ISSN: 2320-5407 Int. J. Adv. Res.

12(04), 01-17

Journal Homepage: -www.journalijar.com

Article DOI:10.21474/IJAR01/18512
DOI URL: http://dx.doi.org/10.21474/IJAR01/18512

RESEARCH ARTICLE
FORECASTING THE CONSUMPTION OF ELECTRICAL ENERGY ON THE CEB ELECTRICAL
NETWORK IN LOMÉ IN TOGO BY APPROACHES: SIMPLE LINEAR REGRESSION, RECURRENT
NEURAL NETWORKS AND GENETIC ALGORITHMS

Komla Kpomonè Apaloo Bara1,2

1. Department of Electrical Engineering, Polytechnic School of Lomé (EPL), University of Lomé, Togo.
2. Engineering Sciences Research Laboratory (LARSI), Polytechnic School of Lomé (EPL), University of Lomé,
Togo.
……………………………………………………………………………………………………....
Manuscript Info Abstract
……………………. ………………………………………………………………
Manuscript History Forecasting the electrical power to be consumed requires planning
Received: 05 February 2024 production on several levels. In this work we used data from the
Final Accepted: 08 March 2024 Electric Community of Benin. Temperature, relative humidity, wind
Published: April 2024
speed, normal direct irradiance, precipitation and diffuse radiation are
Key words:-
the meteorological variables that made it possible to analyze the
Consumption, Electrical Energy, Genetic forecasts. The objective is to do learning with genetic algorithms,
Algorithms, Meteorological Variables, LSTM recurrent neural networks and simple linear regression after a
Recurrent Neural Networks. Simple characterization and a correlation study and then to submit the results to
Linear Regression
performance evaluation criteria. The results of the characterization
made it possible to understand that certain variables are significant and
influence the consumption of electrical energy. The study of the
correlation gives 94% between direct normal irradiance and diffuse
irradiance. Both give with the temperature 67% for one and 68% for the
other. Regarding modeling, the results are bad with genetic algorithms
if we take into account the correlation coefficient (R² = 28.84%), good
with simple linear regression (R² = 69.08%) and very interesting for
networks of recurrent neurons where we find: MAE = 0.11; MSE =
0.02; MAPE = 18.50%; RMSE = 13.09%; RRMSE = 18.25% and R² =
96.11%. Given these results, we deduce that short- and long-term
memory recurrent neural networks (LSTM) are very well suited to
predicting the electrical power consumed on the CEB electrical
network.

Copy Right, IJAR, 2024, All rights reserved.

……………………………………………………………………………………………………....
Introduction:-
The information and communication techniques that revolutionized the world were only possible thanks to the
development of the electricity sector. The latter is now present in all areas of activity. Electricity coverage is part of
the criteria for classifying countries in terms of their development in the world as well as on an industrial level, [1-
3]. Primary energy sources (fossil, renewable) capable of being transformed into electrical energy, to power
industries, exist today everywhere and in several forms, [4-9]. Considering that fossil sources contribute enormously
to the destruction of nature, international conferences are focusing on the subject by advising the production of
electrical energy based on renewable sources, [11, 12]. This is to reduce greenhouse gases, [13,14], which create
climate disruption, [15-16], such as heatwaves, melting glaciers, unexpected floods, etc., [17-19].

Corresponding Author:- Apaloo Bara Komla Kpomoonè 1

Adress:- Department of Electrical Engineering; Polytechnic School of Lomé, University of
Lomé.
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

However, comfort and the search for ease in daily activities mean that the demand for electrical energy continues to
grow exponentially, [20-21], but the storage of electricity is insignificant. The important observation that the
quantity of electricity produced throughout the world cannot be stored for later use deserves our serious attention.
To do this, it must be consumed as it is produced. It will then be necessary to find a balance between production and
consumption in order to avoid waste of available primary energy sources.

Considering all this, energy production companies do everything to provide only the instantaneous electricity
necessary for consumption, [22-23]. Knowing that electricity consumption is very uncertain from one region to
another and from household to household, the work has provided short- and long-term solutions, [24-25]. From
there, many studies around the world have focused on the subject [26-31]. The results of this research have
contributed to the reliability of electricity networks by eliminating energy waste by adapting production to
consumption. It is time to seek to apply it here (in Togo), particularly in our electrical networks. Togo, a country in
humid and coastal West Africa, also has visions of major electrification projects. In the country, energy transport is
provided by the Community Electricity of Benin (CEB), [32], and distribution by the Compagnie d’Energie
Electrique du Togo (CEET), [33]. Thus, to see the national development plan of the country, [34], drawn up by the
government, with regard to electricity coverage, it is necessary that the same solutions are also applied for the
adequacy between production and consumption taking into account energy poverty. Which becomes a complex
problem to solve.

Solving complex problems today involves systems modeling [35-40]. Among the means of modeling, Artificial
Intelligence has become the best following statistical and stochastic methods, [39], [41-42]. Through its algorithms
(Artificial Neural Network, [25], |40], Neuro-fuzzy Inference System, [40], Vast Margin Separator, [40],
Autoregressive Integreted Moving Average, [24], [43], etc. .), it finds almost real models starting from random or
concrete phenomena with even a few sometimes insignificant errors, [25], [40], [43], [44], [45]. Given that weather
forecasts have become more and more precise, it would be wise to use them in forecasting electrical energy needs in
order to monitor their similarity.

