AUIQ Technical Engineering Science

Volume 1 Issue 2 Article 8

2024

Development of a Hybrid Intelligence Algorithm to Estimate the

Derivative Weight of Dawakin Tofa Clay for Heat Storage
Abubakar D. Maiwada
Materials Science and Engineering Department, King Fahd University of Petroleum and Minerals, Dhahran,
Saudi Arabia, [email protected]

Abdullahi A. Adamu
Mechanical Engineering Department, Bayero University, Kano, Nigeria, [email protected]

Umar D. Maiwada
Umaru Musa ‘Yar’adua University, Katsina State, [email protected]

Sani I. Abba
Department of Civil Engineering, Prince Mohammad Bin Fahd University, Al Khobar, 31952, Saudi Arabia,
[email protected]

Follow this and additional works at: https://ates.alayen.edu.iq/home

Part of the Engineering Commons

Recommended Citation
Maiwada, Abubakar D.; Adamu, Abdullahi A.; Maiwada, Umar D.; and Abba, Sani I. (2024) "Development of
a Hybrid Intelligence Algorithm to Estimate the Derivative Weight of Dawakin Tofa Clay for Heat Storage,"
AUIQ Technical Engineering Science: Vol. 1: Iss. 2, Article 8.
DOI: https://doi.org/10.70645/3078-3437.1017

This Research Article is brought to you for free and open access by AUIQ Technical Engineering Science. It has
been accepted for inclusion in AUIQ Technical Engineering Science by an authorized editor of AUIQ Technical
Engineering Science.
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108
Scan the QR to view
the full-text article on
the journal website

ORIGINAL STUDY

Development of a Hybrid Intelligence Algorithm to

Estimate the Derivative Weight of Dawakin Tofa
Clay for Heat Storage

Abubakar D. Maiwada a,*, Abdullahi A. Adamu b , Umar D. Maiwada c , Sani I. Abba d

a Materials Science and Engineering Department, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
b Mechanical Engineering Department, Bayero University, Kano, Nigeria
c Umaru Musa ‘Yar’adua University, Katsina State, Nigeria
d Department of Civil Engineering, Prince Mohammad Bin Fahd University, Al Khobar, 31952, Saudi Arabia

ABSTRACT

The accurate prediction of thermogravimetric properties is critical for evaluating the suitability of natural materials
like Dawakin Tofa clay for heat storage applications, but traditional linear models often fail to capture the complex,
non-linear relationships inherent in such datasets. This study develops a hybrid intelligence framework integrating
Bilateral Neural Network (BNN), Kernel Support Vector Machine (KSVM), Step-Wise Linear Regression (SWLR), and
Robust Linear Regression (RLR) to predict the derivative weight of Dawakin Tofa clay based on 5,030 experimentally
obtained instances. Comprehensive data preprocessing, including normalization, feature selection, and dataset splitting
(80% training and 20% testing), ensured high-quality inputs for the models. The results demonstrated that non-linear
models significantly outperformed linear approaches, with BNN achieving a coefficient of determination R2 of 0.999, a
Mean Absolute Error (MAE) of 0.004377, and a Mean Absolute Percentage Error (MAPE) of 9.6% on the testing dataset.
Similarly, KSVM achieved an R2 of 0.999, MAE of 0.012134, and MAPE of 26.7%, indicating its robust predictive
capabilities. In contrast, linear models performed poorly, with SWLR and RLR yielding R2 values of 0.03 and -0.41,
respectively, and unacceptably high MAPE values of 612% and 53.5%. The findings underscore the limitations of linear
models in predicting complex thermogravimetric behaviors and highlight the transformative potential of advanced
machine learning techniques like BNN and KSVM. Furthermore, these results align with global sustainability efforts,
including SDG 7 and 12, by optimizing the use of locally available, eco-friendly materials for energy storage. This study
provides a replicable framework for leveraging artificial intelligence to enhance material characterization, offering a
significant step toward developing sustainable energy solutions.

Keywords: Dawakin Tofa clay, Thermogravimetric analysis, Machine learning models, Energy storage materials, Sustain-
able development goals

1. Introduction in the context of renewable energy sources [1].

Among the materials explored for TES, clays have
The increasing global demand for efficient energy gained attention due to their favorable thermal prop-
storage systems has driven significant advancements erties, widespread availability, and cost-effectiveness.
in the development of thermal energy storage (TES) Dawakin Tofa clay, a naturally abundant clay found
technologies, which play a crucial role in enhanc- in Nigeria, is a promising candidate for heat stor-
ing the sustainability of energy systems, particularly age applications [2]. However, optimizing its thermal

Received 1 December 2024; accepted 8 December 2024.

Available online 17 December 2024

* Corresponding author.
E-mail addresses: [email protected] (A. D. Maiwada), [email protected] (A. A. Adamu), [email protected] (U. D.
Maiwada), [email protected] (S. I. Abba).

https://doi.org/10.70645/3078-3437.1017
3078-3437/© 2024 Al-Ayen Iraqi University. This is an open-access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/).
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 95

