Article

Deep Learning Algorithms to Predict Output Electrical Power of

an Industrial Steam Turbine
Kossai Fakir 1, *, Chouaib Ennawaoui 1,2 and Mahmoud El Mouden 1

1 Laboratory of Engineering Sciences for Energy (LabSIPE), University Research Center (CUR) in Renewable
Energies & Intelligent Systems for Energy (EnR&SIE), National School of Applied Sciences, Chouaib Doukkali
University, El Jadida 24000, Morocco
2 Green Tech Institute (GTI), Mohammed VI Polytechnic University (UM6P), Benguerir 43150, Morocco
* Correspondence: [email protected]

Abstract: Among the levers carried in the era of Industry 4.0, there is that of using Artificial Intel-
ligence models to serve the energy interests of industrial companies. The aim of this paper is to
estimate the active electrical power generated by industrial units that self-produce electricity. To
do this, we conduct a case study of the historical data of the variables influencing this parameter to
support the construction of three analytical models three analytical models based on Deep Learning
algorithms, which are Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN),
as well as the hybrid CNN algorithm coupled with LSTM (CNN-LSTM). Subsequently, and thanks
to the evaluation of the created models through three mathematical metrics which are Root Mean
Square Error (RMSE), Mean Square Error (MSE), and the variance score (R-squared), we were able to
make a comparative study between these models. According to the results of this comparison, we
attested that the hybrid model is the one that gives the best prediction results, with the following
findings: the variance score was about 98.29%, the value of RMSE was exactly 0.1199 MW, and for
MSE the error was equal to 0.0143 MW. The obtained results confirm the reliability of the hybrid
model, which can help industrial managers save energy by acting upstream of the process parameters
Citation: Fakir, K.; Ennawaoui, C.; El influencing the target variable and avoiding substantial energy bills.
Mouden, M. Deep Learning
Algorithms to Predict Output
Keywords: electrical power; steam turbine; deep learning; artificial intelligence; industry 4.0
Electrical Power of an Industrial
Steam Turbine. Appl. Syst. Innov.
2022, 5, 123. https://doi.org/
10.3390/asi5060123
1. Introduction
Academic Editor: Jay Lee
Nowadays, the fields of electrical energy production have become of vital importance
Received: 23 November 2022 in industrial factories characterized by their energy-intensive aspect, in particular those
Accepted: 6 December 2022 whose manufacturing processes allow the release of thermal energy. This importance is
Published: 8 December 2022 seen mainly in the possibility of achieving the self-sufficiency in electrical energy, so as to
cover partially or totally the internal consumption of the plant.
Publisher’s Note: MDPI stays neutral
The industrial method most commonly used in this sense is to recover the heat released
with regard to jurisdictional claims in
by these processes into tubular exchangers by evaporating the feed water in the thermal
published maps and institutional affil-
iations.
boilers. Then, the steam produced is used under pressure to drive a turbine coupled to an
electricity generator.
Nevertheless, the prediction of produced electricity has now become a key factor in
reducing the energy bills of factories, which is why it is necessary to construct Artificial
Copyright: © 2022 by the authors. Intelligence (AI) models to control and master upstream the factors that create undesirable
Licensee MDPI, Basel, Switzerland. reductions in or disruptions to electrical power.
This article is an open access article The structure of this work is as follows:
distributed under the terms and • Section 2 presents state-of-the-art techniques previously used by scientific researchers
conditions of the Creative Commons
to predict electrical power;
Attribution (CC BY) license (https://
• Section 3 discusses the three methods we employed for prediction, starting with a
creativecommons.org/licenses/by/
technical framework of our case study.
4.0/).

Appl. Syst. Innov. 2022, 5, 123. https://doi.org/10.3390/asi5060123 https://www.mdpi.com/journal/asi

Appl. Syst. Innov. 2022, 5, 123 2 of 12

• Section 4 presents the results of the developed models including a comparison of their
performance.
• Section 5 discusses and interprets the obtained results.
• Section 6 provides the conclusion, which highlights the benefits of these contributions,
as well as some perspectives.

2. Related Works Linked to the Prediction of Electrical Power

In this section, we present a summary of the most recent research works related to the
subject. This summary is shown in Table 1 with various case studies.

Table 1. Summary of recent related works.