The objective of this work is to exploit certain meteorological variables such as temperature, wind speed, relative
humidity, normal direct irradiance, precipitation and diffuse radiation as input variables then the power consumed on
certain CEB sites in Lomé to predict its need for electrical energy. The meteorological variables will be extracted
from the site https://open-meteo.com, an API (Industrial Programmable Logic Controller) which is an open source
weather forecast and offers free access for non-commercial use. We will then exploit simple linear regression,
Genetic Algorithms and Recurrent Neural Networks to learn the models. The goal of this work is to submit the
results to the performance evaluation criteria most used in the literature in order to observe the behavior of the
models to check whether they are favorable or not. Among these criteria, we will use here the mean absolute error
(MAE), the mean square error (MSE), the mean absolute error in percentage (MAPE), the square root of the mean
absolute error (RMSE), the square root of the relative root mean square error (RRMSE) and the Correlation
Coefficient R2. At the end of the work, the conclusions will allow us to confirm or refute, if the algorithms are well
chosen or from the comparisons, which are adapted to the chosen variables a good forecast.

Materials and Methods:-

To develop this work, we will use CEB operating data. The powers consumed during the years from 2019 to 2021,
and from January to December will be used. Generally, this data is recorded automatically on the network through
an Excel file, the appearance of which is shown in Figure 1. Given that the power consumed data is not recorded
over and over with the meteorological variables and that this power consumed is not recorded for a single location
but in a space (Lomé), we will look for the meteorological variables on the site https://open-meteo.com. This is an
Open-Meteo, an open source weather API that provides free access for non-commercial use. On this site, only the
following variables will be used among others:
1. daily gridded dates: YEARS months and days of datatime type;
2. temperature: T (°C) of float type;
3. relative humidity: H (%) of float type;
4. wind speed: V (km/h) of type int64;
5. normal direct irradiance: DNI (W/m²) of float type;
6. precipitation: P (mm) of float type;
7. diffuse radiation: RD (W/m²) type int64.

2
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

First a graphical characterization will be carried out on all the variables. Parameters such as: annual mean, mode,
median, standard deviation, maximum, minimum, asymmetry coefficient and flattening coefficient will be
calculated for each variable in order to judge their correlation. This part will be materialized graphically by a heat
map of correlation between all the variables. Then, we will use Simple Linear Regression, Genetic Algorithms and
Recurrent Neural Networks to perform the learning. The forecast results will be subject to performance evaluation
metrics.

Simple linear regression [25], [46]

We only introduce here the notion of linear model through its most basic expression.After having explained the
necessary hypotheses and the terms of the model, we will present the notions of estimates of the model parameters
and forecasting using the confidence interval.Finally, particular attention will be paid to influential values. In
modeling if we note Y, the real random variable to be explained (endogenous, dependent or response variable) and
X, the explanatory variable or fixed effect (exogenous);the model amounts to assuming that, on average, E(Y), is an
affine function of X. Writing the model implicitly assumes a prior notion of causality in the sense that Y depends
onsymmetrical.Relations (1) or (2) can be considered.

E (Y )  f ( X )  0  1 X (1)

Where

Y  0  1 X   (2)

We will assume for simplicity that X is deterministic. Otherwise, if X is random, the model is then written
conditionally on the observations of relation (3) and leads to the same estimates.

X : E (Y / X  x)  0  1x (3)

The assumptions relating to this model are as follows:

1. the error distribution  is independent of X where X is fixed,
2. the error is centered and of constant variance (homoscedasticity): i  1,..., n E ( i )  0 , Var ( i )   2 ,

3. 0 and 1 are constant, no break in the model,

4. Additional hypothesis for inferences:  N (0, )
2

The estimation of the parameters 0 , 1 and  are obtained by maximizing the likelihood, under the hypothesis that
2

the errors are Gaussian, or by minimizing the sum of the squares of the differences between observations and model
(least squares). The two approaches lead to the same estimates while maximum likelihood leads to better properties
of the estimators. For a sequence of observations f ( xi , yi ) , {( xi , yi )i  1,..., n}
, the least squares criterion is written
by the expression of the relation (4):
n
Min   yi  0  1 xi 
2

0 , 1
i 1 (4)

We pose:

1 n
 xi
1 n
x
n i 1
y  yi
n i 1
; ;
2

 
2
1 n
s 
2
x  xi  x
n  1 i 1
s 2
y
1 n

 yi  y
n  1 i 1

;
S xy
sxy 
1 n

 xi  x yi  y
n  1 i 1
  r
SxS y
;

Least squares are minimized by relations (5) and (6):

3
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

S xy
b1 
S x2 (5)

b0  y  b1 x (6)
Which are the realizations of the estimators 0 and 1 . We show that these estimators are unbiased and have
minimum variance among the linear function estimators yi (resp. among all the estimators in the Gaussian case).
Each value of X corresponds to the estimated or adjusted value of Y: yi  b0  b1xi The calculated or estimated
residuals are presented by the relation (7) and the variance  is estimated by the residual variation (8):
2

ei  yi  yi (7)

1 n 2
2   ei
n  2 i 1 (8)

Genetic algorithms [47-50]

Genetic algorithms are based on the idea of biological evolution and use the concept of genetics to solve
optimization problems. Figure 1 illustrates the operating principle of the genetic algorithm. In order to take full
advantage of the potential of a genetic algorithm, we followed each step of its flowchart methodically and
efficiently.

