ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.

1051/itmconf/20246901009
MAIH2024

Energy Efficiency in Smart Buildings through Prediction

modeling and Optimization Using a Modified Whale
Optimization Algorithm

Nasima EL ASSRI1*, Mohammed ENNEJJAR1, Mohammed Ali JALLAL1, 3, Samira CHABAA1,2, Abdelouhab
ZEROUAL1

1I2SP Research Team, Physics department, Faculty of Sciences Semlalia, Cadi Ayyad University, Marrakesh, Morocco
2
Industrial Engineering Department, National School of Applied Sciences, Ibn Zohr University, Agadir, Morocco
3Univ. Grenoble Alpes, CEA, Liten, Campus Ines, 73375, Le Bourget du Lac, France

Abstract. This paper presents a comparative study on the prediction of energy consumption in buildings
using machine learning techniques. The dataset encompasses a diverse range of buildings with 8 input
features and one output variable, representing the energy consumption. The primary focus is on evaluating
the performance of two prominent and widely-used machine learning algorithms: Artificial Neural Networks
(ANN) and Random Forest (RF). The results indicate a promising predictive capacity of both models,
showcasing their effectiveness in capturing intricate patterns within the dataset. In the case of ANN, the
Root Mean Squared Error (RMSE) is reported at 3.806, demonstrating the model's ability to approximate
the true energy consumption values. Furthermore, the Random Forest model exhibits enhanced predictive
accuracy, as reflected by a lower RMSE of 1.392. In addition to predictive analysis, this study utilizes a
Modified Whale Optimization Algorithm (MWOA) to optimize energy consumption. The MWOA helps to
identify the associated input values that lead to the lowest possible energy consumption, providing valuable
insights for energy-efficient building design. The implications of this research extend to the broader field of
sustainable architecture and urban planning, paving the way for more informed decisions aimed at reducing
energy consumption and fostering environmental sustainability.

1 Introduction leveraging a dataset that encompasses a diverse range of

building types and energy use profiles, we compare the
As global energy consumption continues to rise, there is performance of these models in terms of predictive
a growing emphasis on improving energy efficiency, accuracy. The goal is to determine which model
particularly within the built environment [1], [2]. provides better approximations of true energy
Buildings account for a significant portion of overall consumption and to assess their potential for informing
energy use, making them a crucial area of focus for energy-efficient design and operation in the building
achieving sustainability goals. Reducing energy sector.
consumption in buildings not only mitigates Furthermore, this study utilizes a Modified Whale
environmental impacts but also contributes to the long- Optimization Algorithm (MWOA) to optimize and
term economic and operational efficiency of urban minimize energy consumption. This approach aims to
spaces[3]. identify the associated input values that lead to the
In recent years, the application of machine learning lowest possible energy consumption, providing valuable
techniques has garnered considerable attention in insights for energy-efficient building design and
various fields, including energy prediction [4]. These operation. The significance of this research lies in its
techniques offer the potential to uncover complex potential contributions to sustainable architecture and
patterns in data and provide more accurate predictions, urban planning. By offering a data-driven approach to
enabling decision-makers to optimize energy use. In the energy management, this work can inform decision-
context of building energy consumption, advanced making and contribute to the broader objective of
machine learning models can offer insights into how environmental sustainability[6].
various factors, such as building characteristics, The findings from this study hold significant
occupancy, and weather conditions, influence energy implications for the building industry and the broader
use [5]. pursuit of environmental sustainability. By providing
This study aims to explore the effectiveness of two insights into the effectiveness of different machine
widely recognized machine learning algorithms: learning algorithms in predicting energy consumption,
Artificial Neural Networks and Random Forest in this research can inform the design and operation of
predicting energy consumption in buildings. By energy-efficient buildings, ultimately contributing to the

* Corresponding author: [email protected]

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution
License 4.0 (https://creativecommons.org/licenses/by/4.0/).
ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.1051/itmconf/20246901009
MAIH2024

reduction of overall energy usage and environmental

impact [7].
This study is focused on the prediction and the
optimization of building energy consumption using
machine learning techniques. The structure of this paper
is as follows: Section 1 provides an introduction to the
study, while Section 2 details the methods employed,
including subsections on data description, Artificial
Neural Networks (ANN), Random Forest (RF), and the
Modified Whale Optimization Algorithm (MWOA).
Section 4 presents the results and discussion of the
findings, followed by Section 5, which concludes the
study and discusses the implications for energy-efficient
building design.