Application Primary Energy

Side of the Case Study

Combined Cycle

Industrial Steam
Consumption
Prediction Technique

Production
Team Year Ref.
Followed

Biogas
Water
Wind

Solar
ARIMA (Auto-Regressive Integrated Moving Average);
Khan 2019 X X [1]
SVM (Support Vector Machine).
SVR (Support Vector Regression).
Zafirakis 2019 X X [2]
ANN (Artificial Neural Network).
GRU (Gated Recurrent Units).
Sabri 2021 X X [3]
CNN (Convolutional Neural Network)
ANN (Artificial Neural Network).
Saleel 2021 X X [4]
DNN (Deep Neural Network).
Kons-
2021 LSTM (Long Short-Term memory). X X [5]
tantinou
Hybrid model LSTM-GPR (Gaussian Process
Wang 2021 X X [6]
Regression), Bayesian Network.
Heydari 2021 Fuzzy-GMDH model optimized by Grey Wolf. X X [7]
Liu 2021 CNN (Convolutional Neural Network). X X [8]
Al Decision Tree (DT), Generalized Linear (GL),
2021 X X [9]
Rayess Gradient-Boosted Trees (GBT), and Random Forest (RF).
FBP(Facebook Prophet).
Bendiek 2021 X X [10]
SVM(Support Vector Machine).
NOFGHW (Novel Optimized Fractional Grey
Zhou 2022 X [11]
Holt–Winters).
Shohan 2022 Hybrid model LSTM-NP (Neural Prophet). X [12]
Rossi 2022 MLR (Multiple Linear Regression). X X [13]
Wang 2022 LSTM improved by EMD-PCA-RF. X X [14]

It should be noted that the research works on electrical power prediction cited above
mainly focused on consumption rather than production.
Similarly, although they used various algorithms, they were more oriented toward
renewable sources than fossil energies. We suggest that some of the works required more
research efforts to optimize their uses, especially in the case of steam power plants.
In the present work, we aim to shed light on the prediction of power at the production
level and the level of a steam power plant associated with an industrial exothermal process
linked to the combustion of sulfur as a raw material.
 process linked to the combustion of sulfur as a raw material.

3. Materials and Methods

This section focuses on the framing of the problem of our case study, the exploration
of the data required to develop the solution, as well as the presentation of the techniques
Appl. Syst. Innov. 2022, 5, 123 3 of 12
used for the prediction of electrical power production.

3.1. Industrial Process Description

3. Materials and Methods
The application considered is an industrial process in a thermal power plant for the
This section
production focusesenergy,
of electrical on the framing
which isof the problem
associated withofaour case study,
sulfuric the exploration
acid production line,
of the data required to
as presented in (Figure 1). develop the solution, as well as the presentation of the techniques
usedThefor the prediction
plant of electrical
is continuously power
supplied production.
with high-pressure steam due to the exothermic
nature of the sulfuric line process. Therefore, electricity is produced through the
3.1. Industrial Process Description
mechanical drive of a turbine coupled with an alternator.
The application
Then, considered
the local electrical is an industrial
network process
ensures the in aself-supply
internal thermal power plant for
of energy, the
which
production of electrical energy, which is associated with a sulfuric acid production
could also be connected to the distribution grid of the public operator to ensure an line, as
presented in (Figure 1).
electrical exchange.

Figure 1. Overview of the process related to the co-production of sulfuric acid and electricity.
Figure 1. Overview of the process related to the co-production of sulfuric acid and electricity.

3.2. Dataset Presentation

The plant is continuously supplied with high-pressure steam due to the exothermic
Before
nature of thepreparing theprocess.
sulfuric line dataset,Therefore,
it should electricity
be noted that the construction
is produced of the
through the models
mechanical
was
drivecarried out oncoupled
of a turbine a platform
withusing Python as a computing tool.
an alternator.
Then,
To the local
achieve the electrical
objective network ensures
of the study, we the internal
collected self-supply
numerical dataoffrom
energy, which
a factory
could also
using the be
sameconnected to the distribution
aforementioned gridprocess.
production of the public operator
The data to ensure
covered an electri-
a period of 6
cal exchange.
months, with 1 value recorded every five 5 min as the sampling frequency, i.e., a total of
51,840 data points.
3.2. Dataset Presentation
Before preparing the dataset, it should be noted that the construction of the models
was carried out on a platform using Python as a computing tool.
To achieve the objective of the study, we collected numerical data from a factory using
the same aforementioned production process. The data covered a period of 6 months, with
1 value recorded every five 5 min as the sampling frequency, i.e., a total of 51,840 data points.
According to the analysis of the process history, there are five main parameters related
to steam that impact the variation in the electrical power generation in megawatts (MW),
which are presented in Table 2.
Appl. Syst. Innov. 2022, 5, 123 4 of 12

Table 2. Main feature parameters that influenced the variation in power in the studied power plant.

