Energy 240 (2022) 122692

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Prediction and optimization of heating and cooling loads in a

residential building based on multi-layer perceptron neural network
and different optimization algorithms
Yuanjin Xu a, Fei Li b, c, *, Armin Asgari d
a
Institute of Mathematical Geology and Remote Sensing Geology, School of Earth Resources, China University of Geosciences, 388 Lumo Road, Wuhan,
430074, China
b
Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, 100101, China
c
University of Chinese Academy of Sciences, Beijing, 100049, China
d
Department of Mechanical Engineering, University of Tabriz, Tabriz, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Since cooling and heating loads are regarded as significant parameters to examine the energy perfor-
Received 14 August 2021 mance of buildings, the need to predict and analyze them for the residential buildings seems to be
Received in revised form undeniable. Hence, the present paper wants to optimize the multi-layer perceptron neural network using
9 November 2021
several optimization methods to predict the heating and cooling of energy-efficient buildings. The
Accepted 19 November 2021
Available online 22 November 2021
dataset used in this study consists of eight independent factors: surface area, wall area, roof area, relative
compactness, overall height, orientation, glazing area, and glazing area distribution. To prove the reli-
ability and accuracy of the obtained results, test and training data are also considered. According to the
Keywords:
Biogeography-based optimization
statistical results, biogeography-based optimization has the highest value of R2 and the lowest values of
Cooling and heating load RMSD, normalized RMSD, and MAE in both training data and test data for cooling and heating loads.
Evolutionary algorithms Hence, the forecasting accuracy of the proposed MLP neural network optimized with the BBO optimi-
Energy-efficient buildings zation algorithm with the RMSD, R2, and MAE of 2.82, 0.920, 2.15 in the training phase of heating load
Neural network and with the RMSD, R2, and MAE of 3.18, 0.887, 2.97 in the training phase of the cooling load is much
better than those of the other models.
© 2021 Elsevier Ltd. All rights reserved.

1. Introduction considered two criteria in the building sector [5]. These criteria can
be adjusted through heating, ventilation, and air conditioning
Increased greenhouse gas (GHG) emissions have led to the (HVAC) systems. Various parameters such as building characteris-
melting of Arctic and Antarctic ice reservoirs, posing severe envi- tics and climate conditions are considered to predict cooling and
ronmental concerns to the planet earth [1]. Currently, the con- heating capacities. Because of the complex and non-linear rela-
struction industry is the major contributor to GHG emissions, tionship between these influencing parameters and building en-
accounting for up to 50% of global GHG emissions [2]. With the ergy performance, precisely predicting building energy
worldwide trend toward promoting sustainable development, en- consumption is a sophisticated task.
ergy efficiency in buildings has been identified as a major problem According to the explanations mentioned above, the issue of
in many countries and regions [3]. consuming high sustainable energy is considerably worth paying
Predicting and classifying building energy usage is critical for attention to. Through the test of the energy performance of build-
enhancing decision-making about energy consumption and CO2 ings (EPB) with optimal modeling of the HVAC, the high sustainable
emissions reductions, as it enables the development of smart and energy consumption is obtained. Concerning a high level of energy
sustainable designs [4]. Cooling load (CL) and heating load (HL) are consumption globally, the lack of energy is likely to become a
fundamental problem in the future. Some scientists attribute the
rise in energy consumption to people's tendency to live with high
* Corresponding author. Institute of Geographic Sciences and Natural Resources standards [6]. Therefore, the need to develop a variety of evaluation
Research, Chinese Academy of Sciences, Beijing, 100101, China. systems and numerical methods to suggest an optimum predictor
E-mail address: [email protected] (F. Li).

https://doi.org/10.1016/j.energy.2021.122692
0360-5442/© 2021 Elsevier Ltd. All rights reserved.
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