properties, specifically the derivative weight re- (SVMs), and decision trees can model the relation-
lated to heat storage capacity, presents a significant ships between input variables (e.g., clay composition,
challenge that requires advanced computational ap- temperature, pressure) and output variables (e.g.,
proaches. The shift towards renewable energy sources derivative weight, heat capacity). These models can
necessitates the development of efficient and reliable be further refined using optimization techniques such
energy storage systems to manage the intermittent as genetic algorithms (GAs), particle swarm optimiza-
nature of resources such as solar and wind. TES sys- tion (PSO), and simulated annealing (SA) to improve
tems are particularly valuable in this regard, as they their predictive accuracy and generalization capabil-
can store excess thermal energy generated during ities [9].
peak periods and release it when needed, contributing The study develops a ML algorithm to estimate
to grid stability and energy security [3]. the derivative weight of Dawakin Tofa clay, thereby
Clays, including Dawakin Tofa clay, have been enhancing its application potential in TES systems.
identified as potential materials for TES due to their The successful development of a hybrid intelligence
high heat capacity, thermal conductivity, and sta- algorithm for estimating the derivative weight of
bility at elevated temperatures [4]. However, the Dawakin Tofa clay holds significant potential for
thermal properties of clays can vary significantly de- various applications. Primarily, it can improve the
pending on factors such as composition, processing, efficiency and performance of clay-based TES systems
and environmental conditions. Therefore, accurately by providing accurate predictions of the material’s
estimating the derivative weight, which is closely heat storage capacity under varying conditions. This,
related to the heat storage capacity, is essential in turn, can lead to more effective design and op-
for optimizing the performance of clay-based TES timization of TES systems, facilitating the broader
systems [5]. Traditional methods of estimating the adoption of renewable energy technologies and re-
derivative weight of materials often rely on experi- ducing reliance on fossil fuels. Moreover, the insights
mental measurements and empirical models, which derived from the algorithm can inform the selection
can be time-consuming, resource-intensive, and may and treatment of clays for specific TES applications,
not fully capture the complex interactions between enabling the development of customized materials
different variables. In contrast, hybrid intelligence with tailored thermal properties. By integrating AI,
algorithms provide a more efficient and accurate ap- ML, and optimization techniques, this approach offers
proach by leveraging AI and ML techniques to analyze a more efficient and accurate method for predicting
large datasets, identify patterns, and make predic- the thermal properties of clays, thereby enhancing
tions based on complex, non-linear relationships. their potential for use in TES systems. The insights
These algorithms can be trained on experimental data gained from this study contribute to the development
to develop predictive models that estimate the deriva- of more efficient and sustainable energy storage tech-
tive weight of Dawakin Tofa clay under various condi- nologies, supporting the transition to a low-carbon
tions, providing valuable insights into its thermal be- economy and addressing the growing demand for
havior and potential for heat storage applications [6]. renewable energy [10]. Furthermore, the methodol-
Hybrid intelligence algorithms, which integrate ogy developed in this study has broader applicability
various computational techniques, including artifi- in materials science and engineering, making it a
cial intelligence (AI), machine learning (ML), and valuable contribution to the field of computational
traditional optimization methods, have emerged as materials science. Evolving a hybrid intelligence al-
powerful tools in the field of materials science and gorithm to estimate the derivative weight of Dawakin
engineering. These algorithms have demonstrated Tofa clay involves several key steps. Initially, relevant
great potential in predicting, optimizing, and ana- data must be gathered through experimental studies
lyzing complex material properties, thus accelerating or simulations, including measurements of the clay’s
the development of advanced materials for vari- thermal properties under different conditions. This
ous applications [7]. Hybrid intelligence refers to data forms the foundation for training the machine
the integration of multiple computational techniques learning models. Subsequently, feature selection and
to address complex problems that are beyond the engineering techniques are employed to identify the
capabilities of individual methods. In the context most relevant input variables and transform the data
of materials science, hybrid intelligence algorithms into a suitable format for model training [11, 12].
combine AI and ML techniques with traditional op- Beyond thermal energy storage, the methodology
timization methods to enhance the accuracy and developed in this study can be extended to other areas
efficiency of material property predictions [8]. For of materials science and engineering, where accurate
instance, machine learning models such as artificial prediction and optimization of material properties are
neural networks (ANNs), support vector machines crucial [13]. AI algorithms can be applied to optimize
96 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

the composition and processing of composite mate- multiple AI models in this study offer a robust frame-
rials, predict the mechanical properties of alloys, or work for identifying the most effective predictive
model the behavior of polymers under different en- tools, which can be extended to similar applications
vironmental conditions. The interdisciplinary nature across material science and energy storage domains.
of this approach, which combines materials science, This work also contributes to global sustainability
computational intelligence, and optimization, opens goals by positioning Dawakin Tofa clay as a natural,
new avenues for research and innovation across a cost-effective, and environmentally friendly candi-
wide range of fields [14]. The research highlighted date for energy storage solutions, thereby promoting
significant as it advances the understanding and uti- the development of sustainable and efficient thermal
lization of natural materials for sustainable energy energy systems. Moreover, the proposed hybrid intel-
applications. By accurately predicting the derivative ligence algorithm serves as a pioneering framework
weight of Dawakin Tofa clay using a novel hybrid for future research, enabling the application of AI-
intelligence algorithm, the study provides valuable driven modeling to predict and optimize the thermal
insights into the material’s thermogravimetric behav- properties of other natural materials, ultimately ex-
ior, a critical factor for assessing its suitability in heat panding its applicability and impact across diverse
storage systems. This work supports the development fields.
of cost-effective, natural, and eco-friendly energy Correct prediction of thermogravimetric properties
storage solutions, addressing the global demand for is important for checking if natural materials like
sustainable and renewable energy technologies. The Dawakin Tofa clay can be used for heat storage, which
research also holds practical implications for energy helps in developing sustainable energy technologies.
storage technologies, where thermal stability and effi- Old linear models often do not capture the complex
ciency are paramount. By identifying and optimizing and non-linear relationships in these datasets, indi-
the properties of natural materials like Dawakin Tofa cating a need for better methods. This study is aimed
clay, the study aligns with sustainable development at creating a hybrid intelligence framework that in-
goals and supports the transition to green energy tegrates BNN, KSVM, SWLR, and RLR to predict the
systems. Furthermore, the hybrid intelligence algo- weight changes of Dawakin Tofa clay using 5,030 sets
rithm developed in this study has the potential to of experimental data. The topic is very relevant due
revolutionize predictive modeling across various do- to the global demand for sustainable and low-cost en-
mains, extending its applicability beyond heat storage ergy solutions, where materials such as Dawakin Tofa
to include other fields such as construction ma- clay offer eco-friendly and locally sourced options for
terials, catalysis, and environmental management. thermal energy storage. The machine learning models
Ultimately, this work offers a significant step forward were chosen for their particular strengths: BNN, a
in the integration of AI with material science, con- deep learning structure, was selected for its skill in
tributing to both scientific knowledge and practical learning complex non-linear patterns and performing
solutions for global energy challenges. well with high-dimensional data. KSVM was included
This study presents a groundbreaking contribution because it can manage non-linear separations through
to the field by introducing a novel hybrid intelli- kernel transformations, giving reliable predictions
gence algorithm that integrates BNN, SWLR, RLR, and even with noisy data. Linear models, SWLR and RLR,
KSVM to predict the derivative weight of Dawakin were used as references to show the limits of tra-
Tofa clay for heat storage applications. This inno- ditional methods when facing complex thermogravi-
vative multi-model approach harnesses the unique metric behaviors. Detailed data preprocessing steps,
strengths of each AI technique, resulting in enhanced including normalization, feature selection, and split-
predictive accuracy and reliability, which has not ting the dataset (70% for training and 30% for test-
been previously demonstrated for this specific clay ing), ensured that the models received quality inputs.
material. As the first study to explore the thermo-
gravimetric properties of Dawakin Tofa clay using
advanced computational intelligence, it fills a signif- 2. Proposed intelligent methods
icant research gap and lays the foundation for its
potential role in energy storage systems. By providing The proposed intelligent methods aim to lever-
an in-depth analysis of the thermal behavior of natu- age advanced machine learning algorithms to predict
ral clays, particularly through the accurate prediction the derivative weight of Dawakin Tofa clay, a key
of derivative weight, the research advances material thermogravimetric property essential for evaluating
characterization techniques critical for evaluating the its suitability as a heat storage material. The study
performance of heat storage materials. Furthermore, utilized a dataset of 5,030 instances, obtained ex-
the comprehensive comparison and optimization of perimentally under controlled laboratory conditions,
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 97