Mathematical Physical
Parameter Category
Symbol Measuring Unit
Electrical power
Target Pelec MW
at the output of the generator
Pressure losses
Feature ∆PFB mmH2 O
between furnace and boiler
Steam temperature Feature Temp ◦C

High pressure of driving steam Feature PHP bar

Flow rate of steam
Feature QmB t/h
(at the boiler output)
Flow rate of steam
Feature QmS t/h
(at the collector output)

The rated capacity of the alternator dedicated to this application is 58 MW, which also
has a rated apparent power of 68 MVA.

3.3. Correlation Analysis

The accuracy of the models depends significantly on the correlation between the
variables used. Thus, it is necessary to evaluate the correlation of different inputs with
the outputs of power production. The calculation of these coefficients before constructing
the models can give us a clear idea about the highly correlated parameters and weakly
correlated parameters.
In mathematics and statistics, covariance is a measure of the relationship between
two random variables, whereas correlation is a measure of the strength of the relationship
between the variables. In other words, correlation is the scaled measure of covariance.
The mathematical formulas for covariance (1) and correlation (2) are as follows:

¯ ¯
   
1 N
N ∑ i=1
cov(X, Y) = Xi − X × Yi − Y (1)

cov(X, Y)
corr(X, Y) = (2)
σX × σY
where
Xi is the values of the X variable.
Yi is the values of the Y variable.
¯
X is the mean value (average) of the X variable.
¯
Y is the mean value (average) of the Y variable.
N is the number of data points.
σX is the standard deviation of the X variable.
σY is the standard deviation of the Y variable.
It was expected in this work that the dependency relationships between several
variables can be represented by a correlation matrix in Python.

3.4. Considered Deep Learning Algorithms

3.4.1. Long Short-Term Memory (LSTM)
Before explaining LSTM, it is important to understand recurrent neural networks
(RNN) given the close relationship between them. The structure of the RNN consists
of an input layer, one or more hidden layers, and an output layer. RNNs have chain-
like structures of repeating modules, which are used as memory for storing important
information from previous processing steps [15].
Appl. Syst. Innov. 2022, 5, 123 5 of 12

LSTM is an evolution of RNNs and was introduced to eliminate the drawbacks of

RNNs related to vanishing / exploding gradients and rectify the problems of the short
memory linked to RNNs by adding complementary interactions per cell.

3.4.2. Convolutional Neural Network (CNN)

There are two types of convolutional neural networks, biological neural networks and
artificial neural networks. This work mainly discusses artificial neural networks.
A CNN-based artificial neural network is a modeling method that promotes data
and is similar in form to the synaptic links of the human brain. It is composed of several
neurons; the output of the previous neuron can serve as the input of the next neuron.
To sum up, the structural diagram of the CNN algorithm is composed of an “Input
Layer” that is connected to an “Output Layer” through three steps: “the Convolution Layer
n◦ 1”, “the Convolution Layer n◦ 2”, and a “Hidden Layer” [16].

3.4.3. CNN-LSTM Hybrid Model

A CNN (convolutional neural network) model consists of three layers: an input layer,
a hidden layer, and an output layer. The input of a three-dimensional array is usually fed
into a convolutional layer, where the dimensions are represented by the height, weight,
and number of channels [17].
Both CNN and LSTM models have specific features. Thus, a hybrid CNN–LSTM
DL model was considered in this study, which includes the advantages of both CNN and
LSTM models.

3.5. Assessment Strategy for the Models

Since we are working on a regression problem linked to archived historical data, we
evaluated the studied models using the following metrics: the RMSE (root mean square
error) [18] governed by Equation (3), the MSE (mean square error) [19] described by
Equation (4), and R-squared (explained variance score) expressed by Equation (5).
s
 2
1 n ^
N ∑i=1 i
RMSE= y − y (3)

 2
1 n ^
MSE= ∑i=1 yi −y (4)
N
 2
^
∑ni=1 yi −y
R2 =1−  2 (5)
n ¯
∑i=1 i y − y

where
yi is the actual value of y of the target variable (measured value of electrical power).
ŷ is the predicted value of y of the target variable (predicted value of electrical power).
y is the mean value of y of the target variable (mean value of electrical power).
N is the number of samples related to the prediction.

4. Results
Before presenting the results, we note that for each model, the used dataset was
divided into two sub-samples. The first was the learning sample (80% of the scope of the
dataset) and the second was the validation sample (20% of the scope of the dataset).
Each model was built on the training sample and validated on the test sample with a
performance score.
Appl. Syst. Innov. 2022, 5, 123 6 of 12

4.1. Correlation Check between Variables

Appl.
Appl.Syst.
Syst.Innov.
Innov.2022,
2022,5,5,123
123 6 6ofof1212
By performing a dependency analysis between the studied parameters, we generated
the corresponding correlation matrix in Figure 2, which is based on the heat map concept.