2 Methodology

2.1 Data Description

The dataset comprises a comprehensive set of 3840 Fig. 1. The used Neural network architecture.
records, and its primary objective is to investigate the
intricate relationships between the energy consumption
of residential buildings and eight distinct input variables 2.3 Random Forest (RF)
[8]. These input variables encompass essential features
Random Forest (RF) represents a versatile and robust
of the buildings, including: Building Size (x1), Floor
ensemble learning technique widely employed in
Height (x2), Glazing Area (x3), Wall Area (x4),
predictive modeling[13]. This method leverages the
window-to-wall ratio (WWR) (x5), Win Glazing U-
strength of multiple decision trees to enhance prediction
value (x6), Roof U-value (x7), and External Wall U-
accuracy and mitigate overfitting. In the context of
value (x8). Each record within the dataset is
forecasting building energy consumption, Random
characterized by these input variables, with the
Forest offers a formidable approach to capture intricate
corresponding energy consumption output variable
relationships among various input parameters[14]. The
denoted as BEC (MWh). The inclusion of these specific
ensemble nature of Random Forest involves
features allows for a comprehensive analysis of the
constructing multiple decision trees during training and
impact of various architectural and structural elements
aggregating their predictions to arrive at a final output.
on the energy consumption patterns of residential
Mathematically, the prediction BEC for a specific
buildings. The IES simulation software was utilized to
instance can be expressed as:
create the simulated buildings[9].
1
𝐵𝐵𝐵𝐵𝐵𝐵 = ∑𝑁𝑁
𝑖𝑖=1 𝑓𝑓𝑖𝑖 (𝑋𝑋) (2)
2.2 Artificial Neural Networks (ANNs) 𝑁𝑁

ANNs stand as dynamic computational structures Where, N represents the number of trees in the
inspired by the human brain's neural networks[10]. forest, and 𝑓𝑓𝑖𝑖 (𝑋𝑋) denotes the prediction of the ith
These models comprise layers with interconnected decision tree based on the input features X.
nodes, or neurons, which collectively process and learn The versatility and efficiency of Random Forest lie
complex patterns from input data[11]. In the context of in its ability to handle complex, high-dimensional
building energy consumption prediction, we harness the datasets, making it well-suited for modeling the
formidable capabilities of ANN to decipher intricate multifaceted relationships inherent in building energy
relationships among various building parameters. The consumption.
ANN is particularly adept at capturing nonlinear
dependencies between inputs, making it well-suited for 2.4 Modified Whale Optimization Algorithm
the multifaceted nature of building energy systems. The (MWOA)
predictive power of the network is expressed
mathematically through the following equation [12], The Modified Whale Optimization Algorithm (MWOA)
where 𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖 represents the output of the ith neuron: is a nature-inspired optimization technique that
enhances the original Whale Optimization Algorithm
𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖 = 𝑓𝑓(∑𝑛𝑛𝑗𝑗=1 𝑤𝑤𝑖𝑖𝑖𝑖 ∗ 𝑥𝑥𝑗𝑗 + 𝑏𝑏𝑖𝑖 ) (1) (WOA) by improving its exploration and exploitation
capabilities[15], [16]. MWOA mimics the hunting
Where, 𝑤𝑤𝑖𝑖𝑖𝑖 denotes the weight between the ith neuron behavior of humpback whales and operates as follows
and the jth input, 𝑥𝑥𝑗𝑗 is the jth input, 𝑏𝑏𝑖𝑖 is the bias term, [17]:
and 𝑓𝑓 is the activation function. Figure 1 illustrates the
configuration of a multi-layered feedforward 1. Initialization: A population of candidate
backpropagation network employed in this study. solutions 𝑋𝑋𝑖𝑖 (where i=1,2,…,N) is randomly

