This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Artificial Hummingbird Optimization Algorithm

with Hierarchical Deep Learning for Traffic
Management in Intelligent Transportation
Systems
Abdulrahman Alruban1,*, Hanan Abdullah Mengash 2, Majdy M. Eltahir3, Nabil Sharaf
Almalki4, Ahmed Mahmud5 and Mohammed Assiri6
1
Department of Information Technology, College of Computer and Information Sciences, Majmaah University, Al Majmaah, 11952, Saudi Arabia.
a.alruban@mu.edu.sa
2
Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box
84428, Riyadh 11671, Saudi Arabia. hamengash@pnu.edu.sa
3
Department of Information Systems, College of Science & Art at Mahayil, King Khalid University, Saudi Arabia. meltahir@kku.edu.sa
4
Department of Special Education, College of Education, King Saud University, Riyadh 12372, Saudi Arabia. nalmalki@ksu.edu.sa
5
Research Center, Future University in Egypt, New Cairo 11835, Egypt. mahmoud.a@fueegypt.education
6
Department of Computer Science, College of Sciences and Humanities- Aflaj, Prince Sattam bin Abdulaziz University, Aflaj 16273, Saudi Arabia.
m.assiri@psau.edu.sa

Corresponding author: Abdulrahman Alruban (e-mail: a.alruban@mu.edu.sa).

“The authors extend their appreciation to the Deanship of Scientific Research at King Khalid University for funding this work through large group Research
Project under grant number (RGP2/ 158 /44).”

ABSTRACT Intelligent Transportation Systems (ITS) make use of advanced technologies to optimize
interurban and urban traffic, reduce congestion and enhance overall traffic flow. Deep learning (DL)
approaches can be widely used for traffic flow monitoring in the ITS. This manuscript introduces the Artificial
Hummingbird Optimization Algorithm with Hierarchical Deep Learning for Traffic Management (AHOA-
HDLTM) technique in the ITS environment. The purpose of the AHOA-HDLTM technique is to predict
traffic flow levels in smart cities, enabling effective traffic management. Primarily, the AHOA-HDLTM
model involves data preprocessing and an Improved Salp Swarm Algorithm (ISSA) for feature selection. For
the prediction of traffic flow, the Hierarchical Extreme Learning Machine (HELM) model can be used. The
HELM extracts complex features and patterns, with an additional Artificial Hummingbird Optimization
Algorithm (AHOA)-based hyperparameter selection process to enhance predictive outcomes. The simulation
result analysis under different traffic data demonstrates the better performance of the AHOA-HDLTM
technique over existing models. The hierarchical structure of the HELM model along with AHOA-based
hyperparameter tuning helps to accomplish enhanced prediction performance. The AHOA-HDLTM
technique presents a robust solution for traffic management in ITS, showcasing enhanced performance in
forecasting traffic patterns and congestion. The AHOA-HDLTM technique can be used in various smart cities
and urban regions. Its abilities in real-time traffic flow prediction can be helpful in the design of efficient,
sustainable, and resilient transportation networks.

INDEX TERMS Smart cities; Intelligent transportation system; Deep learning; Traffic management;
Feature selection

I. INTRODUCTION challenge, intelligent transportation systems (ITS) have been

In urban areas, high population growth leads to a growth in developed for the solution. Integrating advanced
carbon emissions, air pollution, and traffic congestion which methodologies like the Internet of Things (IoT), artificial
impact the environment negatively but also delay economic intelligence (AI), and machine learning (ML) can improve
growth and the life quality of people [1]. To overcome this