Fig. 1. Proposed AI Models used in this study.

ensuring its reliability and representativeness of the predictions from these models were integrated using
material’s thermal behavior. To prepare the data a weighted ensemble method, where model weights
for analysis, extensive preprocessing steps were con- were determined based on their cross-validation per-
ducted. Initially, data cleaning was performed to formance to maximize overall predictive accuracy.
address missing values and outliers, ensuring the The proposed approach was rigorously evaluated
dataset’s integrity and consistency. Numerical fea- using the testing dataset, employing performance
tures, such as temperature, heating rate, and initial metrics such as MAE, RMSE, and R2 . This compre-
weight, were normalized to a [0, 1] scale to im- hensive framework not only delivers highly accurate
prove the performance of models sensitive to feature predictions of the derivative weight but also sets
scaling, such as KSVM. Feature selection techniques, a benchmark for utilizing hybrid machine learn-
including correlation analysis and recursive feature ing algorithms in material characterization, offering
elimination RFE, were employed to isolate the most valuable insights for energy storage applications and
relevant predictors of derivative weight, reducing di- paving the way for future research into sustainable
mensionality and enhancing model interpretability. energy solutions.
The dataset was then split into training (80%) and
testing (20%) subsets, with cross-validation applied 2.1. Step Wise Linear Regression (SWLR)
to optimize hyperparameters and minimize overfit-
ting. Fig. 1 shows the proposed AI models employed SWLR is a method used to build a regression
in the study. The hybrid algorithm integrates four ma- model by systematically adding or removing predic-
chine learning models: BNN, SWLR, RLR, and KSVM, tors based on specific criteria. R2 shows how well
each contributing unique strengths to the prediction the model explains changes in the dependent vari-
task. The BNN, designed with two hidden layers and able. High values near 1 mean the model performs
ReLU activation functions, captured complex non- well. MAE computes the average size of errors be-
linear relationships, while SWLR iteratively selected tween actual and predicted values. Lower values
the most significant features, ensuring model sim- show better accuracy of the model. RMSE is the
plicity. The RLR addressed experimental noise and square root of the average of squared errors between
outliers effectively, providing robust predictions, and actual and predicted values. Lower values indicate
KSVM, with an optimized radial basis function RBF better accuracy, as larger errors have more impact
kernel, captured intricate patterns in the data. The because they are squared. AIC (Akaike Information
98 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

Fig. 2. Architecture of BNN model.

Criterion) assesses the model by weighing its fit it well-suited for tasks that involve comparing or
against its complexity, discouraging unnecessary pre- correlating two inputs [18]. The structure of a BNN
dictors. A smaller AIC value suggests a better model. typically involves two identical subnetworks as seen
BIC (Bayesian Information Criterion) is like AIC but in Fig. 2, each handling one of the input data streams.
has a stronger penalty for adding predictors, prefer- These subnetworks share the same parameters, ensur-
ring simpler models. Lower BIC values mean a better ing that both inputs are processed in the same way.
trade-off between simplicity and fit.The goal is to This is particularly useful for tasks such as similarity
identify the most relevant variables, improving both learning, where the goal is to determine the degree of
prediction accuracy and interpretability. The process similarity between two inputs, like image matching or
can start with either no predictors (forward selec- sentence pair comparison [19].
tion) or all predictors (backward elimination) [15]. After the inputs pass through their respective sub-
In forward selection, predictors are added one by one, networks, the resulting feature representations are
with the predictor that most improves the model fit combined, often through a distance metric (e.g.,
being included. In backward elimination, the model Euclidean distance) or concatenation, followed by
begins with all predictors, and the least significant further layers for classification or prediction. The
one is removed at each step [16]. Criteria for adding shared-parameter design helps BNNs generalize well
or removing predictors often include p-values, Akaike to new, unseen data. By comparing two data points in
Information Criterion (AIC), or Bayesian Information parallel and learning the relationships between them,
Criterion (BIC), ensuring a balance between model fit BNNs have proven effective in various domains such
and complexity. The process continues until adding as image recognition, natural language processing,
or removing predictors no longer improves the model and recommendation systems [20]. This architecture
according to the chosen criteria [17]. While stepwise is particularly useful in tasks where the focus is on un-
regression is useful for simplifying models, it can lead derstanding the interaction or relationship between
to overfitting and may overlook interactions between two entities [21].
variables, so model validation on independent data is
recommended. 2.3. Robust Linear Regression (RLR)

2.2. Bilateral Neural Network (BNN) RLR is an alternative to traditional linear regres-
sion that is designed to be less sensitive to outliers
BNN is a type of neural network architecture that and violations of assumptions, such as normality
processes two sets of input data in parallel, making and homoscedasticity. In standard linear regression,
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 99

outliers can disproportionately influence the esti- is used. This method implicitly maps the original
mated regression coefficients, leading to misleading data into a higher-dimensional space where a linear
results [22]. Robust regression addresses this by us- separation is possible, without explicitly computing
ing techniques that reduce the influence of outliers, the transformation. The kernel function calculates
ensuring that the model remains stable even when the the inner product of data points in this higher-
data contains anomalies [23]. One common method dimensional space, enabling the SVM to classify
of robust regression is M-estimation, which assigns data in the original space. Common kernel functions
lower weights to observations with large residuals, include the polynomial kernel, RBF, and sigmoid
diminishing the impact of outliers. Other techniques kernel. These kernels allow the SVM to model com-
include Least Absolute Deviations (LAD) regression, plex relationships and patterns, making it effective
which minimizes the sum of absolute residuals rather in tasks such as image classification, bioinformatics,
than squared residuals, and RANSAC (Random Sam- and text categorization [27]. Kernel SVM is powerful
ple Consensus), which iteratively fits the model to because it can find non-linear decision boundaries
random subsets of the data to identify a model while maintaining computational efficiency, as the
that best fits the majority of the data [24]. Robust higher-dimensional space is never explicitly calcu-
regression is valuable in practical applications where lated. However, selecting the right kernel and tuning
data quality is not guaranteed, such as in real-world its parameters are crucial for achieving good perfor-
datasets that may contain errors or extreme values. mance in specific tasks [9].
It provides more reliable estimates and predictions in
such cases compared to traditional linear regression
[25].
3. Result and discussion

2.4. Kernel Support Vector Machine (KSVM) The Bilateral Neural Network (BNN) model was
trained using MATLAB 2024b’s standard configura-
KSVM is an extension of the basic SVM that al- tions. It consisted of three hidden layers with 128
lows it to handle non-linearly separable data as in neurons in the first layer, 64 neurons in the second
Fig. 3. The traditional SVM finds a hyperplane that layer, and 32 neurons in the third layer. The activa-
best separates the data into classes, but this works tion function for the hidden layers was the Rectified
well only when the data is linearly separable [26]. Linear Unit (ReLU), while the output layer used a
To address more complex patterns, the kernel trick sigmoid activation function. The training process was

Fig. 3. Schematic diagram of kernel SVM.