2
ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.1051/itmconf/20246901009
MAIH2024

∑𝑛𝑛
initialized within the search space, where N is 𝑅𝑅2 = 1 − 𝑖𝑖=1(𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 −𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 )
 
(9)
∑𝑛𝑛 ̅̅̅̅̅̅
𝑖𝑖=1(𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 − 𝐵𝐵𝐵𝐵𝐵𝐵 )
the number of solutions.
2. Fitness Evaluation: Each solution's
Where,
performance is assessed using a fitness
 𝑛𝑛 is the number of observations;
function 𝑓𝑓(𝑋𝑋𝑖𝑖 )aimed at minimizing energy
 𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 is the actual (observed) value for the ith
consumption, represented as f(X) → min. observation;
3. Encircling Prey: Candidate solutions update  𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 is the predicted value for the ith
their positions based on the best-performing observation;
solution 𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 using the following equation:  ̅̅̅̅̅̅ is the mean of the observed values
𝐵𝐵𝐵𝐵𝐵𝐵

𝑋𝑋𝑖𝑖𝑛𝑛𝑛𝑛𝑛𝑛 = 𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 − 𝐴𝐴 . |𝐶𝐶. 𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 − 𝑋𝑋𝑖𝑖 | (3) The application of these metrics enables researchers
and practitioners to gauge the effectiveness of building
where A and C are the coefficients that control the energy consumption prediction models and make
exploration behavior. informed decisions about model selection and
refinement.
4. Exploration: MWOA enhances exploration by In order to enhance the accuracy, the data were
introducing random jumps around the best normalized to a standardized range between 0 and 1.
solution, defined as:
3 Results & discussion
𝑋𝑋𝑖𝑖𝑛𝑛𝑛𝑛𝑛𝑛 = 𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 + 𝐴𝐴 . |𝑅𝑅. rand. D | (4)

where R is a random factor, rand is a random number in 3.1 Comparative prediction using RF and ANN
[0,1], and D is the distance from the current position to In this study, a comprehensive examination involves
𝑋𝑋𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 . eight input variables, specifically Building Size (x1),
Floor Height (x2), Glazing Area (x3), Wall Area (x4),
5. Exploitation: The algorithm balances global window-to-wall ratio (WWR) (x5), Window Glazing U-
and local searches to refine solutions value (x6), Roof U-value (x7), and External Wall U-
effectively by adjusting the coefficients A and value (x8). The performance evaluation encompasses
C over iterations. the consideration of all or various combinations of these
6. Termination: The process continues until a input variables. Table 1 presents the following results of
stopping criterion, such as a maximum number the RMSE, CV, MAPE, MAD and R2 for artificial
of iterations, is met. neural networks at 1000 epochs, providing insights into
how the exclusion or inclusion of specific input
By effectively identifying optimal input values for variables influences the predictive performance of the
minimizing energy consumption, MWOA is used to ANN model. The configuration of ANNs is: number of
enhance energy efficiency in buildings through its input variables: 10 (neurons in the hidden layer): 1
mathematical modeling of whale behaviors. (neuron in the output layer).