tool for building energy consumption became the main issue in a speed and ease of implementation [20]. Artificial neural network
large number of studies [7,8]. (ANN) is regarded as one of the most robust and practical ap-
Building energy prediction models may be divided into two proaches in a vast range of engineering simulations such as dam
categories: first-principle models and data-driven models. Ther- break [23], bearing capacity [24], and landslide geo-hazard [24].
modynamic equations are often used in first-principle models. Also, ANN is known as the trustworthy method in building energy
Building energy consumption may be evaluated using compre- management [25,26].
hensive environmental and building design factors such as the For instance, Escandon et al. [27] considered an ANN for thermal
weather profile of the outdoor environment, floor layout, thermal comfort prediction according to real-world data measured in-situ
characteristics of building envelops, and operation schedules of and introduced it as an efficient tool for testing retrofitting mea-
energy equipment units. These operational characteristics, on the sures. Fan et al. [28] aimed to predict the short-term building en-
other hand, are often difficult to acquire. Furthermore, due to the ergy through the recurrent ANN-based methods. Pino-Mejías et al.
complexity of building layouts, particularly in modern structures, [29] utilized linear regression and ANN to calculate three notions of
first-principle models will be computationally costly. Data-driven heating and cooling energy demands, consumption of energy, and
models, particularly diverse neural networks, are, on the other carbon dioxide emissions in office buildings. The results obtained in
hand, commonly used in forecasting building energy usage. The this study indicated that, due to the greater precision of its pre-
information used to train data-driven models is often historical dictions, the ANN outperforms the other methods. Also, the ANN's
measurements of the external environment and energy use. The ability to predict the diurnal cooling energy was examined by Deb
purpose of the training process is to discover the associated pa- et al. [26]. The ANN was proved to be useful in calculating the daily
rameters of data-driven models and to examine the complex non- energy usage employing the records of five previous days.
linear connection between multiple input and output datasets Other than ANNs, a variety of AI tools such as support vector
[9,10]. machine (SVM) regression, neuro-fuzzy systems, and random for-
The energy performance of buildings can be evaluated through est have been employed to deal with EPB problems [30]. Li et al.
modeling and simulation (first-principle models), and, nowadays, [31] performed a comparative study to calculate the cooling load of
researchers are seeking to present the most practical and a building per hour. In this research, a comparison was made be-
straightforward methods in this regard [11]. Indoor environmental tween the performance of the SVM and three conventional types of
input parameters, including indoor air quality, thermal environ- ANN, including general regression neural network (GRNN), back-
ment, lighting, and acoustics, have been examined [12]. The major propagation neural network (BPNN), and radial basis function
aim of the study was to consider these parameters to design and neural network (RBFNN). According to the acquired error criteria,
evaluate the energy performance. The values of the indoor envi- the SVM and GRNN would propose more appropriate computa-
ronment employed for design were based on the required standard tional abilities than the RBNN and RBFNN as prediction tools.
consisting of long-term assessment in the indoor environment. Moreover, a hybrid approach based on SVM, wavelet transform, and
Also, the principles and philosophy considered in this standard Partial Least Squares regression (PLS) was developed to simulate
were illustrated through an example. Tronchin and Fabbri [13] the dynamic cooling and heating load in an office building in
studied the energy performance building evaluation in Mediterra- Tianjin, China, by Zhao and Liu [32]. In order to gain the best inputs
nean countries using software simulations and operating rating compound, after sensitivity analysis, a multiple-horizon calculation
simulation. Hence, three different models for EPB software calcu- was performed. The obtained results implied that the proposed
lations have been analyzed and compared to quantify their gap model could precisely illuminate the concept of the problem.
with the real energy consumptions. In 2014, energy consumption Allahyari et al. [33] developed an ANN-based model for energy
and renewable energies usage in Iran were investigated by Heravi management in office buildings in Bandar Abbas, Iran. The neural
and Qaemi [14]; the purpose of this study was to evaluate the network results for various inputs revealed that 5-cm-thick brick
design and construction measures concerning building energy ef- wall insulation, joist ceiling with 5-cm-thick insulation, and triple-
ficiency. The results obtained in the study imply the fact that pas- layered glass were the best among the scenarios introduced to
sive solar energy is one of the most advantageous renewable decrease energy utilization and CO2 emission. Baheri et al. [34]
energy systems in urban regions and buildings of Iran. Also, Tavares optimized the energy efficiency of a residential complex in the
et al. [15] attempted to illuminate the benefits of Electrochromic warm and dry climate of Iran, employing a hybrid model of the
(EC) windows in the energy savings of the buildings situated in the genetic algorithm and artificial neural network. The authors
Mediterranean climates. The ESP-r version 11.10 building simula- declared that the results could help construct or reconstruct
tion program was employed to model the EC window dynamic energy-efficient residential buildings in Iran. Fan et al. [35] pro-
behavior and perform various control strategies. The results proved posed a transfer learning-based model to predict 24 h ahead
that when the EC windows were utilized in the west façade, the building energy demand. Based on their obtained results, the pro-
most effective performance was obtained, while the south façade posed transfer learning-based method could decrease about 15% to
has no considerable benefits. 78% of prediction errors compared to standalone models. A stacking
One of the main drawbacks of the first-principle models was model for predicting the building energy consumption was devel-
that certain factors were almost impossible to be adjusted in oped by Wang et al. [36]. The results revealed that the presented
practice. The main challenge with performing simulation and stacking methodology could present better efficiency regarding
modeling is that the EPB cannot be accurately estimated [16]. Also, accuracy, generalization, and robustness. Employing operation data
this method was time-consuming, expensive, and consists of many of real buildings, Fan et al. [37] studied the performance of various
factors. In fact, when the other factors do not change, these ap- deep learning methods in automatically deriving high-quality
proaches are useful to calculate the influences of a single factor on features for building energy predictions. They declared that the
EPB [17]. In order to overcome the defects of the simulation tools in deep learning-based feature engineering methods are purely data-
examining the EPB, the engineers have presented an inverse (data- driven, which helps to fully automate the building energy predic-
driven) modeling [18]. Accordingly, the influential factors' effect tion process. Ivanko et al. [38] considered the prediction of DHW
like comparative compactness, roof area, and direction is appro- heat use in hotels in Norway for two common situations (1. when
priately assessed for EPB [19e22]. Some scholars suggested the prediction is based on historical DHW heat use and 2. when
employing artificial intelligence (AI) due to its high-performance additional explanatory variables added in the model). Therefore,
2
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

