https://doi.org/10.31449/inf.v48i9.5934 Informatica 48 (2024) 191–204 191 
 
Intelligent Environment Design for Indoor Spaces Based on 
Perception and Behavior Correlation 
Xiaomin Xue 
School of Design, Fujian University of Technology, Fuzhou 350118, China 
E-mail: 13860607040@163.com  
Keywords: perception, behavior, PMV indicators, intelligence, fitness 
Received: March 19, 2024 
With the development of social economy, higher demands have been presented for indoor environment. 
Intelligent design has become a research hotspot. A Genetic Algorithm-Back Propagation Neural 
Network model is constructed based on the association between perception and behavior, combined 
with the Predicted Mean Vote index. Six factors affecting the Predicted Mean Vote indicators are 
analyzed. However, there are inherent flaws in the Back Propagation Neural Network. Therefore, 
combined with Genetic Algorithm, a optimized model is built. It had faster convergence speed than the 
unimproved model. The difference of Predicted Mean Vote was small, with better model fitting effects. 
The overall model error remained around 0.01, with a maximum error of only 0.022. The model had 
higher Accuracy, Precision, and F1-score values compared with other models, with values of 97.89%, 
96.15%, and 0.896. From the results, it has better generalization ability, which can accurately predict 
indoor temperature, achieving intelligent control. The model proposed in the study achieves intelligent 
design of indoor comfort by controlling temperature, providing a reliable foundation for further 
improving indoor intelligence in subsequent research. 
Povzetek: Članek obravnava inteligentno zasnovo notranjih prostorov na osnovi korelacije med 
zaznavanjem in vedenjem. Avtorica predstavi model genetskega algoritma in povratne nevronske mreže, 
ki vključuje kazalnik PMV in omogoča hitrejšo konvergenco in natančnejše napovedi.
1 Introduction 
With the rapid development of artificial intelligence and 
Internet of Things technology, intelligent indoor 
environment design has become a hot research field. The 
demand for comfortable, energy-saving, and intelligent 
indoor environments is constantly increasing, which 
requires advanced technology [1-3]. In the design of 
indoor intelligent environments, the Predicted Mean Vote 
(PMV) is a crucial reference standard for evaluating 
indoor comfort, which comprehensively evaluates factors 
such as humidity, constant temperature, and air 
circulation in the indoor environment. However, due to 
the complexity of indoor environments, a single factor 
often cannot comprehensively evaluate indoor comfort. 
Therefore, multiple factors should be comprehensively 
considered to predict PMV. A Genetic Algorithm-Back 
Propagation Neural Network (GA-BPNN) prediction 
method based on perception and behavior correlation is 
constructed to improve the accuracy and intelligent 
regulation ability of PMV indicators in indoor intelligent 
environment design. By analyzing the influencing factors 
of PMV indicators and combining GA and BPNN, indoor 
temperature is intelligently predicted. The research has 
four parts. The first summarizes the research on 
intelligent design and BPNN, and analyze the research 
results. The second part is to construct the model and 
introduce the improved method. The third part is to verify 
the performance through comparative experiments. The 
fourth part is to summarize the experimental results, point 
out the shortcomings in the research, and propose future 
research directions. The study will provide a feasible 
solution for the intelligent design of indoor comfort, 
providing effective theoretical and methodological 
support for improving indoor intelligent environment 
design. 
2 Related works 
Many scholars have conducted relevant research on 
intelligent design [4-6]. Wang et al. designed a new car 
following model for intelligent transportation smoking. 
Based on the dual speed difference model and the 
dynamic characteristics of traffic flow, the front distance 
memory and rear-view effect were analyzed. The linear 
stability theory was used to describe the evolution 
characteristic equation of traffic flow density waves. It 
improved the traffic flow stability and eliminated traffic 
congestion [7]. Cheng et al. built a behavior prediction 
model in intelligent butler design. Limited hardware 
devices were used to collect sufficient information. A 
Quadro-W learning was developed to predict behavior. 
This study built a model based on the obtained Quadro-W 
information. The results indicated that the proposed 
model could effectively predict the initial environment 
according to environmental changes, improving the 
flexibility of the system [8]. Huang et al. proposed a new 