Table 1. Performance metrics with different sets of input

2.5 Evaluation Metrics variables using ANN.
In the evaluation of building energy consumption (BEC) Input RMSE CV MAPE MAD R²
prediction models, various metrics are employed to variables
assess the accuracy and performance of the models [18]. x1, x2, x3, 3.806 6.899 6.662 2.898 0.970
These metrics provide valuable insights into the x4, x5, x6,
predictive capabilities of the models and their ability to x7, x8
generalize to unseen data. The commonly used x3, x6, x8 4.333 7.853 7.461 3.192 0.961
evaluation metrics include Root Mean Squared Error All \ {x3} 7.895 14.310 13.026 6.341 0.873
(RMSE), Coefficient of Variation (CV), Mean Absolute All \ {x7} 4.363 7.908 7.721 3.547 0.961
Percentage Error (MAPE), Mean Absolute Deviation
(MAD), and the coefficient of determination (R²) [19]. The findings indicate that the model, which takes into
account all eight input variables, consistently yields
∑𝑛𝑛
𝑖𝑖=1(𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 −𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 )² superior results across all performance metrics.
 𝑅𝑅𝑀𝑀𝑀𝑀𝑀𝑀 = √  
(5)
𝑛𝑛 Furthermore, a difference was observed between the
networks that considered all input variables and those
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
 𝐶𝐶𝐶𝐶 (%) = ̅̅̅̅̅̅
× 100% 
(6) that utilized a reduced set of inputs. Figure 2 evaluates
𝐵𝐵𝐵𝐵𝐵𝐵
the metrics when excluding Glazing Area (x3). The
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 (%) =
1
∑𝑛𝑛𝑖𝑖=1 |
𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 −𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
| ∗ 100%(7)
model's performance significantly degrades, especially
𝑛𝑛 𝐵𝐵𝐵𝐵𝐵𝐵𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 in terms of RMSE, CV, MAPE, and MAD, while R²
1
decreases, indicating that the variable (x3) plays a
𝑀𝑀𝑀𝑀𝑀𝑀 = ∑𝑛𝑛𝑖𝑖=1|𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 − 𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 | 
(8) crucial role in predicting energy consumption.
𝑛𝑛

3
ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.1051/itmconf/20246901009
MAIH2024

(a) (b)
Fig. 2. Results from ANN models developed (a) with all variables as inputs and (b) without Glazing Area (x3) value.

Figure 3 and Table 2 present a thorough evaluation MAD (12.747), and a lower R² (0.505). Conversely, a
of Random Forest models with varying maximum forest with a maximum depth of 15 demonstrates better
depths (Max depth). The obtained results in Table 2 performance, featuring lower values of RMSE (1.392),
demonstrate that deeper trees in the forest lead to CV (2.523%), MAPE (1.356%), MAD (0.571), and a
improved performance. Specifically, a maximum depth higher R² (0.996).
of 1 results in relatively higher errors, characterized by
larger values in RMSE, CV, MAPE, MAD, and a lower Table 2. Random forest models with different max depth
R². As the maximum depth increases, there is a notable (MD) parameter.
enhancement in model performance, with lower errors
and a higher R². However, the results indicate that the RF RMSE CV MAPE MAD R²
models
forest's performance starts to deteriorate beyond a MD =1 15.589 28.256 30.893 12.747 0.505
maximum depth of 15. This suggests that increasing the MD = 5 4.239 7.683 7.703 3.338 0.963
maximum depth beyond a certain point may not MD = 10 1.537 2.786 1.838 0.782 0.995
significantly enhance model performance for the given MD =15 1.392 2.523 1.356 0.571 0.996
dataset. Table 2 further illustrates these findings by MD = 30 1.407 2.551 1.358 0.573 0.9959
showcasing the performance of forests with maximum MD =50 1.407 2.551 1.358 0.573 0.9959
depths. A forest with a maximum depth of 1 results in MD =100 1.407 2.551 1.358 0.573 0.9959
an under-fit model, characterized by higher values of
RMSE (15.589), CV (28.256%), MAPE (30.893%),

* Corresponding author: [email protected]

4
ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.1051/itmconf/20246901009
MAIH2024

(a) (b)
Fig. 3. RF models with different maximum depths: (a) MD = 15, (b) MD = 1.

