Enhancing Intelligent Mobility: LSTM Neural Networks for
Short-Term Passenger Flow Prediction in Rail Transit
Jiajun Zhang
1,*
and Jingrong Zhang
2
1
College of Artificial Intelligence, Tianjin University of Science and Technology, Tianjin, 300457, China
2
School of Transportation Engineering, Dalian Jiaotong University, Dalian, 116028, China
Keywords: URS, Short-Term Passenger Flow Prediction, LSTM, Traditional Time-Series Prediction Models.
Abstract: In Urban Rail Systems (URS), traffic flow prediction has a long history. However, due to the inherent high
non-linearity and randomness of transportation systems, it remains a challenging issue. Based on the
passenger flow data from subway stations in Seoul, South Korea, this study aims to conduct short-term
passenger flow predictions within the Seoul Metropolitan Subway system in South Korea using the Long
Short-Term Memory (LSTM) network model, thereby verifying the effectiveness and accuracy of the LSTM
model in urban subway passenger flow prediction. The model aids in facilitating early safety warnings and
evacuations for passenger flow. According to the results, the LSTM model is more accurate for short-term
passenger flow prediction in a high traffic station ("Geongdeok Station") and for a stable station ("Jamwon
Station"). Compared to traditional time-series prediction models, LSTM shows superior forecasting
capabilities. The findings and methodology in this study could serve as references and lessons for other
researchers in similar fields.
1 INTRODUCTION
With the accelerated progress of urbanization, rail
transit is become one of the indispensable means of
transportation in most of the cities. and as the rapid
development of urbanization and the continuous
increasing of population, the passenger flow of URS
is showing an accelerated trend of increase. Accurate
short-term prediction of passenger flow at rail transit
stations play an import role in the optimization of
operation management and the improvement of
transportation efficiency (Wu et al 2019). Zhao et al,
proposed that the application of short-term traffic
flow prediction in subway systems can plan subway
departure intervals rationally based on the prediction
results and serve as a reference for subway and light
rail departure intervals (Zhao et al 2019).
Regarding existing traffic prediction models,
these models can be categorised into two main types:
parametric models and non-parametric models.
Parametric models use particular assumptions to
estimate parameters, providing simplicity,
interpretability, and computing efficiency. However,
they may struggle with complex non-linear
relationships and accuracy if assumptions are not met.
Non-parametric models, on the other hand, are more
flexible and adaptable, capable of capturing complex
data patterns. However, they often require more
training data and have more complex structures,
potentially leading to overfitting. The classic
parametric models include the historical average
model, regression analysis method, Bayesian
methods, etc (Tang et al 2021 & Sun and Wei 2017).
Common non-parametric models include the K-
nearest neighbours approach, support vector
machines, neural networks, and others (Cai et al 2016
& Liyanage et al 2022).
Deep learning has achieved significant
advancements in various domains, including image
identification, natural language processing, speech
recognition, and recommendation systems. This has
substantially accelerated the progress and use of
artificial intelligence technologies. Armando
Fandango et al. proposed the application of iterative
strategies in constructing Recurrent Neural Network
(RNN) models for short-term traffic flow prediction
(Fandango and Wiegand 2018). Although RNN
models are applied to traffic passenger flow
prediction because of their ability to process
sequential data, they have some drawbacks in
practical applications, especially the problems of
gradient vanishing or explosion, which affect the
410
Zhang, J. and Zhang, J.
Enhancing Intelligent Mobility: LSTM Neural Networks for Short-Term Passenger Flow Prediction in Rail Transit.
DOI: 10.5220/0012797600004547
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Data Science and Engineering (ICDSE 2024), pages 410-415
ISBN: 978-989-758-690-3
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