Beginning
j  j 1

Load the population

base

Evaluation

Operation
(Crossover,
Mutation)

No
Complete treatment

Yes

Results

End

Figure 1:- Functional flowchart of the genetic algorithm.

4
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Genetic algorithm flowchart:

The main formula used in genetic algorithms is the fitness function which measures the quality of each candidate
solution. Indeed, AGs work on the maximization of an objective function. In many problems, the objective is to
minimize a given function. We thus transform this objective function into an adaptation function defined by relation
(9).
f a ( x)  f max ( x)  f ( x) (9)
Where : f max ( x)  max( f ( x))
In the case where the objective function to be maximized takes negative values, relation (9) is replaced by relation
(10).

f a ( x)  f ( x)  f min ( x) (10)

where : f min ( x)  min( f ( x)) .

The adaptation function has intrinsic characteristics that must be exploited to best effect when adapting to the
relevant research domain. This allowed us to take into account the evaluation strategy, the selection method, and the
type of operation which takes into account mutation and crossing. Because of this characteristic we adapt it here to
the forecasting problem despite it being an optimization algorithm. The operation of the genetic algorithm also arises
from its flowchart structure, as is the case for any algorithm.

Short and long-term memory (LSTM) recurrent neural network [51]

Recurrent Neural Networks (RNN) are powerful machine learning models that allow you to analyze sequences of
data, such as text, speech or time series. These networks allow machines to “remember” past information and use it
to make real-time decisions. In this article we will discover how RNNs work. RNNs sometimes fail to resolve
certain problems; For this they need to have a stronger memory. We do this by making the neurons more complex.
In particular, we use LSTM (Long Short-Term Memory).

Long Short-
Term Memory (LSTM), more explicitly a recurrent neural network with short-term and long-term memory, is a
particular type of recurrent neural network (RNN). In an RNN, each neuron or processing unit is capable of
maintaining the internal state or memory of retaining information associated with the previous input. This feature is
particularly important in many time series applications. The main idea behind this type of neural network
architecture is the consideration of time. The name of this neural network is derived from the fact that these types of
networks operate recursively. An operation is performed for each element of a sequence whose output depends on
the current input and previous operations. This is accomplished by reusing an output from the network at time t with
the network input at time t + 1 (i.e. the output from the previous step is combined with the new input in the new
floor). These cycles allow the existence of information from one stage to the next.

LSTM is considered the most suitable model for forecasting stochastic and noisy data such as time-dependent
consumption data. This use is consistent with the ability to deduce peaks and oscillations from consumption curves
as a function of time.

Figure 2:- Internal architecture of a unit of the LSTM short-term and long-term memory recurrent neural network
[51].

5
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Performance evaluation criteria [24], [25], [40], [45]

Statistical measurements that represent key performance indicators provide information on the performance and
behavior of a given model. Among these measures, we will use the mean absolute error (MAE), the mean square
error (MSE), the root mean square error (RMSE), the Normalized root mean square error (NRMSE), the mean error
absolute percentage (MAPE), the correlation coefficient (R²). RMSE is the most widely used metric for comparing
different classification models, giving a simple and transparent quantitative measure of the difference between the
input and the target [24], [35], [40]. MAPE is one of the most used metrics to compare forecast models. The
NRMSE [40] is calculated as the RMSE divided by the range of the observed values, expressed as a percentage. The
range of the observed values is the difference between the maximum and minimum values of the observed data. R²
studies the correlation between two random (or statistical) variables; that is to say the intensity of the connection that
can exist between these variables. The connection sought is an affine relation. It is best when it is close to 100%.
The mathematical representations of these different indicators are given by equations (11), (12), (13), (14), (15),
(16):
1 N (11)
MAE 
N
p
j 1
pi  pmi
2

MSE 
1 N

 p pi  pmi
N j 1
 (12)

1 N p pi  p pmi (13)
MAPE   p 100
N j 1 pi

2
(14)
RMSE 
1 N
 p p  pmi
N j 1 i
 
2
1 N

 p pi  pmi
N j 1
 (15)
RRMSE  N

p j 1
mi

 p  
N
 pm p * pmi  pmm
(16)
pi
j 1
R 
2

N   2
   *   p 
N

  p pi  pm p
2
mi  pmm 
 j 1 j 1 
Where:
 p p is the predicted power;
i

 pmi is the measured power;

 pmp is the predicted average power;
 pmm is the average power measured;
 N is the number of points sampled.

Results:-
The results of the characterization are presented as follows: Table 1 summarizes the statistical variables of each
parameter by year and the figures going from 3 to 10 show a graphical view of the evolution of each variable during
the study period.

6
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Figure 3:- Energy production from 2019 to 2021 on an hourly basis.

Figure 4:- Temperature variation from 2019 to 2021 on an hourly basis.

7
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Figure 5:- Variation in relative humidity from 2019 to 2021 on an hourly basis.

Figure 6:- Variation in wind speed from 2019 to 2021 on an hourly basis.

Figure 7:- Variation in direct irradiance from 2019 to 2021 on an hourly basis.

8
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Figure 8:- Variation in precipitation from 2019 to 2021 on an hourly basis.

Figure 9:- Variation in diffuse radiation from 2019 to 2021 on an hourly basis.

Figure 10:- Hourly, daily and weekly power averages from 2019 to 2021.