the best DHW heat use model from various prediction techniques 2.1. Multi-layer perceptron (MLP)
was selected.
In 1943, McCulloch and Pitts introduced the concept of ANN for
the first time, and they are considered a pioneer in this field [41]. As
1.1. Main novelties and contributions they have such a remarkable capability for non-linear mapping of
parameters, in previous years, they have seen many ANNs
In general, the literature review shows that, so far, many studies employed in many statistical analyses. Also, MLP is equipped with
have been considered on EPB, ANNs, and AI benefits in EPB evalu- appropriate flexibility in structure and proper representational
ation. However, a closer look at these studies reveals a number of capability and has access to a wide range of data samples [42].
gaps and shortcomings for intelligent models' usage like ANNs. Hence, it is regarded as the most practical and widely used ANN
Hybrid optimization algorithms were presented, and Multi-Layer compared to the other types. The backpropagation algorithm is
Perceptron (MLP) was used to optimize the models by Moayedi utilized for training MLPs, and they are known as feedforward
and Hayati [24]. Actually, metaheuristic algorithms have been neural tools and general approximators [43]. Through the use of
widely employed to optimize the models. For instance, Tien Bui computational units termed “neurons,” the MLPs are approxi-
et al. [39] and Bahiraei et al. [40] have examined the efficiency of mately able to anticipate each input-output planx. Fig. 1 outlines
several algorithms such as Genetic Algorithm (GA), Imperialist the overall structure of MLP considered in this study, according to
Competition Algorithm (ICA), and Elephant Herding Optimization which input and output layers, as well as a hidden layer, constitute
(EHO). Nevertheless, in no study, two parameters of heating and this structure. Besides, there is a strong connection between the
cooling in the building have been considered to train a neural neurons of the adjacent layer(s) [44]. In order to bear the structure
network through the method proposed in this study. It is also worth of MLP in mind with detail, Fig. 1 is worth devoting attention to.
mentioning that the results obtained in this paper have never been
presented elsewhere. Therefore, the present study aims to optimize
2.2. The used algorithms
the parameters of a neural network using experimental data. The
heating and cooling of a building have been considered, and the
The algorithms considered in this paper are described here, and
artificial neural network has been trained based on these param-
the aim is to find the best one in terms of accuracy, authenticity, and
eters. In this research, the neural network is of multi-layer per-
expedition.
ceptron MLP type. Furthermore, six algorithms, namely
Biogeography-Based Optimization (BBO), GA, Particle Swarm
Optimization (PSO), Population-Based Incremental Learning (PBIL), 2.2.1. Genetic algorithm (GA)
Evolution Strategy (ES), and Ant Colony Optimization algorithm In general, the constrained and unconstrained optimization
(ACO), are employed to optimize the optimization parameters of problems can be solved through a GA method based on mimicking
the MLP, and their performances are compared to each other ac- biological evolution in a natural selection process. This algorithm
cording to the criteria of Root Mean Square Deviation (RMSD), can be used without any specific programming efforts, but the so-
Normalized Root Mean Square error Deviation (NRMSD), the co- lution must be encoded as strings of parameters. One of the major
efficient of determination (R2), and Mean Absolute Error (MAE). advantages of this algorithm is its remarkable capability to tackle a
Finally, with respect to the obtained results, BBO was the most specific optimization domain. With respect to the benefits of this
beneficial and feasible algorithm in this analysis regarding accuracy algorithm, the MLP is optimized based on a GA method.
and authenticity. In addition, a detailed comparison between the
experimental results and the analytical results was performed to 2.2.2. Particle swarm optimization (PSO)
validate the results obtained in this paper. So far, most early studies, as well as current work, have exploited
PSO [45,46]. This method for the first time, in 1995, was introduced
[47]. This type of optimization is inspired by animals herding
1.2. Paper organization
behavior and is regarded as a new and well-known stochastic
optimization approach among the others. As is known, a leader
The rest of this paper was organized as follows: Information on
always tries to head the groups, and the members have to follow
the approach, parameters, and algorithms considered in this study
the leader. The same rule is applied in PSO, and the aim is to specify
is given in Section 2. Then, the third section illustrates the data used
the best solution. Particles are called the possible solution in this
to train the neural network and outlines the statistical analysis. The
method, and their task is to seek solutions by following the most
results are discussed and analyzed in the fourth section. Finally, the
suitable member. Overall, the particles' location and expedition to
main conclusions and a brief description of the current paper are
find the optimal solutions are updated based on each particle's best
provided in Section 5.
previously determined location and the global best location of the
swarm.
2. Methodology
2.2.3. Population-Based Incremental Learning algorithm (PBIL)
As mentioned earlier, two main parameters in the EPB, namely Shumeet Baluja [48] introduced a PBIL for the first time in 1994.
cooling and heating, are considered first. Then, the multi-layer The Genetic Algorithm and competitive learning mechanisms
perceptron MLP is trained according to these parameters. The constitute this algorithm, which is also considered a new Evolu-
experimental data are inserted into this ANN type, and this neural tionary Algorithm(EA) [49]. Furthermore, the PBIL algorithm is
network is trained based on the cooling and heating parameters. known as an optimization algorithm in computer science and
Furthermore, five optimization methods are employed to optimize machine learning. One of the significant advantages of this type of
the parameters of MLP. Finally, to reveal the authenticity and algorithm is its simple implementation compared to the standard
practicality of the models, a comparison of the obtained results and genetic algorithm and sometimes outperforms a standard genetic
experimental results was performed. algorithm in terms of the results' accuracy [50].

3
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Fig. 1. The overall structure of the MLP associated with the parameters of this research.