192 Informatica 48 (2024) 191–204                                                                      X. Xue 
intelligent reflection/refraction system. The intelligent 
surface of high-speed vehicles assisted users in 
communication, but the rapidly changing channels in 
communication posed great challenges. The results 
indicated that the proposed system could effectively work 
in high mobile communication scenarios and improve the 
reliability of the output [9]. Zhong et al. proposed the 
Lego modeling method in the intelligent robot design. 
Then a Deep Transfer Deterministic Policy Gradient 
(DT-DPG) was proposed to optimize the efficiency. The 
results showed that the proposed method could accurately 
describe indoor layout and channel status. The training 
efficiency was improved by 30%, which exceeded the 
DPG [10].  
Many scholars have carried out corresponding research 
on BPNN. Yu et al. designed a calibration method based 
on BPNN in multi-body dissipative particle dynamics. 
The study selected three simulation parameters: 
Attraction and repulsion coefficients A, B, and cutoff 
radius for repulsion, as well as four static and dynamic 
properties. The study compared three sampling methods. 
The proposed method was the most effective in reducing 
mean square error, which could accurately and effectively 
determine simulation parameters [11]. Sui et al. proposed 
a BPNN modeling method in the field of architecture. 
Virtual reality technology was used to construct models 
for evaluating building envelope structures and analyzing 
energy-saving building designs. Building models with 
different enclosure structures were constructed. In the 
identification of heat transfer coefficients in building 
envelope structures, the error of BPNN was less than 5%. 
It could effectively save energy, providing a theoretical 
basis for energy-saving design and evaluation of 
buildings [12]. Zhang et al. used BPNN to predict fuel 
ratios in the metallurgical field. The study first analyzed 
the 55 important production parameters in blast furnace 
operation. Finally, a comparative experiment was 
conducted using BPNN and k-nearest neighbor algorithm. 
From the results, the BPNN could keep the error within 
2% and achieve an accuracy of 93.02% [13]. Xu et al. 
developed a compensation method based on BPNN to 
solve nonlinear magnetic interference. To obtain 
sufficient data, a three-dimensional Helmholtz coil was 
used to recover the magnetic signal. It could lower the 
Root Mean Square Errors (RMSEs) of the north, east, and 
vertical components, as well as the total strength to 23.35, 
23.58, 27.42, and 29.72nT [14].  
Although there have been many achievements in 
intelligent design and BPNN in previous studies, the 
BPNN in intelligent design have inherent flaws [15-16]. 
Therefore, the GA is used to improve BPNN, effectively 
solving problems such as large errors and susceptibility to 
local minima. It is expected to update the intelligence of 
indoor environments, promoting the development and 
application of intelligent environmental design. The 
related works table is shown in table 1. 
 
Table 1: The related works table 
Study Method Technology Accuracy (%) 
Wang et al. Two-speed difference model Linear stability theory 94.16 
Cheng et al. Quadro-W Learning Intelligent butler design 93.33 
Huang et al. Intelligent reflex Refraction system 95.12 
Zhong et al. Lego modeling DT-DPG optimization 92.76 
Yu et al. BPNN calibration method Three types of simulation parameters 94.19 
Sui et al. BPNN modeling method Virtual reality technology 96.34 
Zhang et al. BPNN prediction method Blast furnace operating parameters 93.02 
Xu et al. BPNN compensation method Three-dimensional Helmholtz coil 96.74 
Research method GA-BPNN Perceived behavioral correlation 97.89 
 
3 Construction of GA-BPNN based 
on perception and behavior 
correlation 
Based on perceived behavioral correlation, six factors 
affecting human comfort indicators are analyzed. Then a 
model is constructed and optimized. 
3.1 BPNN based on perception and behavior 
Perception is the process in which a person interprets and 
gradually forms consciousness based on their own 
experience after perceiving external things. As an internal 
form of expression, it is caused by the connection 
between humans and the outside world. Behavior 
includes the internal feelings and external language 
actions that humans possess. The collaborative  
 
mechanism of perception and behavior considers humans 
as users of the environment. The characteristics and needs 
of humans directly affect the spatial characteristics 
exhibited in behavioral activities. There is a mutually 
penetrating and interconnected relationship among people, 
behavior, and space. In intelligent environment design, 
human thermal comfort is a crucial aspect. Thermal 
comfort refers to the thermal comfort state towards the 
environment, which is neither cold nor hot. It is applied 
to describe the satisfaction with indoor thermal 
environment. The widely used currently is the PMV, 
which is included in international standards. PMV covers 
almost all environmental factors that affect comfort. The 
calculation method for PMV is shown in formula (1). 

Intelligent Environment Design for Indoor Spaces Based on…         Informatica 48 (2024) 191–204   193 
0.036
[0.303 0.028]
M
PMV e L
−
=+ (1) 
In formula (1), M represents the metabolic rate, 
(
2
/ Wm ). L represents the human heat load. The PMV 
indicators describe the discrepancy in the real heat 
dissipation rate and the required heat dissipation to 
achieve a comfortable state under a given thermal 
environment. The calculation method for L is shown in 
formula (2). 
84
( ) 3.05 [5.733 0.007( ) ]
0.4( 58.15) 0.0173 (5087 )
0.0014(34 ) 3.96 10 ( 273)
()
a
a
a cl cl
cl c cl a
L M W M W P
W M M P
t f t
f h t t
−
= − −  − − −
− − − − −
− − −  +
−−
 (2) 
In formula (2), W represents the mechanical work, 
(
2
/ Wm ). 
a
P represents the partial pressure of air and 
water vapor, ( KPa ). 
a
t represents the air temperature, 
( C  ). 
cl
f represents the clothing area coefficient. 
cl
t 
represents the outer surface temperature of clothes, ( C  ). 
c
h represents the convective heat transfer coefficient, as 
shown in formula (3). 
0.25 0.25
0.25
2.38( ) 2.38( ) 12.1
12.1 2.38( ) 12.1
cl a cl a a
c
a cl a a
t t t t v
h
v t t v

− − 

=

−


 (3) 
In formula (3), 
a
v represents the air velocity. If the 
external work done by a person is zero, PMV can be 
expressed as formula (4). 
 ( , , , , , )
a r cl a
PMV f M t t I v  =   (4) 
Formula (4) represents the optimal comfortable 
environment for the human body. If formula (4) is not 
satisfied, it indicates that the environment is not in the 
optimal state, but it does not indicate that the 
environment has reached an uncomfortable level. The 
indicator that can represent the state of any set of 
environmental variables using a formula is called the 
average predictive response PMV. This indicator is 
divided into seven levels. The meaning of each level is 
represented in Figure 1. 
 