From table 3, the comparison between ANN and RF The choice between these two algorithms often
reveals that RF consistently outperforms ANN in depends on the nature of the problem, data availability,
predicting energy consumption, both in full and reduced and desired performance. RF is commonly favored for
models. RF achieves a lower RMSE (1.407 for the full its robustness to noisy data, ease of tuning, and ability
model and 2.363 for the reduced model) compared to to handle nonlinear relationships. On the other hand,
ANN (4.308 and 4.049, respectively), and it also shows ANN is highly powerful for modeling complex
better performance across CV, MAPE, MAD, and R². relationships but typically requires more data and
While ANN experiences a slight improvement in expertise for proper configuration and optimization.
accuracy when using important variables, RF sees a
decrease in performance with reduced variables. Despite
3.2 Optimization using a Modified Whale
this, RF remains more accurate and robust than ANN in
Optimization Algorithm (MWOA)
all metrics, making it the superior choice for this dataset.
Optimizing design and operational parameters is
Table 3. Prediction errors comparison of full and reduced essential for improving energy efficiency in smart
models.
buildings. These buildings, equipped with advanced
Model RMSE CV MAPE MAD R²
management systems, require the efficient use of
ANN all 4.308 7.808 6.888 3.336 0.962 resources to minimize energy consumption while
ANN 4.049 7.340 6.882 2.969 0.966 maintaining optimal indoor comfort. By adjusting
important factors such as building size, glazing area, or U-values
variables of walls and windows, energy losses can be reduced
RF all 1.407 2.551 1.358 0.573 0.9959 while adhering to architectural constraints.
RF 2.363 4.283 2.666 1.133 0.9886 Optimization becomes even more crucial in the
important context of increasing regulations regarding building
variables energy performance, aiming to reduce carbon footprints
while improving energy efficiency.
In the field of energy consumption prediction for The goal of this optimization process is to minimize
buildings, several studies have highlighted the the total energy consumption while adhering to the
effectiveness of AI and machine learning models. A constraints related to the building design parameters,
previous study compared the performance of an such as building size, floor height, glazing area, and the
Artificial Neural Network (ANN) with the Random thermal performance of windows, roofs, and walls.
Forest (RF) algorithm for predicting HVAC energy The Modified Whale Optimization Algorithm
consumption in a hotel in Madrid, Spain[20]. The study (MWOA) was chosen for this study due to its ability to
concluded that the ANN slightly outperformed RF, with find optimal solutions in complex, multidimensional
a lower RMSE (4.97 for ANN versus 6.10 for RF). search spaces. This algorithm, inspired by the hunting
However, it also emphasized the ease of use and behavior of humpback whales, effectively balances
parameter tuning of the RF model, especially when exploration (global search across the solution space) and
dealing with complex categorical variables and exploitation (local search around promising solutions).
multidimensional data, which are common in buildings. It is particularly well-suited for optimization problems
In our study, we observed similar results, with both with nonlinear relationships between variables, as seen
ANN and RF demonstrating comparable predictive in building energy efficiency models. The parameters
abilities. However, in contrast to the Madrid hotel study, used for the MWOA are summarized in Table 4 below.
our findings show that the RF model outperformed the
ANN in terms of accuracy, with a lower RMSE (1.392 Table 4. Parameters used for the modified whale
compared to 3.806 for ANN). This difference could be optimization algorithm (MWOA).
attributed to the nature of the dataset used in our study,
which might align better with RF’s ability to capture Parameter Description Va
nonlinear and noisy relationships between the input lue
features and energy consumption. Population size Number of candidate solutions 20
(represented as whales)

5
ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.1051/itmconf/20246901009
MAIH2024

Maximum Number of iterations to allow 20 Acknowledgment

iterations the algorithm to converge This study was supported by National Center for
Number of input Number of input variables 8 Scientific and Technical Research (CNRST) in Morocco
variables (dim) corresponding to the building
design parameters
under Grant number: 7UCA2021.