2.2.4. Evolution strategy optimization (ES) 3. Statistical analysis and the used data
The ES optimization is regarded as one of the most important
optimization methods in computer science as it is useful to opti- As stated earlier, this research aims to determine the best type of
mize the MLP based on the parameters of heating and cooling. optimization according to which the MLP optimization parameters
Besides, this optimization is related to evolutionary computation or are optimized with the highest level of accuracy and authenticity.
artificial evolution techniques. Hence, Fig. 2 illustrates the steps of solving this research problem,
and the details are outlined as a flowchart in this figure. It is worth
mentioning that the steps considered to analyze the optimization
techniques presented in this paper are the same for all of them.
2.2.5. Ant colony optimization algorithm (ACO)
However, BBO optimization is shown as an example of them, and
The ant colony algorithm is used to specify optimal paths and is
each optimization technique, such as ACO and ES, can be consid-
operated based on the ants' behavior for seeking food. In the
ered instead of BBO in this flowchart.
beginning, the ants wander everywhere randomly and make an
immediate attempt to go back to the colony once they find a source
3.1. The used data
of food. On their way, they leave some pheromones showing the
path of the food. The ACO algorithm is typically utilized in com-
Considering the elementary cube (3:5
3:5
3:5), twelve
puter science and operations research and is also known as a
building shapes were generated where each building form is
probabilistic method for solving computational problems. Through
composed of 18 elements (elementary cubes). All the buildings
this algorithm, the shortest path between two particular points is
have the same volume, which is 771:75 m3 , but different surface
found in the graph.
areas and dimensions. The materials used for each of the 18 ele-
ments are the same for all building forms. For all building shapes,
the materials utilized for each of the 18 components are the same.
2.2.6. Biogeography-based optimization (BBO) The newest and most prevalent materials in the building con-
BBO refers to an EA that optimizes a function through stochas- struction sector, as well as the lowest U-value: walls (1:78 W=m2 K),
tically and iteratively improving candidate solutions associated floors (0:86 W=m2 KÞ; roofs (0:50 W=m2 K), windows
with a given measure of quality or fitness function. Actually, BBO is 2
(2:26 W=m K), were used to make the sample buildings. These
regarded as one of the novel metaheuristics techniques that were buildings were each simulated as a residential building in Athens,
presented by Simon [51] for the first time and had a remarkable Greece. The shape of each building type is shown in Fig. 3.
performance to solve continuous problems. Biogeography is known The simulation assumes that the buildings are located in Athens,
as science learning regarding the distribution of some particular Greece, are residential with a maximum of seven occupants, and
species based on geographic conditions [52]. The species migrate are used for sedentary purposes (70 W). The following re-
from one island to another, rise, and become extinct according to quirements were established for interior design: 0:6 clothing, 60%
some principles applied in this method. humidity, 0:30 m=s air speed, 300 Lux illumination intensity.
4
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Internal gains were set at sensible (5) and latent (2 W=m2 ), while
the infiltration rate was set to 0:5 for an air change rate with a wind
sensitivity of 0:25. Thermal characteristics were determined using a
mixed mode with a 95% efficiency, a thermostat range of 19e24  C,
and 15e20 hours of operation on weekdays and 10e20 hours on
weekends.
Clearly, valid and enough data has been regarded as the essen-
tial tool to make the most use of the techniques proposed in this
paper. The dataset source employed for the intelligent models is
based on the study conducted by Tsanas and Xifara [53].
Each input parameter corresponds to a different attribute of the
structure. For example, the RC indication identifies the kind of
structure, and the RC-value of each structure is displayed in Fig. 3.
The RC is determined as follows [54]:

RC ¼ 6V 2=3 A1 (1)

where V and A are the volume and surface area of the building,
respectively.
The research models the structure in two different configura-
tions, with and without glazing. In a glazing system, twelve
building designs, three glazing area with five glazing area distri-
butions, and four orientations to come up with 12
3
5
4 ¼
720 samples. For the four orientations without glazing, twelve
different building types were considered. Therefore, in total
12
3
5
4 þ 12
4 ¼ 768 buildings were studied.
The parameters of HL and CL are considered for the residential
buildings named Y1 and Y2 and are affected by eight significant
factors that are as follows: relative compactness, which represents
the ratio of the surface area to the volume in the building [55],

Table 1
The values and conditions assumed to create the database.

Factor Values and states

Number of the buildings 12

The buildings area 771:75 m2
Number of the elements 18
Fig. 2. The flowchart of the analysis considered to present the best type of Number of the orientations 4
optimization. Glazing area 0%; 10%; 25% and 40% of the floor area

Fig. 3. Shapes of the buildings.

5
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

surface area, roof area, wall area, orientation, overall height, glazing and CL of a residential building have been simulated. The BBO, GA,
area which is known as the total area of the glazing calculated PSO, PBIL, ES, and ACO methods have been presented and coded
through the rough opening consisting of the glazing, frame, and through the programming language of MATLAB version 18. Also,
sash [54]; also, glazing area distribution of a residential building 768 samples, with eight independent factors mentioned earlier that
named X1, X2, …, X8. affect HL and CL values, constituted the dataset employed to train
According to Table 1, twelve buildings, four glazing areas with the proposed network. Among these samples, 614 (80% of all data)
five distribution scenarios, and four orientations are considered were considered to train the networks, and the rest (20% of all data)
under the study. to assess the generalization potential of the BBO, GA, PSO, PBIL, ES,
It should be mentioned that the distribution scenarios are and ACO methods. In addition, three accuracy criteria of R2, MAE,
divided into five conditions: uniform (25% glazing on each side), and RMSE were analyzed to compare the mentioned models. A
south (55% on the south and 15% on each other side), north (55% on considerable body of literature exists on these criteria [23,57]. In
the north and 15% on each other side), east (55% on the east and 15% order to obtain the optimal structure of the models, an extensive
on each other side), and west (55% on the west and 15% on each trial and error process was conducted. Then, the prediction results
other side). of each model for the values of HL and CL were presented in detail.
Table 2 outlines the main criteria employed for the statistical
analysis of the dataset that is as follows: the data averages, standard 4.1. Network optimization
error, median, mode, standard deviation, sample variance, skew-
ness, and minimum and maximum values. In fact, a wide range of parameters are affected by soft
computing techniques, and regardless of these parameters, the
3.2. The main equations and number of processor units is also important. In order to specify the
best algorithm for optimization, the most efficient performance
In this study, the criteria of R2, MAE, RMSE are employed to and the most suitable structure are needed to be found in the used
evaluate the accuracy of the intelligent models. It is worth algorithms. Therefore, a variety of structures are designed and
mentioning that, so far, seminal contributions have been made by examined in terms of RMSE. A backpropagation-based network,
these criteria, according to which the results have been obtained with one hidden layer and ten distinctive numbers of neurons, is
correctly. Eqs. (2)e(4) represent these criteria compared to each considered for the MLP. According to the sensitivity analysis results,
other to illustrate the performance of the employed models [56]. the MLP network containing four nodes in its hidden layer can
" , # deliver the most accurate approximation for the CL and HL
N 
X 2 X
N  2 parameters.
2
R ¼1  Yipredicted  Yiobserved Yiobserved  Y observed
i¼1 i¼1
(2)