PMV value
Predictive sensation
Cold
Cool
A little cooler
Moderation
Slightly warmer
Warm
Hot
-3
-2
-1
0
1
2
3
 
Figure 1: Seven level of PMV thermal comfort 
 
Figure 1 is the seven level indicators of PMV thermal 
comfort. The PMV equation contains exponential terms, 
piece-wise functions, parameters, and multiple coupled 
factors. The calculation process involves multiple 
iterations of nonlinear equations. Multiple factors affect 
human thermal comfort, mainly including physical and 
human factors, as displayed in Figure 2. 
 

194 Informatica 48 (2024) 191–204                                                                      X. Xue 
Physical factor
Human factor
Ｍetabolism
Metabolic rate
Room 
temperature
Air velocity
Relative 
humidity
Mean 
radiation 
temperature
 
Figure 2: Factors affecting thermal comfort of human body 
 
Figure 2 shows the specific factors. Physical factors refer 
to the surrounding environment of the human body, 
including relative humidity, open-air temperature, air 
flow rate, and mean radiation temperature. Human factors 
are related to different states of the human body, 
including metabolic rate, and clothing thermal resistance. 
The above factors comprehensively affect the human 
body. When they are in the optimal arrangement and 
combination, the most comfortable state is achieved. The 
first problem that must be solved in the intelligent design 
of indoor temperature is to solve the PMV indicators. 
Generally speaking, the BPNN is used to calculate index 
values [17-18]. BPNN is a supervised learning algorithm 
that simulates the neural network structure of the human 
brain for pattern recognition and prediction tasks. The 
model consists of an input layer, a Hidden Layer (HL), 
and an Output Layer (OL), each of which is composed of 
multiple neurons. These neurons are connected through 
weighted connections and undergo nonlinear 
transformations through activation functions to handle 
complex nonlinear relationships. The core of BPNN lies 
in its learning mechanism, which adjusts the weights and 
thresholds in the network through back propagation 
algorithms. During the learning, the network first receives 
input data and calculates the output result through 
forward propagation. Subsequently, the network 
calculates the error between the predicted output and the 
actual target value, and transmits the error signal back to 
the network through back propagation to adjust weights 
and thresholds, thereby minimizing the prediction error. 
The calculation is displayed in formula (5). 
 
1
, 1,2,...
n
k
j ij i j
i
s w a j p 
=
= − =

   (5) 
In formula (5), 
ij
w
 represents the connection weights 
within ( 1,1) − . 
j

 represents the threshold. 
k
i
a 
represents the input sample. 
j
s
 represents the input of 
each unit. 
j
s
 is applied to calculate the output of each 
unit in the middle layer, as shown in formula (6). 
 
( ), 1,2...
jj
b f s j p ==
   (6) 
In formula (6), () f  represents the transfer function. 
j
b
 
represents the output of each unit in the middle layer. The 
output of each unit in the OL can be calculated through 
the output of the middle layer and the connection weights, 
as shown in formula (7). 
 
1
, 1,2,...,
p
t jt j t
j
L v b t q 
=
= − =

   (7) 
In formula (7), 
t
L represents the output of each unit in 
the OL. 
jt
v
 represents the connection weight. 
t
 
represents the threshold. The output calculation of each 
unit in the OL is used to calculate the response of each 
unit in the OL, as shown in formula (8). 
 ( ), 1 ,2,...,
tt
C f L t q ==   (8) 
In formula (8), 
t
C represents the response of each unit 
in the OL. Based on the real output and the objective 
vector, the generalization error of each unit in the OL can 
be calculated, as shown in formula (9). 
( ) (1 ), 1,2,...,
kk
t t t t t
d y C C C t q = −   − = (9) 
In formula (9), 
k
t
d represents the generalized error of 

Intelligent Environment Design for Indoor Spaces Based on…         Informatica 48 (2024) 191–204   195 
each unit in the OL. 
k
t
y represents the objective vector. 
Based on the connection weights, the generalization error 
of the OL, the output of the middle layer, and the 
generalization error in the middle layer can be calculated, 
as displayed in formula (10). 
 
1
(1 )
q
kk
j t jt j j
t
e d v b b
=

=  −



 (10) 
In formula (10), 
k
j
e represents the generalized error in 
the middle layer. Formulas (5)-(10) are forward 
propagation process, followed by back propagation 
process. By adjusting the generalization error and each 
unit's output in the OL, as well as the generalization error 
and each unit's output in the middle layer, the connection 
weights and thresholds are corrected. Relying on the 
thermal comfort index, the prediction model is shown in 
Figure 3. 
 
Room temperature
PMV value
Air velocity
Relative humidity
Mean radiation temperature
Ｍetabolism
Clothing thermal 
resistance
 
Figure 3: BPNN structure for predicting PMV 
 
Figure 3 displays the BPNN structure for predicting PMV 
indicators. The input of the model is the six factors of 
operating thermal comfort index. The output is the PMV 
indicators. The network structure contains a neuron, with 
a BPNN input layer of 6 dimensions and an OL of 1 
dimension. 
 
 
 
 
3.2 Optimization of BPNN model combined 
with GA 
There are defects in BPNN, such as long training time, 
large errors, and susceptibility to local minima. Therefore, 
GA is used to optimize it. GA is widely used in model 
optimization, because GA has better global search ability 
in discontinuous spaces [19-20]. GA follows the survival 
of the fittest. Genetic operators perform selection, 
crossover, and mutation to create a new group of 
solutions, ultimately completing optimization. The 
process of GA is shown in Figure 4. 
 

