International Journal of Intelligent Networks 5 (2024) 92–100

Contents lists available at ScienceDirect

International Journal of Intelligent Networks

journal homepage: www.keaipublishing.com/en/journals/
international-journal-of-intelligent-networks

Research on intelligent vehicle Traffic Flow control algorithm based on

data mining
Lihua Cheng a, *, Ke Sun b
a
Department of Computer and Software, Chengdu Jincheng College, Chengdu, Sichuan, 610000, China
b
Pudong Development Bank, Chengdu, Sichuan, 610000, China

A R T I C L E I N F O A B S T R A C T

Keywords: Traffic Congestion (TC) is increasing due to urban growth and vehicle numbers, rendering the development of
Machine learning cities and people’s well-being difficult. Traffic Prediction (TP) and control systems have been required to
Dynamic zone segmentation improve Traffic Flow (TF) and reduce TC because standard methods are unsuitable. The paper proposes an
Traffic prediction
innovative method for traffic control using the Dynamic Zone Segmentation Algorithm (DZSA) to solve this
Traffic congestion
Long short-term memory
significant issue. The algorithm uses real-time data and road conditions to partition city traffic into manageable
Bayesian structural time series units, enhancing the adaptability and accuracy of Traffic Prediction (TP) performance. Applying DZSA, the
recommended Long Short-Term Memory + Bayesian Structural Time Series (LSTM + BSTS) learning model
optimizes TP by integrating the best features of conventional and Machine Learning (ML) methods. The model
optimized quality performance when experimentally tested against other benchmark models using metrics like
Mean Absolute Error, Mean Absolute Scaled Error, Accuracy Percent, Root Mean Squared Error, and Mean
Absolute Percent Error. The recommended model, LSTM + BSTS, shows a minimal error rate of 6.68%, indicating
its success.

1. Introduction increased congestion. As metropolitan areas develop, there is a pressing

need for a more sophisticated and flexible Traffic Management System
The primary problem that impacts infrastructure and the quality of (TMS) [3].
life for people is Traffic Congestion (TC) in metropolitan regions Traffic Predicting (TP) is a crucial step in intelligent TF analysis,
worldwide. Decreased productivity, higher fuel prices, and pollution- enabling control mechanisms to proactively adjust parameters like
related damage the global economy hundreds of billions of dollars signal timings, reroute traffic, and issue real-time advisories to drivers.
annually. Modern urban Traffic Flow (TF) is complicated and constantly This proactive system not only resolves existing congestion but also
changing, making traditional Traffic Control Systems (TCS) dependent prevents the formation of new ones, enhancing road infrastructure use
on predetermined timers unsustainable [1]. The outdated Traffic Flow and addressing smart city traffic challenges [4].
Control (TFC) systems are inefficient, leading to bottlenecks and The hybrid method integrates Long Short-Term Memory (LSTM) and
increased congestion. As vehicle numbers increase, intelligent, adaptive Bayesian Structural Time Series (BSTS) to improve the accuracy of
solutions are urgently needed. Data mining and Machine Learning (ML) Traffic Flow Prediction (TFP). LSTM captures temporal dependencies,
technologies can provide real-time, adaptive solutions that optimize while BSTS provides a Bayesian network for ambiguity value and
TFC, reduce congestion, and contribute to sustainable smart city envi­ robustness. This synergy aims to capture complicated patterns in his­
ronments [2]. torical traffic data, enhancing the model’s success. This work context
Conventional TCS face significant drawbacks in managing modern provides more details on the implementation and success of the model.
smart city TF due to their reliance on static timing plans derived from To effectively predict and manage Traffic Congestion (TC), different
past traffic patterns. These designs often fail to adapt to changing traffic methods have been employed, ranging from traditional time-series
conditions, leading to inefficiencies during high-demand periods or analysis to advanced ML algorithms like Random Forests (RF), Sup­
unpredicted measures. The lack of real-time adjustment services also port Vector Machines (SVM), and Neural Networks (NN). While time-
delays their ability to manage TF, resulting in poor performance and series models are beneficial for analyzing historical data to predict

* Corresponding author.
E-mail addresses: [email protected] (L. Cheng), [email protected] (K. Sun).

https://doi.org/10.1016/j.ijin.2024.02.004
Received 22 December 2023; Received in revised form 2 February 2024; Accepted 3 February 2024
Available online 7 February 2024
2666-6030/© 2024 The Authors. Published by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