100 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

Table 1. Predictive results of derivative weight.

Training Testing
Models R2 RMSE MAPE MSE MAE R2 RMSE MAPE MSE MAE
SWLR 0.03 0.3147 inf % 9.90E-02 0.24754 0.03 0.31064 612% 9.65E-02 0.24437
BNN 0.999 0.003776 inf % 1.43E-05 0.002547 0.999 0.006348 9.60% 4.03E-05 0.004377
RLR -0.39 0.37704 inf % 1.42E-01 0.20077 −0.41 0.37348 53.50% 0.13948 0.20194
KSVM 0.999 0.013706 inf % 0.000188 0.012207 0.999 0.01361 26.70% 0.000185 0.012134

conducted over 100 epochs using the Adam optimizer ize, making it highly suitable for this task. Similarly,
with a default learning rate of 0.001. the KSVM achieved exceptional results, also reporting
The Kernel Support Vector Machine (KSVM) em- an R2 of 0.999 for both datasets. The RMSE values
ployed a Radial Basis Function (RBF) kernel, which for KSVM, at 0.013706 during training and 0.01361
is the default kernel function in MATLAB’s fitcsvm. during testing, were slightly higher than those of
The regularization parameter (CCC) was set to 1, and BNN but still indicate excellent accuracy. Despite the
the kernel scale parameter (γ \gammaγ ) was auto- slightly higher MAPE of 26.7% for KSVM in testing,
matically determined based on the data distribution. the low MAE of 0.012134 demonstrates its robustness
Model optimization was performed over a maximum and reliability. In contrast, RLR performed poorly,
of 1,000 iterations. with a negative R2 of −0.39 in training and −0.41
The Stepwise Linear Regression (SWLR) model in testing, signifying that the model performed worse
implemented backward elimination for feature selec- than a simple mean-based prediction. Its RMSE and
tion. A significance level of 0.05 was used as the MAE values remained high, with a notable RMSE of
threshold for including or excluding features during 0.37704 in training and 0.37348 in testing, alongside
the regression process. The model used MATLAB’s a MAPE of 53.5% in testing, highlighting its inabil-
default configuration for stepwise regression. ity to accurately model the data. The consistently
The Robust Linear Regression (RLR) model applied poor performance of SWLR and RLR indicates their
the Huber loss function for robust fitting. The fit- incapability to handle the non-linear and complex
ting process involved iterative optimization, with a relationships present in the dataset. These findings
maximum of 500 iterations specified in MATLAB’s emphasize the need for robust, non-linear models
standard fitlm settings. This configuration ensured re- like BNN and KSVM, which significantly outper-
silience to outliers during the regression process. formed the linear methods. Although BNN slightly
The predictive performance of the models pre- outperformed KSVM in testing accuracy metrics, both
sented in Table 1 highlights significant differences models displayed excellent generalization and pre-
in their ability to estimate the derivative weight of dictive capabilities, making them highly suitable for
Dawakin Tofa clay. The SWLR model demonstrated estimating the derivative weight of Dawakin Tofa
consistently poor performance across both training clay. The “inf %” reported for MAPE in some mod-
and testing datasets, with an R2 value of 0.03, in- els suggests data-specific challenges, possibly arising
dicating that it fails to explain the variance in the from very small or zero values in the derivative
data. This shortcoming is further reflected in the weight, which may have skewed percentage-based
high RMSE values of 0.3147 and 0.31064 for training error metrics. Overall, these results highlight the
and testing, respectively, and the MAE values, which necessity of employing advanced machine learning
remain similarly high at 0.24754 and 0.24437. The techniques, such as BNN and KSVM, to capture the
extreme MAPE value of 612% during testing illus- intricate relationships in thermogravimetric data and
trates the model’s inability to generalize effectively, achieve highly accurate predictions.
likely due to its linear nature, which limits its ability The comparison of R2 values during the test-
to capture the complex relationships in the dataset. ing phase highlights significant differences in model
On the other hand, BNN exhibited near-perfect per- performance, with BNN and KSVM achieving excel-
formance, achieving an R2 value of 0.999 for both lent R2 scores of 0.999, serving as the benchmark
training and testing datasets, indicating a complete for evaluating the other models. In comparison,
explanation of variance. The RMSE values for BNN SWLR demonstrated a substantially lower R2 value
are exceptionally low, at 0.003776 for training and of 0.03, representing a 97% reduction in explanatory
0.006348 for testing, with correspondingly low MSE power compared to the benchmark. This indicates
and MAE values, such as 1.43E-05 and 0.004377, that SWLR is incapable of capturing the com-
respectively. The MAPE of 9.6% in testing further plex, non-linear relationships inherent in the data.
confirms the model’s outstanding ability to general- RLR performed even worse, with an R2 of −0.41,
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 101