196 Informatica 48 (2024) 191–204                                                                      X. Xue 
l ＝1
Initiate
The initial population is 
randomly selected
Chromosome 
coding
Mutation
Intersect
Select
Finish
Calculate the fitness of the individual
　　Termination condition 
discrimination
l＝l ＋1
Ｙ
Ｎ
 
Figure 4: GA basic flow chart 
 
Figure 4 shows the basic process of GA. The initial 
population of GA is generated through encoding, which is 
a prerequisite for GA solving and affects the cross 
mutation. GA first initializes the population, randomly 
generates string structured data, and uses it as the initial 
point for iteration. At the same time, the maximum 
evolutionary iteration is set and a random method is 
applied to generate the initial population. Then, the 
individual fitness function is calculated, which affects the 
optimization process and convergence speed. Afterwards, 
individuals with higher fitness are selected to generate the 
new generation of new populations. According to a 
probability of 0.5-1.0, two individuals are selected to 
cross, that is, to exchange the corresponding gene 
combinations on the individuals. After crossing, some 
individual positions are changed with a small probability. 
Finally, when the fitness stabilizes and achieves the 
global optimum, or after achieving the specified number 
of iterations, the operation stops. Based on the input data 
from BPNN, the GA-BPNN prediction model for PMV 
indicators is established, as displayed in Figure 5. 
 

Intelligent Environment Design for Indoor Spaces Based on…         Informatica 48 (2024) 191–204   197 
Initial BPNN weight, threshold
GA encodes the initial value
Initiate
Determine the network topology
Calculation error
Get the optimal weight, threshold
The weight and threshold are updated
Satisfy constraint condition
Simulation prediction, get the result
Finish
BPNN training error as fitness value
Select
Calculate the fitness of the individual
Intersect
Mutation
　　Termination condition 
discrimination
Input data
Data preprocessing
Ｙ
Ｎ
Ｙ
Ｎ
  
Figure 5: GA-BPNN flowchart 
 
Figure 5 is a schematic diagram of optimized BPNN 
model based on GA. The entire process starts from the 
input data and ensures that the data format is suitable for 
model training through data preprocessing, such as 
normalization. Next, the network topology of BPNN is 
determined, including the number of layers and the 
number of neurons in each layer. The added GA encodes 
the initial weights and thresholds of BPNN to form the 
initial population. These encoded individuals evaluate 
fitness by calculating BPNN training errors, with smaller 
errors indicating higher fitness. Based on the fitness, GA 
performs selection operations to select the best 
individuals for crossover and mutation, in order to 
generate new offspring populations, enhance the model's 
generalization ability, and avoid falling into local optima. 
During the iteration process, newly generated individuals 
are evaluated based on their fitness, and the weights and 
thresholds of BPNN are updated accordingly. This 
process is repeated until the preset termination conditions 
are met, such as reaching the maximum number of 
iterations or fitness reaching the predetermined threshold. 
In addition, the process also includes checking the 
constraints of the solution to ensure that the found 
solution not only has high fitness but also satisfies all the 
constraints of the problem. Finally, the optimized BPNN 
model is used for simulation prediction, resulting in more 
accurate prediction results. The relationship between the 
length of the encoding string and BPNN is expressed as 
formula (11). 
 S m h h n h n =  +  + +   (11) 
In formula (11), mh  represents the encoding length of 
the connection weight in the input layer and HL. hn  
represents the encoding length of the weight between the 
HL and the OL. h represents the encoding length of the 
HL threshold. 
n
 represents the encoding length of the 
OL threshold. According to the BPNN model, the input 
layer has 6 parameters, and the HL is 10. The OL has a 
PMV value. Therefore, the encoding length is calculated, 
and 6 10 10 1 10 1 81 S =  +  + + = . The initial population 
consists of N randomly generated numerical particles. 

198 Informatica 48 (2024) 191–204                                                                      X. Xue 
The population size refers to the number of individuals 
included. Fitness measures the individual superiority or 
inferiority in GA. Individual fitness values have an 
impact on the probability they leave behind. Individuals 
with higher fitness are closer to the optimal solution. 
BPNN is actually the process of finding the optimal 
solution. Nonlinear transformation generates output 
values. The actual and output value errors are transmitted 
in reverse. The weights and thresholds are modified to 
ultimately output the weights and thresholds that achieve 
the minimum error. The mean square error of BPNN is 
displayed in formula (12). 
 
 
2
11
1
( ) ( )
2
q m
oo
ko
E d k y k
m
==
=−

  (12) 
Based on the above analysis, the reciprocal mean square 
error of BPNN can serve as the fitness function of GA. It 
can distinguish individual strengths and weaknesses. The 
fitness function of GA is expressed as formula (13). 
 
2
11
11
1
( ) ( )
2
q m
oo
ko
fitness
E
d k y k
m
==
==
−

 (13) 
In formula (13), fitness represents the fitness function. 
A higher fitness function value indicates good individual 
fitness and superior algorithm performance, while the 
opposite indicates poorer performance. The main step of 
crossover in GA is to randomly pair particles with high 
fitness in the population. Genes are crossed to form new 
individuals. Crossover operation is shown in formula 
(14). 
 