9
 ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Figure 11:- Correlation table between the study variables.

The results of the statistical characterization are summarized in Table 1 and Table 2 shows the summary of the
performances obtained by the algorithms

Table 1:- Summary of the characterization of the study variables.

Years Variables Mean Mode Std Median Min Max Skewness Kurtosis
Power : (MW) 115.579 88.77 20.549 115.285 4.14 171.37 -0.096 -0.099
Temperature : T (°C) 27.624 27.7 27.4 2.120 23.3 33.2 0.393 -0.638
Relative humidity : (%) 83.514 91 8.035 86.0 50 97 -0.890 0.273
2019 Wind speed : (km/h) s 4.88 1.425 4.6 0.0 8.5 -0.381 -0.153
Normal direct irradiance : (W/m²) 195.166 0.0 251.656 0.0 0.0 865.0 0.894 -0.677
Precipitation : (mm) 0.1258 0 0.514 0.0 0 11.1 8.400 98.196
Diffuse radiation: (W/m²) 140.646 0 208.040 0.0 0 827 1.322 0.455
Power : (MW) 110.987 104.2 7.531 111.3 4.4 154.2 -0.2798 -0.141
Temperature : T (°C) 27.481 27.2 2.258 27.3 21.2 33.9 0.304 -0.426
Relative humidity : (%) 82.422 91 9.962 85.0 27 98 -1.734 4.713
2020 Wind speed : (km/h) 4.717 5.02 1.430 4.85 0.14 8.9 -0.384 -0.017
Normal direct irradiance : (W/m²) 190.314 0.0 248.746 0.0 0.0 879.1 0.932 -0.598
Precipitation : (mm) 0.162 0 0.718 0.0 0 15.5 10.016 135.165
Diffuse radiation: (W/m²) 137.182 0 205.545 0.0 0 854 1.366 0.597
Power : (MW) 124.640 130.0 19.334 126.8 2.4 174.1 -0.564 0.663
Temperature : T (°C) 27.726 27.9 1.985 27.5 22.8 32.7 0.273 -0.696
Relative humidity : (%) 82.308 88 8.129 84.0 32 95 -0.972 1.342
2021 Wind speed : (km/h) 4.580 4.88 1.388 4.71 0.22 8.39 -0.393 -0.118
Normal direct irradiance : (W/m²) 202.542 0.0 263.322 0.0 0.0 902.9 0.928 -0.592
Precipitation : (mm) 0.095 0 0.4756 0.0 0 12.6 11.211 183.569
Diffuse radiation: (W/m²) 145.195 0 216.739 0.0 0 874 1.375 0.662

The learning results across the performance criteria are gathered in Table 2 and the graphical evolution of the test
results in relation to learning are observed through the figures ranging from 12, 13 and 14.

10
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Table 2:- Learning results obtained by performance criteria.

Performance evaluation criteria
MAPE RMSE RRMSE R2
Algorithms Used MAE MSE
(en %) (en %) (en %) (en %)
Simple Linear Regression 12.01 233.74 17.45 20.14 17.21 69.08
Genetic Algorithms 27.98 141668.92 17.45 20.14 17.21 68.08
Recurrent Neural Networks 0.11 0.02 18.50 13.09 18.25 96.11

Figure 12:- State consumption prediction with the recurrent neural network.

Figure 13:- State of prediction of electricity consumption by genetic algorithms.

11
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

Figure 14:- Data prediction by simple linear regression.

Analysis and Discussions:-

The “Energy Futures 2050” describe possible paths to transform the electricity system with a view to achieving
carbon neutrality. Over the time horizons considered, the evolution of the electricity mix and the climate reinforce
dependence from the balance of the system to climatic hazards and constitutes a crucial element to take into account
in planning the energy system. Of all meteorological variables, this work takes into account temperature, relative
humidity, wind speed, normal direct irradiance and precipitation to predict the active power to be consumed in a
one-hour interval. The variables in the study are observed over a period of three years from 2019 to 2021. We
observe that the highest temperature is 33.9°. The distribution of its values is stable and Figure 4 shows us that it is
almost uniform each year. Observation confirmed through table XX where its average value is 27.624° in 2019;
27.481° in 2020 and 27.726° in 2021 then reinforced by the asymmetry coefficients which remains around 0.
Through its statistical behavior, we can declare it well chosen for predicting the power to consume. The same
observation can be made regarding relative humidity (figure 5). However, if we observe this figure correctly, we
find extremely low values occasionally. This is 27% in 2020 and 32% in 2021 while the minimum in 2019 is 50%.
Also its average is 83.514% in 2019; 82.422% in 2020 and 82.308 in 2021 showing its stability on a statistical level
and capable of being considered for the prediction of electricity consumption; but the minimum values observed
mean that an algorithm is needed to properly carry out the study of the forecast. Likewise, the wind speed also
responds to the observations made in relation to the forecast of the power to be consumed because its statistical
values are as follows: average 08 km/h in 2019; 4,717 km/h and 4,580 km/h. The minimum values are: zero in 2019;
0.14 km/h in 2020 and 0.22 in km/h. The skewness is almost around zero, a normalized value showing the validity
of the observation. On the other hand, climate change, the reality of which is no longer in doubt today and which is
likely to continue over the coming decades, we must question the resilience of the electricity system's infrastructure
to its effects. In order to propose robust electricity mix options in the face of the effects of climate change, the
prospective of the energy system must therefore necessarily integrate these effects into the study of the functioning
of the system in the long term. We can mention normal direct irradiance, precipitation and relative humidity. For
direct normal irradiance, we have: 195.166 w/m² in 2019; 190,314 w/m² in 2020 and 202,542 w/m² as average. The
values are zero in mode, median and minimum showing its very random statistical nature. Precipitation is sometimes
zero in mode, median and minimum observable in Table 1 in 2020 against maximums which go up to 15.5 mm in
2020; 11.1 mm in 2019 and 12.6 mm in 2021. We also find the observed skewness: 10.016 in 2020 shows the
unequal statistical distribution not leading to a conclusion for a forecast based on statistics. Observations also made
on diffuse radiation where the averages are: 140.646 w/m² in 2019; 137.182 w/m² in 2020 and 145.195 w/m² against
the zero mode, median and minimum in 2021. It is the same for the years 2019 and 2020.