N 
X 
 
MAE ¼ ð1 = NÞ Yiobserved  Yipredicted  (3)
I¼1

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N h i2
u X
RMSE ¼ tð1=NÞ Yiobserved  Yipredicted (4)
I¼1

Concerning the equations above, Yiobserved and Yipredicted represent

the actual and predicted values of HL or CL of the efficient building,
respectively. Also, the number of distances is shown by the term N,
and Yobserved denotes the average of the actual values of HL (or CL).

4. Results and discussion

In this research, the MLP neural network has been optimized

using five algorithms mentioned in the second section, and the HL
Fig. 4. RMSE and iteration number for the BBO, GA, PSO, PBIL, ES, and ACO.

Table 2
The Statistical analysis of energy-efficient design on ten features of the efficient building.

Characteristics No. of conceivable factors Descriptive indices

Mean Standard error Median Mode Standard deviation Sample variance Skewness Min. Max.

Relative compactness 12 0:76 0:00 0:75 0:98 0:11 0:01 0:50 0:62 0:98
Surface area 12 671:71 3:18 673:75 514:50 88:09 7759:16  0:13 514:50 808:50
Wall area 7 318:50 1:57 318:50 294:00 43:63 1903:27 0:53 245:00 416:50
Roof area 4 176:60 1.63 183.75 220.50 45:17 2039:96  0:16 110:25 220:50
Overall height 2 5.25 0.06 5.25 7.00 1:75 3:07 0:00 3:50 7:00
Orientation 4 3.50 0.04 3.50 2.00 1:12 1:25 0:00 2:00 5:00
Glazing area 4 0.23 0.00 0.25 0.10 0:13 0:02  0:06 0:00 0:40
Glazing area distribution 6 2.81 0.06 3.00 1.00 1:55 2:41  0:09 0:00 5:00
Heating load 586 22.31 0.36 18.95 15.16 10:09 101:81 0:36 6:01 43:10
Cooling load 636 24.59 0.34 22.08 21.33 9:51 90:50 0:40 10:90 48:03

6
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Fig. 4 represents the RMSE for each algorithm based on iteration errors shown in Figs. 5 and 6. In this case, the model's capability for
number according to which BBO has the best performance since its prediction is clearly specified using these measurements.
RMSE follows a decreasing trend as iteration number increases. The error analysis was conducted for both parameters, and the
According to Fig. 4, the BBO has a remarkable performance in BBO, GA, PSO, PBIL, ES, and ACO were evaluated in terms of accu-
terms of RMSE and iteration number, while the ACO is not racy. According to Figs. 5 and 6, the error in the prediction ability of
acceptable at all in this regard. Hence, the BBO has the highest BBO is almost zero.
convergence compared to the others, and this feature implies the For the sake of validation, a careful comparison was made be-
fact that the BBO outperforms the other mentioned algorithms. tween the results of the trained data and experimental data. Hence,
R2 was considered for the trained data and test data according to CL
and HL parameters. These results have been highlighted in Figs. 7
4.2. Model evaluation and 8.
Overall, a detailed look at Figs. 7 and 8 reveal the fact that the
Out of 768 samples, the 80:20 ratio is considered to select the BBO has excellent performance for prediction. On the other hand,
training and testing datasets randomly. The models were used with there is a good agreement between the analytical results and
their optimal structure, and a careful comparison was made be- experimental, and this implies the authenticity of the training
tween the outputs and the target data to assess the integrity of the process of the MLP through the mentioned algorithms. It is note-
applied models. The statistical indices of R2, RMSE, and MAE were worthy that the R2 criterion for the ACO indicates that this algo-
also analyzed. Figs. 5 and 6 indicate a graphical comparison be- rithm is not capable of predicting the target data and optimizing
tween the predicted and actual values of HL and CL. The frequency MLP parameters with high accuracy and has the lowest acceptable
of each error value was outlined through the histogram of the

Fig. 5. Histogram diagram and error limitation for HL prediction through the BBO, GA, PSO, PBIL, ES, and ACO.

7
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Fig. 6. Histogram diagram and error limitation for CL prediction through the BBO, GA, PSO, PBIL, ES, and ACO.

performance. Furthermore, the GA algorithm can be a good alter- 4.3. Models comparison
native for BBO and provide the most accurate results and pre-
dictions after BBO. The gap between the exact results and the Tables 3 and 4 give the complete information regarding the
results obtained through the ES and PBIL is also considerable. values of the criteria obtained through BBO, GA, PSO, PBIL, ES, and
Therefore, ES and PBIL are not reliable optimizations for MLP ACO for HL and CL parameters, respectively.
and HL, and CL parameters. As it is obvious, the less RMSD, the more accuracy of the opti-
Figs. 9 and 10 illustrate the values of the experimental data and mization model. With respect to the results outlined in Tables 3 and
the trained data per sample number obtained through the BBO, GA, 4, the BBO has the highest accuracy since it has the least value of
PSO, PBIL, ES, and ACO for HL and CL parameters. RMSD compared to other models. The value of RMSD based on
As stated earlier, the main reason for employing training data training data for the HL parameter is 2.82, and it is 3.47 according to
and test data as well as comparing them is to prove the accuracy the test data. Concerning the CL parameters, this value for training
and reliability of the analysis conducted in this research. Accord- and test data is 3.18 and 3.31, respectively. Clearly, the ACO has the
ingly, the data of BBO, GA, PSO, PBIL, ES, and ACO for HL and CL lowest accurate performance and is not regarded as a suitable
parameters are compared in detail, and their distance from the alternative for the optimization models. The NRMSD is the second
target data is the main criterion to represent their authenticity. criterion that has been considered to evaluate the performance of
the models in this research. As shown in Tables 3 and 4, there is a
minor difference between the values of NRMSD among the models

