https://doi.org/10.31449/inf.v48i13.6159 Informatica 48 (2024) 31–50 31 
Passenger Flow Prediction of Tourist Attractions by Integrating 
Differential Evolution and GWO 
 
Jie Yu 
Modern Service College, Zhengzhou Tourism College, Zhengzhou 451464, China 
E-mail: gracefish_83@163.com 
Keywords: time series, GWO algorithm, LSTM network, differential evolution algorithm, traffic prediction, travel 
Received: May 7, 2024 
To further improve the experience of visiting tourist attractions and promote their long-term healthy 
development, this study analyzes the short-term passenger flow prediction of tourist attractions. The 
traditional long and short-term memory network is selected as the basis of the prediction framework. 
Then the grey wolf optimization algorithm is used to optimize the hyper-parameters of the long and 
short-term memory network, and the differential evolution algorithm is used to improve the 
shortcomings of easily falling into the local optimum. Finally, a passenger flow prediction model is 
constructed based on intelligent optimization and deep learning. The experiment outcomes denote that 
the differential evolution improvement strategy designed in the study is beneficial for improving the 
global optimization of the grey wolf evolutionary algorithm. The average optimization values of 
different test functions are closest to the global minimum, effectively improving the population fitness. 
In the parameter optimization, the maximum value of hypervolume can reach 0.91. The minimum value 
of the inverse generation distance converges to 0.09, and the quality of the Pareto front solution is 
relatively high. The Spacing and Spread values are both above 0.8, indicating better diversity in the 
solution set. The improved prediction model has the lowest values in terms of mean absolute 
percentage error, root mean square error, and mean absolute error, and the minimum value of the error 
is only 0.0693. The maximum R2 value can reach 0.945, indicating good prediction accuracy and 
goodness of fit. The prediction results of this prediction model have high accuracy in predicting 
passenger flow at different time periods. The accuracy and F1 values are close to 0.95 and the 
precision and recall are higher than 0.90 on different datasets. This study enriches the theoretical 
basis for optimizing and improving traditional time series models, improves the accuracy of predicting 
tourist flow in tourist attractions, and helps promote the healthy development of the tourism industry. 
Povzetek: Predstavljen je sistem za napovedovanje potniškega toka turističnih znamenitosti, ki 
združuje diferencialno evolucijo in optimizacijski algoritem sivega volka (GWO) z LSTM mrežo. Model 
dosega visoko točnost napovedi s povprečnim R² do 0,945, kar prispeva k izboljšanju načrtovanja 
turističnih obiskov in razvoju turizma.
1 Introduction 
With the transformation and upgrading of the national 
industrial structure, the tourism industry takes a pivotal 
part in economic development. The tourism industry has 
emerged as a significant contributor to regional economic 
development, bringing in a large amount of economic 
income while also driving the development of other 
related industries [1]. But with the booming development 
of the tourism, the passenger flow of tourist attractions 
(TAs) continues to increase. Moreover, travelers are 
mostly concentrated at holiday nodes, leading to an 
imbalance in the allocation of tourism resources in terms 
of time and space. These types of problems seriously 
interfere with the safety management of scenic spots, 
affect the travel experience of tourists, and bring negative 
impacts to the development of TAs [2-3]. Thus, it is 
significant to forecast the passenger flow data of TAs  
 
through effective methods, help scenic spots to make 
reasonable planning and adjustments in resource 
allocation, personnel allocation, service quality, etc., 
formulate reasonable flow limiting measures to avoid 
overcrowding and resource waste, and promote the 
long-term sustainable development of the tourism [4]. At 
present, research on predicting the passenger flow of TAs 
mostly comes from the analysis of annual and monthly 
tourism data, and the prediction results have a strong lag, 
making it difficult to provide effective and reasonable 
guidance for scenic area management and tourist travel 
[5-6]. Today, in the context of "Internet plus", advanced 
computer, big data and other technologies and 
information systems have greatly promoted the upgrading 
and transformation of intelligence and informatization of 
tourism. However, the management of smart tourism 
services is still in the exploratory stage, and the 
construction of technology and infrastructure still needs 

32   Informatica 48 (2024) 31–50                                                                        J. Yu 
to be further strengthened [7-8]. In view of this, to further 
enhance the intelligence and scientificity of tourism 
industry management, deep learning algorithms are 
introduced into the problem of tourism passenger flow 
prediction (PFP), and a corresponding framework is built 
grounded on long and short-term memory networks 
(LSTM). Moreover, the grey wolf optimization (GWO) 
algorithm is introduced to optimize the hyperparameters 
of LSTM. The introduction of differential evolution (DE) 
algorithm during the optimization improves the 
shortcomings of GWO. 
The innovation of the research is mainly reflected in 
three aspects. Firstly, it further improves the level of 
intelligent management in the tourism industry. Secondly, 
it enriches the theoretical research level of time series 
prediction models and introduces intelligent optimization 
algorithms (IOAs) to achieve parameter optimization. 
Thirdly, it expands the application fields and potential of 
IOAs. The research is broken into four parts. Part 1 
illustrates the current research conditions of prediction 
problems and IOAs. Part 2 constructs a PFP model with 
LSTM. Part 3 conducts testing experiments on the 
effectiveness of the prediction model. The fourth part 
summarizes the main conclusions of the study and future 
work. This study is expected to strengthen the 
intelligence level of the entire tourism industry 
management and enrich the travel experience of tourists. 
2 Related works 
Predictive research is related to decision-making 
problems in many fields and is an important research 
direction in computer and artificial intelligence, receiving 
widespread attention from scholars. At present, the traffic 
prediction methods for TAs are mostly limited to raw 
traffic data and road networks. Gao et al. 
comprehensively considered the urban road network, 
scenic spot popularity, accessibility, and traffic volume, 
and used multiple historical data sources to learn the 
traffic dependence of multiple scenic spots through 
multi-graph convolutional networks and gated recurrent 
units for TAs traffic volume prediction. The experimental 
results indicated that the model can utilize integrated data 
for traffic flow prediction [9]. Holiday tourism 
requirement predication is a critical part of the tourism 
transportation system’s planning and management. Li et 
al. designed an improved spatiotemporal related LSTM 
model to predicate tourism demand by combining the 
spatiotemporal related history of passenger flow, weather, 
time, Internet search index and other data. Empirical 
analysis and verification of actual passenger flow data in 
suburban TAs in Beijing indicated that in contrast with 
other conventional prediction models, this model has 
better ability to capture spatiotemporal correlation of 
traffic flow and higher prediction accuracy [10]. The 
existing solutions for PFP are mostly based on regional 
prediction, with limited prediction capabilities. Sáenz et 
al. constructed a prediction framework based on graph 
neural networks, which integrates multi-source 
heterogeneous tourism data related to national mobility 
and infrastructure characteristics. The graph neural 
network was integrated into the graph model, and the 
experiment findings indicated that the model’s F1 value 
was higher than 0.7 [11]. Short-term PFP is the key to the 
operation and scheduling decisions of urban rail transit. 
To avoid potential threats caused by massive passenger 
flows, Xu conducted PFP research on the subway system 
and designed a multi-stage urban rail transit short-term 
PFP model that integrates convolutional neural networks, 
LSTM, support vector machines, and wavelet transform. 
The experimental outcomes indicated that the integrated 
model is superior to the baseline model in accuracy and 
efficiency, and the values of various error evaluation 
indicators are better than classical prediction techniques 
[12]. 
The spatial clustering of tourist quantity sequences 
limits the accuracy of tourist flow prediction models. Xue 
et al. designed a multimodal deep learning method for 
predicting hourly tourist flow in scenic spots based on 
spatial aggregation. This method effectively utilized 
search engine data to extract daily features of scenic spots, 
and utilized social media and tourist quantity data to 
explore spatial aggregation relationships. The empirical 
results showed that this method is superior to existing 
advanced baseline models, with a significance level of 
1%. Compared with the optimal baseline model, this 
method achieved a maximum reduction of 50.0% in error 
values [13]. To provide better personalized itinerary 
management for TAs and effectively avoid peak travel 
hours, Sun et al. designed a convolutional block attention 
module with deep learning for predicting tourism demand. 
This method first extracted a passenger flow grid graph 
from mobile signaling data, then constructed a 
convolutional block attention module using a 
multi-channel spatiotemporal grid graph constructed from 
multiple continuous passenger flow grid graphs, and 
finally predicted the demand for TAs. The findings of the 
experiment indicated that the mean absolute percentage 
error (MAPE) of this method is 8.11%, which is superior 
to that of other deep learning models [14]. 
IOA is a type of solving algorithm based on natural 
phenomena and biological inspiration, and different 
intelligent algorithms are broadly applied in different 
fields. Lu et al. conducted research on the optimization 
analysis of genetic algorithm, particle swarm 
optimization (PSO) algorithm, GWO algorithm, and 
sparrow search algorithm (SSA), in support vector 
regression. The optimized machine learning method was 
used to forecast the geological displacement generated by 
tunnel excavation. The fitness function analysis results 
showed that the GWO algorithm and SSA have high 
accuracy, efficiency, and stability [15]. Wang et al. 
designed a new multi-objective marine predator 
combination strategy for wind speed combination 
probability prediction. The experiment outcomes 
indicated that this method overcomes the deficiencies of 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 33 
traditional multi-objective optimization algorithms and 
effectively measured and minimized the uncertainty in 
the prediction process [16]. To predict the uniaxial 
compressive strength of the novel rubber sand concrete 
material, Mei et al. selected PSO, fruit fly optimization, 
lion swarm optimization, and SSA to enhance the 
backpropagation neural network. The verification results 
of six performance evaluation indicators showed that the 
lion swarm optimization algorithm had the best 
performance [17]. Abdullaev et al. optimized the 
customer churn prediction model based on bidirectional 
LSTM using the chicken swarm optimization algorithm. 
Experiment outcomes indicated that the optimized 
prediction model performs better than other prediction 
techniques [18]. To improve the problem of poor 
trajectory accuracy and susceptibility to local optima in 
inland ship trajectory prediction models, Zheng Y et al. 
improved the SSA using sine chaotic mapping and 
utilized it to strengthen the weight and threshold of the 
prediction model. Experiment outcomes indicated that 
this method improves the prediction model’s accuracy 
and stability [19]. 
In summary, both domestically and internationally, 
the PFP in TAs mainly relies on traditional time series 
models, machine learning algorithms, or deep learning 
algorithms, and the prediction problem has made good 
research progress. The main methodology, dataset, 
assessment indicators and key findings of the study are 
shown in Table 1. However, these classic time series 
prediction methods require a large amount of 
computation and a longer training time. There is still 
some room for improvement in the model's predictors. At 
present, IOAs have been widely used for optimizing 
various model parameters or solution spaces, the use of 
IOAs to improve the performance of prediction models 
has become a popular trend, and the problem of tourism 
traffic prediction needs to be solved. Moreover, research 
also attempts to introduce IOAs to raise the performance 
of traditional time series prediction models. 
 