12
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

The analysis was continued further by studying the correlation between the variables. In fact, correlation exclusively
measures the more or less linear character of the cloud of points, in other words the quality of the linear relationship
or the degree of linear dependence between the variables. Its properties, [53], [54], are as follows:
 -1 ≤ r ≤ 1 (or |r| ≤ 1)
 r is a dimensionless indicator;it is not sensitive to the units of each of the variables
 |r| = 1 is equivalent to the existence of an exact linear relationship
 If the variables are not independent r = 0, but the converse is not true in general: non-correlation is not
necessarily independence, while independence necessarily leads to non-correlation
 note finally that the correlation is not transitive

Figures 15:-Show the different distributions of the data for assigned values of the correlations.

a- Correlation = 1 b- Correlation close to 1

c- Correlation = 0 d- Correlation < 0

Figure 15:- Different distributions of data depending on correlations.

Figure 16, thermal correlation table shows the different values of the links between the variables. Between direct
normal irradiance and diffuse irradiance, it is 94% showing a good correlation between them. Both always have
good values with temperature and vary between 67% for one and 68% for the other. It is not at all good between
precipitation and normal direct irradiation (04%) then temperature and precipitation (06.7%)

Taking into account observations, to name just a few, led us to explore some algorithms in this work such as genetic
algorithms, recurrent neural networks and simple linear regression. The results of the studies, in order to properly
judge them, are subject to several performance evaluation criteria. Here, we cite: MAE, MSE, MAPE, RMSE,
RRMSE and R². We observe in this regard that the results are not at all good for genetic algorithms because the
MSE goes up to 141668.92, a very high value for the square of an error. For simple linear regression, it is not as
favorable especially since we find that its value 233.74, still large. On the other hand, for recurrent neural networks
we find an extremely interesting result because the MSE is only 0.02. Thus, the modeling results are presented as:
MAE = 27.98; MSE = 141668.92; MAPE = 12.88%; RMSE = 16.14%; RRMSE = 14.32% and R² = 28.84% for
genetic algorithms. We have for simple linear regression, MAE = 12.01; MSE = 233.74; MAPE = 17.45%; RMSE =
20.14%; RRMSE = 17.21% and R² = 69.08%. Finally, for recurrent neural networks, we find: MAE = 0.11; MSE =
0.02; MAPE = 18.50%; RMSE = 13.09%; RRMSE = 18.25% and R² = 96.11%. Given these results, we deduce that
short- and long-term memory recurrent neural networks (LSTM) are very well suited to predicting the electrical
power consumed on the CEB electrical network.

Conclusion:-
The fight against global warming requires the electrification of energy users and the abandonment of carbon-based
production sources. It is therefore necessary that the production and supply of electrical energy be carried out while
respecting short-term efficiency constraints, but also by providing reliable signals to guide the investments of
producers to balance production with consumption, not only under normal macroeconomic conditions but also by
using new forecasting methods in order to find a reliable solution. Today, it is artificial intelligence that provides
solutions to complex situations. In the work combined in this manuscript, we used hourly consumption data
collected by the CEB over the years 2019 to 2021 to predict the electrical energy needs in the network. To carry out
this study, we used meteorological variables whose forecasts have become increasingly precise. These are

13
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

temperature (T in °C), relative humidity (H in %), wind speed (V in km/h), normal direct irradiance (DNI in W/m²),
precipitation (P in mm) and diffuse radiation: (RD in W/m²). As methods, we first carried out a characterization of
all the variables by calculating the statistical parameters (mean, mode, standard deviation, median, minimum,
maximum, skewness and kurtosis). Then we carried out a correlation study on all the variables in order to monitor
their relevance on the electrical energy consumption to be predicted. Genetic algorithms, simple linear regression
and LSTM recurrent neural networks made it possible to learn the models. Learning outcomes are subject to
performance evaluation criteria such as: MAE, MSE, MAPE, RMSE, RRMSE and R².