future traffic conditions, they often fail to factor in dynamic variables predicting specific kinds of TF. The authors [12] introduced a new
such as weather, special events, or road construction. On the other hand, spatial-temporal graph neural network to capture spatial and temporal
ML algorithms provide a more adaptive approach that can consider TF patterns. The model features a learnable positional attention mech­
multiple variables simultaneously, delivering precise forecasts in real- anism for collecting data from neighbouring roads and a sequential
time. However, even these advanced models have limitations component to understand local and global time-related traffic varia­
regarding traffic patterns’ complexity and constantly changing nature tions. Tests on multiple real-world traffic datasets validate the frame­
[5]. work’s effectiveness.
Hybrid modelling is more practical than TFP because it combines the The authors [13] introduce the Relative Position Congestion Tensor
metrics of numerous methods. Traditional models face real-world traffic (RPCT) for forecasting TC. The model employs spatiotemporal tensors
complexities, but hybrid models offer a more robust tool for TP. They constructed from congestion matrices of road nodes and utilizes a con­
integrate statistical methods with ML algorithms, providing a nuanced volutional LSTM network for future predictions. Experimentally, the
understanding of historical and real-time data. Hybrid models are approach significantly surpasses a range of baseline models in predicting
beneficial for segmented or grid-based traffic information, as they can congestion hotspots. The study demonstrates the model’s capture of
focus on individual sections, enhancing the accuracy of their TP. This temporal and spatial traffic dynamics. Propose a Neuro-fuzzy-based
granular motivation captures localized traffic patterns, road conditions, forecasting model for TC to reduce vehicular delays at urban in­
and other contextual factors, leading to more accurate predictions and tersections. The model integrates data from IoT sensors deployed across
smarter TC adoptions [6]. roads and processes it through a neuro-fuzzy engine on a cloud server.
The above challenges motivated me to propose a novel approach to The approach shows remarkable accuracy, reaching 98.72% in valida­
TMS through the Dynamic Zone Segmentation (DZS) algorithm, and the tion and 99.214% in training, outperforming existing state-of-the-art
idea is to segment the traffic zone into more manageable, smaller sec­ models. The study holds significant implications for enhancing intelli­
tions based on road and GPS information. The algorithm dynamically gent city TMS.
adjusts the sizes of these zones based on real-time traffic volume and The authors [14] introduce the MDCGCN model to improve
road hierarchy. Furthermore, the segmentation does not merely rely on medium-TFP and long-term TFP. The model addresses the complexities
zone centroids as nodes and incorporates other significant elements such of spatial-temporal correlations in traffic by utilizing a three-component
as major intersections, traffic signals, and public transport stops [7]. The structure: recent, daily, and weekly periods. Each component contains
hybrid LSTM + BSTS learning model TFP levels within each zone once it an adaptive mechanism and data correlation block to refine multi-sensor
identifies and starts these dynamic zones. The proposed model was data quality and capture dynamic traffic patterns. Tested on two real
experimented with against other baselines for different metrics, and the datasets, MDCGCN outperforms existing methods in prediction accu­
results have shown that the proposed model has fared well against other racy. The authors [15] proposed a Deep Spatial and Temporal Network
baseline models. Model (DSGCN) for more accurate Traffic Congestion Prediction (TCP).
The paper is structured as follows: Section 2 presents the research The research employs grid-based segmentation of traffic areas, using the
literature review, Section 3.0 offers the recommended model, Section 4 grid centroids as nodes in an adjacency matrix. Graph Convolutional
shows the results of the experiments, and Section 5 provides a Neural Networks (GCNN) capture spatial correlations, while a dual-layer
conclusion. feature model (DSTM) assesses temporal relationships. Tests on real
PeMS datasets confirm that DSGCN surpasses baseline models in pre­
2. Literature review dictive accuracy.
[16] made a significant contribution by introducing a Grey Predic­
This work [8] critically reviewed 165 articles on data mining and ML tion Model (GPM) using tensor decomposition, considering the
technologies in TMS. The authors identify a lack of standardized ap­ multi-modal characteristics of TF data and its inherent uncertainties,
proaches, offering valuable insights for researchers, traffic software particularly in the short-term TFP scenario. Their work introduced a
companies, and government officials. Their work, one of the most GPM using tensor decomposition. Assumed their multi-modal nature,
expansive reviews to date, underscores the need for innovative TMS. The the research primarily emphasizes transforming TF data into tensors and
study is a comprehensive resource and catalyst for future research. The applying the tensor decomposition algorithm to the Verhulst model, a
study challenges intelligent traffic control regarding data complexity classic GPM. This adaptation proved beneficial for short-term TFP, as
and real-time requirements. The researchers propose a cloud-based their results displayed that their model excelled in extracting the intri­
control system that uses Deep Learning (DL) to predict short-term TF cate multi-modal correlations of the TF data. Their method surpassed
and congestion. Their model also employs intelligent optimization al­ five other GPMs in comparative analyses, providing real-time dynamic
gorithms for real-time control, with simulation results confirming its information crucial for Intelligent Transportation System (ITS) plan­
effectiveness. ning, control, and optimization.
A hybrid Convolutional LSTM (ConvLSTM) model was proposed [9]
to improve travel time, Origin-Destination (OD) and TFP. Their model 3. Proposed dynamic Zone Segmentation (DZS) algorithm and
employed a base approach to represent OD. Time slots segment the day, traffic node analysis
facilitating spatial and temporal OD flow predictions. Using accurate
taxi data, the study validates that the model performs better than others The road network is divided into dynamic zones defined by
in predicting city-wide taxi OD flow. In Ref. [10], the study introduces geographical and traffic parameters. These zones will serve as inde­
an optimized short-term TFP model using Support Vector Regression pendent units for subsequent traffic analyses, ensuring a comprehensive
(SVR). Experimental results show a low error rate of 3.22% with sig­ understanding that extends beyond major highways and arterial roads.
nificant improvements in prediction accuracy during peak hours. The The model partitions the target area into dynamic zones based on traffic
study also enhances pedestrian counting methods, hinting at further size and road pyramid, enhancing adaptability and practicality. Zone
optimizations in feature calculations for even more accurate results. centroids are used as nodes, along with significant connections, traffic
The authors [11] introduce a hybrid approach, Time-Series Analysis signals, and public transport stops, for a more comprehensive network
and Supervised-Learning (TSA-SL), for Short-Term TFP. Using the illustration.
Fourier Transform to model periodic behaviors and correlations as input Implementing the proposed DZS algorithm and hybrid LSTM + BSTS
features, they develop three hybrid models. The authors tested the model in real-life urban settings will lead to better traffic planning,
method on a large Electronic Registration Identification (ERI) dataset adaptive traffic segmentation, and TP. The DZS algorithm provides a
and found that it outperforms traditional models, particularly in dynamic approach to traffic segmentation, allowing for more accurate

93
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

predictions and adaptable TC methods. The hybrid LSTM + BSTS model 1 Initialize Dynamic Zones: Partition the target area into initial zones
leverages the metrics of classic and ML approaches, contributing to more based on road information and GPS boundaries.
reliable TP. 2 Adjust Zone Sizes: Use traffic volume data to adjust the sizes of the
As such, the network graph created using zone and node data is zones dynamically.
represented as EQU [1]. 3 Node Characterization
G = (V, W, M) (1)
• Identify different types of nodes:
where: V be the set of all nodes, where each node v has a type T and • Centroids of zones (C)
geographical coordinates (Latitude, Longitude), EQU [2] • Significant intersections (I)
• Traffic signals (S)
V = {(v1 , T1 , lat1 , lon1 ), (v2 , T2 , lat2 , lon2 ), …, (vn , Tn , latn , lonn )} (2) • Public transport stops (P)
Here, T could be one of the following:
4 Create Node Set V Using:
• C for Centroid V = {(v1 , T1 , Lat1 , Lon1 ), (v2 , T2 , Lat2 , Lon2 ), …, (vn , Tn , Latn , Lonn )}
• I for Intersection
• S for Traffic Signal
• P for Public Transport Stop 5 Calculate Weights for Node Links:

The set W contains weighted links between the nodes. Each link wij • For each pair of nodes i and :
between node i and node j has a weight w calculated as EQU [3]. • Calculate wij = dij × α(Ti , Tj )
( )
wij = dij × α Ti , Tj (3) 6 Define Scaling Factor (T i , T j ) : α(Ti , Tj ) is based on the types of
nodes i and j
Here dij is the GPS distance between i and j, and α(Ti , Tj ) is a scaling
7 Create Weighted Network Graph G = (V, W, M)
factor based on the types of nodes i and j, defined as EQU [4].
⎧
• W contains weighted links wij
( ) ⎨ 1, if both Ti and Tj are centroids
α Ti , Tj = 1.2, if one of Ti or Tj is an intersection or traffic signal (4) • M contains multi-dimensional vectors →
m ij = [dij , sij , rij ]
⎩
1.5, if one of Ti or Tj is a public transport stop
8 Calculate Multi-Parametric Distance Metric dij dij =
This factor is used to scale the weight wij of the link between i and j. A
novel distance metric that goes beyond simple GPS distance is
• α(Ti , Tj ) × GeoDist (Lat1 , Lon1 , Lat2 , Lon2 )
employed. This metric incorporates a scaling factor, making it more
adaptive to real-time traffic dynamics. The distance metric dij between
9 Calculate Great-Circle Distance (GeoDist): Use the great-circle
two nodes i and j is defined as follows: EQU [5]
distance between two points based on latitude and longitude.
( )
dij = α Ti , Tj × GeoDist(lat1 ,lon1 ,lat2 ,lon2 ) (5) 10 Finalize Network Graph: Use the weighted network graph ’G’
for further TC analyses.
Here, α(Ti , Tj ) is the scaling factor based on the types of nodes being A. Data Extraction and Preprocessing
considered. GeoDist represents the great circle distance between two
points on the Earth’s surface, measured using their latitude and In contrast to traditional TC models that often concentrate solely on
longitude. specific types of roads, like highways or streets, this approach employs a
GeoDist(lat1 ,lon1 ,lat2 ,lon2 ) = R × arccos [cos (lat1 ) × cos (lat2 ) × cos (lon1 comprehensive Grid-Based Modelling (GBM) technique. A GBM divides
the entire traffic network, including intersections and other key features,
− lon2 ) + sin (lat1 ) × sin (lat2 )] (6)
into independent grids, each representing a distinct area. A 22.8 km road
segment stretching from Dongfeng Avenue Junction in the Luyuan
where
District to Shiji Plaza in Changchun, Jilin, China, is chosen for this GBM.
Fig. 1 illustrates the selected location and its DZS. This stretch includes
• R is the Earth’s radius, approximately 6371 km.
several major intersections, which serve as critical nodes in the grid-
• arccos is the arccosine function.
based model. Focusing on this localized area aims to capture a micro­
cosm of traffic patterns with broader implications for TCP in the city’s
Short distances are approximated by this great-circle distance EQU
transportation network. The primary data source is GPS data from May
[6], which assumes the Earth is a perfect sphere. The distance ’d’ will be
1, 2022, to April 30, 2023, from vehicles that traverse this targeted road
in the same unit as R, typically in km. The following algorithm presents
stretch and its intersections. This data provides a multi-faceted, real-
the process for DZS.
time snapshot of traffic conditions.
Algorithm 1. for DZS Input The GPS data fields include:

• Raw GPS data containing latitude, longitude, and other relevant • ID: A unique identifier for each vehicle, essential for tracking and
information aggregating data within individual grids and intersections.
• Traffic volume data • Vehicle_Sim_ID: This serves a similar purpose to the ID but offers the
• Road information possibility of integration with other databases, providing a richer
context.
Output • GPS Speed: The current speed of the vehicle acts as an immediate
indicator of congestion levels within each grid and at intersections.
• Dynamic zones for traffic analysis • GPS_Time: Timestamps are crucial for TSA and enable synchroni­
• Weighted network graph G = (V, W, M) zation with other variables that can impact TF at intersections.

94
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

Fig. 1. Zone segmentation.