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

transportation security, sustainability, and efficiency [2]. ITS II. LITERATURE REVIEW
is the main part of smart cities that can aid in improving the Abdullah et al. [11] designed a bi-directional RNN (BRNN)
employment of transportation infrastructure, enhance mobility employing GRU. The method employs a BRNN to fake and
for citizens, and decrease traffic congestion. Traffic forecasts predict traffic congestion in the smart areas. This study
offer experts a period to propose resource provisions to recommends a data-driven model by utilizing BRNN for
safeguard a safe journey [3]. By congestion, the worth of street traffic organization that employs real-time information from
networks is restricted. The decreases outcome in direct as well devices and connected mechanisms to control traffic
as indirect charges for the public. The congestion effects on efficiently. In [12], an enhanced solution is designed, which
the social structure and economic system have been integrates a novel target tracing and moving vehicle counting
researched widely [4]. Prediction of traffic flow is regarded as technique and a better LSTM system for the forecast of traffic
one of the time-series-based issues because the future cost of flow. Specifically, the MultiNetV3 framework and DCN V2
traffic flow has been assessed based on the previous data from convolution kernel are mainly employed to exchange
one or more locations. The volume of data develops the big YOLOv4’s traditional convolutional kernel and backbone
data idea in the transportation field from a wide range of network to recognize multiple targets tracking and moving
sources. Depending upon the sort of data gathered, dissimilar separately.
artificial intelligence (AI) techniques are utilized to value the In [13], an effectual approach for OD (Origin–Destination)
overload parameters [5]. There are many methods for matrix forecast depends on traffic info employing DL. The
analyzing and storing so big data have a risky effect on the model using DNN with the LSTM or autoencoder layers is
prediction task. The development of big data computing categorized by comparatively high resistance and accuracy to
amenities provides chances for attaining exactness in traffic miss data from temporary measurement points that are located
data prediction [6]. in the urban road network. In [14], a new DL design called
In recent times, AI and MI field has stemmed from the Ensemble Attention-based Graph Time Convolutional
capability to expect traffic congestion. Recently, this study Networks (EAGTCN) is designed. At the initial stage, by
area has extended owing to the arrival of enormous spatial blocks, the global spatial pattern is captured which are
information from static sensor networks [7]. Numerous traffic attached by a three-dimensional ensemble attention layer as
factors have been estimated to anticipate traffic congestion, well as a Graph Convolution Network (GCN). In [15], a
especially short-term congestion issues. Most researchers on temporal threshold mechanism and soft spatial is designed. To
expecting traffic congestion use previous data. Whereas, a few fill out the absent weather info spatial interpolation methods
researchers have expected congestion issues in the real period were utilized. Hybrid DL techniques such as CNN-LSTM and
[8]. Congestion forecasting is one of the most challenging LSTM are also proposed. The hybrid method eliminates the
issues to resolve from the view of modifiability when features of spatiotemporal and later employs these features as
compared to traffic flow forecast in non-congested situations. memory. This method forecasts the traffic stream variances
However, the Traffic supervisors implemented relief actions based on past features as well as time-based input. Qi et al.
to the alarm system. The shallow design technique has been [16] presented a DL model depending on a spatiotemporal
utilized for the prediction task in most cases. This framework GCN. This model consists of four processes. First, creating a
offers effective solutions for the small size of data only but biased adjacency matrix utilizing the Gaussian similarity
does not support big traffic data [9]. In recent days, a Deep function. Second, accumulating a feature matrix. Lastly
learning (DL) framework such as deep belief network (DBN), creating a spatiotemporal GCN depend on a DL design (i.e. T-
DNN, recurrent neural networks (RNN), and CNN utilized in GCN).
several difficult applications such as video analysis, large data Yang et al. [17] present a traffic flow dependency and
set for image, natural language procedure and also in several dynamics-based DL-aided method (TD2-DL). Exactly, the
data mining procedure [10]. graph theory-based model is proposed to classify the local
This manuscript presents an Artificial Hummingbird temporal-spatial traffic dependency. Next, the LSTM model is
Optimization Algorithm with Hierarchical Deep Learning for used. Last, an EKF is applied for integrating the expected
Traffic Management (AHOA-HDLTM) technique in the ITS traffic speed that is projected by v-CTM joined with the
environment. The AHOA-HDLTM technique initially LSTM and the field traffic information. So, the FNN technique
performs data preprocessing and improved salp swarm has been mainly developed. In [18], a PSO-BiLSTM method
algorithm (ISSA) based feature selection. For traffic flow depends on the integration of BiLSTM neural network and
monitoring, the AHOA-HDLTM method applies a PSO is proposed. The PSO model searches for the optimal
hierarchical ELM (HELM) model. At last, the predictive result parameters on a total scale that has been utilized and then non-
of the HELM approach can be upgraded by the use of an linear adjustable inertial weights are taken rather than linear
AHOA-based hyperparameter selection process. The weights. In addition to that, the BiLSTM prediction technique
enhanced performance of the AHOA-HDLTM method can be of the network is enhanced by employing the PSO method.
assured using a detailed simulation analysis of diverse traffic
data.