 
Table 1: Summary of existing research reviews 
Literature Main methods Dataset 
Evaluation 
indicators 
Main conclusions 
[9] 
Graph convolutional 
network, gated 
recurrent unit 
Integrate geographic 
data and historical 
traffic data 
Accuracy 
The prediction 
accuracy is better 
than several 
classical and recent 
methods 
[10] 
An improved 
spatiotemporal 
correlation LSTM 
model 
Historical data of 
the tourist flows, 
and the auxiliary 
data including 
meteorological data, 
temporal data, and 
Internet search 
index 
Prediction accuracy 
Better ability to 
capture 
spatiotemporal 
correlation of flow 
with higher 
prediction accuracy 
[11] 
Graph neural 
network 
Heterogeneous 
tourism data 
F1 value F1 value above 0.7 
[12] 
Integrated 
convolutional neural 
network, LSTM, 
support vector 
machine and 
wavelet transform 
Historical data of 
urban subway 
system 
MAPE, RMSE, 
mean square error, 
R
2
, and accuracy 
The prediction 
model is better than 
the baseline model 
in accuracy and 
efficiency, and the 
evaluation index is 
better than the 
classical prediction 
technology 

34   Informatica 48 (2024) 31–50                                                                        J. Yu 
[13] 
Multimodal deep 
learning methods 
Real-time tourist 
quantity dataset for 
different scenic 
spots in Beijing 
MAPE, RMSE 
The Diebold 
Mariano test is 
superior to 
state-of-the-art 
baseline models at 
the 1% level, with 
significantly 
reduced error values 
[14] 
Deep learning 
model based on 
convolutional block 
attention module 
Beijing and Xiamen 
mobile signaling 
data 
MAPE 
The MAPE of this 
method is 8.11%, 
which is better than 
other deep learning 
models 
[15] 
Genetic algorithm, 
PSO, GWO and 
SSA optimize the 
machine learning 
prediction model 
Soil pressure 
balance shield 
tunnel excavation 
construction data 
Pearson correlation 
coefficient 
The GWO 
algorithm and SSA 
have high accuracy, 
efficiency, and 
stability 
[16] 
Combined 
probability 
prediction of wind 
speed based on 
multi-target marine 
predator 
combination 
strategy 
Two wind speed 
data sets 
Prediction accuracy 
Improve wind speed 
prediction accuracy, 
effectively measure 
and minimize 
forecast 
uncertainties 
[17] 
PSO, fruit fly 
optimization, lion 
swarm optimization, 
and SSA to enhance 
the backpropagation 
neural network 
Uniaxial 
compression test 
data from RSC 
laboratory 
RMSE, correlation 
coefficient, 
coefficient of 
determination, 
MAE, mean square 
error, and sum of 
squares of error 
Lion swarm 
optimization 
enhance the 
backpropagation 
neural network 
better than the other 
three mixed models 
[18] 
Bidirectional LSTM 
prediction model 
based on chicken 
flock optimization 
algorithm 
Customer 
information data set 
/ 
The performance is 
better than other 
prediction 
techniques 
[19] 
Sinusoidal chaotic 
mapping improved 
SSA to optimize the 
prediction model 
Data of ship 
automatic 
identification 
system in the 
Yangtze River 
Accuracy and 
stability 
This method 
improves the 
accuracy and 
stability of the 
prediction model 
 
3 A tourist flow prediction model for 
tourist attractions based on 
improved LSTM 
Predicting the passenger flow of TAs helps optimize 
tourism products and services, improve tourist 
satisfaction and experience. The study selects LSTM 
from deep learning algorithms to construct a prediction 
model framework, and introduces IOAs to optimize the 
model parameters. 
 
 
 
 
 
 
 
 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 35 
 
 
 
 
3.1 Distribution characteristics of passenger 
flow in tourist attractions and construction 
of LSTM prediction framework 
 
The PFP of TAs can be broken into medium and 
long-term PFP and short-term PFP. Medium and 
long-term passenger flow refers to the number of tourists 
over a relatively long period of time, usually evaluated 
based on the passenger flow of statistical years or 
quarters. It is influenced by factors such as economic 
conditions, scenic area facilities and service levels, and 
tourism policies. Short-term passenger flow refers to the 
passenger flow within a relatively short time range, 
usually daily, weekly, or monthly. It has many 
influencing factors and fluctuates greatly [20-21]. The 
study focuses on analyzing PFP, and the process of 
analyzing passenger flow characteristics and building a 
prediction framework is shown in Figure 1. 
 