reflecting a 141% lower value than the benchmark. able innovation in energy systems, offering a scalable
A negative R2 not only signifies poor performance solution to global energy and environmental chal-
but also suggests that RLR performs worse than a lenges.
naive prediction based on the mean of the data. The testing phase results for the models, as evalu-
This stark contrast highlights the inadequacy of lin- ated by MAPE and MAE, reveal significant differences
ear models like SWLR and RLR in handling the in their predictive accuracy and generalization ca-
intricate thermogravimetric properties of Dawakin pabilities. For BNN, the MAPE is 9.6%, indicating a
Tofa clay. On the other hand, both BNN and KSVM high level of accuracy in percentage terms, as it de-
achieved R2 values of 0.999, indicating their ability to viates minimally from the true values, making it the
fully capture and generalize the relationships within most reliable model in terms of proportional error.
the testing dataset without any loss in explanatory Its corresponding MAE of 0.004377 further supports
power. The 0% difference between BNN and KSVM its precision, as the average absolute deviation be-
underscores their equally exceptional predictive ca- tween the predicted and actual derivative weight
pabilities. These results emphasize the transformative values is remarkably small. KSVM also performed
potential of advanced, non-linear models like BNN well, with a MAPE of 26.7%, which, while slightly
and KSVM in analyzing complex datasets, particularly higher than BNN, is still within an acceptable range
for material characterization and thermal property for practical applications. Its MAE of 0.012134 is low,
prediction. The stark percentage differences further demonstrating that it consistently produces predic-
illustrate the critical limitations of traditional lin- tions close to the actual values, though marginally
ear models in applications requiring high accuracy less precise than BNN. In contrast, SWLR shows a
and nuanced pattern recognition. The findings re- MAPE of 612%, a clear indication of its inability
inforce the importance of leveraging sophisticated to generalize to the testing data, as its percentage-
AI techniques to ensure reliable predictions, paving based errors are excessively high. Its MAE of 0.24437
the way for their integration into advanced energy further reflects significant absolute deviations, high-
storage research. The application of machine learn- lighting its poor suitability for the prediction task.
ing models like BNN and KSVM for predicting the RLR, while performing better than SWLR, still demon-
thermogravimetric properties of Dawakin Tofa clay strates suboptimal results with a MAPE of 53.5%,
carries significant environmental implications, align- which indicates considerable proportional error. Its
ing with global sustainability efforts such as the UN’s MAE of 0.20194 corroborates the model’s limited
Sustainable Development Goals (SDGs) and the EPA’s predictive capability, as it deviates significantly from
guidelines. These models enable the precise charac- the true values on average. These results emphasize
terization of natural materials for use in sustainable the superiority of non-linear models like BNN and
energy storage systems, directly supporting SDG 7 KSVM for this task, as their ability to capture complex
(Affordable and Clean Energy) by promoting cost- relationships in the data leads to significantly lower
effective and renewable energy technologies through errors compared to the linear models, SWLR and RLR,
the use of locally available, low-cost materials like which fail to handle the intricacies of the dataset
clay. By reducing the need for extensive experimental effectively. This analysis underscores the critical im-
trials, the models also align with SDG 12 (Respon- portance of advanced machine learning techniques
sible Consumption and Production) by conserving for ensuring accurate and reliable predictions in ma-
resources and minimizing material waste, and SDG terial characterization.
13 (Climate Action) by facilitating the transition The conclusions drawn from this study, particu-
to renewable energy systems that mitigate green- larly the superior performance of non-linear machine
house gas emissions. In line with the EPA’s principles learning models such as BNN and KSVM in predicting
of sustainable materials management, this approach the thermogravimetric properties of Dawakin Tofa
prioritizes the efficient and environmentally respon- clay, are consistent with trends observed in exist-
sible use of natural resources, reducing reliance on ing literature. Non-linear models have been widely
energy-intensive synthetic materials and lowering the acknowledged for their ability to capture complex,
environmental footprint. Moreover, the use of clay non-linear relationships in datasets where traditional
for heat storage minimizes industrial emissions and linear models often fail. For instance, studies on pre-
resource depletion, supporting cleaner air and wa- dictive modeling in material science, such as those
ter, protecting ecosystems, and promoting a circular focusing on thermal property prediction, have consis-
economy through recyclability and safe disposal. By tently highlighted the limitations of linear regression
integrating predictive analytics into material char- approaches like SWLR and RLR due to their inability
acterization, this work exemplifies how advanced to account for intricate feature interactions, which
AI-driven methodologies can contribute to sustain- aligns with the poor performance of these models in
102 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

this study. Similarly, the findings affirm the effective- natural clays, have highlighted the environmental
ness of neural networks and kernel-based models like and economic benefits of leveraging advanced al-
KSVM in achieving high prediction accuracy, as re- gorithms to optimize material characterization. This
ported in prior studies on thermogravimetric analysis study contributes to this narrative by demonstrating
and energy storage materials. Literature on the ap- how machine learning can facilitate the identifica-
plication of machine learning in predicting material tion of cost-effective and environmentally friendly
properties, such as thermal conductivity, mechanical materials like Dawakin Tofa clay, supporting global
strength, and degradation behavior, often emphasizes efforts in line with the SDGs. The results of this study
the superior generalization capabilities of these ad- are consistent with existing literature in validating
vanced models, particularly when handling diverse the effectiveness of non-linear, AI-driven approaches
and high-dimensional datasets. For example, recent over traditional linear models for complex material
works have demonstrated that neural networks out- property predictions. The findings further enrich the
perform linear and tree-based methods in predicting scientific discourse by providing empirical evidence
the thermal performance of composites and clays, of the capabilities of BNN and KSVM, particularly in
supporting this study’s conclusion regarding BNN’s the context of thermogravimetric analysis and sus-
exceptional accuracy and reliability. tainable energy materials. This strengthens the case
Moreover, the observed limitations of RLR and for widespread adoption of such advanced methods
SWLR in this study, particularly their inability to in future research and industrial applications.
generalize and the presence of high MAPE and MAE Fig. 4 shows a bar chart that compares training
values, mirror findings in the literature that identify performance metrics. The blue bars show MAE, which
linear models as inadequate for datasets character- has moderate error values, meaning the average pre-
ized by noise, outliers, or non-linear dependencies. diction errors in the datasets or models are low. The
Existing studies have also shown that models like orange bars, which represent MSE, are always lower
KSVM, which utilize non-linear kernels such as the than the green bars (RMSE), as RMSE is calculated
RBF, are better suited for such tasks, as they effec- as the square root of MSE. This difference shows that
tively map input data to higher-dimensional spaces RMSE reacts more strongly to larger errors. The RMSE
where linear separability is achievable. Further, the values, shown by the green bars, are significantly
study’s conclusion aligns with the growing body of higher than MAE and MSE, indicating possible out-
literature advocating for the integration of AI-driven liers or greater variance in the prediction errors.
methods in sustainable energy and material science Fig. 5 illustrates a bar chart comparing the test-
research. Studies on predictive modeling of energy ing performance measures. The blue bars represent
storage materials, such as phase change materials and MAE and show medium error levels across the data

.40
.35
MAE-Training
.30 MSE-Training
RMSE-Training
.25
.20
.15
.10
.05
.00
SWLR BNN RLR KSVM
Fig. 4. Bar plot for training.
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 103

.40
.35
MAE-Testing
.30 MSE-Testing
RMSE-Testing
.25
.20
.15
.10
.05
.00
SWLR BNN RLR KSVM
Fig. 5. Bar plot for testing.