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

In [19], a traffic flow recognition technique that depends on

DL on the edge node has been presented. Initially, the authors
present a vehicle recognition technique that depends on the
YOLOv3 method trained with a great volume of traffic data.
Yu et al. [20] present a DL-based traffic safety performance
for a mixture of autonomous and manual vehicles from a 5G-
enabled ITS. In [21], cloud-assisted IoT-ITS (CIoT-ITS) was
presented to overcome traffic management problem. At this
point, the IoT sensor combined camera has been installed in
the entire traffic signal corner for monitoring the vehicle’s
flow. Chen et al. [22] examine energy-efficient and
regenerative energy recovery systems for sustainable ITS
utilizing Artificial societies, Computational experiments, and
Parallel execution (ACP) structure.

III. THE PROPOSED MODEL

In this manuscript, a novel AHOA-HDLTM method is
established in the ITS environment. The main purpose of the
AHOA-HDLTM system is to predict the level of traffic flow
in the ITS, enabling proper traffic management in smart cities.
To accomplish this, the AHOA-HDLTM technique
encompasses data preprocessing, ISSA-based FS, HELM-
based classification, and AHOA-based parameter tuning. Fig.
1 shows the workflow of the AHOA-HDLTM technique.

A. DATA PREPROCESSING
Min-max normalized called feature scaling, is a generally
employed method in traffic flow monitoring and data pre-
processing. It changes and scales the range of numerical data
to a certain range, usually between zero and one, making it FIGURE 1. Workflow of AHOA-HDLTM algorithm
appropriate for many ML and data analysis tasks. In the
context of traffic flow monitoring, min-max normalized can 𝐹𝑗 + 𝑐1 ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 )𝑐2 + 𝑙𝑏𝑗 ) , 𝑐3 ≥ 𝑟
be executed for numerical features like traffic speed, volume, 𝑥𝑗1 = { (1)
or occupancy. 𝐹𝑗 + 𝑐1 ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 )𝑐2 + 𝑙𝑏𝑗 ) , 𝑐3 < 𝑟
Where the up and low boundaries of the searching space are
B. FEATURE SELECTION USING ISSA 𝑢𝑏j 𝑎𝑛𝑑 𝑙𝑏j , the uniformly distributed random integer within
The ISSA is used for electing an optimum subset of [0,1] are r, 𝑐2 and 𝑐3 parameters that are embedded in the SSA.
features. SSA is a recent bio-inspired technique based on the The 𝑐1 parameter is then evaluated by Eq. (2):
2
foraging and navigation of salp in the ocean [23]. A leader and 𝑐1 = 2𝑒 −(4𝑡/𝑇) (2)
followers are the two groups performed in SSA. The solution Where the existing and the maximal iterations that should
candidate is based on the salp location while foraging and be initially set are 𝑡 and 𝑇. The follower location of salps is
navigation, usually for optimization problems is treated as a evaluated by Eq. (3):
2D matrix known as 𝑥. The present optimum solution is 1
known as a food source, 𝐹𝑗 , and the leader location of salp, 𝑥𝑗1 𝑥𝑗𝑖 = (𝑥𝑗𝑖 + 𝑥𝑗𝑖−1 )𝑖 ≥ 2 (3)
2
is updated according to Eq. (1): The salp follower will be completely directed by the leader
in the original SSA to discover the optimum solution
according to the region of possible solution is evaluated
through the change among the upper and lower limitations of
the control variable. Once the problem dimension is
considered smaller, then it performs well. Nonetheless, if the
search for a possible solution is large, then the exploitation
process becomes ineffective. In resolving the optimization
problem, the balance between exploitation and exploration
needs to be decided wisely, note that the exploration capability
of SSA is quite restricted. In this work, two enhancements will