Figure
Figure2.
Figure 2.2.Correlation
Correlationmatrix
Correlation matrixof
matrix ofofstudied
studiedparameters.
studied parameters.
parameters.

4.2.
4.2. Theoretical
4.2.Theoretical Interpretation
TheoreticalInterpretation ofofCorrelation
Interpretationof Correlation Results
CorrelationResults
Results
As
As seen
Asseen ininthe
seenin theabove
the abovematrix,
above matrix,the
matrix, theparameter
the parameter«««Steam
parameter SteamMass
Steam Massflow
Mass flowrate
flow rateQ
rate QQ » had
mS» »had
mS
mS the
hadthe
the
highest
highest dependency
dependency relationship
relationship with
with the
the target
target parameter
parameter
highest dependency relationship with the target parameter (Pelec). (P
(Pelec
elec ).
).
We
Wedemonstrated
We demonstrated
demonstrated theoretically
theoreticallythatthat
theoretically thisthis
that strong dependence
thisstrong
strong was justified
dependence
dependence was by a physical
wasjustified
justified by
byaa
link between
physical
physicallink these
linkbetweentwo parameters,
betweenthese
thesetwo as seen
twoparameters, in Figure
parameters,asasseen 3.
seenininFigure
Figure3.3.

Figure
Figure3.3.Simplified
Simplifieddiagram
Simplified diagramshowing
showingthe
themost
mostcorrelated
correlatedfeature
featureparameter
parameterofofelectrical power.
electricalpower.
power.

By
Byapplying
applyingthe
thefirst
firstlaw
lawofofthermodynamics
thermodynamics[20] [20]totothe
theHP
HPsteam
steamturbine
turbinefollowing
following
the
theRankine
Rankinecycle
cyclemodel,
model,the theinternal energyisisexpressed
internalenergy expressedasasfollows:
follows:
∆𝐇
∆𝐇
∆H 𝐇𝐏
HP ++∆E
𝐇𝐏+ ∆𝐄
∆𝐄 𝐊+
K𝐊+ +∆𝐄
∆E 𝐏𝐏=
P∆𝐄
= Q=𝐐
+𝐐+
W+𝐖𝐖 (6)
(6)
(6)
where
where
where
∆𝐇
∆𝐇𝐇𝐏𝐇𝐏isisthe
thevariation
variationininenthalpy
enthalpy(kJ).
(kJ).
∆H HP is the variation in enthalpy (kJ).
∆𝐄
∆𝐄𝐊𝐊 is the
is thevariation
variationininkinetic
kineticenergy
energy (kJ).
(kJ).
∆𝐄
∆𝐄 is the
𝐏 𝐏 is the variation in potential energy(kJ).
variation in potential energy (kJ).
𝐐𝐐isisthe
thecalorific
calorificenergy
energy(kJ).
(kJ).
𝐖𝐖isisthethework
workcarried
carriedout
outononthe
thesystem
system(kJ).
(kJ).
Appl. Syst. Innov. 2022, 5, 123 7 of 12

∆EK is the variation in kinetic energy (kJ).

∆EP is the variation in potential energy (kJ).
Q is the calorific energy (kJ).
W is the work carried out on the system (kJ).
By considering the following hypotheses of the Rankine cycle:
- the transformation is adiabatic (Q = 0);
- the linear velocity is constant (∆EK = 0);
- the altitude is constant (∆EP = 0).
we obtain
∆HHP = W (7)
Likewise, by multiplying the work by the mass flow rate of the steam QmS , we obtain
the useful power of the steam Pstm :

Pstm = W. QmS = ∆HHP .QmS (8)

Then,
∆HHP .QmS
Pstm = (9)
m
where
Pstm is the power of the steam (kW).
∆HHP is the variation of enthalpy (kJ).
QmS is the mass flow rate of steam at collector output (t/h).
QmS is the mass of the steam (tons).
Otherwise, we calculate the global efficiency of the steam turbine [21] coupled with
the generator:
ηG = ηtur . ηgen (10)
Pmec
ηtur = (11)
Pstm + ΣPL1
Pelec
ηgen = (12)
Pmec + ΣPL2
where
ηG is the global efficiency of the turbine generator.
ηtur is the efficiency of the turbine.
ηgen is the efficiency of the generator.
ΣPL1 is the global losses at the level of the turbine (kW).
ΣPL2 is the global losses at the level of the generator (kW).
Considering the fact that losses tend to zero, we obtain