The results made it possible to understand that certain variables (temperature, wind speed and precipitation) are
significant and have a good statistical distribution confirming that they have an influence on the consumption of
electrical energy in relation to their statistical characteristics. Which is not the case for the others (relative humidity,
direct normal irradiance and diffuse irradiance). The study of the correlation confirmed this behavior of the
modeling by its values. Between direct normal irradiance and diffuse irradiance, it is 94%. Both give with the
temperature 67% for one and 68% for the other. Contrary, between precipitation and normal direct irradiation we
have 04% then between temperature and precipitation we find 06.7%. Furthermore, the results of the modeling are
as follows: MAE = 27.98; MSE = 141668.92; MAPE = 12.88%; RMSE = 16.14%; RRMSE = 14.32% and R² =
28.84% for genetic algorithms. We have for simple linear regression, MAE = 12.01; MSE = 233.74; MAPE =
17.45%; RMSE = 20.14%; RRMSE = 17.21% and R² = 69.08%. Finally, for recurrent neural networks, we find:
MAE = 0.11; MSE = 0.02; MAPE = 18.50%; RMSE = 13.09%; RRMSE = 18.25% and R² = 96.11%. Given these
results, we deduce that short- and long-term memory recurrent neural networks (LSTM) are very well suited to
predicting the electrical power consumed on the CEB electrical network. It would be necessary to test these results
using data from the years felt (from 2022 to 2023) and explore other algorithms to confirm and/or refute the choice
made in this work.

References:-
1. Jacques Attali : Rapport de la Commission pour la libération de la croissance française Xème Édition, La
Documentation française, 2008
2. Claude E. Delisle, Jean-Pierre Revéret : L'évaluation des impacts sur l'environnement : processus, acteurs et
pratique ...De Pierre André, 2ème édition, Ecole polytechnique de montréal, 2003
3. De Paul Bairoch, Bouda Etemad, Structure par produits des exportations du tiers-monde : 1830 – 1937 Genèse
librairie Droz, 11, rue Massot, 1985
4. Rabiou Bounou Madja1 and al. : Evaluations de la biomasse fourragere foliaire et de la composition chimique
des feuilles d’un ligneux fourrager sahelien : maerua.crassifolia forssk, , Int. J. Adv. Res. 11(12), ISSN : 2320-
5407, 879-889, December 2023
5. Maria Rodriguez Gamez and al : Renewable Energy Sources and Local Development, International Journal of
Social Sciences and Humanities Available online, Vol. 1 No. 2, pages: 10~19 e-ISSN: 2550-7001, p-ISSN:
2550-701X , August 2017, http://dx.doi.org/10.21744/ijssh.v1i2.31
6. Leon Freris David Infield : Les énergies renouvelables pour la production d’électricité, 2e ÉDITION, © Dunod,
11, rue Paul Bert, 92240 Malakoff www.dunod.com ISBN 978-2-10-082454-0 Paris, 2021
7. Brockway, P.E., Owen, A., Brand-Correa, L.I. et al. Estimation of global final-stage energy-return-on-
investment for fossil fuels with comparison to renewable energy sources. Nat Energy4, 612–621 (2019).
https://doi.org/10.1038/s41560-019-0425-z, July 2019
8. Vladimir S. Arutyunov and Georgiy V. Lisichkin : Energy resources of the 21st century: problems and forecasts.
Can renewable energy sources replace fossil fuels© 2017 Russian Academy of Sciences, Turpion Ltd and IOP
Publishing Ltd Russian Chemical Reviews, Volume 86, Number 8 Russ. Chem. Rev.86 777 DOI
10.1070/RCR4723, 2017
9. Melle MEKKI MOUNIRA, Récupération des déperditions d’énergie dans les complexes industriels et leur
conversion en énergie électrique exploitable, THESE de DOCTORAT en Sciences Spécialité : Electrotechnique
Option : Réseaux électrique, UNIVERSITE BADJI MOKHTAR ANNABA, Faculté des sciences de l’ingénieur
Département d’électrotechnique, 2014
10. Abdalfettah Aljedk SURVEY OF SOLAR ENERGY USAGES & ITS APPLICATIONS IN LIBYA,Int. J. Adv.
Res. 8(12), 346-351, SSN: 2320-5407, December 2020
11. La COP 27, Conférence des Nations unies sur les changements climatiques 2022, 6 nov - 20 nov 2022, Sharm
el-Sheikh Climate Change Conference - November 2022