Filtration 
Weather Season Holiday 
Passenger traffic data
Data Cleaning
Feature selection and 
feature dimensionality 
reduction
PCA
Data  normalization
Divide the training set 
and test set
Training the LSTM 
model
Fulfillment of expectations?
Adjustment 
parameters
Test  set
Trained 
LSTM
Test  model
Comparative 
results analysis
N
Y
 
Figure 1: Flow of passenger traffic characterization and predication framework construction 
 
The study summarizes the influencing factors and 
distribution characteristics of tourist flow in TAs by 
reviewing relevant materials, literature, and yearbook 
reports related to scenic area management and passenger 
flow analysis. According to research, the main factors 
affecting tourist flow in TAs include weather, season and 
holiday influences, public opinion, and word-of-mouth. 
Factors such as weather comfort, rainfall, and 
temperature can affect the travel comfort of tourists. 
Under the current vacation system in China, holidays 
usually have a significant influence on the passenger flow 
of TAs. In addition, the promotion and publicity activities, 
positive public opinion, and good reputation of the scenic 
area will also attract tourists to travel. Overall analysis 
shows that the distribution characteristics of passenger 
flow in TAs are influenced by various factors, showing 
significant fluctuations during peak and valley periods. 
On weekends and holidays, the increase in passenger 
flow is more pronounced. Therefore, the passenger flow 
data has non-linear characteristics and obvious 
periodicity, showing differences between centralized 
distribution and decentralized distribution. 
The influencing factors of passenger flow are 
collected, they are transformed into trainable predictive 
model features, and feature selection and dimensionality 
reduction are performed to reduce redundant information 
between data, reduce data dimensions, and improve the 
model's generalization ability and interpretability. The 
feature selection method adopted in the study is the 
filtering method, which filters features by setting a 
threshold. The method employed for the reduction of 
dimensionality is principal component analysis (PCA). 
PCA applies the covariance matrix of data to map the 
original features to a new feature space, preserving the 
original information and reducing data correlation. The 
expression of the covariance matrix is shown in equation 
(1). In equation (1), 
m
 represents the feature dimension 
of the sample. 
x
 represents matrix elements. 
1
1 m
T
ii
i
xx
m
=
=

  (1) 
The study uses recurrent neural networks (RNNs) in 
deep learning as the technical foundation for prediction 
models. Traditional artificial neural networks rely on 
nonlinear mapping relationships to achieve predictive 
outputs, lacking the ability to capture the relationship 
between data and time. RNN is suitable for processing 
sequential data, memorizing the information learned 
during training, and applying it to the learning and 
computation of the current task. The network structure of 
RNN is denoted in Figure 2 [22]. 
 

36   Informatica 48 (2024) 31–50                                                                        J. Yu 
1
2
3
4
5
6
7
8
9
  Input layer
   Hidden layers
  Output layer
1
x
2
x
3
x
41
w
2
y
1
y
  Input layer
Hidden layers
Output layer
Circular layer
W
Carry  out
W W W
U U U
W
V V V
( ) ot
( ) o t 1 − ( ) o t 1 +
( ) st ( ) s t 1 +
( ) s t 1 −
84
w
 
Figure 2: Schematic diagram of recurrent neural networks 
 
After inputting the sequence data into the RNN, the 
hidden state is iteratively calculated at each time step, and 
the output is predicted with the hidden state. The nodes 
between the hidden layers (HLs) are interconnected and 
jointly determine the output of the next moment. The 
operation process of RNN is shown in equation (2), 
wherein ( ) ht and ( ) ot represent the outputs of the 
HL and output layer (OL), respectively. t represents a 
time series. () st means the memory of the sample at 
time t . ( ) f and ( ) g represent nonlinear activation 
functions. ( ) g , W , V represent the weight matrices 
of input layer, HL, and OL, respectively. 
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
-- h t Ux t 1 Ws t 1
s t f h t
o t g Vs t

=+

=


=

 (2) 
The complexity of the structural composition of 
RNNs increases over time, and information is lost during 
transmission. RNNs face problems such as vanishing or 
exploding gradients. To avoid such phenomena, the 
improved structure LSTM of RNN is used in the study 
[23]. LSTM replaces the hidden neurons of RNN with 
structures with LSTM, including memory cells, input gate 
(IG), output gate (OG), and forget gate (FG). The 
different parts control the memory and forgetting of 
information through learning parameters, thereby 
determining the updating of cell states and the calculation 
of hidden states. Among them, the FG determines the 
information forgotten from the cell state, the IG 
determines the addition of input data to the cell state, and 
the OG calculates the hidden state based on the current 
input and cell state [24]. Compared to RNN, LSTM can 
capture and retain important information, better handle 
long-term dependencies and long-range correlations in 
sequences. The schematic diagram of LSTM structure is 
denoted in Figure 3. 
 
  
tanh
tanh
Concat 
t
X
t
f
t
i
t c
t
O
1 t
c
− 1 t
c
−
t
c
t
c
t
h
t
h
1 t
h
−
 
Figure 3: Schematic diagram of LSTM network 
 
 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 37 
The expression for the internal state 
t
S and HL 
state 
t
y of LSTM is shown in equation (3). In equation 
(3), 
t
f , 
t
i , and 
t
O correspondingly denote FG, IG, and 
OG. 
t S
 denotes the intermediate state of the current 
internal state. 
( )
1
    
tanh   
t
t t t t
t t t
S f S i S
y O S
−

=  + 


=


 (3) 
The expression for the FG is shown in equation (4). 
In equation (4), 
f
W
 means the weight of the FG. 
f
b
 
means forgetting gate bias. 

 represents the activation 
function. The FG can effectively filter out abnormal and 
redundant data. 
  ( )
1
,
t f t t f
f W y x b 
−
=+ (4) 
The IG determines the addition of input data to the 
cell state, and the calculation process is denoted in 
equation (5). In equation (5), 
i
W , 
i
b , and 
o
W , 
o
b 
correspond to the weight matrix and bias of the IG and 
OG. 
  ( )
  ( )
1
1
,
,
t i t t i
t o t t o
i W y x b
O W y x b


−
−
=+


=+


 (5) 
 
3.2 Design of improved LSTM prediction 
model based on DE and GWO 
Traditional LSTM networks involve the selection of 
many parameters, including the amount of HL nodes, 
initial learning rate, iteration times, etc. When the amount 
of HL nodes is small, the model cannot capture detailed 
patterns and relationships, which is prone to underfitting. 
When there are many HL nodes, it may lead to overfitting 
of the model. The size of the learning rate influences the 
speed and direction of network parameter updates, and 
affects the convergence ability of the model. The number 
of iterations is related to the model's learning of data 
features, which affects the model's fitting ability. In 
summary, the LSTM model involves many parameters, 
and the selection of parameters requires repeated training 
and adjustment based on actual tasks and data. To find 
reasonable and effective LSTM network parameters, the 
study chose to introduce GWO, an IOA, for 
hyperparameter optimization. Three hyperparameters, the 
number of HL nodes, the batch size and the initial value 
of the learning rate of the LSTM network, were selected 
for optimization. The range of the number of HL nodes 
was set to [1, 250], the range of the batch size was set to 
[1, 64], and the range of the initial value of the learning 
rate was set to [0.001,0.5]. The hyperparameters obtained 
from intelligent optimization were input into the test set 
for validation, and the hyperparameters were considered 
optimal when the mean square error of the prediction 
model was minimized. The total number of parameters of 
LSTM network included IG, OG, FG and candidate state, 
and the parameters of LSTM were simplified into two 
matrices, which map the input and output respectively, 
one of which has the dimension of * hidden input , and 
the other has the dimension of * hidden hidden , and the 
total number of parameters was namely 
4 * * hidden input hidden hidden hidden ++ （ ） . The input 
feature dimension of the LSTM network is defined as 
n
, 
the length of the input sequence as T , the dimension of 
the hidden state as 
m
, and the time complexity of the 
LSTM network as 
2
) 44 ( O Th Tnh + . The introduction of 
the GWO algorithm will somewhat increase the 
computational complexity of the LSTM network. 
The GWO algorithm is a natural heuristic algorithm 
that simulates the predatory social behavior of grey wolf 
groups. According to the mechanism of hunting 
cooperation in grey wolf groups, it achieves the goal of 
problem optimization and solution. GWO has been 
widely used in neural network optimization, various 
scheduling, control, and combination problem solving. 
Compared to other optimization algorithms, GWO 
possesses the following attributes: simplicity, ease of 
implementation, robust global search capabilities, and 
rapid convergence speed. [25-26]. 
Usually, there is a strict hierarchical system in a grey 
wolf pack, which is generally divided into four levels: 
first-level 

, second-level 

, third-level  , and 
fourth-level 

. Based on the hierarchy of the grey wolf 
pack, high-level grey wolves have absolute dominance 
over low-level grey wolves. When designing GWO 
models, it is necessary to first construct a hierarchical 
model of the grey wolf pack. It calculates the fitness 
values of individual populations separately, and labels the 
grey wolf level based on the fitness size. Grey wolf 

 
is the leader of all wolf packs, which is responsible for 
commanding activities such as hunting, gathering, and 
escaping. Grey wolf 

 is responsible for assisting grey 
wolf 

 in decision-making, supervising and leading 
other groups to implement actions, and providing 
feedback on suggested information to grey wolf 

. 
Grey wolf  and 

 are responsible for obeying and 
executing orders, safeguarding the safety of the wolf pack, 
and ensuring the balance within the grey wolf pack. The 
hierarchical system and hunting mechanism of grey 
wolves are shown in Figure 4 [27-28]. 
 