8
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Fig. 8. The results of R2 for the trained data of CL and HL based on the BBO, GA, PSO,
PBIL, ES, and ACO.
Fig. 7. The results of R2 for the test data of CL and HL based on the BBO, GA, PSO, PBIL,
ES, and ACO.
most values of R2 for CL and HL training and experimental data.
Thus, this model is reliable in terms of accuracy and precise pre-
for CL and HL parameters, both in training data and in test data. For diction. In addition to the aforementioned explanations, the results
instance, there is no difference between the NRMSD value of BBO of MAE indicate the BBO's excellent performance to optimize the
and GA in the test data of the CL parameter. A lower NRMSD value optimization parameters of MLP. According to Table 3, the MAE
means more accuracy. Hence, the BBO is able to predict the results value for the ES based on the training data is equal to 8.43 and
with the minimum error as it has the lowest value of NRMSD in all consequently is considered the lowest reliable optimization model
conditions. Furthermore, the BBO is also proved to be the best in this regard. On the other hand, the ACO has the most value of
optimization model according to the results of R2, and it has the MAE, 8.79, in the test data of the HL parameter. Nevertheless, the

9
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Fig. 9. The values of the trained and test data per sample number for HL.

Fig. 10. The values of the trained and test data per sample number for CL.

10
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

Table 3 Credit author statement

Obtained RMSE, R2 and MAE for models (HL parameer).

Models Data Yuanjin Xu: Conceptualization, Investigation, Methodology,

Training Testing
Software, Validation, Writing - review & editing, Language review.
Fei Li: Conceptualization, Supervision, Formal analysis, Methodol-
RMSD NRMSD R2 MAE RMSD NRMSD R2 MAE
ogy, Investigation, Writing - review & editing, Language review.
BBO 2.82 0.076 0.920 2.15 3.47 0.114 0.944 2.35 Armin Asgari: Conceptualization, Investigation, Methodology, Su-
GA 3.01 0.081 0.909 2.25 3.84 0.126 0.932 2.73
pervision, Resources, Data curation.
PSO 6.46 0.174 0.597 4.73 6.66 0.218 0.623 3.72
PBIL 8.11 0.219 0.733 6.90 4.55 0.149 0.855 5.58
ES 9.98 0.269 0.648 8.43 6.78 0.222 0.740 6.65 Declaration of competing interest
ACO 10.17 0.306 0.624 6.67 4.84 0.159 0.842 8.79

The authors declare that they have no known competing

financial interests or personal relationships that could have
Table 4 appeared to influence the work reported in this paper.
Obtained RMSE, R2 and MAE for model (CL parameter).

Models Data Acknowledgements

Training Testing

RMSD NRMSD R2 MAE RMSD NRMSD R2 MAE This work was supported by the Fundamental Research Funds
for the Central Universities, China University of Geosciences
BBO 3.18 0.086 0.887 2.97 3.31 0.100 0.900 2.46
GA 3.72 0.100 0.847 3.29 3.29 0.100 0.893 2.41
(Wuhan) (No.CUGQY1911). This study was supported by the Key
PSO 5.14 0.139 0.771 5.07 4.15 0.126 0.826 3.01 Research Program of the Innovation Academy for Green Manufac-
PBIL 6.76 0.182 0.738 3.58 5.10 0.154 0.832 4.17 ture, Chinese Academy of Sciences (No. IGAM-2019-A16-1), and the
ES 8.32 0.224 0.679 5.08 5.80 0.176 0.835 4.49 Key Research Program of the Alliance of International Science
ACO 11.26 0.303 0.265 3.86 8.92 0.270 0.360 6.65
Organizations(ANSO-CRKP-2020-02).