• GPS_Longitude and GPS_Latitude: These GPS coordinates place are excluded from further analysis.
each vehicle within a specific grid or intersection, facilitating local­
ized congestion modelling. C Geographical Scope-Based Data Filtration for Targeted Traffic Zone
Analysis
Specialized vendor APIs were used to source the GPS data, meeting
the high-frequency, real-time model requirements. These APIs can Next in the methodology is filtering GPS data points that fall outside
handle the substantial data needed for the study, approximately 19 the predefined geographical scope of the target area, Changchun. The
million data points per day, with a data-sending frequency of about 60 rationale behind this data cleansing step is to enhance the precision and
seconds. This level of temporal resolution is critical for accurate TCP relevance of the traffic zone division within the city. Including data
both within the grid and at intersections. Therefore, we tailored the points outside the target region could introduce noise and inaccuracies,
selected API to provide updates at this high frequency, ensuring that the diluting the quality of the predictive models and analysis. The study
model remains current and allows real-time adjustments in traffic con­ strictly confines its geographical scope to Changchun, in Jilin Province.
trol measures. Specifically, data points that fall within the latitude range of 43∘ 83″N to
43∘ 84″, and the longitude range of 125∘ 18″ to 125∘ 40″ are alone
B. Dynamic Multi-Level Anomaly Detection in GPS Traffic Data considered. After obtaining the raw GPS data, the first step is identifying
the GPS coordinates, latitude, and longitude associated with each data
This method identifies and removes unreliable GPS data points that point. A Boundary-Checking Algorithm (BCA) verifies whether each
can compromise the validity of the TCP model. A multi-level thresh­ data point falls within Changchun’s specified latitude and longitude
olding technique and a variable-focused evaluation function filter out ranges. The algorithm immediately excludes any data point outside
anomalies. The evaluation function g(pi , pj ) considers speed (V), accel­ these boundaries from the dataset. After filtering, a random sample of
eration (A), and change (D) in direction between two data points pi and the dataset is manually inspected to confirm the efficacy of the BCA.
pj , EQU [7] Then, this study uses the filtered dataset, which no longer contains out-
{ of-scope data points, for the subsequent phases of traffic zone division
( ) − 1, if V(pi ,pj ) > δV or A(pi ,pj ) > δA or D(pi ,pj ) > δD
g pi , pj = (7) and TCP. The data filtering process explicitly tailors the analysis and
1, otherwise
conclusions to Changchun’s traffic conditions and challenges.
where:
D LSTM for TCP in Segmented Zones
• V(pi ,pj ) is the speed between pi and pj in km/h.
LSTM is a type of Recurrent Neural Network (RNN). While tradi­
• A(pi ,pj ) is the acceleration between pi and pj in m/s2 .
tional RNNs are capable of handling sequences, they often suffer from
• D(pi ,pj ) is the change in direction in degrees.
learning long-range dependencies due to the vanishing gradient prob­
lem. Designing LSTMs to overcome this issue makes them highly effec­
The speed threshold δV , acceleration threshold δA , and direction
tive for tasks that require understanding context over long sequences.
change threshold δD EQU is dynamically determined based on real-time
This has led to their widespread adoption in various fields. In the context
traffic conditions and road type [8].
( ) of the proposed study, which centres on dynamically segmented data,
δV , δA , δD = fdt RT(pi ,pj ) , TC(pi ,pj ) (8) the LSTM model’s capacity to capture long-term dependencies proves
invaluable. TC inherently exhibits temporal patterns influenced by
various factors. The data source consists of high-frequency GPS data,
• RT(pi ,pj ) Road type is a categorical variable representing the type of offering a detailed, real-time view of traffic conditions at intervals of
road between pi and pj (e.g., highway, arterial road, local street). approximately 60 seconds. This time-series data aligns seamlessly with
• TC(pi ,pj ) Traffic Condition: a categorical variable representing the the LSTM model, allowing the model to grasp intricate temporal re­
current traffic condition between pi and pj (in km/h). lationships that simpler models might miss.
• The function fdynamicThreshold is the threshold defined as a weighted The construction of the feature vector for the LSTM model will utilize
sum of predefined maximum values for each road type and real-time the existing GPS data fields. At every time step ‘t’, the feature vector Ft is
traffic conditions. generated in the following manner: EQU [10].
Ft = [Nt , St , Tt , Lat t , Lon t ] (10)
The composite anomaly indicator Zpi for each data point pi is defined
as EQU [9]. where:
∑
i+r
( )
Zpi = g pi , pj (9) • Nt is the number of unique vehicle IDs in the grid at time ’t’, serving
j=i− r as a proxy for traffic volume.
Here, r is a fixed window size set to 5, chosen based on preliminary • St is the average GPS speed across all vehicles in the grid at time ’t’.
data analysis. Data points with Zpi < 0 are identified as abnormal and

95
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

• Tt is the GPSTime; this timestamp is decomposed into auxiliary zones.