38   Informatica 48 (2024) 31–50                                                                        J. Yu 




D

D

D

C
C
C
Candidate   
wolf
a
a
a
Move
 
Figure 4: Grey wolf hierarchy and hunting mechanisms 
 
As shown in Figure 4, the population system of grey 
wolves takes a pivotal part in the hunting process. Grey 
wolves search and track their prey through group 
activities, and then surround them from different 
directions. Grey wolf 

 commands grey wolf 

 and 
 to launch an attack on their prey, while grey wolf 

 
is responsible for containment, achieving the 
transformation and movement of the entire encirclement. 
The process of the group surrounding the prey is shown 
in equation (6), wherein D indicates the distance 
between the grey wolf and the prey. C represents a 
constant, oscillation factor. 
( )
( ) ,
pt
X X t
 represents the 
positions of prey and grey wolf at the t iteration. 
( )
( )
pt
D C X X t =−
  (6) 
The calculation of the oscillation factor C is 
indicated in equation (7), where 
2
r represents a random 
constant between [0,1]. 
2
2 Cr =  (7) 
The position update of the grey wolf is indicated in 
equation (8), where A represents the convergence 
factor. 
( ) ( ) 1
p
X t X t A D + = − (8) 
The calculation of the convergence factor is 
indicated in equation (9), where 
a
 represents a linear 
decrease from 2 to 0 as the iterations rises. 
_ Max it
 
represents the max number of iterations set. 
1
r 
represents a random constant between [0,1]. 
1
2
22
_
A ar a
t
a
Max it
=− 


=−


  (9) 
The hunting phase involves updating the position of 

 using the positions of grey wolf 

, 

, and  , and 
the updating process of 

 is denoted in Equation (10). 
( )
( )
( )
( )
( )
( )
1
2
3
t
t
t
D C X X t
D C X X t
D C X X t
 



=−


=−


=−

 (10) 
The position vectors of wolf packs of different levels 
are calculated as denoted in equation (11). 
11
22
33
X X A D
X X A D
X X A D



=− 

=−


=−

  (11) 
Equations (10) and (11) can be used to calculate the 
distance between an individual and the optimal three 
wolves, and thus comprehensively determine the 
direction in which the individual moves towards the prey. 
It implements local or global search based on the size of 
the convergence factor. The process of optimizing LSTM 
model parameters using GWO is shown in Figure 5. 
 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 39 
Weather Season Holiday 
Passenger traffic data
Data processing and 
division of training and 
test sets
Initialize LSTM 
parameters
Training set data 
input
Generation of an 
initial 
population of 
gray wolves
Train the 
network and 
obtain fitness 
values to update 
the            ranks 
of gray wolves
  
Updated gray 
wolf locations
Whether the maximum 
number of iterations has been 
reached
Optimizing 
LSTM models
Validation of 
test sets
N
Y
preys
D 
D 
D 
Move
 
Figure 5: Schematic diagram of the process of optimizing LSTM model parameters by GWO 
 
Although GWO has many advantages and achieves 
hyperparameter optimization of LSTM models, the GWO 
algorithm is inclined to get stuck in local optima in the 
later phrase of parameter optimization. Therefore, the DE 
algorithm is introduced to optimize the GWO algorithm 
in this study. DE is an evolutionary algorithm with fast 
convergence speed and high accuracy, which utilizes the 
differences between individuals in the population to 
explore the search space and search for the optimal 
solution. The DE algorithm originates from the 
improvement of genetic annealing algorithm. The 
algorithm gradually promotes the individuals in the 
population through continuous iteration, making them 
gradually approach the optimal solution. The working 
mechanism of the DE algorithm is indicated in Figure 6 
[29-30]. 
 
Mutated 
individual
Differential 
vector
( ) ( ) ( ) ( ) ( )
1
ij aj bj cj
v t x t F x t x t + = +  −
Candidate  
individual 2
Candidate  
individual 1
Target  
individual
i
u
i
u
i
x
i
v
( )
aj
xt
( )
bj
xt
( )
cj
xt
0
x
1
x
Start
population  initialization
Satisfaction of 
termination conditions
variant operation
crossover operation
Output optimal 
individuals
Y
N
Selection of operation
Boundary condition 
processing
Calculate the objective 
function
 
Figure 6: Differential evolution algorithm working mechanism and flow 
 
Firstly, a certain number of individuals are produced 
as the initial population, and the expression for each 
individual is shown in equation (12). In equation (12), the 
i th individual of the population is represented as 
( ) ( ) ( ) ( ) ( )  
12
, ,..., ,...,
i i i ij iD
x t x t x t x t x t = , where 
j
 
represents the parameter index and   1,2,..., jD = . 
,
low up
jj
xx represent the boundary conditions of individual 
population. t represents evolutionary algebra. 
( )   ( )
0 0,1
low up low
ij j j j
x x rand x x = +  − (12) 

40   Informatica 48 (2024) 31–50                                                                        J. Yu 
The population dominant individuals are calculated 
based on the fitness function, and then mutation 
operations are performed within the individuals [31]. 
New individuals are generated from three individuals 
through differential strategy, and the calculation is 
denoted in equation (13). In equation (13), ( )
ij
vt 
represents the mutated individual. ( )
aj
xt , ( )
bj
xt , and 
( )
cj
xt are three different individuals. ( )
cj
xt stands for 
mutation operator. 
( ) ( ) ( ) ( ) ( )
1
ij aj bj cj
v t x t F x t x t + = +  − (13) 
In the process of evolution, to increase the diversity 
of the population, new individuals are crossed with 
existing individuals, as shown in equation (14). In 
equation (14), ( ) 1
ij
ut + represents the variable of 
variation. 
R
C represents the crossover operator. 
rand
j 
represents a random dimension. 
( )
( ) ( )
( ) ( )
1     0,1  
1
     0,1  
ij R rand
ij
ij R rand
v t if rand C or j j
ut
x t if rand C or j j
+  = 

+=

=


(14) 
Finally, the selection operation is conducted 
according to the greedy criterion, and the calculation 
process is shown in equation (15) [32]. In equation (15), 
f represents the fitness function. 
( )
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )
1     1
1
     1
i i i
i
i i i
u t if f u t f x t
xt
x t if f u t f x t
 + + 

+=

+


(15) 
After completing the encirclement, hunting, and 
attack behavior of the grey wolf population according to 
the GWO algorithm, the DE algorithm is applied to 
search for the best wolf location, complete the wolf pack 
location update, and make the GWO algorithm jump out 
of the local optimal solution. The process of improving 
the GWO algorithm by mixing the DE algorithm is 
shown in Figure 7. 
 