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

be suggested for enhancing the exploitation-exploration

abilities. The improved version is performed in Eq. (1), but the
low boundary has been removed to enhance the process’s
capability to use resources that are formulated by Eq. (4):
𝐹𝑗 + 𝑐1 ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 )𝑐2 ) , 𝑐3 ≥ 𝑟
𝑥𝑗1 = { (4)
𝐹𝑗 + 𝑐1 ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 )𝑐2 ) , 𝑐3 < 𝑟
The second enhancement is executed to enhance the
exploration ability, where follower expression in Eq. (3) is
transformed into:
𝑥𝑗𝑖 = (𝑝𝑥𝑗𝑖 + 𝑞𝑥𝑗𝑖−1 )𝑖 ≥ 2 (5)
In Eq. (5), 𝑝 is a random integer [0,1] and 𝑞 is evaluated as
(1‐ 𝑝). FIGURE 2. ELM structure

The fitness function (FF) employed in the ISSA To trained data 𝑋 = [𝑥1 , 𝑥2 , … , 𝑥𝑁 ]𝑇 ∈ ℝ𝑁×𝑑 that contains
methodology has been presented to take a balance amongst the 𝑁 instances with all the samples are a 𝑑‐dimensional vector,
count of preferred features from the entire outcome (lesser) 𝑇 = (𝑡1 , 𝑡2 , … 𝑡𝑁 )𝑇 ∈ ℝ𝑁×1 refers to the equivalent target
and classification outcome (higher) acquired through applying matrix. Simple ELM purposes for obtaining the resultant
chosen feature, Eq. (6) refers to the FF for measuring weight by decreasing the amount of predictive error and the
outcomes. norm of resultant weights:
|𝑅| min‖𝐻𝛽 − 𝑇‖2 + 𝜆‖𝛽‖2 (10)
𝐹𝑖𝑡𝑛𝑒𝑠𝑠 = 𝛼𝛾𝑅 (𝐷) + 𝛽 (6) whereas 𝐻 = [ℎ1 , ℎ2 , … , ℎ𝑁 ]𝑇 ∈ ℝ𝑁×𝐿 inferred the
|𝐶|
In Eq. (6), 𝛾𝑅 (𝐷) stands for the classifier error rate of the arbitrary latent representation matrix of the input 𝑋. It can be
offered classifier. |𝑅| implies the cardinality of the chosen attained by Eq. (8) or (9) with 𝑎𝑗 parameter, 𝑏𝑗 arbitrarily
subset and |𝐶| denotes complete feature counts in the created and ℎ𝑖 = (𝐺1 (𝑥𝑖 ), 𝐺2 (𝑥𝑖 ), … 𝐺𝐿 (𝑥𝑖 ))𝑇 , 𝑖 = 1, 2, … 𝑁.
database, parameters 𝛼 and 𝛽 are equivalent to the impact of 𝜆 implies regularized hyperparameter.
classification algorithm quality and subset length. 𝛽 ∗ = (𝐻𝑇 𝐻 + 𝜆𝐼)−1 𝐻𝑇 𝑇 (11)
Once the hidden neuron counts 𝐿 is lesser than the count of
C. HELM BASED PREDICTION trained data 𝑁. When the hidden neuron counts are superior to
To forecast the traffic flow, the HELM approach is utilized. the amount of trained data, a better performance is:
ELM is initially presented to train SLFNs [24]. A prominent 𝛽 ∗ = 𝐻𝑇 (𝐻𝐻 𝑇 + 𝜆𝐼)−1 𝑇 (12)