14
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

12. COP28 refers to the 28th session of the Conference of the Parties to the United Nations Framework Convention
on Climate Change (UNFCCC), which convenes from 30 November to 12 December, 2023 in the United Arab
Emirates
13. Nguyen, B.T., Dong, H.P., Le, L.B. et al.Contrasting emissions of carbon-based greenhouse gases from two
paddy soils under submerged conditions as affected by biochar addition. Environ Earth Sci 82, 177 (2023).
https://doi.org/10.1007/s12665-023-10874-7, 27 March 2023
14. Osama Asanousi Lamma, The impact of recycling in preserving the environment, International Journal of
Applied Research; 7 (1 1 ): 297 -3 02, IJAR 2021; 7(11): 297-302, 2021
15. Changes proposed to be made to the underlying assessment (except for the
Technical Summary, which is separately submitted as IPCC-XVII/Doc. 3c, Add. 1), for consistency with the
approved Summary for Policymakers (ref. IPCC-XVII/Doc. 3c), and accepted by the Working Group at its
Sixth Session (Geneva, 13-16, February 2001)
16. Ravindranath, N.H., Sathaye, J.A. Climate Change and Developing Countries. In: Climate Change and
Developing Countries. Advances in Global Change Research, vol 11. Springer, Dordrecht. (2002).
https://doi.org/10.1007/0-306-47980-X_9
17. Ghulam Rasul1, Qin Dahe2, Q. Z. Chaudhry, GLOBAL WARMING AND MELTING GLACIERS ALONG
SOUTHERN SLOPES OF HKH RANGES, Pakistan Journal of Meteorology, Vol. 5 Issue 9 : July 2008
18. Margreth Keiler , Jasper Knight and Stephan Harrison : Climate change and geomorphological hazards in the
eastern European Alps ,Published:28 May 2010 https://doi.org/10.1098/rsta.2010.0047
19. Ghulam Rasul, Q. Z. Chaudhry, Arif Mahmood, Glaciers and Glacial Lakes under Changing Climate in
Pakistan, Pakistan Journal of Meteorology, Vol. 8, Issue 15
20. Douglas Le Comte : International Weather Highlights 2021 : Heat Waves, and Wildfires Epic Floods,:, Pages
24-31 | Published online: 21 Apr 2022, https://doi.org/10.1080/00431672.2022.2042131
21. Antoine Boubault, La production mondiale d'électricité : une empreinte-matière en transition, Article produit
pour la rubrique Ecomine du site Minéralinfo, http://www.mineralinfo.fr/ecomine/production-mondiale-
delectricite-empreinte-matiere-en-transition, 21/12/2018
22. Li Zhang, Tim Brown, G. Scott Samuelsen, Fuel reduction and electricity consumption impact of different
charging scenarios for plug-in hybrid electric vehicles, Journal of Power Sources, Volume 196, Issue 15, Pages
6559-6566, ISSN 0378-7753, 2011, https://doi.org/10.1016/j.jpowsour.2011.03.003
23. Kris R Voorspools, William D D'haeseleer, The influence of the instantaneous fuel mix for electricity
generation on the corresponding emissions, Energy, Volume 25, Issue 11, Pages 1119-1138, ISSN 0360-5442,
2000, https://doi.org/10.1016/S0360-5442(00)00029-3.
24. Agbessi, A. P and al., 2023.Peak Electrical Energy Consumption Prediction by ARIMA, LSTM, GRU,
ARIMA-LSTM and ARIMA-GRU Approaches Energies 2023, Advanced Optimization and Forecasting
Methods in Power Engineering), 16(12), 4739; 15 June 2023,https://doi.org/10.3390/en16124739
25. APALOO BARA Komla Kpomonè, Multiplayer Perceptron and Simple Regression Linear Approaches to
Predict Photovoltaic Active Power Plant: Case Study, International Journal of Research and Review Vol. 10;
Issue: 12; E -ISSN: 2349-9788 ; P-ISSN: 2454-2237,December 2023
26. Mamadou Seck, and Ange-Lionel Toba : Powering West Africa: modelling sustainable long-term electricity
planning, International Journal of Critical Infrastructures, Vol. 15, No. 3 pp 260-284 March 25, 2019
https://doi.org/10.1504/IJCIS.2019.100430
27. N. Tippayawong, A. Promwungkwa, P. Rerkkriangkrai, Long-term operation of a small biogas/diesel dual-fuel
engine for on-farm electricity generation, iosystems Engineering, Volume 98, Issue 1, 2007, Pages 26-32, ISSN
1537-5110, https://doi.org/10.1016/j.biosystemseng. 2007.06.013
28. Bin Xu, Ping-An Zhong, Zachary Stanko, Yunfa Zhao, William W.-G. Yeh : Water Ressources Research , A
multiobjective short-term optimal operation model for a cascade system of reservoirs considering the impact on
long-term energy production First published: 15 April 2015 https://doi.org/10.1002/2014WR015964
29. Shuang Han, Yan-hui Qiao, Jie Yan, Yong-qian Liu, Li Li, Zheng Wang, Mid-to-long term wind and
photovoltaic power generation prediction based on copula function and long short term memory network,
Applied Energy, Volume 239, 2019, Pages 181-191, ISSN 0306-2619,
https://doi.org/10.1016/j.apenergy.2019.01.193.
30. António Cardoso Marques, José Alberto Fuinhas, Diogo Santos Pereira, The dynamics of the short and long-run
effects of public policies supporting renewable energy: A comparative study of installed capacity and electricity
generation, Economic Analysis and Policy, Volume 63, 2019, Pages 188-206, ISSN 0313-5926,
https://doi.org/10.1016/j.eap.2019.06.004.