model explains the variance in the training data. It

appears more stable but at a lower level compared
to the R metric across all models. The blue line (R
Training), which indicates the correlation coefficient,
shows more variation among models, reflecting the
strength of the linear relationship between predicted
and actual values in the training set. The chart points
out the differences in model performance, with some
models (like SWLR and KSVM) being stronger in one
metric versus the other. This aids in assessing how
effectively the models capture patterns in the training
data.
Fig. 7 is also a radar plot that compares R-Squared
(Testing) and R (Testing) for four models: SWLR,
KSVM, RLR, and BNN. The red line (R-Squared
Testing) indicates how much variance each model
Fig. 6. Radar plot for training.
explains during testing. It usually has higher values
than the blue line for most models, showing that
sets/models, indicating average prediction errors in the models are reasonably good at explaining the
tests. The orange bars for MSE are usually lower than variability in the testing data. On the other hand,
the green bars, since RMSE (green) is the square root the blue line (R Testing), which represents the cor-
of MSE and raises the effect of bigger errors. The relation coefficient, tends to have lower and more
RMSE values, shown by the green bars, are much fluctuating values across the models. This suggests
higher than both MAE and MSE, showing the effect of that there are weaker linear connections between pre-
outliers or high variability in prediction errors during dicted and actual results in the testing phase. The plot
tests. This indicates that RMSE is more sensitive to un- suggests that the models, especially SWLR and BNN,
usual prediction errors than MAE and MSE, offering do not perform consistently during testing, displaying
better details on error size during model assessment. lower correlation coefficients and moderate variance
Fig. 6 is a radar chart that compares two mea- explanation. This points to a need for better model
surements: R-Squared (Training) and R (Training) generalization for the testing data.
across four models: SWLR, KSVM, RLR, and BNN. The Fig. 8 presents box plots comparing the observed
red line (R-Squared Training) shows how well each weight (%/m) with the predicted weights from four
104 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

best model for this data, whereas SWLR needs work to

fit the observed weight distribution. KSVM and BNN
have some potential but show variability.
Fig. 9 illustrates histograms of the observed weights
alongside the weights predicted by the KSVM, RLR,
SWLR, and BNN models. The distribution of the ob-
served weights appears uniform, indicating an even
spread of values across the range without pronounced
skewness or clustering. In the case of KSVM, the pre-
dictions display a pronounced peak at one end of
the range, suggesting a tendency to predict values
within a narrow interval, which may introduce bias.
Similarly, the RLR model yields a peaked distribution,
reflecting limited variability and a focus on a re-
stricted range of values. In contrast, the SWLR model
demonstrates a more balanced distribution; however,
it exhibits a noticeable concentration of predictions
at one end, indicating a propensity to underestimate
Fig. 7. Radar plot for testing.
weights. The BNN model, on the other hand, gener-
ates a broader distribution with both a peak and a
models. From the Observed Weight, the data looks tail, indicating higher variability and a wider range
consistent, having a small range and a balanced of predictions compared to the other models.
spread. RLR: gives predictions that are closest to the
actual data, meaning it has better accuracy and fit.
KSVM and BNN models however show more changes 4. Discussion on charts and graphs
in predictions, often underpredicting. SWLR model
has the biggest difference, with serious underpredic- The first chart is a box plot comparing the ob-
tion and variability. RLR among all seems to be the served derivative weight of Dawakin Tofa clay with

Fig. 8. Box Plot for observed weight and the predicted models.
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 105

80
1,200
70
1,000
60
800
50
Frequency

Frequency
40 600

30 400
20 200
10
-.33 -.32 -.31 -.30 -.29 -.28 -.27 -.26 0
-.55 -.50 -.45 -.40 -.35 -.30 -.25 -.20 -.15 -.10 -.05 .00
Predicted SWLR Predicted BNN
1,200
80
1,000 76
800 72
68
Frequency

600

Frequency
64
400 60
200 56
52
0
-.55 -.50 -.45 -.40 -.35 -.30 -.25 -.20 -.15 -.10 -.05 .00 .05 48
-.025 -.023 -.021 -.019 -.017 -.015 -.013
Predicted KSVM
Predicted RLR

200

160

120
Frequency

80

40

0
-.060 -.056 -.052 -.048 -.044 -.040 -.036 -.032 -.028
Derivative Weight (%/m)

Fig. 9. Histogram of the observed weight and the predicted models.

the predicted values from the four models: KSVM, reinforcing their reliability in generalizing to new
RLR, SWLR, and BNN. The observed weight shows data. In contrast, both RLR and SWLR show poor
a distribution ranging from −0.03 to −0.06. Among performance, with very low R2 and R values during
the models, KSVM and BNN predictions closely fol- testing. This implies that these models are unable to
low the observed values, indicating a higher level capture the variance in the data effectively and are
of accuracy with narrower ranges and less deviation. less suited for making accurate predictions in this
On the other hand, RLR and SWLR predictions show context.
broader ranges, particularly in the case of SWLR, The histograms, which likely represent the error
which exhibits significant deviation. This suggests distributions for each model, further illustrate the dis-
that RLR and SWLR struggle to capture the accurate parity in performance. BNN and KSVM demonstrate
relationship between the input features and the target tighter, more consistent error distributions, suggest-
output. ing their predictions are both accurate and stable,
The radar charts display R2 and R values for train- with fewer significant outliers. Conversely, RLR and
ing and testing phases across the four models. These SWLR show much broader error distributions, indi-
charts highlight the superior performance of BNN and cating higher and more erratic prediction errors. The
KSVM, as both models exhibit perfect or near-perfect presence of outliers and widely spread errors in these
R2 and R values, indicating an excellent fit. The models reflects their lower precision and inability to
testing results for these models also remain strong, generalize effectively.
106 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