feature of ELM is that the parameters among the hidden and H‐ELM is acquired in 2 phases. Primarily, similar to typical
input layers (HL) can be arbitrarily created and the only free DL approaches, 𝐾 unsupervised ELM-SAEs are fixed for
parameters, which require that optimizer are the resultant learning 𝐾 layers of latent representation 𝑅𝑛𝑑 the learned
weights among the HL and the resultant state. Fig. 2 illustrates representation (𝑅) of the preceding element has been utilized
the infrastructure of ELM. To provide a training set as input of the next element.
{(𝑥𝑖 , 𝑡𝑖 )|𝑥𝑖 ∈ ℝ𝑑 , 𝑡𝑖 ∈ ℝ, 𝑖 = 1, … , 𝑁}, whereas 𝑥𝑖 implies 𝑅𝑖 = 𝑠(𝑅𝑖−1 (𝛽𝑖∗ )𝑇 ) (13)
the trained data and 𝑡𝑖 denotes the target. SLFNs network In which, 𝑅𝑖 denotes the learned representation of 𝑖 𝑡ℎ HL
function with 𝐿 hidden node (HN) is defined as: (the significant representation extracted by 𝑖 𝑡ℎ ELM sparse
𝐿 AE), 𝑅𝑖−1 implies the learned representation of 𝑖-𝑙 th HL and
𝑓𝐿 (𝑥) = ∑ 𝐺𝑗 (𝑥) ∗ 𝛽𝑗 , 𝛽𝑗 ∈ ℝ (7) it is additionally utilized as the input of 𝑖-1 component. 𝛽𝑖∗
𝑗=1
denotes the better-recreated matrix of 𝑖 𝑡ℎ ELM sparse AE that
In which, 𝛽𝑗 implies the weighted linking of the 𝑗𝑡ℎ HN to obtains 𝑅𝑖−1 as input, and 𝑠 denotes the activation function of
the output node and 𝐺𝑗 represents the 𝑗𝑡ℎ HN resultant HLs. When the feature in the preceding HL has been extracted,
function. To additive nodes with activation function 𝑔, 𝐺𝑗 is the weighted matrix of existing HL is set and requires not to
expressed as: be adjusted.
𝐺𝑗 (𝑥) = 𝑔(𝑎𝑗 ⋅ 𝑥 + 𝑏𝑗 ), 𝑎𝑗 ∈ ℝ𝑑 , bj ∈ ℝ (8)
D. AHOA-BASED PARAMETER TUNING
whereas 𝑎𝑗 refers to the weighted vector linking the input
Finally, the AHOA approach can be executed for the
layer to 𝑗𝑡ℎ HN and 𝑏𝑗 is the biases of 𝑗𝑡ℎ HN, To the RBF
parameter tuning process. AHOA performs the optimizer by
node with initiation function 𝑔, 𝐺j is represented as:
the natural behavior of hummingbirds [25]. It primarily
‖𝑥 − 𝑎𝑗 ‖
𝐺𝑗 (𝑥) = 𝑔 ( ) , a j ∈ ℝ𝑑 , 𝑏𝑗 ∈ ℝ+ (9) contains 3 components namely visit table, food sources, and
𝑏𝑗 population (No. of hummingbirds). While handling the
whereas 𝑎𝑗 and 𝑏𝑗 denote the centre and effect factor of the foraging state, it again consists of three diverse approaches
𝑗𝑡ℎ RBF node. migrating foraging, guided foraging, and territorial foraging.
The mathematical formula can be discussed in the following.