stability of model training and the accuracy of
prediction (Lv et al 2015). LSTM is a unique iteration
of the Recurrent Neural Network (RNN) model,
making it more effective than conventional RNN
models. The LSTM model incorporates forget gates,
input gates, and output gates to effectively capture the
long-term dependencies in time series data, resulting
in enhanced prediction accuracy. Furthermore, due to
the LSTM model's superior ability to handle high-
dimensional data and extract valuable information,
numerous researchers employ deep learning
techniques to analyse high-dimensional
spatiotemporal data. Shi et al. conducted a
comparative analysis of the efficacy of random
forests, Back Propagation (BP) neural networks, and
LSTM models in predicting rail transit flow. They
concluded that the LSTM model exhibited superior
fitting results and demonstrated more comprehensive
predictive capabilities and higher average prediction
accuracy compared to the other models (Shi et al
2020).
This paper mainly utilizes the LSTM model to
predict urban subway passenger flow, reaffirming the
accuracy and irreplaceability of the LSTM model in
the field of short-term passenger flow prediction, and
providing more accurate passenger flow information
for subway operation managers, offering a scientific
basis for decision-making.
2 METHOD AND DATE
2.1 Data Source and Preprocessing
The dataset employed in this study consists of
passenger flow data from the Seoul Metropolitan
Subway system, spanning from 2015 to 2018. This
dataset is publicly accessible on the Kaggle website.
It records the number of passengers entering and
exiting each of the 275 stations of the Seoul subway
from 5 AM to midnight daily, with a time resolution
of one hour. The format of the original data is
presented as shown in Table 1.
Data preprocessing is a crucial initial step in data
analysis, aiming to ensure the quality and consistency
of the data for effective analysis and modeling. The
Jupyter Notebook development tool and the Python
programming language were used in this paper to
preprocess the data. The pandas library was used for
data cleaning, while the glob library was used to look
for data file directories and paths. By removing
duplicate entries, confirming data types, looking for
missing values, and doing statistical analysis, the
dataset's accuracy and integrity were verified. There
were no missing or duplicate records discovered, and
there were no outliers in the number of passengers for
any given period.
Because different features differ significantly
from one another, it is easy for little data to be missed
during training. Thus, for analysis to be effective, all
data must be normalized.
𝐱
norm
=
𝐱𝐱
min
𝐱
max
𝐱
min
(1)
Where, x_"norm" represents the normalized
value. x is the sample value, x_"min" is the minimum
value, and x_"max" is the maximum value.
Normalization, by unifying the data scale, enhances
the learning efficiency of the algorithm and reduces
the risk of gradient vanishing or explosion, thereby
improving the performance and stability of short-term
prediction models. Normalization, through
standardizing the data scale, not only increases the
efficiency of algorithm learning but also minimizes
the risk of gradient disappearance or explosion,
consequently enhancing the effectiveness and
steadiness of short-term forecast models.
2.2 Cluster Analysis
By evaluating the correlation between the silhouette
coefficient of the dataset and the number of clusters,
it can be deduced that the ideal number of clusters is
either 2 or 4, as depicted in Fig. 1. From a research
perspective, two clusters often lack significant
research value; therefore, four clusters were selected
for subsequent clustering.
The 3D visualization following clustering clearly
indicates that the clustering effect on the original data
is not pronounced. Therefore, the data was subjected
to PCA (Principal Component Analysis) for
dimensionality reduction before clustering, as shown
in Fig. 2.
Table 1: Setting Word’s margins.
USE_DT station_code station_name division 05~06 …… 23~24
2018/11/1 0:00 2530 Gongdeok in 74 …… 342
2018/3/26 0:00 309 Jichuk Out 19 …… 2
Enhancing Intelligent Mobility: LSTM Neural Networks for Short-Term Passenger Flow Prediction in Rail Transit
411

Figure 1: Initial Clustering Results of K-means (Picture
credit: Original).
Figure 2: Clustering Results after Dimensionality
Reduction Using K-means (Picture credit: Original).
Fig. 3 is a heatmap of the clustering results for 20
selected stations on the Seoul Metro line in South
Korea. For clearer visualization, the passenger flow is
categorized into four classes, ranging from high to
low. It is observable that, spatially, the passenger flow
radiates from areas of higher to lower density. There
is a strong correlation in passenger flow among these
stations.
Figure 3: Spatial Heat Map (Picture credit: Original).
2.3 LSTM Model
The LSTM model is a sophisticated recurrent neural
network architecture that stands out due to its internal
structures, capable of maintaining a long-term flow of
information when processing sequential data. The
core component of the LSTM model is the cell state,
often metaphorically referred to as a 'highway' for
information transfer between network layers. It
remains almost unchanged, effectively preserving
long-term information continuity. To precisely
control the flow of information, the LSTM model is
equipped with three meticulously designed gate
mechanisms: the forget gate determines which
irrelevant information to discard from the cell state;
the input gate controls the entry of new information;
and the output gate decides what information to
output based on the cell state. The intricate weight
and bias parameters acquired through the network's
training process govern these gates, enabling the
LSTM model to excel and exhibit remarkable
adaptability in various sequence prediction tasks,
particularly those involving extended time intervals
and delays. LSTM, an RNN variation, solves the
RNN's long-term reliance issue. Fig. 4 shows a
visualization of the procedure (Yousfi 2017).
Figure 4: LSTM Flowchart (Picture credit: Original).
The formula for the forget gate in an LSTM is:
𝑓