References
BBO generally has the lowest value of MAE compared to the other
models both in training data (MAE for HL and CL is 2.15 and 2.97 [1] Hassan HS, Abdel-Gawwad HA, Va squez-García SR, Israde-Alca ntara I, Flores-
respectively) and test data (MAE for HL and CL is 2.35 and 2.46 Ramirez N, Rico JL, et al. Cleaner production of one-part white geopolymer
respectively). Finally, it can be concluded that the whole statistical cement using pre-treated wood biomass ash and diatomite. J Clean Prod
2019;209:1420e8.
results represent the remarkable benefits of BBO for the prediction [2] Liu G, Yang H, Fu Y, Mao C, Xu P, Hong J, et al. Cyber-physical system-based
and optimization of MLP's optimization parameters. real-time monitoring and visualization of greenhouse gas emissions of pre-
fabricated construction. J Clean Prod 2020;246:119059.
[3] Ye Z, Cheng K, Hsu S-C, Wei H-H, Cheung CM. Identifying critical building-
oriented features in city-block-level building energy consumption: a data-
5. Summary driven machine learning approach. Appl Energy 2021;301:117453.
[4] Bourdeau M, qiang Zhai X, Nefzaoui E, Guo X, Chatellier P. Modeling and
In summary, the present paper aims to determine the optimal forecasting building energy consumption: a review of data-driven techniques.
Sustain Cities Soc 2019;48:101533.
optimization model to train the MLP and optimize the optimization [5] McQuiston FC, Parker JD, Spitler JD. Heating, ventilating, and air conditioning:
parameters of this neural network. The experimental data is analysis and design. John Wiley & Sons; 2004.
inserted into the trained and optimized MLP. The data is also [6] Turhan C, Kazanasmaz T, Uygun IE, Ekmen KE, Akkurt GG. Comparative study
of a building energy performance software (KEP-IYTE-ESS) and ANN-based
trained based on cooling and heating parameters that are the input
building heat load estimation. Energy Build 2014;85:115e25.
parameters of the energy performance of buildings. The optimi- [7] Dogan T, Shoeboxer Reinhart C. An algorithm for abstracted rapid multi-zone
zation models consist of BBO, GA, PSO, PBIL, ES, and ACO. These urban building energy model generation and simulation. Energy Build
2017;140:140e53.
models are assessed based on the four criteria, which are as fol-
[8] Saffari M, de Gracia A, Ushak S, Cabeza LF. Passive cooling of buildings with
lows: RMSD, NRMSD, R2, and MAE. Besides, the analytical results of phase change materials using whole-building energy simulation tools: a re-
these models for HL and CL are considered according to both view. Renew Sustain Energy Rev 2017;80:1239e55.
training and test data to prove the reliability of the findings. Since [9] Wei Y, Zhang X, Shi Y, Xia L, Pan S, Wu J, et al. A review of data-driven ap-
proaches for prediction and classification of building energy consumption.
the BBO has the highest value of R2 and lowest values of RMSD, Renew Sustain Energy Rev 2018;82:1027e47.
NRMSD, and MAE, it was revealed that this model generally out- [10] Amasyali K, El-Gohary NM. A review of data-driven building energy con-
performs other optimization models mentioned in this paper and sumption prediction studies. Renew Sustain Energy Rev 2018;81:1192e205.
[11] Zhao H, Magoule s F. A review on the prediction of building energy con-
has the highest convergency. It is noteworthy that BBO has a sumption. Renew Sustain Energy Rev 2012;16:3586e92.
remarkable capability of prediction since its results are strikingly [12] Olesen BW. The philosophy behind EN15251: indoor environmental criteria
close to target data. Furthermore, a careful comparison was made for design and calculation of energy performance of buildings. Energy Build
2007;39:740e9.
between the results of training data and test data. Finally, a good [13] Tronchin L, Fabbri K. Energy performance building evaluation in Mediterra-
agreement has been observed between the results of training data nean countries: comparison between software simulations and operating
and experimental data. Hence, the forecasting accuracy of the rating simulation. Energy Build 2008;40:1176e87.
[14] Heravi G, Qaemi M. Energy performance of buildings: the evaluation of design
proposed MLP neural network optimized with the BBO optimiza- and construction measures concerning building energy efficiency in Iran.
tion algorithm with the RMSD, R2, and MAE of 2.82, 0.920, 2.15 in Energy Build 2014;75:456e64.
the training phase of heating load and with the RMSD, R2, and MAE [15] Tavares PF, Gaspar AR, Martins AG, Frontini F. Evaluation of electrochromic
windows impact in the energy performance of buildings in Mediterranean
of 3.18, 0.887, 2.97 in the training phase of the cooling load is much
climates. Energy Pol 2014;67:68e81.
better than those of the other models. [16] Park JS, Lee SJ, Kim KH, Kwon KW, Jeong J-W. Estimating thermal performance
In future work, investigating the other optimization models in and energy saving potential of residential buildings using utility bills. Energy
this regard might be useful to enhance the capability of prediction Build 2016;110:23e30.
[17] Yezioro A, Dong B, Leite F. An applied artificial intelligence approach towards
tools. Looking forward, further attempts could prove quite benefi- assessing building performance simulation tools. Energy Build 2008;40:
cial to the literature. 612e20.