Pelec
Pstm = (13)
ηtur . ηgen

According to Equations (9) and (13), we deduce

Pelec = K .Qms (14)

where K is a positive coefficient that is equal to

∆HHP .ηtur . ηgen

K = (15)
m
From Equation (14), we observe that there is a direct proportional relationship between
the target variable (Pelec ) and its most correlated feature (Qms ).
Thus, the correlation of 99% found automatically through Python is physically justified.
Appl. Syst. Innov. 2022, 5, 123 8 of 12
Appl. Syst. Innov. 2022, 5, 123 8 of 12
Appl. Syst. Innov. 2022, 5, 123 8 of 12
Appl. Syst. Innov. 2022, 5, 123 8 of 12
4.3. Results of Electrical Power Prediction Using LSTM Model
4.3. Results of Electrical Power Prediction Using LSTM Model
•4.3. Results
Training and validation
of Electrical phase: Using LSTM Model
Power Prediction
• Results
4.3. Training and validation
of Electrical phase: Using LSTM Model
Power Prediction
• The resultand
Training of this phase phase:
validation is described in the Figure 4 below.
•The Training
result of this
and phase is described
validation phase: in the Figure 4 below.
The result of this phase is described in the Figure 4 below.
The result of this phase is described in the Figure 4 below.

Figure4.4.LSTM
Figure LSTM loss
loss function
function behavior
behavioraccording
accordingtotothe
thenumber
numberofof
epochs.
epochs.
Figure 4. LSTM loss function behavior according to the number of epochs.
Figure 4. LSTM loss function behavior according to the number of epochs.
•• Testingphase:
Testing phase:
• Testing phase:
•TheThe
result
Testingofphase:
resultthis
of phase is described
this phase in thein
is described Figure 5 below.
the Figure 5 below.
The result of this phase is described in the Figure 5 below.
The result of this phase is described in the Figure 5 below.

Figure 5. LSTM prediction results versus actual results.

Figure 5.
5. LSTM prediction
predictionresults
resultsversus
versusactual
actualresults.
Figure 5. LSTM
Figure LSTM prediction results versus actual results.
results.
4.4. Results of Electrical Power Prediction Using CNN Model
4.4.
4.4. Results
Results of Electrical
ElectricalPower
PowerPrediction
PredictionUsing
UsingCNN
CNN Model
Model
• Results
4.4. Training and validation
of Electrical phase: Using CNN Model
Power Prediction
•• Training
Training and validation
validationphase:
phase:
•The Training
result of this
and phase is described
validation phase: in the Figure 6 below.
The result of this
ofphase is described in the in
Figure 6 below.
The The result
result of this this phase
phase is described
is described the Figure
in the Figure 6 below.
6 below.

Figure 6. CNN loss function behavior according to the number of epochs.

Figure 6. CNN loss function behavior according to the number of epochs.
Figure 6.
Figure 6. CNN
CNN loss
loss function
functionbehavior
behavioraccording
accordingtotothe number
the of of
number epochs.
epochs.
Appl. Syst. Innov. 2022, 5, 123 9 of 12
Appl. Syst. Innov. 2022, 5, 123 9 of 12
Appl. Syst.
Appl. Syst. Innov.
Innov. 2022,
2022, 5,
5, 123
123 99 of
of 12
12

• Testing phase:
•• Testing
Testing phase:
phase:
• TheTesting
resultphase:
of this phase is described in the Figure 7 below.
The result
The result of
result of this
this phase
of this phase is
is described
described in
in the
the Figure
Figure 777 below.
below.
The phase is described in the Figure below.

Figure7.7.CNN
Figure CNNprediction
predictionresults
resultsversus
versusactual
actualresults.
results.
Figure 7.
Figure 7. CNN
CNN prediction
prediction results
results versus
versus actual
actual results.
results.
4.5.
4.5. Results ofofElectrical Power Prediction Using CNN-LSTM Hybrid Model
4.5. Results
4.5. Results of Electrical
Results of Electrical Power
Electrical Power Prediction
Power Prediction Using
Prediction Using CNN-LSTM
Using CNN-LSTM Hybrid
CNN-LSTM Hybrid Model
Hybrid Model
Model
•• Training
Training and
and validation
validation phase:
phase:
•• Training and
Training and validation
validation phase:
phase:
The The result
result of of this
this phasephase
is is described
described in the Figure 8 below.
The result
The result of
of this
this phase
phase is described in
is described in the
in the Figure
the Figure 888 below.
Figure below.
below.

Figure8.8.Hybrid
Figure HybridCNN-LSTM
CNN-LSTMloss
lossfunction
functionbehavior
behavioraccording
accordingtotothe
thenumber
numberofofepochs.
epochs.
Figure 8.
Figure 8. Hybrid
Hybrid CNN-LSTM
CNN-LSTM loss
loss function
function behavior
behavior according
according toto the
the number
number of
of epochs.
epochs.