Initialize LSTM 
parameters
Setting the fitness value of 
the parent gray wolf 
population
Find the gray wolf           
 based on fitness values
Output prediction results
Start 
Division  of training 
and test sets
Setting the DE-GWO 
Initial Parameters
Randomly generated 
and mutated gray wolf 
populations
Updating the location and 
parameters of individuals in 
the parent population
Crossover operations 
between parents and mutant 
populations to search for the 
location of optimal optimal 
wolf individuals
The maximum 
number of iterations 
  
N
Y
DE-GWO optimizes 
LSTM parameters
Data processing 
 
Figure 7: Flowchart of DE algorithm mixed with improved GWO algorithm 
 
4 Performance testing and 
application effect analysis of 
tourism passenger flow prediction 
model 
To identify the performance of the hybrid improved 
GWO algorithm and DE-GWO-LSTM PFP model 
designed in the research, a series of performance testing 
experiments and prediction application effect analysis 
experiments were conducted. 
 
 
 
4.1 Performance testing of improved GWO 
parameter optimization algorithm 
Theoretical analysis can to some extent demonstrate the 
global search capability of the DE hybrid improved GWO 
algorithm, but quantitative performance analysis is still 
needed. The performance of the study was validated 
using different testing functions. The study selected 5 sets 
of testing functions, namely the unimodal benchmark 
testing functions Sphere, Quartic, multimodal Rastigin, 
Ackley, and Generalized Penalized testing functions. 
Fire-fly Algorithm (FA), Whale Optimization Algorithm 
(WOA), and traditional GWO were selected for 
comparison. It set the max number of iterations for the 
experiment to 1000, the initial population size to 60, and 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 41 
the optimization times for each function to 40. The 
optimization outcomes for various test functions are 
indicated in Table 2. In Table 2, for the two unimodal test 
functions, the average optimization result of DE-GWO 
was closer to the global minimum value, with values of 
6.134E-12 and 6.189E-10, respectively. The standard 
deviation value was relatively small, and the improved 
GWO algorithm had better stability than the traditional 
GWO algorithm and the other two IOAs on the unimodal 
test function. On three different multimodal test functions, 
the partial optimization values of different optimization 
algorithms increased, but compared to each other, the 
improved GWO algorithm’s optimization value was still 
the smallest. In 40 optimization tests, the convergence 
rate of the improved GWO algorithm reached 100%. The 
improvement strategy of research design enabled GWO 
to effectively explore space, avoiding the occurrence of 
local optima and solving the problem of multimodal test 
functions having multiple local minima. 
 
 
Table 2: Comparison of test function optimization results 
Test 
function 
Algorithm 
Global 
minimum 
Variable 
dimension 
Average 
convergence 
value 
Convergence 
times 
Standard 
deviation 
value 
Sphere 
FA 0 30 6.156E-3 16 0.5687 
WOA 0 30 2.371E-5 18 0.2671 
GWO 0 30 7.684E-6 26 0.0061 
DE-GWO 0 30 6.134E-12 40 0.0001 
Quartic 
FA 0 10 2.644E-4 20 0.6412 
WOA 0 10 6.157E-3 19 0.3941 
GWO 0 10 4.198E-6 23 0.0267 
DE-GWO 0 10 6.189E-10 40 0.0035 
Rastrigin 
FA 0 10 4.197E-3 13 0.7616 
WOA 0 10 1.684E-5 18 0.7613 
GWO 0 10 6.167E-4 26 0.0646 
DE-GWO 0 10 8.168E-9 40 0.0000 
Ackley 
FA 0 10 6.164E-2 16 0.6971 
WOA 0 10 1.649E-4 21 0.7364 
GWO 0 10 5.167E-5 33 0.3764 
DE-GWO 0 10 6.169E-7 40 0.0046 
Generalized 
Penalized 
FA 0 10 1.649E-2 23 0.6791 
WOA 0 10 1.174E-3 24 0.5314 
GWO 0 10 3.197E-4 29 0.3181 
DE-GWO 0 10 6.318E-6 40 0.0364 
 
The GWO algorithm’s optimization ability before 
and after improvement was compared and analyzed, 
population fitness was utilized as the evaluation indicator, 
and the analysis outcomes are indicated in Figure 8. In 
Figure 8 (a), the traditional GWO algorithm had a faster 
evolutionary speed in the early phrase and a slower 
evolutionary speed in the later phrase. It gradually 
converged in the middle and later phrases of iteration, 
approaching the optimal population fitness around 115  
 
generations. In Figure 8 (b), the hybrid optimized 
DE-GWO algorithm approached the optimal population 
fitness curve in the early stages of iteration. It can be seen 
that the DE improvement strategy increased the 
population fitness ability of the GWO algorithm. With the 
data analysis in Table 1, the DE-GWO algorithm’s the 
global optimal solution search ability significantly 
improved. 
 

42   Informatica 48 (2024) 31–50                                                                        J. Yu 
Average distance
20 40 60 100 120 140 160 80 0 180
(a) Before improvement
Generation 
200 20 40 60 100 120 140 160 80 0 180
(b)  After improvement
Generation 
200
Best value
Average target value 
Best value
Average target value 
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Average distance
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
 
Figure 8: Comparison of population evolution before and after GWO algorithm improvement 
 
4.2 Performance testing of DE-GWO-LSTM 
prediction model 
Firstly, the optimization ability of DE-GWO for LSTM 
hyperparameters was evaluated, and Hyper volume (HV) 
and Inverse Generational Distance (IGD) were selected as 
evaluation indicators. The experiment outcomes are 
indicated in Figure 9. HV and IGD are convergence 
indicators for evaluating algorithms. HV can measure the 
volume occupied by the solution set in the target space. It 
can be demonstrated that a higher value of HV leads to 
enhanced performance and convergence of the solution 
set. IGD measures the distance between the approximate 
Pareto front generated by the algorithm and the true 
Pareto front. A reduction in the value of IGD is indicative 
of an enhanced convergence of the algorithm. In Figure 9 
(a), the HV indicator curve of DE-GWO was the highest, 
with a maximum value of 0.91. It was better than 
traditional GWO by 0.74, while the HV value of the FA 
algorithm was the smallest, only 0.42. As shown in 
Figure 9 (b), the IGD curve of DE-GWO converged to 
the minimum value of around 0.09, which was 0.21 lower 
than the FA algorithm and 0.05 lower than the GWO 
algorithm. Overall, the DE-GWO algorithm has good 
convergence in the LSTM parameter optimization 
process, and the generated Pareto frontier solutions cover 
a large number of true frontier solutions. 
 
5
Iterations 
10
Hypervolume Indicator
15 20 25 30 40 45
0.1
0.0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 35 50
(a) HV
5
Iterations 
10
Inverted Generational Distance
15 20 25 30 40 45
0.1
0.0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
35 50
(b) IGD
0
FA
WOA
DE-GWO
GWO
FA
WOA
DE-GWO
GWO
 
Figure 9: Comparison of HV and IGD for different optimization algorithms 
 
Simultaneously it evaluated the diversity of the 
solution set of the DE-GWO algorithm, as the diversity of 
parameters affects the accuracy of the algorithm's 
prediction. The experimental results of Spacing and 
Spread are shown in Figure 10. Spacing and Spread 
indicators measured the minimum and maximum 
Euclidean distance between all solutions, respectively. 
The larger the values of the two indicators, the more 
dispersed and diverse the solutions in the solution set. As 
shown in Figure 10 (a), the Spacing curve of the 
DE-GWO algorithm was at its highest level and 
ultimately stabilized above the value level of 0.8. As 
shown in Figure 10 (b), the Spread curve of the DE-GWO 
algorithm was at the highest level, while the Spread 
values of the other three algorithms were all below 0.7. 
The DE-GWO achieved good application results in 
optimizing LSTM parameters. 
 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 43 
5
Iterations 
10
Spacing
15 20 25 30 40 45
0.1
0.0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 35 50
(a) Spacing experiment results
5
Iterations 
10
Spread
15 20 25 30 40 45
0.1
0.0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
35 50
(b) Spread experiment results
0
DE-GWO
FA
WOA
GWO
DE-GWO
FA
WOA
GWO
 