= σ𝑊

⋅

ℎ

, 𝑥


𝑏

 (2)
The forget gate is responsible for selecting and
discarding specific information from the cell state.
The formulas for the input gate and the candidate
values in an LSTM are:
𝑖

= σ

𝑊

⋅

ℎ

, 𝑥


𝑏


(3)
𝐶


=tanh

𝑊

⋅

ℎ

, 𝑥


𝑏


(4)
The input gate determines the selection of fresh
information to be stored in the cell state, whereas the
candidate values encompass the potential information
that can be incorporated into the cell state.
ICDSE 2024 - International Conference on Data Science and Engineering
412

The formula for the cell state update in an LSTM is:
𝐶

= 𝑓

∗𝐶

𝑖

∗𝐶


(5)
The cell state is updated through the process of
discarding previous information and incorporating
new information.
The formulas for the output gate and the output
value in an LSTM are:
𝑜

= σ

𝑊

⋅

ℎ

, 𝑥


𝑏


(6)
ℎ

= 𝑜

∗ tanh

𝐶


(7)
The output gate determines the specific portion of
the cell state that will be emitted, and the resulting
output value is the refined representation of the cell
state regulated by this gate.
The core of an LSTM network is its cell state
C

,which is updated through a gating mechanism.
These gates control the flow of information in and
out, including the forget gatef

, the input gatei

,and
the output gateo

.W

,W

,W

,W

represent the weight
matrices for the forget gate, input gate, output gate,
and cell candidate values respectively, each
determining how information flows and updates
within the model.b

,b

,b

,b

are the bias terms for the
forget gate, input gate, output gate, and the cell state
candidate vector, used to adjust the activation level of
each gate and cell candidate vector. The LSTM
equations employ the sigmoid function to control the
selective transmission of information and the
hyperbolic tangent function to modulate the value
range of the cell state and output. The LSTM
equations describe how the long short-term memory
units dynamically regulate the long-term storage and
forgetting of information by combining current input
with past state information, using a complex gating
mechanism, thereby solving the vanishing gradient
problem found in traditional recurrent neural
networks while maintaining the temporal
dependencies of sequential data.
3 RESULTS AND DISCUSSION
3.1 Temporal Segmentation
The total daily passenger traffic of the Seoul subway
in South Korea is depicted in Fig. 5. It is observed to
exhibit a bimodal pattern, with peaks occurring
between 8-10 AM and 6-7 PM. This aligns with the
commuting times for most people and the school
hours for students, consistent with real-world
scenarios.
Figure 5: Seoul Subway daily station passenger flow
(Picture credit: Original).
3.2 Temporal Segmentation
As shown in Fig. 6, the horizontal axis represents the
daily passenger flow in December 2018, with each
day serving as an observation point. The observed
results indicate a significant periodic pattern in
passenger flow, particularly evident in the troughs
observed on December 2nd, 9th, and 16th, with a
strict seven-day interval between these dates. This
periodic pattern is likely influenced by the changes
between weekends and weekdays.
Figure 6: Seoul Subway one-month passenger flow (Picture
credit: Original).
3.3 Spatial Characteristics
Fig. 7 demonstrates the significant differences in
passenger traffic across different subway stations.
This highlights the necessity of analyzing and
comparing the daily passenger flow distribution at
various subway stations.
Due to the significant differences in passenger
flow distribution characteristics among various
subway stations, this study has selected two
representative stations for in-depth analysis.
"Geongdeo Station" (transfer station) and "Jamwon
Station" (non-transfer station).
Enhancing Intelligent Mobility: LSTM Neural Networks for Short-Term Passenger Flow Prediction in Rail Transit
413