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

Step1: Initialization: Let ℎ be a number of birds and prey During this case, the AHOA has been deployed for
for the optimizer. It can be established randomly. determining the hyperparameter contained in the HELM
𝑓𝑗 = 𝐿𝐵 + 𝑛 ⋅ (𝑈𝐵 − 𝐿𝐵), 𝑗 = 1,2, … , ℎ (14) algorithm. The MSE has assumed that the main purpose is
Where 𝑛 is the random number range from zero to one. written as:
𝐿 𝑀
Here, the visit table can be described as the visit level of food 1 2
sources for respective birds. The visit level proves more 𝑀𝑆𝐸 = ∑ ∑(𝑦𝑗𝑖 − 𝑑𝑗𝑖 ) (23)
𝑇
important. Hereafter, each bird attacks the food via the visit 𝑗=1 𝑖=1
table. Whereas 𝑀 and 𝐿 denote the outcome value of state and
The food source is initialized in the visit table shown below: data, 𝑦𝑗𝑖 and 𝑑𝑗𝑖 denote the gained and suitable magnitudes for
0 𝑖𝑓 𝑗 ≠ 𝑘 𝑡ℎ𝑒 𝑗𝑡ℎ unit from the outcome state of the network in time 𝑡
𝑠𝑡𝑖,𝑘 = { , j, 𝑘 = 1,2, … , ℎ (15) correspondingly.
𝑛𝑢𝑙𝑙 𝑗 = 𝑘
Now, 𝑛𝑢𝑙𝑙 represents the food captured by hummingbirds,
the 0 value signifies that 𝑘 𝑡ℎ food is visited by 𝑗𝑡ℎ bird. IV. PERFORMANCE VALIDATION
Step2: Guided searching:‐ This behavior can be stimulated The AHOA-HDLTM approach has been validated utilizing
by the visit stage of the food source. When the food is targeted, the traffic data comprising every 30s raw sensor information
the hummingbird flies to feed it. The flying or flight nature for 30 days. The traffic information gathered in the 1st 10 days
relies on 3 directions namely axial, omnidirectional, and has been utilized as a training database the database
diagonal flights. comprising the data for the remaining 20 days can be
𝑗 1 if 𝑗 = 𝑟𝑑𝑗([1, 𝑧]), 𝑗 = 1, … , 𝑧 employed as a testing database. During this research, the group
𝐺𝑎𝑓 = { (16) of data comprised of 15min of combined data in vehicles per
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑖𝑓 = 𝐷(𝑘), ∈ [1, ], 15 minutes (veh per 15 minutes). Therefore, 96 data groups
can be accessible every day. Earlier in the computation, the
𝑗 1 𝐷 = 𝑟𝑑𝑝 ( ),
𝐺𝑑𝑓 = 𝑗 group of data was normalised and the information was
𝑖 ∈ [2, (𝑛1 ⋅ (𝑧 − 2)) + 1] rendered from the range of zero to one.
{0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 Table 1 and Fig. 3 provide the overall predictive outcomes
= 1, … , 𝑧 (17)
𝑗 of the AHOA-HDLTM technique in terms of MAPE. The
𝐺𝑜𝑓 = 1 𝑗 = 1, … , 𝑧 (18) outcomes imply that the AHOA-HDLTM technique exhibits
Where 𝑟𝑑𝑗([1, 𝑧]) and 𝑝(𝑖) respectively are the randomly effectual traffic flow prediction results under several
generated integer and permutation integer. Furthermore, the iterations. It is also noticed that the MAPE values are
random value is taken from 0 to 1 using the parameter 𝑛1 . A decreased with a rise in iterations. It is identified that the
candidate food source can be obtained by these flights. AHOA-HDLTM technique attains the best fitness of 0.23
The prey can be modernized from old to another location. average fitness of 0.94 and worst fitness of 2.61 under iteration
𝑔𝑗 (𝑏 + 1) = 𝑓𝑗,𝑡𝑔 (𝑏) + 𝛼 ⋅ 𝐺 ⋅ (𝑓𝑗 (𝑏) − 𝑓𝑗,𝑡𝑔 (𝑏)) (19) 100.
TABLE I
In Eq. (19), the prey position and targeted prey are stated as MAPE OUTCOME OF AHOA-HDLTM SYSTEM AT DISTINCT ITERATIONS
𝑓𝑖 (𝑏) and 𝑓𝑖,𝑡𝑔 (𝑏), correspondingly. Next, the guided feature MAPE (%)
as 𝛼 ∼ 𝑅(0,1) shows the uniform distribution function with No. of Best Average Worst
mean 0 and standard deviation (SD) 1. Iterations Fitness Fitness Fitness
0 1.50 4.98 7.47
𝑓𝑗 (𝑏) 𝐹𝑇 (𝑓𝑗 (𝑏)) ≤ 𝐹𝑇 (𝑔𝑗 (𝑏 + 1))
𝑓𝑗 (𝑏 + 1) = { (20) 10 0.68 2.69 7.02
𝑔𝑗 (𝑏 + 1) 𝐹𝑇 (𝑓𝑗 (𝑏)) > 𝐹𝑇 (𝑔𝑗 (𝑏 + 1)) 20 0.29 0.84 4.52
Step3: Territorial searching: ‐When the target prey has been 30 0.47 0.90 4.21
visited that can consumed already, afterwards it can visit novel 40 0.53 1.30 4.52
or other food sources in their territory. The new location is 50 0.34 1.06 3.89
produced by means of Eq. (21) based on the respective area. 60 0.42 0.56 3.38
𝑔𝑗 (𝑏 + 1) = 𝑓𝑗 (𝑏) + 𝛽 ⋅ 𝐺 ⋅ 𝑓𝑗 (𝑏) (21) 70 0.32 1.20 3.17
Where 𝛽 shows the territorial factor as ∼ 𝑅(0,1). 80 0.39 0.99 2.66
Step4: Migration searching: The repeated visiting lacks the 90 0.24 0.27 2.77
food source in the territorial area. A migration co-efficient has
100 0.23 0.94 2.61
been considered the worst nectar‐refilling rate. It transfers
towards a novel target. Therefore, it assumes random
generation of food sources by worse nectar.
𝑓𝑤𝑟𝑠𝑡 = 𝐿𝐵 + 𝑛. (𝑈𝐵— 𝐿𝐵) (22)
As a result, the optimum solution is attained from AHOA
which is used for parameter selection.