Figure 10: Comparison of solution diversity of different algorithms 
 
Time series prediction models LSTM, GWO-LSTM, 
and Grey Forecast Model (GM) were selected for 
performance comparison. The NYC-Taxi, Air-Travel, 
Public-Transportation, and Retail datasets were utilized 
as the PFP dataset. NYC-Taxi included the travel records 
of New York City taxis. Air-Travel contained flight 
information and passenger data from different airports. 
Public-Transportation included passenger flow data for 
public transportation systems such as subways, buses, etc. 
Retail included both retail store sales records and 
customer visit records, which can be used for predicting 
and analyzing traffic in corresponding fields. The 
experimental data was broken into training and testing 
sets in an 8:2 ratio. 
The MAPE, Root Mean Square Error (RMSE), 
Mean Absolute Error (MAE), and R2 indicators of 
different prediction models were compared, and the 
experiment findings are indicated in Table 3. MAPE, 
RMSE, and MAE are all indices that determine the 
deviation between predicted and actual values. The three 
indices comprehensively assess the model’s the 
predictive ability, and smaller error values indicate better 
model performance. R2 can measure the degree to which 
a predictive model interprets the dependent variable and 
is applied to evaluate the goodness of fit of the model. 
The closer the R2 is to 1, the greater the model's 
explanatory power for variation in the dependent variable 
and the better the model's predictive effectiveness. In 
Table 3, the three types of error indicators of the 
improved DE-GWO-LSTM model were the smallest, 
while those indicators of the traditional LSTM and the 
GM models had larger values, roughly floating above the 
0.4 value level. The performance improvement of the 
improved GWO-LSTM model was not significant 
compared to the baseline model, and the error value still 
fluctuated around 0.3 level. The R2 values of the 
DE-GWO-LSTM model were all above the 0.85 level, 
with a maximum value of 0.9450. Compared to 
traditional LSTM models, it improved by 46.14%. This 
indicated that the model can accurately fit the 
characteristics of historical data and use data features at 
different time scales to complete data prediction. Overall, 
it can be seen that parameter optimization helped the grey 
wolf population break out of local optima, improving 
prediction accuracy and global search ability. 
 
 
Table 3: Comparison of predictive performance of different prediction models 
Model NYC-Taxi Air-Travel Public-Transportation Retail 
LSTM 
MAE 0.4861 0.4012 0.5109 
RMSE 0.4392 0.3694 0.4971 
R
2
 0.6466 0.6711 0.7130 
MAPE 0.4316 0.4160 0.3915 
GWO-LSTM 
MAE 0.3106 0.3067 0.3406 
RMSE 0.2906 0.3941 0.3166 
R
2
 0.7169 0.7613 0.7164 
MAPE 0.3641 0.3296 0.3064 
GM 
MAE 0.4216 0.4613 0.4136 
RMSE 0.4035 0.4067 0.3914 
R
2
 0.6812 0.6913 0.7066 
MAPE 0.3916 0.4036 0.4162 
DE-GWO-LSTM 
MAE 0.0693 0.0946 0.1067 
RMSE 0.1306 0.1649 0.1741 
R
2
 0.9450 0.9426 0.8697 
MAPE 0.1642 0.0952 0.2234 

44   Informatica 48 (2024) 31–50                                                                        J. Yu 
The results of the time efficiency comparison of 
different prediction models are shown in Figure 11. In 
Fig. 11, the prediction efficiency of different prediction 
models on different datasets was significantly different, 
but the computational efficiency of DE-GWO-LSTM 
model was higher in all datasets, indicating that the 
hyperparameter seeking optimization of LSTM by GWO 
algorithm after optimization by DE algorithm improved 
the efficiency of prediction model, which is comparable 
to the prediction time of simple baseline model. 
 
5
Data sample(×100)
10
Time (s)
15 20 25 30 40 45
2
0
4
6
8
10
12
14
16
18
20
0 35 50
(a) NYC-Taxi
5
Data sample(×100)
10
Time (s)
15 20 25 30 40 45
2
0
4
6
8
10
12
14
16
18
20
35 50
(b) Air-Travel
0
5
Data sample(×100)
10
Time (s)
15 20 25 30 40 45
2
0
4
6
8
10
12
14
16
18
20
0 35 50
(c) Public-Transportation
5
Data sample(×100)
10
Time (s)
15 20 25 30 40 45
2
0
4
6
8
10
12
14
16
18
20
35 50
(d) Public-Transportation
0
GM
LSTM
DE-GWO-LSTM
GWO-LSTM
GM
LSTM
DE-GWO-LSTM
GWO-LSTM
GM
LSTM
DE-GWO-LSTM
GWO-LSTM
GM
LSTM
DE-GWO-LSTM
GWO-LSTM
 
Figure 11: Comparison of prediction efficiency of different models 
 
4.3 Analysis of the application effect of 
DE-GWO-LSTM passenger flow prediction 
model 
Relevant data of a TA in China from January 1, 2023 to 
December 31, 2023 was selected as the experimental 
dataset, with the data from January 1, 2023 to May 31, 
2023 as the training set. The data was divided into 
weekdays and holidays. The predication results of 
different time periods and overall is denoted in Figure 12. 
From sub-figures 12 (a) and (b), the prediction model 
designed in the study had a generally consistent trend 
with the fluctuation of actual passenger flow data for 
weekdays and holidays, but there were errors in 
estimating the number of passenger flows at some times. 
From Figure 12 (c), the prediction model designed in the 
study was more accurate in predicting the overall 
passenger flow over a six-month period, with a higher 
accuracy in predicting the trend of passenger flow and the 
number of passengers. 
 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 45 
Number of people
40 60 100 120 140 160
0
5000
80 180
(a) Weekdays  prediction results
Days
20 0
10000
15000
20000
25000
30000
35000
40000
45000
50000
Number of people
20 30 50 60 70 80
0
10000
40 90
(b) Holidays prediction results
Days
10 0
20000
30000
40000
50000
60000
70000
80000
90000
100000
Real
Prediction 
Real
Prediction 
Number of people
7 8 10 11 12
0
10000
9
(c) Overall prediction results
Month 
6 5
20000
30000
40000
50000
60000
70000
80000
90000
100000
Real
Prediction 
 
Figure 12: Passenger flow prediction results for different time periods 
 
The accuracy, precision, recall, F1 values, and error 
stability of the prediction model on the training and 
testing sets were evaluated and analyzed. The experiment 
outcomes are indicated in Figure 13. From Figures 13 (a) 
and (b), those values of the DE-GWO-LSTM prediction 
model were relatively close on the test and training sets, 
with an accuracy close to 0.95 and a precision and recall 
higher than 0.90. The combined use of four indicators 
comprehensively evaluated the prediction model’s overall 
effectiveness, meeting the actual needs of tourist flow 
prediction in scenic areas. As shown in Figure 13 (c), the 
model had good predictive stability. The prediction error 
for different data samples always maintained a relatively 
stable small range fluctuation, with a stability higher than 
0.90. This indicated that the model can accurately predict 
passenger flow by capturing seasonal and cyclical 
fluctuations. 
 

46   Informatica 48 (2024) 31–50                                                                        J. Yu 
Number of iterations
10 20 30 40 50 60 70 80 90 100
Indicator values
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Accuracy training set
Accuracy test set
Precision training set
(a) Accuracy and Precision
Precision test set
Number of iterations
10 20 30 40 50 60 70 80 90 100
Indicator values
(b)Recall and F1
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
0
Recall  training set
Recall test set
F1 training set
F1 test set
0
Sample size ( ×10)
10 20 30 40 50 60 70 80 90 100
Stability
0.80
0.84
0.86
0.88
0.90
0.92
0.94
0.96
0.98
1.00
(c) Stability of mean error
0
 
Figure 13: Comprehensive performance evaluation of the prediction model 
 
Finally, the PFP model designed for the study was 
applied to the simulation of TAs. A questionnaire survey 
was conducted to analyze and evaluate the economic and 
social benefits, management convenience, and tourist 
experience of the prediction model. Quarterly evaluation 
and follow-up would be conducted, and the statistical 
outcomes are indicated in Figure 14. As shown in Figure 
14, during the follow-up period on a quarterly basis, the 
satisfaction rating curves of the three stakeholders 
towards the predictive model showed an upward trend, 
achieving good application feedback. This model greatly 
facilitated the management of scenic spots and prepared 
tourists’ reception scientifically and reasonably based on 
the predicted results. At the same time, it maximized the 
travel experience for tourists. 
 