(1 )
(1 )
kj kj lj
lj lj kj
a a b a b
a a b a b
= − +
= − +
      (14) 
In formula (14), 
k
a and 
i
a represent different 
chromosomes that undergo crossover operations at 
position 
j
. b represents a random number between 0 
and 10. Mutation operation is carried out after crossover 
to alter individual genes and prevent the optimization 
process from converging in the immature stage. The 
mutation form is shown in formula (15). 
max
min
( ) ( ) 0.5
( ) ( ) 0.5
ij ij
ij
ij ij
a a a f g r
a
a a a f g r
+ −   

=

+ −  


 (15) 
In formula (15), 
ij
a
 represents the 
j
-th gene of the 
i -th individual. 
min
a represents the minimum lower 
bound of 
ij
a
. 
max
a represents the maximum upper 
bound value of 
ij
a
.
2 max
( ) (1 / ) f g r g G =− . 
2
r 
represents a randomly generated number. 
max
G 
represents the maximum iteration times. 
g
 represents 
the population iteration. r refers to a random number 
between 0 and 1. 
4 Analysis of GA-BPNN model based 
on perception and behavior 
correlation 
The study analyzes the proposed model in two parts. 
Firstly, the BPNN model is compared to verify its 
effectiveness. Then it is compared with other models to 
verify its superiority. 
4.1 Optimization Analysis of GA-BPNN 
Model 
To test the proposed model, the topology parameters of 
the BPNN remain unchanged. The maximum iteration 
time for GA is 200, and the population size is 50. The 
fitness function takes the derivative of the prediction 
error of the BPNN, with a crossover probability of 0.4 
and a mutation probability of 0.2. The study selects a 
thermal comfort index database from a certain university, 
with a data sample of 2000 groups, which is separated 
into training and testing samples according to a 9:1 ratio. 
Firstly, all data is standardized and normalized. The 
laboratory environment settings are shown in Table 2. 
 
 
Table 2: Laboratory environment setting 
Hardware and software configuration Version model 
CPU Intel(R)Core i7-7700@3.6GHz 
Operating system Ubuntu 18.04 LTS 
CUDA 9.1 
Deep learning frameworks Pytorch1.10 
Python version 3.7 
 
 
 
 
 
 

Intelligent Environment Design for Indoor Spaces Based on…         Informatica 48 (2024) 191–204   199 
Table 2 displays the experimental settings in the 
laboratory. Experiments are conducted on BPNN and 
GA-BPNN. The fitness curve is shown in Figure 6. 
 
Network error
0 20 40 60 80 100 120 140 160 180 200
(a) BPNN training convergence process
10
-4
10
-3
10
-2
10
-1
10
0
10
1
10
2
10
3
10
4
0 20 40 60 80 100 120 140 160 180 200
(b) GA-BPNN training convergence process
10
-4
10
-3
10
-2
10
-1
10
0
10
1
10
2
10
3
10
4
Network error
Training times Training times
 
Figure 6: GA-BPNN and BPNN training convergence process diagram 
 
Figure 6 (a) presents the convergence process of BPNN 
training. Figure 6 (b) presents the convergence process of 
GA-BPNN training. From the Figure, BPNN and 
GA-BPNN gradually converged with increasing training 
times. For the BPNN, it quickly converged in the first 60 
training times, then converged around 100 training times, 
and the network error stabilized at 10-3. For the 
GA-BPNN, the convergence speed was fast in the first 40 
training times, converged around 60 training times, and 
the network error stabilized at 10-3. The BPNN model 
improved by GA has a fast convergence speed. Two 
models are further experimented, as shown in Figure 7. 
 
Output PMV index values
0 20 40 60 80 100 120 140 160 180 200
(a) BPNN comparison of true value and predicted value
-2.5
-2
-1.5
-0.5
-1
0.5
1
1.5
2
True value Predicted value
0 20 40 60 80 100 120 140 160 180 200
(b) GA-BPNN comparison of true value and predicted value
True value Predicted value
-2.5
-2
-1.5
-0.5
-1
0.5
1
1.5
2
Output PMV index values
Sample size
Sample size
 
Figure 7: Comparison of actual and predicted values between GA-BPNN and BPNN 

200 Informatica 48 (2024) 191–204                                                                      X. Xue 
 
Figure 7 (a) displays the actual and predicted values of 
BPNN, which had a significant deviation, with poor 
fitting effect. Figure 7 (b) displays the comparison results 
of GA-BPNN. The difference of PMV was small, and the 
model fitting effect was better. The error curves of the 
two models are shown in Figure 8. 
 
0 20 40 60 80 100 120 140 160 180 200
(a) BPNN prediction error graph
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Prediction error
0 20 40 60 80 100 120 140 160 180 200
(b) GA-BPNN prediction error graph
Prediction error
0
0.005
0.01
0.015
0.02
0.025
Sample size
Sample size
 
Figure 8: GA-BPNN and BPNN error curves 
 
Figure 8 (a) shows the BPNN error curve. From the 
Figure, the BPNN model had a large error in PMV 
prediction, with the error basically remaining around 0.1 
and the maximum error reaching 0.16. Figure 8 (b) shows 
the GA-BPNN error curve. Compared with the BPNN 
model, the GA-BPNN model generally maintained lower 
error, with an error of around 0.01 and a maximum error 
of only 0.022. The PMV had a large number of input 
samples. The BPNN model made the training objectives 
more complex, with longer runtime and lower efficiency. 
There is over fitting during the training process, resulting 
in poor model prediction performance. GA improves the 
BPNN model by optimizing weights and thresholds, 
reducing unnecessary training, and achieving good fitting 
results, which is an effective improvement. 
4.2 Performance analysis of GA-BPNN 
model 
The proposed method is compared with other intelligent 
environment design methods that can be used to combine 
PMV indicators in experiments. The selected comparison 
models include Support Vector Machine (SVM), 
Convolutional Neural Network (CNN), Recurrent Neural 
Network (RNN), and GA-Recurrent Neural Network 
(GA-RNN). The SVM is a supervised learning strategy 
commonly used for data classification and regression 