The optimization process using the MWOA involved References

running the algorithm for 20 iterations. The minimal [1] H. P. Das et al., « Machine Learning for Smart and
energy consumption achieved during each iteration was Energy-Efficient Buildings », 27 novembre 2022,
also recorded. For instance, in the first iteration, the best arXiv: arXiv:2211.14889. doi:
solution found was [145.86, 2.8, 165.6353, 178.5094, 10.48550/arXiv.2211.14889.
14.6358, 4.5807, 0.1823, 0.6899], resulting in a [2] N. El Assri, M. A. Jallal, S. E. El Aoud, S. Chabaa, et
minimum energy consumption value of 33.6525. By the A. Zeroual, « Synergistic Neural Network and Velocity
Pausing Particle Swarm Optimization for Enhanced
final iteration 20, the best solution adjusted to [145.86,
Residential Building Energy Efficiency: A Case Study
2.8, 19.6301, 19.6301, 10.0, 0.97, 0.13, 0.26], yielding in Kuwait », Eng. Technol. Appl. Sci. Res., vol. 14, no
a minimum energy consumption value of 10.2967. 5, p. 17507‑17516, oct. 2024, doi: 10.48084/etasr.8278.
The key findings from the optimization process [3] S. Fathi, R. Srinivasan, A. Fenner, and S. Fathi,
demonstrate a substantial reduction in energy “Machine learning applications in urban building
consumption, from an initial value of 33.6525 units to energy performance forecasting: A systematic review,”
10.2967 units after 20 iterations. This reduction was Renewable and Sustainable Energy Reviews, vol. 133,
achieved through iterative adjustments of critical 2020, Art. no. 110287. doi:
building design parameters, including the glazing area, 10.1016/j.rser.2020.110287.
wall area, window-to-wall ratio, and the U-values of [4] K. Amasyali et N. M. El-Gohary, « A review of data-
driven building energy consumption prediction
windows, roof, and external walls, which improved the studies », Renewable and Sustainable Energy Reviews,
building’s thermal performance. These results highlight vol. 81, p. 1192‑1205, janv. 2018, doi:
the efficacy of the Modified Whale Optimization 10.1016/j.rser.2017.04.095.
Algorithm (MWOA) in optimizing energy efficiency. [5] S. Ardabili, L. Abdolalizadeh, C. Mako, B. Torok, et A.
For architects, engineers, and building designers, this Mosavi, « Systematic Review of Deep Learning and
approach provides a data-driven method for optimizing Machine Learning for Building Energy », Front.
building designs to achieve lower energy consumption, Energy Res., vol. 10, mars 2022, doi:
enabling them to make more informed decisions 10.3389/fenrg.2022.786027.
regarding material selection, building geometry, and [6] T. Hong, Z. Wang, X. Luo, et W. Zhang, « State-of-the-
art on research and applications of machine learning in
thermal insulation. the building life cycle », Energy and Buildings, vol.
212, p. 109831, avr. 2020, doi:
10.1016/j.enbuild.2020.109831.
Conclusion [7] Digital twins based day-ahead integrated energy system
This study highlights the comparative effectiveness of scheduling under load and renewable energy
uncertainties -
Artificial Neural Networks (ANN) and Random Forest [8] Ibrahim, D. M. (2020). A dataset for residential
(RF) algorithms in predicting energy consumption in buildings energy consumption with statistical and
buildings, revealing that RF significantly outperforms machine learning analysis [Data set]. GitHub.
ANN in terms of accuracy, as evidenced by its lower https://github.com/Dr-Dina-M-Ibrahim/A-dataset-for-
RMSE across various metrics. Both models showcased residential-buildings-energy-consumption-with-
strong predictive capabilities, yet RF demonstrated statistical-and-machine-learning-analysi ».
greater robustness when handling complex datasets [9] « IESVE. Integrated Environmental Solutions Virtual
common in energy consumption scenarios, making it a Environment (IESVE). httpswww.iesve.com. »
more reliable choice for practitioners in the field. [10] Pittarello, M.; Scarpa, M.; Ruggeri, A.G.; Gabrielli, L.;
Schibuola, L. Artificial Neural Networks to Optimize
Moreover, the implementation of the Modified Zero Energy Building (ZEB) Projects from the Early
Whale Optimization Algorithm (MWOA) for Design Stages. Appl. Sci. 2021, 11, 5377.
optimizing energy consumption further enhances the https://doi.org/10.3390/app11125377 ».
findings of this research. The MWOA successfully [11] R. Kumar, R. K. Aggarwal, et J. D. Sharma, « Energy
identified optimal design parameters, resulting in a analysis of a building using artificial neural network: A
notable reduction in energy consumption from an initial review », Energy and Buildings, vol. 65, p. 352‑358,
value of 33.6525 to a final minimum of 10.2967. These oct. 2013, doi: 10.1016/j.enbuild.2013.06.007.
results underscore the potential of combining predictive [12] T. Khatib, A. Mohamed, K. Sopian, et M. Mahmoud,
modeling with advanced optimization techniques to « Solar Energy Prediction for Malaysia Using Artificial
Neural Networks », International Journal of
improve energy efficiency in building design, ultimately
Photoenergy, vol. 2012, no 1, p. 419504, 2012, doi:
contributing to more sustainable architectural practices. 10.1155/2012/419504.
Future research should focus on refining these [13] Y. Mo, D. Zhao, et M. Syal, « Effective Features to
methodologies and exploring their applicability in Predict Residential Energy Consumption Using
diverse building contexts. Machine Learning », p. 284‑291, juin 2019, doi:
10.1061/9780784482445.036.
[14] L. N. Nyakundi, « PREDICTING ELECTRICITY
CONSUMPTION IN NORWAY: A COMPARISON

