:

2024,

doi:

10.1016/J.APENERGY.2024.

122971.

[

16] X. Xie, A. K. Parlikad, and R.

S. Puri, “A Neural Ordinary

Differential Equations Based

10.1109/TDLA47668.2020.93

6206.

Approach

for

Demand

2

Forecasting within Power Grid

Digital Twins,” in 2019 IEEE

International Conference on

[

13] J. F. Chen and Q. H. Do,

Forecasting Daily Electricity

Load by Wavelet Neural

Networks Optimized by

Cuckoo Search Algorithm,” in

“

Communications,

and Computing Technologies

for Smart Grids

Control,

51

Abimael-Diaz et al. (2025)

(

SmartGridComm), 2019, pp.

10.1016/J.EGYAI.2024.1003

68.

1

–6.

doi:

0.1109/SmartGridComm.20

9.8909789.

[

21] Z. Jiang, Q. Tan, N. Li, J. Che,

and X. Tan, “A novel BiGRU

[

17] A. B. A. Ferreira, J. B. Leite,

and D. H. P. Salvadeo, “Power

substation load forecasting

multi-step

forecasting approach based

on multi-label integration

wind

power

using

interpretable

random

forest

feature

transformer-based temporal

selection and neural network

clustering,” Energy Convers

Manag, vol. 319, p. 118904,

fusion

neural

networks,”

Electric

Power

Systems

Research, vol. 238, p.

Nov.

2024,

doi:

1

9

11169, Jan. 2025, doi:

0.1016/J.EPSR.2024.11116

.

10.1016/J.ENCONMAN.2024.

118904.

[

22] B. A. Romero-Ushiña, K. A.

[

18] C. ge Wu, X. Fu, and Y. Xia,

Spare part demand

Salme-Montaluisa,

C.

I.

“

Quinatoa-Caiza, and J. L.

Camacho-Diaz, “Control de

convertidores formadores de

red con fuente de voltaje

conectado al sistema eléctrico

de distribución balanceada,”

Revista Científica INGENIAR:

forecasting using PSO trained

Quantile Regression Neural

Network,” Comput Ind Eng,

vol. 200, p. 110841, Feb.

2

1

025, doi:

0.1016/J.CIE.2024.110841.

Ingeniería,

Investigación. ISSN: 2737-

Tecnología

e

[

19] T. Peng et al., “Railway cold

chain freight demand

forecasting with graph neural

6249., vol. 8, no. 15, pp. 146–

167, May 2025, [Online].

networks:

A

novel

Available:

GraphARMA-GRU

model,”

https://www.journalingeniar.or

g/index.php/ingeniar/article/vi

ew/285

Expert Syst Appl, vol. 255, p.

1

3

24693, Dec. 2024, doi:

0.1016/J.ESWA.2024.12469

.

[

20] A. Miraki, P. Parviainen, and

R. Arghandeh, “Electricity

demand at

distribution and household

levels using explainable

forecasting

causal graph neural network,”

Energy and AI, vol. 16, p.

100368, May 2024, doi:

52