••• Testing
Testing phase:
Testingphase:
phase:
• Testing phase:
The result
TheThe
result of
resultthis phase
of this
of this
this is
is described
phase
phase in
in the
is described
described in Figure
the the 99 below.
Figure
Figure 9 below.
below.
The result of phase is described in the Figure 9 below.

Figure 9. Hybrid CNN-LSTM prediction results versus actual results.

Figure9.9.
Figure
Figure 9.Hybrid
HybridCNN-LSTM
Hybrid CNN-LSTMprediction
CNN-LSTM predictionresults
prediction resultsversus
results versusactual
versus actualresults.
actual results.
results.
Appl. Syst. Innov. 2022, 5, 123 10 of 12

4.6. Performance Metrics Comparison of the Constructed Models

The results obtained by the calculation of the evaluation metrics are presented in
(Table 3), which also shows a comparison of the performance achieved by each prediction model.

Table 3. Comparison of the performance metrics during the testing phase.

Model RMSE MSE R-Squared

LSTM 0.2414401 0.05829334 0.98385581
CNN 0.43010269 0.18498833 0.94662498
CNN-LSTM 0.1199 0.014397927 0.982824288

5. Discussion
5.1. Findings
As observed in the correlation analysis in Figure 2, the variables “Steam Flow Rate
at the collector output” and “Steam Flow Rate at the Boiler output” had the strongest
dependencies on the target variable studied with correlations of 99% and 97%, respectively.
They were followed by the parameters “Steam Pressure” and “Pressure Losses”, with
respective dependencies of 87% and 84%. Finally, the variable “Steam Temperature”
correlated with electrical power with a percentage of 73%.
We then deduced that all predefined parameters had a significant correlation with
the target variable. It is therefore appropriate to retain all of them to train and test the
studied models.
On the other hand, and in light of the performance results obtained in (Table 3), the
model based on the long short-term memory (LSTM) algorithm offered a better quality of
prediction of the electrical power parameter. This performance was seen in the R-squared
metric score, which was the highest (≈98.39%). However, the scores of the two metrics
RMSE and MSE, which interpreted the errors, were, respectively, 0.241 MW and 0.058 MW.
Similarly, and by training the model based on the CNN-LSTM algorithm, we were able
to maintain a high R-squared score (≈98.29%) and also minimize the error margins gener-
ated by LSTM to achieve an RMSE of 0.1199 MW and MSE equal to 0.0143 MW. This im-
provement confirmed that the CNN-LSTM hybrid mode is very suitable for power prediction.
As for the convolutional neural network (CNN) algorithm, its performance was less
acceptable given that its score was also high (R2 = 94.66%), except that the margin of error
was greater than that of the two previous models, with an RMSE of 0.4301 MW and an MSE
of 0.1849 MW, which makes this model ranked third in our comparative study.
To allow a more complete analysis of the models, the loss curves were drawn during
the training phase and are presented in Figure 4 for the LSTM model, Figure 6 for the
CNN model, and Figure 8 for the CNN-LSTM model. In this sense, the LSTM algorithm
reached its maximum score (98%) after 100 epochs, the CNN-LSTM model reached a score
of 98% after only 50 epochs, and finally, we recorded a score of 94% for the CNN model
after 200 epochs, which also makes the calculation time of the CNN model much more
important than the previous two models.
According to the respective results obtained in Figures 5, 7 and 9, it is shown that the
three chosen models offer good predictions of the target variable, with an advantageous
prediction accuracy for the CNN-LSTM hybrid model compared to the others.

5.2. Implications of Findings

Given that 35 MW is the nominal consumption of the industrial unit, and as shown in
Figure 10, the energy exchanged between the self-produced electricity of the factory and
the public electrical grid plays a dual role:
(a) It can inject and sell supplementary energy to the external electrical grid when the
produced power exceeds the internal nominal consumption;
 Given that 35 MW is the nominal consumption of the industrial unit, and as shown
in Figure 10, the energy exchanged between the self-produced electricity of the factory
and the public electrical grid plays a dual role:
Appl. Syst. Innov. 2022, 5, 123 11 of 12
(a) It can inject and sell supplementary energy to the external electrical grid when the
produced power exceeds the internal nominal consumption;
(b) It provides the possibility for the industrial unit to obtain electricity during periods
It provides
(b) when the possibility
its thermal for the
power plant industrial
does unitlocal
not cover to obtain electricity
self–supply, i.e., during
when itperiods
is in
when its
deficit. thermal power plant does not cover local self–supply, i.e., when it is in deficit.