11
Y. Xu, F. Li and A. Asgari Energy 240 (2022) 122692

[18] O'Neill Z, O'Neill C. Development of a probabilistic graphical model for pre- 2019;240:35e45.
dicting building energy performance. Appl Energy 2016;164:650e8. [38] Ivanko D, Sørensen ÅL, Nord N. Selecting the model and influencing variables
[19] Catalina T, Virgone J, Blanco E. Development and validation of regression for DHW heat use prediction in hotels in Norway. Energy Build 2020;228:
models to predict monthly heating demand for residential buildings. Energy 110441.
Build 2008;40:1825e32. [39] Tien Bui D, Moayedi H, Anastasios D, Kok Foong L. Predicting heating and
[20] Gao W, Guirao JLG, Basavanagoud B, Wu J. Partial multi-dividing ontology cooling loads in energy-efficient buildings using two hybrid intelligent
learning algorithm. Inf Sci (Ny) 2018;467:35e58. models. Appl Sci 2019;9:3543.
[21] Gao W, Wu H, Siddiqui MK. Baig AQ. Study of biological networks using graph [40] Bahiraei M, Heshmatian S, Goodarzi M, Moayedi H. CFD analysis of employing
theory. Saudi J Biol Sci 2018;25:1212e9. a novel ecofriendly nanofluid in a miniature pin fin heat sink for cooling of
[22] Yu Z, Haghighat F, Fung BCM, Yoshino H. A decision tree method for building electronic components: effect of different configurations. Adv Powder Technol
energy demand modeling. Energy Build 2010;42:1637e46. 2019;30:2503e16.
[23] Seyedashraf O, Mehrabi M, Akhtari AA. Novel approach for dam break flow [41] McCulloch WS, Pitts W. A logical calculus of the ideas immanent in nervous
modeling using computational intelligence. J Hydrol 2018;559:1028e38. activity. Bull Math Biophys 1943;5:115e33.
[24] Moayedi H, Hayati S. Artificial intelligence design charts for predicting friction [42] Fan X, Wang L, Li S. Predicting chaotic coal prices using a multi-layer per-
capacity of driven pile in clay. Neural Comput Appl 2019;31:7429e45. ceptron network model. Resour Pol 2016;50:86e92.
[25] Biswas MAR, Robinson MD, Fumo N. Prediction of residential building energy [43] Zarei T, Behyad R. Predicting the water production of a solar seawater
consumption: a neural network approach. Energy 2016;117:84e92. greenhouse desalination unit using multi-layer perceptron model. Sol Energy
[26] Deb C, Eang LS, Yang J, Santamouris M. Forecasting diurnal cooling energy 2019;177:595e603.
load for institutional buildings using Artificial Neural Networks. Energy Build [44] Heidari M, Shamsi H. Analog programmable neuron and case study on VLSI
2016;121:284e97. implementation of Multi-Layer Perceptron (MLP). Microelectron J 2019;84:
[27] Escando n R, Ascione F, Bianco N, Mauro GM, Su arez R, Sendra JJ. Thermal 36e47.
comfort prediction in a building category: artificial neural network generation [45] Khanna V, Das BK, Bisht D, Singh PK. A three diode model for industrial solar
from calibrated models for a social housing stock in southern Europe. Appl cells and estimation of solar cell parameters using PSO algorithm. Renew
Therm Eng 2019;150:492e505. Energy 2015;78:105e13.
[28] Fan C, Wang J, Gang W, Li S. Assessment of deep recurrent neural network- [46] Liu Z, Zhu P, Chen W, Yang R-J. Improved particle swarm optimization algo-
based strategies for short-term building energy predictions. Appl Energy rithm using design of experiment and data mining techniques. Struct Multi-
2019;236:700e10. discip Optim 2015;52:813e26.
[29] Pino-Mejías R, Pe rez-Fargallo A, Rubio-Bellido C, Pulido-Arcas JA. Comparison [47] Kennedy J, Eberhart R. Particle swarm optimization. Proc. ICNN’95-Int Conf
of linear regression and artificial neural networks models to predict heating Neural Netw 1995;4:1942e8. IEEE.
and cooling energy demand, energy consumption and CO2 emissions. Energy [48] Baluja S. Population-based incremental learning. a method for integrating
2017;118:24e36. genetic search based function optimization and competitive learning. Carne-
[30] Dong B, Cao C, Lee SE. Applying support vector machines to predict building gie-Mellon Univ Pittsburgh Pa Dept Of Computer Science; 1994.
energy consumption in tropical region. Energy Build 2005;37:545e53. [49] Zhang Q, Wu T, Liu B. A population-based incremental learning algorithm
[31] Li Q, Meng Q, Cai J, Yoshino H, Mochida A. Predicting hourly cooling load in with elitist strategy. Third Int. Conf. Nat. Comput. (ICNC 2007;3:583e7. IEEE;
the building: a comparison of support vector machine and different artificial 2007.
neural networks. Energy Convers Manag 2009;50:90e6. [50] Baluja S, Caruana R. Removing the genetics from the standard genetic algo-
[32] Zhao J, Liu X. A hybrid method of dynamic cooling and heating load fore- rithm. Mach. Learn. Proc. 1995:38e46. Elsevier; 1995.
casting for office buildings based on artificial intelligence and regression [51] Simon D. Biogeography-based optimization. IEEE Trans Evol Comput 2008;12:
analysis. Energy Build 2018;174:293e308. 702e13.
[33] Allahyari F, Behbahaninia A, Rahami H, Farahani M, Khadivi S. Development of [52] Huggett RJ. Fundamentals of biogeography. Routledge; 2004.
a model for energy management in office buildings by neural networks (case [53] Tsanas A, Xifara A. Accurate quantitative estimation of energy performance of
study: Bandar Abbas). Int J Environ Sci Technol 2020:1e10. residential buildings using statistical machine learning tools. Energy Build
[34] Baheri A, Najafi M, Azimi A, Aghanajafi C. Analyzing the sensitivity and energy 2012;49:560e7.
efficiency of a residential complex in warm and dry climate of Iran: the case [54] Huang Y, Niu J, Chung T. Comprehensive analysis on thermal and daylighting
study Yazd Province. Int J Environ Sci Technol 2021:1e30. file:///C:/Users/ performance of glazing and shading designs on office building envelope in
hypernet/Downloads/scholar (5. cooling-dominant climates. Appl Energy 2014;134:215e28.
[35] Fan C, Sun Y, Xiao F, Ma J, Lee D, Wang J, et al. Statistical investigations of [55] Ouarghi R, Krarti M. Building shape optimization using neural network and
transfer learning-based methodology for short-term building energy pre- genetic algorithm approach. Ashrae Trans 2006;112.
dictions. Appl Energy 2020;262:114499. [56] Wu Y, Xie P, Dahlak A. Utilization of radial basis function neural network
[36] Wang R, Lu S, Feng W. A novel improved model for building energy con- integrated with different optimization methods for the prediction of water
sumption prediction based on model integration. Appl Energy 2020;262: productivity of seawater greenhouse system. Energy Rep 2021;7:6658e76.
114561. [57] Mosallanezhad M, Moayedi H. Developing hybrid artificial neural network
[37] Fan C, Sun Y, Zhao Y, Song M, Wang J. Deep learning-based feature engi- model for predicting uplift resistance of screw piles. Arab J Geosci 2017;10:
neering methods for improved building energy prediction. Appl Energy 479.

12