4
3
5
6
7
8
9
10
User  satisfaction 
0
Follow -up time(days)
40 80 140 180
Economic and social benefits
Tourism management benefits
20 60 100 120 160
Tourist experience
 
Figure 14: Satisfaction evaluation of predictive model application

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 47 
5 Discussion 
With the continuous development of social economy, 
empowering the tourism industry through smart tourism 
is of crucial significance for developing new quality 
productivity, accelerating the construction of a strong 
tourism country, and promoting the high-quality 
development of tourism. To accelerate the development 
of smart tourism, the study analyzed around the TAs PFP, 
and the study chose LSTM, which is more sensitive to 
spatiotemporal sequences, as the basic framework of the 
prediction model, which is consistent with the technical 
basis of the literature [12], [18]. 
At the same time, the study, in the summary of 
related work, found that the IOA in the process of 
hyper-parameter optimization to improve the 
performance of the prediction model was more obvious. 
Literature [16] uses multi-objective marine predator 
combination strategy to optimize wind speed prediction, 
literature [17] uses PSO, Drosophila optimization, lion 
swarm optimization and SSA to enhance backpropagation 
neural network, and literature [18] uses flock 
optimization algorithm to improve bidirectional LSTM 
prediction model all effectively improve the prediction 
accuracy. In view of this, the study chose the simpler 
GWO algorithm to optimize the hyperparameters of 
LSTM, but the GWO algorithm had the deficiency of 
easily falling into the local optimum. The study 
reintroduces the DE algorithm to optimize the population 
of GWO algorithm, increase the diversity of the 
population, and improve the algorithm's ability of finding 
the optimal solution. Compared with the existing 
advanced research, the study integrated a variety of 
networks and algorithms to improve the performance of 
the prediction model, and at the same time, the model 
complexity was taken into account, and a simpler IOA 
was introduced, which is a better technical strategy. 
Comparing the experimental results of the literature [11], 
it is found that the model designed by the study improves 
the F1 value by 0.25, and the MAPE takes the lowest 
value of 0.0952, which is comparable to the results of the 
literature [14], but the model complexity has been 
significantly reduced. However, the research integrated 
LSTM, GWO and DE algorithms to a certain extent also 
increased the technical complexity, which increased the 
difficulty of using the technology for the tourism 
industry. 
On the whole, the method designed by the study 
realizes the optimal combination of DE and GWO, gives 
full play to the advantages of the two algorithms, and 
improves the accuracy and efficiency of PFP. This helps 
to enrich the theoretical research of IOAs and depth 
studies, and promote the application and development of 
optimization algorithms in the field of time series 
forecasting. The use of this method can help scenic spot 
managers to develop a reasonable resource allocation 
plan, improve service quality, and optimize the visitor 
experience. In future research work, further optimization 
of the algorithm can be considered to improve the 
scalability of the model in the face of different sizes of 
datasets and different types of TAs. Consideration should 
also be given to realizing further extensions of the 
forecasting method to increase the applicability of the 
method in other time series forecasting problems, such as 
weather forecasting, stock forecasting, and so on. 
Moreover, the prediction algorithm is integrated using a 
distributed system in order to realize the real-time 
prediction of tourist volume. 
6 Conclusion 
To improve the intelligent and information-based 
development of tourism and provide tourists with more 
convenient and personalized travel experiences, an LSTM 
PFP model improved by introducing IOAs was studied 
and designed. The experimental results showed that the 
quantitative analysis results further validated the 
theoretical analysis. The average optimization results of 
the unimodal test function were 6.134E-12 and 6.189E-10, 
respectively, which are closest to the global mini value. 
The DE algorithm optimized the global search capability 
of GWO, avoiding the problem of multiple local minima 
in multimodal test functions. Simultaneously it optimized 
the grey wolf population’s the fitness curve. DE-GWO 
showed good optimization results for LSTM 
hyperparameters, with a maximum HV index of 0.91, 
which is better than the traditional GWO's 0.74. The IGD 
curve converged to the minimum value of 0.09, a 
decrease of 0.05 compared to the GWO algorithm. The 
Spacing and Spread curves were both above the 0.8 value 
level, indicating better diversity in the solution set. The 
three error indicators of the DE-GWO-LSTM model on 
different datasets were the smallest, and the R2 value 
increased by up to 46.14%. This model had a relatively 
accurate prediction of overall passenger flow, with good 
stability in accuracy, precision, recall, and F1 values on 
the test and training sets. The prediction error fluctuation 
was small, and it achieved good economic and social 
benefits, which is conducive to scenic area management 
and improves the travel experience of tourists. This study 
is conducive to the intelligent management of tourism, 
but the accuracy of the prediction model in predicting 
sudden changes in passenger flow at peaks or valleys still 
needs to be strengthened. 
References
 
[1] Junfeng Kang, Xingyu Guo, Lei Fang, Xiangrong 
Wang, and Zhengqiu Fan. Integration of Internet 
search data to predict tourism trends using 
spatial-temporal XGBoost composite model. 
International journal of geographical information 
science, 36(2):236-252, 2022. 
https://doi.org/10.1080/13658816.2021.1934476 
[2] Jian-Wu Bi, Chunxiao Li, Hong Xu, and Hui Li. 
Forecasting daily tourism demand for tourist 