Intelligent Environment Design for Indoor Spaces Based on…         Informatica 48 (2024) 191–204   201 
analysis. SVM classifies data by mapping it to a 
high-dimensional space and finding a hyperplane that 
maximizes the interval. The core of the CNN is a network 
structure that includes three layers. It extracts features 
from input data through convolutional operations, reduces 
the feature map dimensionality through pooling 
operations, and finally performs classification or 
regression tasks through fully connected layers. The RNN 
has a recurrent neural network structure, which can 
handle serialized data and preserve previous information. 
The GA-RNN combines GA and RNN. The GA is 
adopted to upgrade the parameters of RNN, which can 
find the optimal RNN structure and hyper-parameters to 
optimize the performance. The study selects Accuracy, 
Precision, F1-score, RMSE, and R-squared as model 
validation indicators. Accuracy is the most intuitive 
performance indicator, which represents the proportion of 
correctly predicted samples in the model to the total 
number of samples. Precision measures the proportion of 
positive classes predicted by the model, that is, how many 
of the predicted positive categories are accurate. The 
F1-score is the harmonic mean of precision and recall 
(accuracy), which strikes a balance between the two. It is 
a comprehensive indicator that considers both precision 
and recall. A high F1-score indicates higher precision and 
recall of the model. RMSE is a commonly used indicator 
to measure the error between model predictions and 
observations. R-squared is the determination coefficient 
that represents the degree to which the model explains the 
variability of observations. The proposed method is 
compared with comparison models in the laboratory. 
Figure 9 displays the results. 
 
SVM
Sample size
RMSE
GP-BPNN
0 20 40 60 80 100 120 140 160 180 200
CNN
RNN
GP-RNN 
10
11
12
13
14
15
9
8
ACC 
90
92
94
96
98
100
88
0 20 40 60 80 100 120 140 160 180 200
GP-BPNN
RNN
CNN
SVM
GP-BPNN
(a)RMSE comparison of multiple algorithms
(b) Accuracy comparison of multiple algorithms
Sample size
 
Figure 9: Comparison of RMSE and Accuracy of various models 
 
Figure 9 (a) displays the RMSE results of multiple 
algorithms. From the Figure, the proposed model had the 
lowest RMSE, while the SVM had the highest, indicating 
that the difference was small, and the predictive ability 
was good. The SVM model had a significant difference, 
with poor predictive ability. Figure 9 (b) shows the 
Accuracy comparison results of multiple algorithms. The 
proposed model had high Accuracy. In addition, the 

202 Informatica 48 (2024) 191–204                                                                      X. Xue 
GA-RNN model also had high Accuracy. Table 3 
presents the results of other indicators. 
 
 
Table 3: Comparison results of multiple models 
Model Accuracy  Precision  F1-score RMSE R-square 
SVM 88.23% 89.71% 0.714 14.29 0.73 
CNN 91.94% 89.97% 0.753 12.54 0.81 
RNN 93.67% 91.02% 0.787 11.60 0.76 
GA-RNN 95.62% 92.36% 0.815 10.73 0.93 
GA-BPNN 97.89% 96.15% 0.896 10.03 0.92 
 
Table 3 compares other indicators of multiple models. 
From the data in the table, the proposed model achieved 
higher Accuracy, Precision, and F1-score, with values of 
97.89%, 96.15%, and 0.896. Compared with the 
GA-BPNN, the GA-RNN presented a decrease in 
Accuracy, Precision, and F1-score. In terms of RMSE, 
the GA-RNN and GA-BPNN had lower values of 10.73  
 
 
and 10.03, respectively. For the R-square, the GA-RNN 
model reached the highest, at 0.93, while the GA-BPNN 
decreased slightly but also maintained a high level. This 
indicates that the GA-BPNN and GA-RNN can 
effectively explain the changes in observed data, thereby 
accurately predicting comfort feelings under different 
environmental conditions. The comparison of network 
fitness is shown in Figure 10. 
 
SVM
Number of iterations
Mean fitness
GP-BPNN
0 20 40 60 80 100 120 140 160 180 200
CNN
RNN
GP-RNN 
0.2
0.3
0.4
0.5
0.6
0.7
0.1
0
0.8
0.9
1
 
Figure 10: Multiple model network fitness curves 
 
Figure 10 shows the fitness curves. The GA-BPNN 
model exhibited significant changes in 0-20 iterations, 
gradually stabilizing at 40 iterations and finally 
stabilizing at 0.26. The GA-RNN model showed 
significant changes in 0-60 iterations. Thereafter, its 
fitness remained stable at 0.24. The fitness of RNN, CNN, 
and SVM was relatively low, which stabilized at 0.12, 
0.13, and 0.17. The GA-BPNN model has good 
generalization ability and better predictive performance. 
The performance of RNN, CNN, and SVM is not ideal 
enough to achieve prediction results. 
5 Conclusion 
A GA-BPNN prediction model based on perception and 
behavior correlation was constructed to achieve 
intelligent environment design in indoor spaces. PMV is a 
crucial component in indoor intelligent environments. 
The factors affecting PMV were analyzed. Then the 
BPNN model was introduced to predict PMV indicators. 
However, the BPNN model has drawbacks such as long 
training time, large errors, and the tendency to fall into 
local minima. Therefore, the GA was introduced to 
improve it, and the GA-BPNN model was constructed. 
The effectiveness of the improvement was verified. Then 
the superiority of the model through comparative 
experiments was proved. The experimental results 
showed that the GA-BPNN converged faster than the 
BPNN. The difference of PMV was small, with better 
model fitting effects. The BPNN had a significant error in 
PMV prediction, with an error of around 0.1 and a 
maximum error of 0.16. The overall error of the 
GA-BPNN remained at a relatively low level, with an 
overall error of around 0.01 and a maximum error of only 
0.022. Compared with other models, the GA-BPNN had 
higher Accuracy, Precision, and F1-score, with values of 