Finally, the bar graphs compare the key error gravimetric behaviors. These comparisons highlight
metrics—MAE, MSE, and RMSE—across the models the transformative potential of machine learning ap-
during both training and testing phases. BNN and proaches, such as BNN and KSVM, in accurately
KSVM again show superior performance with very predicting thermogravimetric properties of natural
low error values, indicating minimal deviation from materials, thereby contributing to the development
the true values and high prediction accuracy. SWLR of sustainable energy solutions.
and RLR, on the other hand, display much higher
error values in both phases, confirming their lack of
reliability in this context. Their high MAE, MSE, and 5. Conclusion
RMSE further emphasize their difficulty in capturing
the correct relationships within the data. This study has demonstrated the transformative
BNN and KSVM clearly outperform RLR and SWLR potential of advanced machine learning models,
in terms of accuracy, error consistency, and gen- specifically BNN and KSVM, for accurately predicting
eralization. These two models consistently demon- the thermogravimetric properties of Dawakin Tofa
strate better alignment with the observed derivative clay, a natural material with significant promise for
weights and show stronger metrics across all perfor- heat storage applications. The results revealed that
mance evaluations. Therefore, for the development of non-linear models outperformed traditional linear
a hybrid intelligence algorithm to estimate the deriva- methods by a wide margin, with BNN and KSVM
tive weight of Dawakin Tofa clay, BNN and KSVM are achieving R2 values of 1 and low error metrics such
the most suitable models, while RLR and SWLR prove as MAPE (9.6% and 26.7%) and MAE (0.004377 and
inadequate due to their high errors and poor fit. 0.012134), respectively, while SWLR and RLR failed
In the study “Deep Learning-Based Modelling of to capture the complexity of the dataset, yielding
Pyrolysis,” Ozcan et al. applied various machine high MAPE values (612% and 53.5%) and poor R2
learning algorithms, including Bi-directional Long scores (0.03 and −0.41). These findings highlight the
Short-Term Memory (LSTM) networks, to predict the limitations of linear models in characterizing non-
thermal behavior of biomass materials during pyroly- linear relationships and validate the use of advanced
sis. Their Bi-directional LSTM model achieved a Mean AI techniques for optimizing resource use and min-
Squared Error (MSE) of 0.0001 and a coefficient of imizing the need for extensive experimental trials.
determination R2 of 0.999, indicating high predictive Future studies should explore integrating other ma-
accuracy. [28] chine learning models, such as ensemble methods
Similarly, in “Prediction of Thermogravimetric or gradient boosting, and expanding the framework
Data in the Thermal Recycling of E-Waste Using Ma- to other natural materials like biomass and agricul-
chine Learning Techniques,” Ali et al. employed RF tural residues to assess their energy storage potential.
and Support Vector Regression (SVR) models to pre- Also, long-term thermal and structural stability stud-
dict thermogravimetric data of e-waste materials. The ies, economic feasibility assessments, and life-cycle
RF model achieved R2 values up to 0.9999, while the environmental impact evaluations will be critical for
SVR model reached R2 values up to 0.9973, demon- validating the real-world applicability of Dawakin
strating excellent predictive performance. [29]. Tofa clay. Larger datasets collected under diverse
In another study, “Machine Learning Backprop- experimental conditions will further enhance the ro-
agation Prediction and Analysis of the Thermal bustness of the predictive framework, ensuring its
Degradation of Poly (Vinyl Alcohol),” Otaru et al. generalizability to broader contexts. Practical recom-
utilized a backpropagation neural network to model mendations include field testing, prototyping, and
the thermal degradation of polyvinyl alcohol. Their integrating the material into decentralized energy
model achieved a correlation coefficient of 0.992 storage systems, particularly in resource-limited set-
between predicted and experimental data, reflecting tings, alongside policy and industry collaborations
strong predictive capability. [30]. to promote eco-friendly material adoption. Despite
Relatively, this study’s Bilateral Neural Network the controlled laboratory focus, the study aligns with
(BNN) and Kernel Support Vector Machine (KSVM) SDG goals and EPA guidelines by fostering sustain-
models both achieved R2 values of 0.999, with Mean ability and advancing affordable energy solutions.
Absolute Errors (MAE) of 0.004377 and 0.012134, By bridging predictive analytics with practical im-
respectively. These results indicate that DW mod- plementation, this research lays a foundation for
els exhibit predictive performance comparable to leveraging natural materials like Dawakin Tofa clay
or exceeding those reported in similar studies, un- in scalable, sustainable energy storage systems that
derscoring the effectiveness of advanced machine contribute to global energy efficiency and climate
learning techniques in modeling complex thermo- resilience.
 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108 107