15
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

31. Tom Brijs, Arne van Stiphout, Sauleh Siddiqui, Ronnie Belmans, Evaluating the role of electricity storage by
considering short-term operation in long-term planning, Sustainable Energy, Grids and Networks, Volume 10,
Pages 104-117, ISSN 2352-4677, 2017,https://doi.org/10.1016/j.segan.2017.04.002.
32. Benin-Togo - Communauté électrique du Bénin (CEB) Energy Production Dispatching and Transport Project –
PPER, 15-jan-2014,
33. Financement AFD FERC CTG 1147, compagnie Energie électrique du Togo (CEET), élaboration du plan
directeur du développement des infrastructures d’Energie électrique du Togo, énergie élaboration du plan
directeur du développement des infrastructures d’Energie électrique du Togo, RAPPORT REF. CLIENT
MARCHE N⁰ 06/PRMP/DG/CEET/2013
34. Plan d’Action National 2020-2024 de la République du Togo pour la Politique d'Intégration du Genre dans
l'Accès À l’Energie de la CEDEAO Version finale Février 2020
35. Sangué Oraléou Djandja and al., ESTIMATION OF SOIL ELECTRICAL RESISTIVITY USING MLP, RBF,
ANFIS AND SVM APPROACHES. Int. J. Adv. Res. 7(5), 48-60, ISSN: 2320-5407, May 2019, DOI:
10.21474/IJAR01/9004, http://dx.doi.org/10.21474/IJAR01/9004
36. Joti Kanoujiya, Abhay Sancheti2 and S. Swami, Prediction of difficult laryngoscopy by ultrasound guided
valuation of anterior neck soft tissue thickness. Int. J. Adv. Res. 7(2), 242-255, ISSN: 2320-5407, February
2019
37. Kirtan Jha, Aalap Doshi and Poojan Patel: Intelligent irrigation system using artificial intelligence and machine
learning: a comprehensive review., Int. J. Adv. Res. 6(10), 1493-1502, ISSN: 2320-5407, October 2018
38. Marie Frapin, Charlotte Roux, Edi Assoumou, Bruno Peuportier: Modelling long-term and short-term temporal
variation and uncertainty of electricity production in the life cycle assessment of buildings, Applied Energy,
Volume 307, 2022, 118141, ISSN 0306-2619, https://doi.org/10.1016/j.apenergy.2021.118141
39. De Shunji Osaki : Applied Stochastic System Modeling, Springer Verlas, 1ère ed. 1992 e-ISBN-13 :978-3-642-
84681-6, DOI :10.1007/978-3-642-84681-6
40. APALOO BARA, K. K. 2020. Modeling the electrical resistivity of soils in tropical areas using imperfect data
analysis methods: study and creation of a tool to help characterize and predict electrical resistivities in Togo,
Science of Science ThesisEngineer, Electrical Engineering Option, University of Lomé, September 17, 2020
41. Kamal Boukhetala, Jean-François Dupuy (Dir.). Modélisation Stochastique et Statistique Book of Proceedings.
hal-02593238, 2019.
42. Sandrine VATON & Thierry CHONAVEL : Modélisation et Simulation Stochastique, Télécom Bretagne,
INFO & SC Mai 2014
43. J.W. DeHeur « ARIMA » processes for hydrometeorological forecasting and simulation, Purdue University
(USA) Détaché au Laboratoire National d'Hydraulique, Chatou, LA HOUILLE BLANCHE / N° 6-1978, 1976
44. Alnufaielafi S. : Type-2 fuzzy supervisor for a stand-alone system using pmsg variable speed wind turbine., Int.
J. Adv. Res. 7(2), 1041-1049, ISSN: 2320-5407, February 2019
45. Sangué Oraléou Djandja and al. : Estimation Of Soil Electrical Resistivity Using MLP, RBF, ANFIS and SVM
Approaches, Int. J. Adv. Res. 7(5), 48-60, ISSN: 2320-5407, DOI: http://dx.doi.org/10.21474/IJAR01/9003
46. Arnaud Guyader : Régression linéaire, Université Rennes 2, tiré du livre du livre de Pierre-André Cornillon et
Eric MatznerLøber, Régression avec R, paru chez Springer en 2010, Année 2012/2013
47. Aleksandr Kulikov, Anton Loskutov, and Dmitriy Bezdushniy , Relay Protection and Automation Algorithms
of Electrical Networks Based on Simulation and Machine Learning Methods , Department of Electric Power
Engineering, Power Supply and Power Electronics, Nizhny Novgorod State Technical University n.a. R.E.
Alekseev, Minin st., 24, 603950 Nizhny Novgorod, Russia, Energies 2022, 15(18), 6525; 7 September 2022 ,
https://doi.org/10.3390/en15186525
48. John McCall, Genetic algorithms for modelling and optimization, Journal of Computational and Applied
Mathematics, Elsevier, Volume 184, Issue 1, Pages 205-222, 1 December 2005,
https://doi.org/10.1016/j.cam.2004.07.034
49. L V Stepanov, A S Koltsov, A V Parinov and A S Dubrovin, Mathematical modeling method based on genetic
algorithm and its applications, Journal of Physics: Conference Series, Volume 1203, International Conference
"Applied Mathematics, Computational Science and Mechanics: Current Problems" 17–19 December 2018,
Voronezh State University, Voronezh, Russian Federation , DOI 10.1088/1742-6596/1203/1/012082
50. Shinichi Kikuchi and al., Dynamic modeling of genetic networks using genetic algorithm and S-system
Bioinformatics, Volume 19, Issue 5, Pages 643–650, 22 March 2003,
https://doi.org/10.1093/bioinformatics/btg027
51. Aida Boudhaouia Miled. Analyse, classification et prédiction de consommation d’eau et d’électricité par des
techniques de machine learning. Intelligence artificielle [cs.AI]. Université de Haute Alsace - Mulhouse, 2022

16
ISSN: 2320-5407 Int. J. Adv. Res. 12(04), 01-17

52. Climat et système électrique : une sensibilité accrue liée au changement climatique et à l’évolution du mix,
futurs énergétiques 2050, principaux résultats, octobre 2021
53. Ségolen Geffray : Corrélation linéaire et régression linéaire simple, IUT Carquefou, Année 2008-2009
54. L. Laurencelle : Le tau et le tau-b de Kendall pour la corrélation de variables ordinales simples ou catégorielles,
Université du Québec à Trois-Rivières, Tutorials in Quantitative Methods for Psychology, Vol. 5(2), p. 51-58,
2009, doi.org/10.20982/tqmp.05.2.p051.

17