Intelligent Environment Design for Indoor Spaces Based on…         Informatica 48 (2024) 191–204   203 
97.89%, 96.15%, and 0.896, respectively. The GA-BPNN 
showed significant changes in fitness during the first 20 
iterations, with the fitness value ultimately stabilizing at 
0.26, indicating good generalization ability and predictive 
performance. The proposed model achieves intelligent 
design of indoor comfort by controlling temperature. 
However, the factors that affect PMV indicators are not 
limited to temperature. In subsequent research, the 
optimal combination of environmental parameters can be 
adjusted to improve indoor intelligence. 
6 Discussion 
The GA-BPNN model based on perception and behavior 
correlation proposed in the study has demonstrated 
advantages in the intelligent indoor environment design. 
By comparing with relevant works in existing literature, 
the contributions and advantages of this study can be 
more clearly identified. Compared with the car following 
model proposed by Wang W J et al. in intelligent 
transportation systems, the designed model focuses on 
predicting the thermal comfort of indoor environments. 
Although both adopt simulation and optimization 
strategies, the research method combines the global 
search ability of GA and the learning ability of BPNN, 
particularly targeting the optimization of PMV indicators, 
which is an innovative application in indoor environment 
design. Compared with the behavior prediction model 
proposed by Cheng et al. in intelligent butler design, the 
proposed model not only predicts user behavior, but also 
further predicts the impact of these behaviors on indoor 
environmental comfort. The BPNN optimized by GA 
shows faster convergence speed and higher accuracy in 
predicting PMV indicators, which is clearly reflected in 
the comparison results of Figure 6 and Figure 7. In 
addition, the designed model outperforms the intelligent 
reflective/refractive surface assisted high mobility 
communication system proposed by Huang Z et al. in 
terms of accuracy, precision, and F1-score. In Table 2, 
the GA-BPNN model has higher reliability and 
effectiveness in predicting indoor environmental comfort, 
with scores of 97.89%, 96.15%, and 0.896 on these 
indicators, respectively. Compared with the BPNN-based 
multi-body dissipative particle dynamics calibration 
method proposed by Yu X et al., the GA-BPNN model 
exhibits stronger robustness and adaptability in dealing 
with nonlinear problems and multivariate optimization. 
Through GA optimization, the model not only avoids 
getting stuck in local minima, but also improves the 
accuracy and efficiency of indoor environment control. 
Overall, the GA-BPNN model proposed in the study 
provides a new solution for intelligent design of indoor 
environments. This model not only improves the 
accuracy of PMV indicator prediction, but also provides 
strong technical support for achieving automation and 
intelligence of indoor comfort through intelligent 
temperature control. 
 
Fundings
 
The research is supported by Fujian Provincial Education 
Science "14th Five-Year Plan" 2021 Annual Project: 
Exploration and Practice of Practice Based Talent 
Cultivation Mode for Environmental Design Majors in 
Universities under the Background of Rural 
Revitalization Strategy (Fund Number: FJJKBK21-167). 
References
 
[1] X. M. Long, Y. J. Chen, and J. Zhou, “Development 
of AR experiment on electric-Thermal effect by 
open framework with Simulation-Based asset and 
User-Defined input,” Artificial Intelligence and 
Applications, vol. 1, no. 1, pp. 52-57, 2023. 
https://doi.org/10.47852/bonviewAIA2202359 
[2] J. He, K. Yu, Y. Shi, Y. Zhou, W. Chen, and K. B. 
Letaief, “Reconfigurable intelligent surface assisted 
massive MIMO with antenna selection,” IEEE 
Transactions on Wireless Communications,  vol. 
21, no. 7, pp. 4769-4783, 2021. 
https://doi.org/10.1109/TWC.2021.3133272 
[3] M. Cucchi, D. Parker, E. Stavrinidou, P. Gkoupidenis, 
and H. Kleemann, “In Liquido computation with 
electrochemical transistors and mixed conductors 
for intelligent bioelectronics,” Advanced Materials, 
vol. 35, no. 15, pp. 2209516-2209530, 2023. 
[4] Z. Feng, B. Clerckx, and Y. Zhao, “Waveform and 
beamforming design for intelligent reflecting 
surface aided wireless power transfer: Single-user 
and multi-user solutions,” IEEE Transactions on 
Wireless Communications, vol. 21, no. 7, pp. 
5346-5361, 2022. 
https://doi.org/10.1109/TWC.2021.3139440 
[5] L. Zhang, X. Yang, H. Liu, H. Zhang, and J. Cheng, 
“Efficient residual shrinkage CNN denoiser design 
for intelligent signal processing: Modulation 
recognition, detection, and decoding,” IEEE Journal 
on Selected Areas in Communications, vol. 40, no. 1, 
pp. 97-111, 2021. 
https://doi.org/10.1109/JSAC.2021.3126074 
[6] X. Lu, W. Liao, Y. Zhang, and Y. Huang, “Intelligent 
structural design of shear wall residence using 
physics‐enhanced generative adversarial networks,” 
Earthquake Engineering & Structural Dynamics, vol. 
51, no. 7, pp. 1657-1676, 2022. 
https://doi.org/10.1002/eqe.3632 
 [7] W. J. Wang, M. H. Ma, S. D. Liang, G. Y.  Ma, and 
Y. S. Wang, “Density waves in an improved 
car-following model under intelligent transportation 
system environment,” Modern Physics Letters B, 
vol. 36, no. 15, pp. 2250014-2250032, 2022. 
https://doi.org/10.1142/S0217984922500142 
[8] S. T. Cheng, C. Hsu, G. J. Horng, and K. T. Tsai, 
“Quadro-W learning for human behavior prediction 
in an evolving environment: a case study of the 
intelligent butler technology,” The Journal of 
Supercomputing, vol. 79, no. 6, pp. 6309-6346, 