Figure 7: One day’s passenger flow at each subway station (Picture credit: Original).
3.4 Results
Given the complexity of the Seoul subway system in
South Korea, which comprises 23 lines, this paper
focuses solely on "Geongdeo Station" and "Jamwon
Station" for study. Gongdeok Station, as a
transportation hub connecting Lines 5 and 6 of the
Seoul Subway, has a high volume of passenger flow
and complex traffic patterns due to its transfer
characteristics. On the other hand, Jamwon Station, a
standard station on Line 3, has a more stable flow of
passengers, reflecting the daily commuter status of
ordinary stations in the Seoul subway system.
The selection of these two stations as research
subjects aims to compare and analyze the passenger
flow characteristics of different types of subway
stations. The data from Gongdeok Station helps to
understand and predict the fluctuation of passenger
flow at transfer stations during different periods.
Meanwhile, Jamwon Station provides a benchmark
for assessing the accuracy of the model in predicting
regular passenger flow.
After a thorough analysis of the graphical data
characteristics, this study selected a specific 8-day
period as the subject for short-term passenger flow
prediction based on LSTM. The parameter
adjustments of the LSTM model used are detailed in
Table 2.
Table 2 : Parameter Tuning.
learning_rate Batch_Size Epochs
0.01 32 100
Figure 8: 2530 station Predicted data and true results of the
test set (Picture credit: Original).
Fig. 8 presents the experimental results for
passenger inflow at "Geongdeo Station"
(station_code: 2530). The horizontal axis represents
the date, while the vertical axis indicates passenger
volume. The time interval spans from 5 a.m. to 24
p.m. each day. The blue star line denotes the predicted
values, and the yellow plus line represents the actual
values. It can be observed that most of the predictions
closely align with the real values, though there are
slight deviations in certain areas.
Figure 9: 328 Station Predicted data and true results of the
test set (Picture credit: Original).
ICDSE 2024 - International Conference on Data Science and Engineering
414

Fig. 9 illustrates the predicted inbound passenger
flow results for "Jamwon Station" (station code: 328).
Similar to the case of "Geongdeo Station," the overall
predictions are accurate.
In this paper, ARIMA and BP models are used as
comparative models. The study utilised two
evaluation metrics, namely Mean Absolute
Percentage Error (MAPE) and Root Mean Square
Error (RMSE), to conduct error analysis. As shown in
Table 3, it is evident that, whether it is MAPE or
RMSE, the performance of the LSTM model is
superior to the other models.
Table 3: MAPE and RMSE of Three Test Models.
Model MAPE/% RMSE
LSTM 9.19 156.67
BP 15.86 228.74
ARIMA 18.43 279.87
4 CONCLUSION
This paper focuses on short-term passenger flow
prediction for specific stations within the Seoul
Metropolitan Subway system in South Korea using
LSTM network model. The LSTM network is capable
of learning time series with long-term dependencies,
offering a flexible framework for employing various
combinations of variables. The study not only
demonstrates the effectiveness of the LSTM model in
forecasting passenger flows at different types of
subway stations but also highlights the advantages of
deep learning in handling high-dimensional
spatiotemporal data. These findings provide more
accurate passenger flow information for subway
operation management and scientific decision-
making support for policymakers. While this research
focuses on the impact of time on traffic flow, it does
not account for other factors such as commercial
zones or weather conditions. Future studies could
consider incorporating these additional elements,
including weather and the presence of commercial
areas, to further enhance the model's predictive
accuracy.
AUTHORS CONTRIBUTION
All the authors contributed equally and their names
were listed in alphabetical order.
REFERENCES
Z. Q. Wu, T. Y. Huang, Y. W. Yan, M. S. Zhao, Urban
Mass Transit. 22, 31 (2019)
H. Zhao, D. M. Zhai, C. H. Shi, Urban Mass Transit 32, 50
(2019)
J. Q. Tang, X. W. Zhong, J. Liu, T. R. Li, J. Chongqing
Jiaotong Univ. (Nat. Sci.) 40, 31 (2021)
B. Sun, M. Wei, Highways Autom. Appl. 4, 20 (2017)
P. Cai, Y. Wang, G. Lu, P. Chen, C. Ding, J. Sun, Transp.
Res. Part C Emerg. Technol. 62, 21 (2016)
S. Liyanage, R. Abduljabbar, H. Dia, P. W. Tsai, J. Urban
Manag. 11, 365 (2022)
A. Fandango, R. P. Wiegand, Towards investigation of
iterative strategy for data mining of short-term traffic
flow with recurrent neural networks, in Proceedings of
the 2nd International Conference on Information
System and Data Mining (2018)
Y. Lv, Y. Duan, W. Kang, Z. Li, F. Y. Wang, IEEE Trans.
Intell. Transp. Syst. 16, 865 (2015)
X. R. Shi, C. H. Wang, D. J. Liu, X. Zhang, B. Zhang,
Electr. Technol. Softw. Eng. 182 (2020)
S. Yousfi, S. A. Berrani, C. Garcia, Pattern Recogn. 64, 245
(2017)
Enhancing Intelligent Mobility: LSTM Neural Networks for Short-Term Passenger Flow Prediction in Rail Transit
415