features such as time-of-day and day-of-week, enabling the model to Due to its high-frequency nature, collected at 60-s intervals, the
capture the cyclical patterns inherent to urban traffic. LSTM model is designed for adaptability. It uses a backpropagation al­
• Latt and Lont The average GPSLatitude and GPSLongitude in the grid gorithm optimized with Stochastic Gradient Descent (SGD) methods for
serve as the grid’s center point. real-time updating. Dynamic updates of the model parameters,
including weights and biases, adapt to changing traffic conditions,
The rationale for selecting these features hinges on their capability to ensuring continuous refinement of predictive accuracy, making it an
convey a multifaceted view of traffic conditions. Nt offers a quantifiable essential tool for TMS and long-term urban planning.
metric for traffic density while St gives an empirical account of how
freely TF, directly indicating congestion and its absence. The temporal E BSTS for Congestion Prediction in Dynamically Segmented Zones
feature Tt adds a layer of complexity by enabling the model to discern
patterns over varied time scales, which is fundamental for capturing the BSTS models are advanced statistical models for predicting and
cyclical nature of TC. Lastly, the coordinates Latt and Lont imbue the scrutinizing TSA, incorporating seasonal variations, trends, and special
model with spatial awareness, allowing for more granular, location- events like holidays, all within the context of Bayesian analysis. This
specific predictions. enables the model to generate predictions with a probability distribu­
tion, offering a spectrum of potential results rather than a singular-point
1 LSTM Model Formulation approximation.
Given a feature vector Ft = [Nt ,St ,Tt ,Latt ,Lont ], the BSTS model aims
The LSTM model’s core is its ability to remember over long se­ to predict the congestion score St . The model is based on the state-space
quences, making it ideal for TCP. Unlike FFNNs, LSTMs maintain a formulation and can be formally described using the following EQU [19]
memory that captures calculated information, ensuring accurate TCP and EQU [20]:
considering traffic current state and temporal dynamics. The LSTM
State : αt+1 = Zt αt + ηt , ηt ∼ N(0, Wt ) (19)
model takes in a sequence of feature vectors Ft− n , Ft− (n− 1) , …, Ft and
outputs a congestion score St for the target zone. The LSTM model
(20)
operates through a series of gates and states, formalized through several Observation : St = Tt αt + ϵt , ϵt ∼ N(0, Ht )
equations. The forget gate ft decides which information from the pre­
Here, αt represents the state vector containing the underlying
ceding cell state Ct− 1 must be retained using the EQU [11].
structural components at time t, Zt is the transition matrix, ηt is the
( )
ft = σ Wf ⋅ [ht− 1 , Ft ] + bf . (11) system noise, Wt is the system covariance matrix, Tt is the design matrix,
and ϵt is the observation noise with variance Ht .
The input gate it updates the cell state with new information, as given In the BSTS model, Bayesian reasoning is essential for calculating the
by EQU [12]. posterior distributions of the fundamental elements. The core principle
it = σ (Wi ⋅ [ht− 1 , Ft ] + bi ). (12) is to use Bayesian methods to gauge these distributions, starting with
setting prior distributions for each component. Specific priors for these
The cell is then updated to a new state C
̃ t using EQU [13]. elements are defined as follows:
̃ t = tanh (WC ⋅ [ht− 1 , Ft ] + bC )
C (13)
•p(μ1 ) : The prior for the initial local level
Combining the previous and new states, modulated by the forget and • p(γ1 ):: The prior for the initial seasonal effect
input gates, produces EQU [14]. •p(β) : The prior for regression coefficients
• p(σ2 ): The prior observation noise variance
̃t.
Ct = ft × Ct− 1 + it × C (14)
The output gate ot then determines what the next hidden state ht With these established priors, the joint posterior distribution for the
should be using EQU [15]. state and parameters conditioned on the observed data can be EQU [21]
as follows:
ot = σ(Wo ⋅ [ht− 1 , Ft ] + bo ) (15)
p(α1:T , S1:T |Y1:T ) ∝ p(Y1:T |α1:T , S1:T ) × p(α1:T , S1:T ) (21)
The new hidden state ht is then calculated as EQU [16]
where Y1:T are the observed data and α1:T and S1:T are the state and
ht = ot × tanh(Ct ) (16)
congestion score sequences, respectively.
Finally, the congestion score St for the target zone is calculated using The BSTS model’s training phase uses a Markov Chain Monte Carlo
EQU [17] (MCMC) method, which differs from the Gradient-Descent-Based Opti­
mization (GDBO) used by LSTM models. The MCMC method takes
St = SoftMax (Ws . ht + bs ) (17) samples from the posterior distribution of the model’s parts. This lets
The LSTM model is trained using historical data to minimize the you use a probabilistic model that takes into account the fact that pre­
difference between TCP scores ‘St’ and actual levels. It adapts to dictions about congestion aren’t always accurate. The RMSE, denoted
incoming data for accurate TCP. The model is fine-tuned using a mathematically as EQU [22], gauges the model’s performance during
comprehensive historical dataset, including high-frequency GPS data, to this probabilistic training.
minimize the RMSE during training, EQU [18]. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√
√1 ∑ N
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ RMSEBSTS = √ (St − ̂ S t )2 (22)
√1 ∑ N N t=1
RMSElstm = √ (St − ̂ S t )2 (18)
N t=1
Here, St is the observed TC score, and ̂ S t is the BSTS predicted
congestion score for each time slice t. The RMSE in this context is
where St is the observed TC score and ̂ S t is the LSTM model’s TFP score especially significant because it captures the dispersion of the sampling
for each time step t. RMSE is a stringent metric that imposes higher errors in the Bayesian framework.
penalties for significant errors, ensuring that the LSTM model accurately Due to the high temporal resolution of the GPS data, the BSTS model
captures the intricacies of TFP dynamics in Changchun’s segmented is configured to adapt iteratively to fluctuating traffic conditions. This

96
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

adaptability differs from LSTM models as it relies on continuous Output:

Bayesian updating. It is exceptionally well-suited for generating real-
time and probabilistically nuanced TCP in the segmented zones. • Hybrid congestion score St,Hybrid for each DSZ at each time step ′t′
Adding LSTM + BSTS to a model can improve time series analysis by
detecting long-lasting dependencies between sequential data. LSTM, a Steps:
RNN, is ideal for tasks like time series prediction, while BSTS offers a
statistical model that breaks down time series into components like 1 Data Preprocessing:
trend and seasonality. This combination can enhance the model’s ability
to capture intricate temporal patterns and integrate probabilistic • Segment the area into grids using the DZS algorithm.
modelling features from the Bayesian model. The DL capabilities of • Generate feature vectors Ft = [Nt ,St ,Tt ,lont ,latt ], for each grid at each
LSTM + BSTS models make this approach applicable in domains like time step t.
financial forecasting and demand prediction. However, the execution
details and reasons for merging these methods may vary depending on 2 Train LSTM Model:
the specific problem area and desired attributes of the model.
• Initialize the LSTM model with parameters Wf ,Wi ,Wo ,WC ,bf ,bi ,bo ,bC
F. Hybrid LSTM + BSTS Model for TCP • Train the LSTM model on historical data to minimize RMSELSTM
• Validate the model and fine-tune the parameters.
After independent training and validation, the LSTM + BSTS models
generate their respective congestion scores, denoted as St,LSTM and 3 Train BSTS Model:
St,BSTS . These scores serve as individual predictions of congestion levels
at each dynamically segmented zone at a given time ’t’. The final step • Initialize the BSTS model with prior distributions for μ1 , γ 1 , β, σ2 .
synthesises these predictions into a single, unified TC score. This hybrid • Use MCMC to sample from the posterior distribution and train the
model aims to capture the strengths of the LSTM + BSTS models. The BSTS model.
hybrid model of the LSTM + BSTS is presented in Fig. 2. • Train the BSTS model on historical data to minimize RMSEBSTS .
The hybrid congestion score St,Hybrid is calculated as follows. • Validate the model and fine-tune the parameters.