204 Informatica 48 (2024) 191–204                                                                      X. Xue 
2023. https://doi.org/10.1007/s11227-022-04899-1 
[9] Z. Huang, B. Zheng, and R. Zhang, “Transforming 
fading channel from fast to slow: Intelligent 
refracting surface aided high-mobility 
communication,” IEEE Transactions on Wireless 
Communications, vol. 21, no. 7, pp. 4989-5003, 
2021. https://doi.org/10.1109/TWC.2021.3135685 
[10] R. Zhong, X. Liu, Y. Liu, Y. Chen, and X. Wang, 
“Path design and resource management for NOMA 
enhanced indoor intelligent robots,” IEEE 
Transactions on Wireless Communications, vol. 10, 
no. 21, pp. 8007-8021, 2022. 
https://doi.org/10.1109/TWC.2022.3163422 
[11] X. Yu, J. Zhao, S. Chen, D. Huang, K. Zhang, and D. 
Cao, “The calibration for many-body dissipative 
particle dynamics by using back-propagation neural 
networks,” Molecular Simulation, vol. 48, no. 11, pp. 
955-964, 2022. 
https://doi.org/10.1080/08927022.2022.2055755 
 [12] Z. Sui, J. Mu, T. Wang, and S. Zhang, “Evaluation 
of energy saving of residential buildings in North 
China using back-propagation neural network and 
virtual reality modeling,” Journal of Energy 
Engineering, vol. 148, no. 3, pp. 4022013-4022025, 
2022. 
https://doi.org/10.1061/(ASCE)EY.1943-7897.0000
832 
[13] L. Zhang, K. Jiao, L. Zhang, J. Zhang, M. Sun, and 
Z. Zhou, and A. Zheng, “Prediction of blast furnace 
fuel ratio based on back‐propagation neural network 
and K‐Nearest neighbor algorithm,” Steel Research 
International, vol. 93, no. 10, pp. 2200215-2200226, 
2022. https://doi.org/10.1002/srin.202200215 
[14] Y. Xu, Z. Liu, Q. Zhang, F. Guan, Z. Yan, B. Huang, 
A. Pan, J. Hu, Z. Chen, Q. Ding, X. Qiu, and Y. 
Tang, “A geomagnetic vector compensation method 
compatible with nonlinear interferences based on 
back propagation network and 3D Helmholtz coil,” 
Review of Scientific Instruments, vol. 94, no. 3, pp. 
35111-35119, 2023. 
https://doi.org/10.1063/5.0138606 
[15] J. Zolgharnein, T. Shariatmanesh, S. D. Farahani, 
“Artificial neural network (ANN) modeling for 
simultaneous removal of a binary mixture of Pb (II) 
and Cu (II) by cobalt hydroxide nano‐flakes,” 
Journal of Chemometrics, vol. 37, no. 4, pp. 
3475-3485, 2023. https://doi.org/10.1002/cem.3475 
[16] X. Yu, J. Zhao, S. Chen, D. Huang, K. Zhang, and D. 
Cao, “The calibration for many-body dissipative 
particle dynamics by using back-propagation neural 
networks,” Molecular Simulation, vol. 48, no. 11, pp. 
955-964, 2022. 
https://doi.org/10.1080/08927022.2022.2055755 
 [17] G. Veselov, A. Tselykh, A. Sharma, and R. Huang, 
“Applications of artificial intelligence in evolution 
of smart cities and societies,” Informatica, 45(5), pp. 
603, 2021. https://doi.org/10.31449/inf.v45i5.3600 
[18] F. Zeng, R. Wan, Y. Xiao, F. Song, C. Peng, and H. 
Liu, “Predicting the Self-Diffusion coefficient of 
liquids based on backpropagation artificial neural 
network: A quantitative Structure–Property 
relationship study,” Industrial & Engineering 
Chemistry Research, vol. 61, no. 48, pp. 
17697-17706, 2022. 
https://doi.org/10.1021/acs.iecr.2c03342 
[19] J. Ding, R. Alroobaea, A. M. Baqasah, A. Althobaiti, 
and R. Miglani, “Big data intelligent collection and 
network failure analysis based on artificial 
intelligence,” Informatica, vol. 46, no. 3, pp. 
253-392, 2022. https://doi.org/10.31449/inf. 
v46i3.3866 
[20] L. Yang, K. Wang, and H. Sun, “A grating lobe 
suppression method for displaced subarrays using 
genetic algorithm,” IET Microwaves, Antennas & 
Propagation, vol. 16, no. 7, pp. 457-464, 2022. 
https://doi.org/10.1049/mia2.12258