Figure 10. Electrical production of a factory self–producing electricity: (a) case of normal production
Figure 10. Electrical production of a factory self–producing electricity: (a) case of normal production
and consumption of electricity, leading to the export of 18 MW, (b) case of undesirable decreasing
and consumption of electricity, leading to the export of 18 MW, (b) case of undesirable decreasing of
of produced electrical power, leading to the import of 13 MW from the public grid.
produced electrical power, leading to the import of 13 MW from the public grid.
For this type electrical energy deficit problem, there is a need to predict the power
For this type electrical energy deficit problem, there is a need to predict the power
produced since in this case, the industrial unit will have to buy electricity, impacting its
produced since in this case, the industrial unit will have to buy electricity, impacting its
energy bill.
energy bill.
In other words, the prediction of the target parameter Pelec can allow industrial
In other words, the prediction of the target parameter Pelec can allow industrial
operators to act upstream at the right time on the processing parameters described in
operators to act upstream at the right time on the processing parameters described in
Table 2, primarily when the power produced is likely to decrease to under 35 MW.
Table 2, primarily when the power produced is likely to decrease to under 35 MW.
6. Conclusions
6. Conclusions and and Perspectives
Perspectives
This study showed
This study showed that that
the the
hybridhybrid
CNN-LSTMCNN-LSTM modelmodel
was thewas mostthe mostalgorithm,
reliable reliable
which made it possible to provide accurate and efficient predictions of the electricalofpower
algorithm, which made it possible to provide accurate and efficient predictions the
electrical power produced in the case of a steam power plant
produced in the case of a steam power plant based on industrial exothermic reactions. This based on industrial
exothermic reactions.
achievement implied an This achievement
impact implied
seen at two an impact seen at two levels.
levels.
First, the prediction of the attenuation linkedto
First, the prediction of the attenuation linked tothe
theelectrical
electricalpower
powerasasthe thetarget
target
parameter takes on an important anticipation characteristic, which
parameter takes on an important anticipation characteristic, which can act as a decision- can act as a decision-
making tool
making tool forfor industrial
industrial managers.
managers. This Thisisisseen seenasasbeing
beingparticularly
particularlyimportant
importantfor for
ensuring the continuous and autonomous electrical feeding of
ensuring the continuous and autonomous electrical feeding of industrial facilities. industrial facilities.
Second, this
Second, this prediction
prediction also
also aims
aims toto determine
determinethe theinput
inputparameters
parametersthat thatinfluence
influence
the decrease in the produced power. Thus, this prediction can
the decrease in the produced power. Thus, this prediction can provide the opportunityprovide the opportunity to
instantly take action in the process in order to prevent the
to instantly take action in the process in order to prevent the occurrence of prolongedoccurrence of prolonged
undesirable variations
undesirable variations in in these
these input
input variables,
variables, that thatmay
mayforce
forcethe theend-user
end-usertotoimport
import
electricity for a long period
electricity for a long period of time.of time.
In terms
In termsofof perspective,
perspective, it isitnow
is necessary
now necessary to use Business
to use Business Intelligence
Intelligence (BI)
(BI) techniques
techniques
to to create adashboard,
create a dynamic dynamic dashboard,
which displayswhich in displays in real-time
real-time the measurement
the measurement curve of
curve of electrical power superimposed on that of the prediction
electrical power superimposed on that of the prediction with a forecast horizon, with a forecast horizon,
as well as
as well
the as themeasurement
real-time real-time measurement
of the input of parameters
the input parameters
to visualize to and
visualize
track and
theirtrack their
variations.
As a result, helping decision makers to take the necessary actions at the right time can
increase the profitability of industrial plants that self-produce electricity.
Appl. Syst. Innov. 2022, 5, 123 12 of 12