wLSTM ⋅St,LSTM + wBSTS ⋅St,BSTS

St,Hybrid = (23) 4 Calculate LSTM and BSTS Weights:
2
In this EQU [23], wLSTM and wBSTS represent optimized weights for • Use a validation set to optimize weights wLSTM , wBSTS by minimizing
the LSTM + BSTS models, respectively. These weights are fine-tuned to RMSEHbrid .
minimize the overall RMSE on a validation set, ensuring that the hybrid
model offers the most accurate real-time congestion predictions. The 5 Real-Time Prediction:
optimization objective for these weights is to minimize the hybrid
RMSE, defined as EQU [24]. • For each new time step :
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ • Generate the feature vector Ft for the grid.
√1 ∑ Use the LSTM model to predict St,LSTM.
N
( )2 •
RMSEHybrid = √ St − St,Hybrid (24)
N t=1 • Use the BSTS model to predict St,BSTS.
• Calculate the hybrid congestion score St,Hybrid using:
The optimization process ensures that St,Hybrid is as close as possible to
wLSTM ⋅St,LSTM + wBSTS ⋅St,BSTS
the observed congestion levels St , benefiting from the predictive power St,Hybrid =
2
of the LSTM + BSTS models. Algorithm 2 presents the process for the
proposed hybrid LSTM + BSTS.
6 Output
Algorithm 2. for Hybrid LSTM + BSTS Algorithm
Input:
• Return St,Hybrid as the final congestion score for each grid at each
time step ’t’.
• Raw GPS data containing
• Historical congestion scores for validation
Table 3 assesses the efficacy of different TP models, such as ARIMA,
• N : Number of time steps for the LSTM sequence
SVR, BSTS, FNN, LSTM, and the hybrid LSTM + BSTS, across three-time
• M : Number of MCMC samples for the BSTS model
intervals: 15 minutes, 30 minutes, and 45 minutes. Examining Mean
Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean
Absolute Percentage Error (MAPE) offers a systematic evaluation of the
accuracy of predictions. ARIMA has high errors, indicating difficulties in
accurately capturing complex traffic patterns. The SVR model improves
performance, while BSTS + FFNN surpasses conventional models. LSTM
enhances accuracy, and the hybrid LSTM + BSTS model consistently
outperforms other models, generating minimal errors across all metrics
and time intervals. Combining LSTM’s Deep Learning (DL) capabilities
with BSTS’s shows significant effectiveness, demonstrating the hybrid
model’s ability to provide accurate and reliable TP, making it a prom­
ising alternative for TMS applications.

4. Experimental analysis

The Linux-based system used for the experiments featured an AMD

Fig. 2. LSTM + BSTS architecture. Ryzen 9 3950X CPU and an NVIDIA Tesla V100 GPU. The software stack

97
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

consisted of Python and TensorFlow 2.5. This study collected traffic data MAPE gives an understanding of errors in terms of percentage,
for one year, from May 1, 2022, to April 30, 2023, resulting in a dataset making it easier to interpret the relative size of the errors compared to
of 242,034 records. This paper partitioned this dataset, utilizing 80% for the actual TC scores.
training the model and setting aside the remaining 20% for testing and
validation. Table 1 shows the parameters used to train the proposed
model. The analysis was conducted for three different time intervals of 4.5. Performance comparison
the dataset: (i) 15 minutes, (ii) 30 minutes, and (iii) 45 minutes. To
assess the effectiveness of the hybrid LSTM-BSTS model for TCP, its To evaluate the effectiveness of the proposed hybrid LSTM + BSTS
performance is evaluated using the three key metrics discussed below: model in TCP, it is compared against a set of baseline models such as:

4.1. Definitions 1. Auto-Regressive Integrated Moving Average (ARIMA)

2. Support Vector Regression (SVR)
St,predicted The predicted congestion score at time t. 3. Feed Forward Neural Network (FNN)
St,actual The actual observed congestion score at time t. 4. Long Short Term Memory (LSTM)
5. Bayesian Structural Time Series (BSTS)
4.2. MAE
The performance analysis results, displayed in Table 2 and Fig. 3 for
The MAE is computed as EQU [25]. the 15-min timeframe of the dataset, reveal that the traditional ARIMA
model exhibits the least effective performance, with an MAE of 37.69, an
1 ∑ RMSE of 46.95, and a MAPE of 29.51%. This indicates a relatively high
N ⃒ ⃒
MAE = ⃒St,predicted − St,actual ⃒ (25)
N t=1 level of error and inaccuracy in its predictions. The SVR model shows
some improvement over ARIMA, but its performance metrics still leave
MAE calculates the average size of the discrepancies between fore­ room for improvement, particularly with an MAE of 30.97 and RMSE of
casted and accurate TC scores without considering their direction. This 42.69.
offers a simple and transparent method for evaluating the model’s pre­ The FFNN + BSTS models perform considerably better, indicating
cision, EQU [26]. their suitability for capturing intricate patterns in high-frequency traffic
data. However, the standout performers are the LSTM and the proposed
4.3. RMSE hybrid LSTM + BSTS models. The LSTM model demonstrates high pre­
diction with an MAE of 10.03, an RMSE of 16.88, and a MAPE of 9.99%.
The RMSE is defined as EQU [25]. Yet, the hybrid LSTM + BSTS model outshines all, achieving the lowest
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ error rates across all three metrics— an MAE of 8.48, an RMSE of 14.13,
√
√1 ∑ N
( )2
RMSE = √ St,predicted − St,actual (26) and a MAPE of just 7.48%.
N t=1 Fig. 4 shows how well the different models could predict the 30-min
dataset. This shows how powerful ML methods are, especially the hybrid
RMSE gives more weight to more significant errors by squaring them LSTM + BSTS model. The traditional ARIMA model shows the highest
before averaging. This makes it worthwhile when significant errors are error rates, with an MAE of 37.84, an RMSE of 47.16, and a MAPE of
particularly undesirable. 29.79%. These figures indicate that the model struggles to capture the
temporal patterns in the traffic data accurately for a 30-min interval.
4.4. MAPE The SVR model fares slightly better but exhibits considerable error, as
indicated by its MAE of 30.48 and RMSE of 42.42.
MAPE is calculated using the EQU (27) 17.66% MAE, 24.08% RMSE, and 13.75 MAPE are all within
N ⃒
100 ∑
⃒
⃒St,predicted − St,actual ⃒ acceptable ranges for the BSTS model. The MAE, RMSE, and MAPE for
MAPE = ⃒
⃒
⃒
⃒ (27) the FNN model are all similar, coming in at 15.52, 23.06, and 11.30%,
N t=1 St,actual
respectively. The true highlight is the contrast between the LSTM and
hybrid LSTM + BSTS models. With an MAE of 10.33, an RMSE of 17.02,
Table 1 and a MAPE of 10.14%, the LSTM model results in excellent results. On
Comprehensive analysis. the other hand, the hybrid LSTM + BSTS model outperforms it with an
Reference Drawbacks Key Insights MAE of 9.68, RMSE of 14.78, and MAPE of 8.09%.
[17] Lack of standardized Need for innovative and
The analysis of the 45-min dataset shown in Fig. 5 provides further
approaches in traffic standardized TMS.
technology. Table 2
[18] Challenges in Intelligent Proposes DL-based cloud control for Hyperparameter for LSTM + BSTS.
Traffic Control (ITC) TF and congestion
[19] Fuzzy model for traffic control Hybrid ConvLSTM model for Model Hyperparameter Values
improved travel and OD predictions
LSTM Sequence Length (NN) 50
[20] TC Optimized short-term TFP model
Batch Size 64
using SVR
Learning Rate 0.001
[21] TC Hybrid TSA-SL approach for Short-
Epochs 100
Term TFP on ERI dataset
LSTM Units 128
[22] ITC Spatial-temporal graph neural
Dropout Rate 0.2
network for TF patterns
Activation Function Sigmoid for gates, Tanh for cell states
[23] RPCT for TFP Significantly surpasses baseline
Loss Function RMSE
models in TCP
[24] Neuro-fuzzy model for TC Remarkable accuracy in reducing BSTS MCMC Samples (M) 1000
vehicular delays at intersections Seasonal Periods 7 for daily, 365 for yearly
[25] MDCGCN for medium and Outperforms existing methods in Prior for μ1 Normal(0, 1)
long-term TFP prediction accuracy Prior for γ1 Normal(0, 1)
[16] DSGCN for more accurate TCP Surpasses baseline models in Prior for β Normal(0, 1)
predictive accuracy Prior for σ2 Inverse-Gamma(1, 1)