48   Informatica 48 (2024) 31–50                                                                        J. Yu 
attractions with big data: an ensemble deep learning 
method. Journal of travel research, 61(8):1719-1737, 
2022. https://doi.org/10.1177/004728752110405 
[3] Xiaoyao Zheng, Baoting Han, and Zhen Ni. 
Tourism route recommendation based on a 
multi-objective evolutionary algorithm using 
two-stage decomposition and pareto layering. 
IEEE/CAA journal of automatica sinica, 
10(2):486-500, 2023. 
https://doi.org/10.1109/JAS.2023.123219 
[4] Hanyoung Go, and Myunghwa Kang. Metaverse 
tourism for sustainable tourism development: 
tourism agenda 2030. Tourism review, 
78(2):381-394, 2023. 
https://doi.org/10.1108/TR-02-2022-0102 
[5] Chulmo Koo, Jookyung Kwon, Namho Chung, and 
Jungkeun Kim. Metaverse tourism: conceptual 
framework and research propositions. Current issues 
in tourism, 26(20):3268-3274, 2023. 
https://doi.org/10.1080/13683500.2022.2122781 
[6] Graham Miller, and Anna Torres-Delgado. 
Measuring sustainable tourism: a state of the art 
review of sustainable tourism indicators. Journal of 
sustainable tourism, 31(7):1483-1496, 2023. 
https://doi.org/10.1080/09669582.2023.2213859 
[7] Dogan Gursoy, Yu Li, and Hakjun Song. ChatGPT 
and the hospitality and tourism industry: an 
overview of current trends and future research 
directions. Journal of hospitality marketing & 
management, 32(5):579-592, 2023. 
https://doi.org/10.1080/19368623.2023.2211993 
[8] Shenghong Wang, Yunyun Yu, Jiaqi Chen, and Jun 
Liu. Impact of climate change on cherry blossom 
viewing tourism: analysis and simulation based on 
Weibo proxy data. Current issues in method and 
practice, 26(5):718-734, 2022. 
https://doi.org/10.1080/13683500.2022.2049711 
[9] Yuan Gao, Yao-Yi Chiang, Xiaoxi Zhang, and Min 
Zhang. Traffic volume prediction for scenic spots 
based on multi-source and heterogeneous data. 
Transactions in GIS, 26(6):2415-2439, 2022. 
https://doi.org/10.1111/tgis.12975 
[10] Wanying Li, Hongzhi Guan, Yan Han, Haiyan Zhu, 
and Ange Wang. Short-term holiday travel demand 
prediction for urban tour transportation: a combined 
model based on STC-LSTM deep learning approach. 
KSCE journal of civil engineering, 26(9):4086-4102, 
2022. https://doi.org/10.1007/s12205-022-2051-8 
[11] Fernando Terroso Sáenz, Francisco Arcas-Tunez, 
and Andres Muñoz. Nation-wide touristic flow 
prediction with graph neural networks and 
heterogeneous open data. Information fusion, 
91(3):582-597, 2023. 
https://doi.org/10.1016/j.inffus.2022.11.005 
[12] Qian Xu. Incorporating CNN-LSTM and SVM with 
wavelet transform methods for tourist passenger 
flow prediction. Soft comput, 28(3)2719-2736, 2024. 
https://doi.org/10.1007/s00500-023-09592-w 
[13] Gang Xue, Shifeng Liu, Long Ren, and Daqing 
Gong. Forecasting hourly attraction tourist volume 
with search engine and social media data for 
decision support. Information processing & 
management, 60(4):103399-103416, 2023. 
https://doi.org/10.1016/j.ipm.2023.103399 
[14] Haodong Sun, Yang Yang, Yanyan Chen, Xiaoming 
Liu, and Jiachen Wang. Tourism demand 
forecasting of multi-attractions with spatiotemporal 
grid: a convolutional block attention module model. 
Information technology & tourism, 25(2):205-233, 
2023. https://doi.org/10.1007/s40558-023-00247-y 
[15] Dechun Lu, Yiding Ma, Fanchao Kong, Caixia Guo, 
Jinbo Miao, and Xiuli Du. Support vector regression 
with heuristic optimization algorithms for predicting 
the ground surface displacement induced by EPB 
shield tunneling. Gondwana research, 123(11):3-15, 
2023. https://doi.org/10.1016/j.gr.2022.07.002 
[16] Jianzhou Wang, Honggang Guo, Zhiwu Li, Aiyi 
Song, and Xinsong Niu. Quantile deep learning 
model and multi-objective opposition elite marine 
predator optimization algorithm for wind speed 
prediction. Applied mathematical modelling, 
115(3):56-79, 2023. 
https://doi.org/10.1016/j.apm.2022.10.052 
[17] Xiancheng Mei, Chuanqi Li, Qian Sheng, Zhen Cui, 
Jian Zhou, and Daniel Dias. Development of a 
hybrid artificial intelligence model to predict the 
uniaxial compressive strength of a new aseismic 
layer made of rubber-sand concrete. Mechanics of 
advanced materials and structures, 
30(11):2185-2202, 2023. 
https://doi.org/10.1080/15376494.2022.2051780 
[18] Ilyоs Abdullaev, Natalia Prodanova, Mohammed 
Altaf Ahmed, E. Laxmi Lydia, Bhanu Shrestha, 
Gyanendra Prasad Joshi, and Woong Cho. 
Leveraging metaheuristics with artificial 
intelligence for customer churn prediction in 
telecom industries. Electron research archive, 
31(8):4443-4458, 2023. 
http://www.aimspress.com/journal/ERA 
[19] Yuanzhou Zheng, Lei Li, Long Qian, Bosheng 
Cheng, Wenbo Hou, and Yuan Zhuang. 
Sine-SSA-BP ship trajectory prediction based on 
chaotic mapping improved sparrow search algorithm. 
Sensors-basel, 23(2):704-722, 2023. 
https://doi.org/10.3390/s23020704 
[20] Faten Khalid Karim, Doaa Sami Khafaga, Marwa M. 
Eid, S. K. Towfek, and Hend K. Alkahtani. A novel 
bio-inspired optimization algorithm design for wind 
power engineering applications time-series 
forecasting. Biomimetics, 8(3):321-334, 2023. 
https://doi.org/10.3390/biomimetics8030321 
[21] Ejay Nsugbe. Toward a self-supervised architecture 
for semen quality prediction using environmental 
and lifestyle factors. Artificial intelligence and 
applications, 1(1):35-42, 2023. 
https://orcid.org/0000-0003-0674-1611 

Passenger Flow Prediction of Tourist Attractions by Integrating… Informatica 48 (2024) 31–50 49 
[22] Tariq Limouni, Reda Yaagoubi, Khalid Bouziane, 
Khalid Guissi, and El Houssain Baali. Accurate one 
step and multistep forecasting of very short-term PV 
power using LSTM-TCN model. Renewable energy, 
205(4):1010-1024, 2023, 
https://doi.org/10.1016/j.renene.2023.01.118 
[23] Bei Sun, Xudong Liu, Jiayuan Wang, Xuezhe Wei, 
Hao Yuan, and Haifeng Dai. Short-term 
performance degradation prediction of a commercial 
vehicle fuel cell system based on CNN and LSTM 
hybrid neural network. International journal of 
hydrogen energy, 48(23):8613-8628, 2023. 
https://doi.org/10.1016/j.ijhydene.2022.12.005 
[24] Xia Liu, Kangyi Wang, and Chongguang Liu. 
Recognition and prediction of nonlinear dynamic 
states of microelectronic devices based on machine 
learning algorithms. Informatica-lithuan, 
48(10):119-132, 2024. 
https://doi.org/10.31449/inf.v48i10.6010 
[25] Irfan Ullah, Kai Liu, Toshiyuki Yamamoto, Md 
Shafiullah, and Arshad Jamal. Grey wolf 
optimizer-based machine learning algorithm to 
predict electric vehicle charging duration time. 
Transportation letters, 15(8):889-906, 2023. 
https://doi.org/10.1080/19427867.2022.2111902 
[26] Hemn Unis Ahmed, Reham R. Mostafa, Ahmed 
Mohammed, Parveen Sihag, and Azad Qadir. 
Support vector regression (SVR) and grey wolf 
optimization (GWO) to predict the compressive 
strength of GGBFS-based geopolymer concrete. 
Neural computing and applications, 
35(3):2909-2926, 2023. 
https://doi.org/10.1007/s00521-022-07724-1 
[27] Osman Altay, and Elif Varol Altay. A novel hybrid 
multilayer perceptron neural network with improved 
grey wolf optimizer. Neural computing and 
applications, 35(1):529-556, 2023. 
https://doi.org/10.1007/s00521-022-07775-4 
[28] Zakariya A. Oraibi, and Safaa Albasri. Efficient 
COVID-19 prediction by merging various deep 
learning architectures. Informatica-lithuan, 
48(5):55-62, 2024. 
https://doi.org/10.31449/inf.v48i5.5424S 
[29] Sanjoy Chakraborty, Apu Kumar Saha, Absalom E. 
Ezugwu, Jeffrey O. Agushaka, Raed Abu Zitar, and 
Laith Abualigah. Differential evolution and its 
applications in image processing problems: a 
comprehensive review. Archives of computational 
methods in engineering, 30(2):985-1040, 2023. 
https://doi.org/10.1007/s11831-022-09825-5 
[30] Anuradha Thakare, Ahmed M. Anter, and Ajith 
Abraham. Seizure disorders recognition model from 
EEG signals using new probabilistic particle swarm 
optimizer and sequential differential evolution. 
Multidimensional systems and signal processing, 
34(2):397-421, 2023. 
https://doi.org/10.1007/s11045-023-00870-2 
[31] Hongliang Zhang, Tong Liu, Xiaojia Ye, Ali Asghar 
Heidari, Guoxi Liang, Huiling Chen, and Zhifang 
Pan. Differential evolution-assisted salp swarm 
algorithm with chaotic structure for real-world 
problems. Engineering with computers, 
39(3):1735-1769, 2023. 
https://doi.org/10.1007/s00366-021-01545-x 
[32] Manoharan Premkumar, Chandrasekaran Kumar, 
Thankkapan Dharma Raj, Somasundaram David 
Thanasingh Sundarsingh Jebaseelan, Pradeep Jangir, 
and Hassan Haes Alhelou. A reliable optimization 
framework using ensembled successive history 
adaptive differential evolutionary algorithm for 
optimal power flow problems. The IET generation 
transmission & distribution, 17(6):1333-1357, 2023. 
https://doi.org/10.1049/gtd2.12738 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

50   Informatica 48 (2024) 31–50                                                                        J. Yu