Author Contributions: Conceptualization, K.F.; Methodology, C.E. and M.E.M.; Software, K.F.;
Validation, C.E. and M.E.M.; Formal analysis, K.F., C.E. and M.E.M.; Investigation, M.E.M. and
C.E.; Resources, K.F., C.E. and M.E.M.; Data curation, K.F.; Writing—original draft preparation, K.F.;
Writing—review and editing, K.F., C.E. and M.E.M.; Visualization, K.F., C.E. and M.E.M.; Supervision,
C.E. and M.E.M.; Project administration, C.E. and M.E.M.; Funding acquisition, K.F. All authors have
read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Mansoor, K.; Tianqi, L.; Farhan, U. A New Hybrid Approach to Forecast Wind Power for Large Scale Wind Turbine Data Using
Deep Learning with TensorFlow Framework and Principal Component Analysis. Energies 2019, 12, 2229.
2. Dimitris, Z.; Georgios, T.; John, K. Forecasting of Wind Power Generation with the Use of Artificial Neural Networks and Support
Vector Regression Models. In Proceedings of the Applied Energy Symposium and Forum, Renewable Energy Integration with
Mini/Microgrids, Rhodes, Greece, 29–30 September 2018.
3. Mohammed, S.; El Hassouni, M. A Novel Deep Learning Approach for Short Term Photovoltaic Power Forecasting Based on
GRU-CNN Model. In E3S Web of Conferences; EDP Sciences: Les Ulis, France, 2022; Volume 336, p. 00064.
4. Ahamed Saleel, C. Forecasting the energy output from a combined cycle thermal power plant using deep learning models. Case
Stud. Therm. Eng. 2021, 28, 101693. [CrossRef]
5. Maria, K.; Stefani, P.; Alexandros, G. Solar Photovoltaic Forecasting of Power Output Using LSTM Networks. Atmosphere 2021,
12, 124. [CrossRef]
6. Ying, W.; Feng, B.; Qing-Song, H.; Sun, L. Short-Term Solar Power Forecasting: A Combined Long Short-Term Memory and
Gaussian Process Regression Method. Sustainability 2021, 13, 3665.
7. Azim, H.; Meysam Majidi, N.; Mehdi, N.; Davide Astiaso, G.; Farshid, K.; Livio, D.; Lina, B. A Combined Fuzzy GMDH Neural
Network and Grey Wolf Optimization Application for Wind Turbine Power Production Forecasting Considering SCADA Data.
Energies 2021, 14, 3459.
8. Tianyang, L.; Zunkai, H.; Tian, L.; Yongxin, Z.; Wang, H.; Songlin, F. Enhancing Wind Turbine Power Forecast via Convolutional
Neural Network. Electronics 2021, 10, 261. [CrossRef]
9. Alrayess, H.; Asli, U. Forecasting the hydroelectric power generation of GCMs using machine learning techniques and deep
learning (Almus Dam, Turkey). Geofizika 2021, 38, 1–90. [CrossRef]
10. Paula, B.; Ahmad, T.; Qammer, H.; Basel, B. Solar Irradiance Forecasting Using a Data-Driven Algorithm and Contextual
Optimisation. Appl. Sci. 2022, 12, 134.
11. Weijie, Z.; Huihui, T.; Huimin, J. Application of a Novel Optimized Fractional Grey Holt-Winters Model in Energy Forecasting.
Sustainability 2022, 14, 3118. [CrossRef]
12. Shohan, M.; Faruque, M.; Simon, Y. Forecasting of Electric Load Using a Hybrid LSTM-Neural Prophet Model. Energies 2022,
15, 2158. [CrossRef]
13. Elena, R.; Isabella, P.; Renato, I. Multilinear Regression Model for Biogas Production Prediction from Dry Anaerobic Digestion of
OFMSW. Sustainability 2022, 14, 4393. [CrossRef]
14. Dongyu, W.; Xiwen, C.; Dongxiao, N. Wind Power Forecasting Based on LSTM Improved by EMD-PCA-RF. Sustainability 2022,
14, 7307. [CrossRef]
15. Xuan-Hien, L.; Hung, V.; Giha, L.; Sungho, J. Application of Long Short-Term Memory (LSTM) Neural Network for Flood
Forecasting. Water 2019, 11, 1387. [CrossRef]
16. Wang, J.; Zihao, L. Research on Face Recognition Based on CNN. In IOP Conference Series: Earth and Environmental Science,
Proceedings of the 2nd International Symposium on Resource Exploration and Environmental Science, Ordos, China, 28–29 April 2018; IOP
Publishing Ltd.: Bristol, UK, 2018; Volume 170, p. 170.
17. Rahim, B.; Taghi Aalami, M.; Adamowski, J. Short-term water quality variable prediction using a hybrid CNN–LSTM deep
learning model. Stoch. Environ. Res. Risk Assess. 2020, 34, 415–433.
18. Reddy, P.C.; Sureshbabu, A. An Adaptive Model for Forecasting Seasonal Rainfall Using Predictive Analytics. Int. J. Intell. Eng.
Syst. 2019, 12, 22–32. [CrossRef]
19. Saengmuang, A.; Sitjongsataporn, S. Convergence and Stability Analysis of Spline Adaptive Filtering based on Adaptive
Averaging Step-size Normalized Least Mean Square Algorithm. Int. J. Intell. Eng. Syst. 2020, 13, 267–277.
20. Loverude, M.E.; Kautz, C.H.; Paula, R.L. Heron: Student understanding of the first law of thermodynamics: Relating work to the
adiabatic compression of an ideal gas. Am. J. Phys. 2002, 70, 137. [CrossRef]
21. Onwuamaeze, P.I. Improving steam turbine efficiency: An appraisal. Res. J. Mech. Oper. 2018, 1, 24–30.
Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.