6
ITM Web of Conferences 69, 01009 (2024) https://doi.org/10.1051/itmconf/20246901009
MAIH2024

OF MACHINE LEARNING MODELS », Master

thesis, Norwegian University of Life Sciences, 2024.
[15] Y. Sun, X. Wang, Y. Chen, et Z. Liu, « A modified
whale optimization algorithm for large-scale global
optimization problems », Expert Systems with
Applications, vol. 114, p. 563‑577, déc. 2018, doi:
10.1016/j.eswa.2018.08.027.
[16] S. Mirjalili et A. Lewis, « The Whale Optimization
Algorithm », Advances in Engineering Software, vol.
95, p. 51‑67, mai 2016, doi:
10.1016/j.advengsoft.2016.01.008.
[17] L. Abualigah et al., « 8 - Whale optimization algorithm:
analysis and full survey », in Metaheuristic
Optimization Algorithms, L. Abualigah, Éd., Morgan
Kaufmann, 2024, p. 105‑115. doi: 10.1016/B978-0-
443-13925-3.00015-7.
[18] R. Wang, S. Lu, et Q. Li, « Multi-criteria
comprehensive study on predictive algorithm of hourly
heating energy consumption for residential buildings »,
Sustainable Cities and Society, vol. 49, p. 101623, août
2019, doi: 10.1016/j.scs.2019.101623.
[19] A. A. Al-Shargabi, A. Almhafdy, D. M. Ibrahim, M.
Alghieth, et F. Chiclana, « Buildings’ energy
consumption prediction models based on buildings’
characteristics: Research trends, taxonomy, and
performance measures », Journal of Building
Engineering, vol. 54, p. 104577, août 2022, doi:
10.1016/j.jobe.2022.104577.
[20] M. W. Ahmad, M. Mourshed, et Y. Rezgui, « Trees vs
Neurons: Comparison between random forest and ANN
for high-resolution prediction of building energy
consumption », Energy and Buildings, vol. 147, p.
77‑89, juill. 2017, doi: 10.1016/j.enbuild.2017.04.038.