98
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

Table 3
Performance comparison for MAE, RMSE and MAPE.
Model 15 min 30 min 45 min

MAE RMSE MAPE MAE RMSE MAPE MAE RMSE MAPE

ARIMA 37.70 46.95 29.51% 37.84 47.17 29.79% 38.11 47.33 29.93%
SVR 30.97 42.70 22.85% 30.48 42.42 22.60% 29.73 41.92 22.33%
BSTS 17.22 23.67 13.54% 17.66 24.09 13.75% 17.88 24.25 14.10%
FNN 14.90 22.92 11.10% 15.52 23.06 11.30% 16.35 23.50 11.44%
LSTM 10.04 16.89 9.99% 10.34 17.03 10.14% 11.12 17.21 10.50%
LSTM-BSTS 8.48 14.14 7.48% 9.68 14.78 8.09% 6.61 10.52 6.68%

the complexities of TC over time. The MAE of the SVR model is 29.73,
and the RMSE is 41.14, indicating that the model still has elevated error
rates.
The MAE of the BSTS model is 17.87, the RMSE is 24.25, and the
MAPE is 14.10%. Similarly, the FFNN model remains steady, boasting an
MAE of 16.34, an RMSE of 23.5, and a MAPE of 11.44%. Nevertheless,
the proposed LSTM + BSTS model continues to outperform the rest. The
LSTM model has an MAE of 11.11, an RMSE of 17.21, and a MAPE of
10.50%. The hybrid LSTM + BSTS model, however, outperforms all
models substantially with an MAE of just 6.61, an RMSE of 10.52, and a
MAPE of 6.68%. When evaluated across different periods, the perfor­
mance boost of the proposed model becomes particularly prominent for
the more extended 45-min prediction window. This period generally
Fig. 3. Performance for 15 min.
adds more layers of complexity and unpredictability to TF. Despite these
inherent challenges, the hybrid model experiences a significant reduc­
tion in error rates, making it an exceptionally sturdy and dependable
instrument for long-term TCP. This highlights the model’s capability to
adapt and generalize well over different time horizons, solidifying its
status as the most effective model for TCP in DZS.

5. Conclusion and future work

This study blocks the persistent and intensifying problem of urban

Traffic Congestion (TC) by employing a novel method that merges
advanced Machine Learning (ML) techniques with conventional Traffic
Management Systems (TMS). The Dynamic Zone Segmentation (DZS)
algorithm effectively partitions complex urban environments into
manageable segments, allowing for a more nuanced comprehension of
real-world traffic patterns. Real-time Traffic Prediction (TP) is reliable
with the hybrid Long Short-Term Memory + Bayesian Structural Time
Series (LSTM + BSTS) learning model, which integrates Machine
Fig. 4. Performance for 30 minutes
Learning (ML) and empirical techniques. It improves baseline models in
Mean Absolute Error, Mean Absolute Scaled Error, Accuracy Percent,
Root Mean Squared Error, and Mean Absolute Percent Error, promising
applications for other industries which require real-time, versatile
decision-making processes. TMS can implement the DZS Algorithm and
hybrid LSTM + BSTS learning model to enhance road implementation,
travel time, and people’s quality of life while minimizing congestion in
urban areas.
Future endeavours will explore further enhancements to the model,
its adaptability to other domains, and the possibility of integration with
Internet of Things (IoT) devices for even more precise, real-time data
collection and analysis.