References 14. J. C. Ince, et al., “Overview of emerging hybrid and compos-

ite materials for space applications,” In Advanced Composites
and Hybrid Materials, vol. 6, no. 4, 2023, Springer Interna-
1. M. I. Khan, F. Asfand, and S. G. Al-Ghamdi, “Progress in
tional Publishing. doi: https://doi.org/10.1007/s42114-023-
research and technological advancements of thermal energy
00678-5.
storage systems for concentrated solar power,” Journal of En-
15. G. Papazafeiropoulos, “Stepwise regression for increasing the
ergy Storage, vol. 55, no. PD, p. 105860, 2022. doi: https:
predictive accuracy of artificial neural networks: Applica-
//doi.org/10.1016/j.est.2022.105860.
tions in benchmark and advanced problems,” Modelling, vol.
2. O. M. Sadek and W. K. Mekhamer, “Ca-montmorillonite clay
5, no. 1, pp. 153–179, 2024. doi: https://doi.org/10.3390/
as thermal energy storage material,” Thermochimica Acta, vol.
modelling5010009.
363, no. 1–2, pp. 47–54, 2000. doi: https://doi.org/10.1016/
16. A. Miller, J. Panneerselvam, and L. Liu, “A review of re-
S0040-6031(00)00598-0.
gression and classification techniques for analysis of common
3. S. E. Etuk, U. W. Robert, O. E. Agbasi, and S. S. Ekpo, “A
and rare variants and gene-environmental factors,” Neurocom-
study on thermophysical properties of clay from Agbani: its
puting, vol. 489, pp. 466–485, 2022. doi: https://doi.org/10.
assessment as potential walling material for naturally-cooled
1016/j.neucom.2021.08.150.
building design,” Epitoanyag - Journal of Silicate Based and
17. E. J. Ward, “A review and comparison of four commonly used
Composite Materials, vol. 74, no. 3, pp. 93–96, 2022. doi:
Bayesian and maximum likelihood model selection tools,”
https://doi.org/10.14382/epitoanyag-jsbcm.2022.15.
Ecological Modelling, vol. 211, no. 1–2, pp. 1–10, 2008. doi:
4. A. R. Sane, et al., “An investigation of the physical, ther-
https://doi.org/10.1016/j.ecolmodel.2007.10.030.
mal and mechanical properties of fired clay/SiC ceramics
18. T. Simons and D. J. Lee, “A review of binarized neural net-
for thermal energy storage,” Journal of Thermal Analysis and
works,” Electronics (Switzerland), vol. 8, no. 6, 2019. doi:
Calorimetry, vol. 140, no. 5, pp. 2087–2096, 2020. doi: https:
https://doi.org/10.3390/electronics8060661.
//doi.org/10.1007/s10973-019-08964-5.
19. O. A. M. F. Alnaggar, et al., “Efficient artificial intelligence
5. D. V. Voronin, E. Ivanov, P. Gushchin, R. Fakhrullin, and V.
approaches for medical image processing in healthcare: com-
Vinokurov, “Clay composites for thermal energy storage: A
prehensive review, taxonomy, and analysis,” In Artificial
review,” Molecules, vol. 25, no. 7, pp. 1–26, 2020. doi: https:
Intelligence Review, vol. 57, no. 8, 2024, Springer Netherlands.
//doi.org/10.3390/molecules25071504.
doi: https://doi.org/10.1007/s10462-024-10814-2.
6. P. Zhao, et al. “A machine learning and CFD modeling hybrid
20. Q. Duan, S. Hu, R. Deng, and Z. Lu, “Combined federated and
approach for predicting real-time heat transfer during coke-
split learning in edge computing for ubiquitous intelligence
making processes,” Fuel, vol. 373, no. May, p. 132273, 2024.
in internet of things: State-of-the-art and future directions,”
doi: https://doi.org/10.1016/j.fuel.2024.132273.
Sensors, vol. 22, no. 16, 2022. doi: https://doi.org/10.3390/
7. B. F. Azevedo, A. M. A. C. Rocha, and A. I. Pereira, “Hybrid
s22165983.
approaches to optimization and machine learning methods:
21. T. Liang, J. Glossner, L. Wang, S. Shi, and X. Zhang, “Prun-
a systematic literature review,” In Machine Learning, vol.
ing and quantization for deep neural network acceleration: A
113, no. 7, 2024, Springer US. doi: https://doi.org/10.1007/
survey,” Neurocomputing, vol. 461, pp. 370–403, 2021. doi:
s10994-023-06467-x.
https://doi.org/10.1016/j.neucom.2021.07.045.
8. V. Gupta, W. k. Liao, A. Choudhary, and A. Agrawal, “Evolu-
22. M. S. Ummah, “Structural equation modeling of
tion of artificial intelligence for application in contemporary
health-related indicators among home-dwelling elderly
materials science,” MRS Communications, vol. 13, no. 5, pp.
focusing on subjective health perceptions,” Sustainability,
754–763, 2023. doi: https://doi.org/10.1557/s43579-023-
vol. 11, no. 1, Article 123, 2019. doi: http://scioteca.
00433-3.
caf.com/bitstream/handle/123456789/1091/RED2017-
9. I. H. Sarker, “Machine learning: Algorithms, real-world appli-
Eng-8ene.pdf?sequence=12&isAllowed=y%0Ahttp:
cations and research directions,” SN Computer Science, vol. 2,
//dx.doi.org/10.1016/j.regsciurbeco.2008.06.005%0Ahttps:
no. 3, pp. 1–21, 2021. doi: https://doi.org/10.1007/s42979-
//www.researchgate.net/publication/305320484_SISTEM_
021-00592-x.
PEMBETUNGAN_TERPUSAT_STRATEGI_MELESTARI.
10. F. Villano, G. M. Mauro, and A. Pedace, “A review on
23. L. Bottmer, C. Croux, and I. Wilms, “Sparse regression for
machine/deep learning techniques applied to building en-
large data sets with outliers,” European Journal of Operational
ergy simulation, optimization and management,” Thermo, vol.
Research, vol. 297, no. 2, pp. 782–794, 2022. doi: https://doi.
4, no. 1, pp. 100–139, 2024. doi: https://doi.org/10.3390/
org/10.1016/j.ejor.2021.05.049.
thermo4010008.
24. D. M. Khan, M. Ali, Z. Ahmad, S. Manzoor, and S. Hussain,
11. A. R. Sane, et al., “Clay/phosphate-based ceramic materials for
“A new efficient redescending M-estimator for robust fitting
high temperature thermal energy storage - Part II: validation
of linear regression models in the presence of outliers,” Math-
of high temperature storage performance at pilot scale,” Solar
ematical Problems in Engineering, vol. 2021, 2021. doi: https:
Energy, vol. 278, no. January, p. 112799, 2019. doi: https:
//doi.org/10.1155/2021/3090537.
//doi.org/10.1016/j.solener.2024.112799.
25. M. E. McNamara, M. Zisser, C. G. Beevers, and J. Shumake,
12. A. Jain, V. Jha, F. Alsaif, B. Ashok, I. Vairavasundaram, and C.
“Not just “big” data: Importance of sample size, measurement
Kavitha, “Machine learning framework using on-road realtime
error, and uninformative predictors for developing prognos-
data for battery SoC level prediction in electric two-wheelers,”
tic models for digital interventions,” Behaviour Research and
Journal of Energy Storage, vol. 97, no. PB, p. 112884, 2024. doi:
Therapy, vol. 153, (June 2021), p. 104086, 2022. doi: https:
https://doi.org/10.1016/j.est.2024.112884.
//doi.org/10.1016/j.brat.2022.104086.
13. G. Sadeghi, “Energy storage on demand: Thermal energy
26. D. Tomar and S. Agarwal, “Twin support vector machine: A
storage development, materials, design, and integration chal-
review from 2007 to 2014,” Egyptian Informatics Journal, vol.
lenges,” Energy Storage Materials, vol. 46, no. January, pp.
16, no. 1, pp. 55–69, 2015. doi: https://doi.org/10.1016/j.eij.
192–222, 2022. doi: https://doi.org/10.1016/j.ensm.2022.
2014.12.003.
01.017.
108 AUIQ TECHNICAL ENGINEERING SCIENCE 2024;1:94–108

27. J. C. Y. Ngu, W. S. Yeo, T. F. Thien, and J. Nandong, “A com- 29. L. Ali, et al. “Prediction of thermogravimetric data in the ther-
prehensive overview of the applications of kernel functions mal recycling of e-waste using machine learning techniques:
and data-driven models in regression and classification tasks A data-driven approach,” ACS Omega, vol. 8, no. 45, pp.
in the context of software sensors,” Applied Soft Computing, 43254–43270, 2023. doi: https://doi.org/10.1021/acsomega.
vol. 164, no. July, p. 111975, 2024. doi: https://doi.org/10. 3c07228.
1016/j.asoc.2024.111975. 30. A. J. Otaru, Z. A. Alhulaybi, and I. Dubdub, “Machine learn-
28. A. Ozcan, A. Kasif, I. V. Sezgin, C. Catal, M. Sanwal, and H. ing backpropagation prediction and analysis of the thermal
Merdun, “Deep learning-based modelling of pyrolysis,” Cluster degradation of poly (vinyl alcohol),” Polymers, vol. 16, no. 3,
Computing, vol. 27, no. 1, pp. 1089–1108, 2024. doi: https: 2024. doi: https://doi.org/10.3390/polym16030437.
//doi.org/10.1007/s10586-023-04096-6.