Data availability

Not Applicable

Funding
Fig. 5. Performance for 45 minutes

NA
evidence of the success of the hybrid LSTM + BSTS model in TCP. As
evidenced by its MAE of 38.11, RMSE of 47.32 and MAPE of 29.93%,
ARIMA consistently exhibits the worst error rates among traditional
models. These results indicate that the model is ineffective at capturing

99
L. Cheng and K. Sun International Journal of Intelligent Networks 5 (2024) 92–100

CRediT authorship contribution statement [11] M. Gollapalli, D. Musleh, N. Ibrahim, M.A. Khan, S. Abbas, A. Atta, A. Omer,
A neuro-fuzzy approach to road traffic congestion prediction, Comput. Mater.
Continua (CMC) 73 (1) (2022).
Lihua Cheng: Writing – original draft, Methodology, Formal anal­ [12] L. Chen, L. Zheng, J. Yang, D. Xia, W. Liu, Short-term traffic flow prediction: from
ysis, Conceptualization. Ke Sun: Writing – review & editing, Supervi­ the perspective of traffic flow decomposition, Neurocomputing 413 (2020)
sion, Project administration, Investigation, Formal analysis. 444–456.
[13] C. Li, P. Xu, Application on traffic flow prediction of machine learning in intelligent
transportation, Neural Comput. Appl. 33 (2021) 613–624.
[14] Y. Qi, Z. Cheng, Research on Traffic Congestion Forecast Based on Deep
Declaration of competing interest Learning—information 14 (2) (2023) 108.
[15] Z. Duan, K. Zhang, Z. Chen, Z. Liu, L. Tang, Y. Yang, Y. Ni, Prediction of city-scale
The authors declare that they have no known competing financial dynamic taxi origin-destination flows using a hybrid deep neural network
combined with travel time, IEEE Access 7 (2019) 127816–127832.
interests or personal relationships that could have appeared to influence [16] M. Shengdong, X. Zhengxian, T. Yixiang, Intelligent traffic control system based on
the work reported in this paper. cloud computing and big data mining, IEEE Trans. Ind. Inf. 15 (12) (2019)
6583–6592.
[17] N.O. Alsrehin, A.F. Klaib, A. Magableh, Intelligent transportation and control
Acknowledgement systems using data mining and machine learning techniques: a comprehensive
study, IEEE Access 7 (2019) 49830–49857.
NA [18] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (8)
(1997) 1735–1780.
[19] M. Tong, H. Duan, X. Luo, Research on short-term traffic flow prediction based on
References the tensor decomposition algorithm, J. Intell. Fuzzy Syst. 40 (3) (2021)
5731–5741.
[20] Uma Maheswari Arumugam, Suganthi Perumal, Trust-based secure and reliable
[1] T. Afrin, N. Yodo, A Survey of road traffic congestion measures towards a
routing protocol of military communication on MANETs, Journal of Machine and
sustainable and resilient transportation system, Sustainability 12 (11) (2020) 4660.
Computing (January 2023) 47–57, https://doi.org/10.53759/7669/
[2] Y. Zhang, K. Shang, Z. Cui, Z. Zhang, F. Zhang, Research on traffic flow prediction
jmc202303006.
at intersections based on DT-TCN-attention, Sensors 26 (2023) 6683, 23 (15).
[21] Madeleine Wang, Yue Dong, A survey on multi-agent system and its applications in
[3] S.A. Kumar, R. Madhumathi, P.R. Chelliah, L. Tao, S. Wang, A novel digital twin-
power system engineering, Journal of Computational Intelligence in Materials
centric approach for driver intention prediction and traffic congestion avoidance,
Science 1 (2023) 1–11, https://doi.org/10.53759/832X/JCIMS202301001.
Journal of Reliable Intelligent Environments 4 (2018) 199–209.
[22] Fauziya Njeru, A review of artificial intelligence and its application in business,
[4] S.A. Sayed, Y. Abdel-Hamid, H.A. Hefny, Artificial intelligence-based traffic flow
Journal of Enterprise and Business Intelligence 3 (1) (January 2023) 44–53,
prediction: a comprehensive review, Journal of Electrical Systems and Inf
https://doi.org/10.53759/5181/JEBI202303005.
Technology 10 (13) (2023).
[23] W. Du, S. Chen, Z. Li, X. Cao, Y. Lv, A spatial-temporal approach for multi-airport
[5] J. Cheng, G. Li, X. Chen, Research on travel time prediction model of freeway based
traffic flow prediction through causality graphs, in: IEEE Transactions on
on gradient boosting decision tree, IEEE Access 7 (2018) 7466–7480.
Intelligent Transportation Systems, vol. 25, Jan. 2024, pp. 532–544, https://doi.
[6] T. Ahmad, H. Chen, A review on machine learning forecasting growth trends and
org/10.1109/TITS.2023.3308903, 1.
their real-time applications in different energy systems, Sustain. Cities Soc. 54
[24] X. Ouyang, Y. Yang, W. Zhou, Y. Zhang, H. Wang, W. Huang, CityTrans: domain-
(2020) 102010.
adversarial training with knowledge transfer for spatio-temporal prediction across
[7] W. Duan, H. Rao, W. Huang, X. He, Minimalist Traffic Prediction: Linear Layer Is
cities, in: IEEE Transactions on Knowledge and Data Engineering, vol. 36, Jan.
All You Need, 2023 arXiv preprint arXiv:2308.10276.
2024, pp. 62–76, https://doi.org/10.1109/TKDE.2023.3283520, 1.
[8] N. Van Thanh, T.T.M. Linh, Real-time trajectory planning for autonomous vehicles
[25] Z. Yang, D. Li, W. Nai, L. Liu, J. Sun, X. Lv, Compressible non-Newtonian fluid
in dynamic traffic environments: a survey of modern algorithms and predictive
based road traffic flow equation solved by physical-informed rational neural
techniques, Journal of Intelligent Connectivity and Emerging Technologies 7 (12)
network, in: IEEE Access, vol. 12, 2024, pp. 12992–13009, https://doi.org/
(2022) 1–25.
10.1109/ACCESS.2024.3356173.
[9] W. Li, X. Wang, Y. Zhang, Q. Wu, Traffic flow prediction over multi-sensor data
[26] B. Liu, C.-T. Lam, B.K. Ng, X. Yuan, S.K. Im, A graph-based framework for traffic
correlation with graph convolution network, Neurocomputing 427 (2021) 50–63.
forecasting and congestion detection using online images from multiple cameras,
[10] M. Bai, Y. Lin, M. Ma, P. Wang, L. Duan, PrePCT: traffic congestion prediction in
in: IEEE Access, vol. 12, 2024, pp. 3756–3767, https://doi.org/10.1109/
smart cities with relative position congestion tensor, Neurocomputing (444) (2021)
ACCESS.2023.3349034.
147–157.

100