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

FIGURE 3. MAPE outcome of AHOA-HDLTM system under various

iterations
Table 2 and Fig. 4 represent the actual vs. predicted traffic
flow of the AHOA-HDLTM technique under varying runs and
time indexes. The obtained values pointed out that the AHOA-
HDLTM technique effectually predicted the flow of traffic
under all runs. It is also noticed that the AHOA-HDLTM
technique accomplishes effectual outcomes under all time
indices. FIGURE 4. Traffic flow outcome of AHOA-HDLTM algorithm (a-e) Runs
TABLE II 1-5
TRAFFIC FLOW OUTCOME OF AHOA-HDLTM ALGORITHM UNDER Table 3 illustrates the comparative outcome of the AHOA-
VARIOUS RUNS
Traffic Flow HDLTM technique with recent techniques with respect to
Predicted Root Mean Square Error (RMSE), Mean Absolute Error
Time Index Actual
Run_1 Run_2 Run_3 Run_4 Run_5 (MAE), and equal coefficient (EC) [26].
0 193 229 233 247 236 268 TABLE III
10 139 132 132 119 93 90 COMPARATIVE OUTCOME OF AHOA-HDLTM SYSTEM WITH RECENT
MODELS UNDER ALL LAGS
20 887 899 863 867 924 919
30 967 968 978 984 930 938 No. Hybri AST2F AHOA-
40 726 699 661 656 655 623 PSOLS FFOLS IAROE
of d LS- P OH- HDLT
-SVM -SVM L-TFMS
50 774 803 838 799 805 791 Lag SVM DBN M
60 960 975 995 1010 989 953 s
70 863 872 833 836 882 893 Root-Mean-Square Error
80 970 1024 1039 1031 1014 1053 Lag 33.88
90 1041 1042 1042 1021 965 955 =1 40.748 39.520 3 28.316 21.916 20.226
Lag 30.85
100 615 639 604 614 564 547
=2 43.887 42.146 1 24.706 19.016 17.306
Lag 23.87
=3 48.880 46.371 2 17.614 17.574 15.684
Lag 29.31
=4 52.487 50.548 1 24.281 24.001 22.261
Lag 29.79
=5 56.702 54.532 8 24.531 19.141 17.321
Mean Absolute Error
Lag 23.21
=1 26.101 23.401 3 16.907 15.517 13.777
Lag 20.41
=2 33.292 26.541 0 13.586 13.056 11.186
Lag 24.84
=3 40.864 32.838 9 19.323 17.413 15.693
Lag 28.51
=4 48.276 38.580 4 21.660 22.700 20.880
Lag 25.90
=5 53.663 46.022 5 21.191 19.331 17.531
Equal Coefficient
Lag
=1 95.85 95.96 98.08 98.51 99.24 99.42

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

Lag FPOHDBN, and IAROEL-TFMS systems gain lesser EC of

=2 95.71 95.92 98.03 98.67 99.13 99.28
Lag
95.05%, 95.25%, 97.40%, 98.78%, and 99.47%,
=3 95.48 95.58 98.71 98.95 99.21 99.37 correspondingly.
Lag
=4 95.28 95.41 98.56 98.78 99.32 99.59
Lag
=5 95.05 95.25 97.40 98.78 99.47 99.71
Fig. 5 represents the RMSE inspection of the AHOA-
HDLTM technique with recent approaches. The outcomes
indicate that the PSOLS-SVM and FFOLS-SVM systems are
depicted as ineffective performance with maximal RMSE
values. Besides, the hybrid LSSVM model has attained a
somewhat lesser RMSE value whereas the AST2-FPOHDBN
and IAROEL-TFMS algorithms have attained considerable
RMSE values. However, the AHOA-HDLTM system reports
effectual outcomes with minimal RMSE values of 20.226,
17.306, 15.684, 22.261, and 17.321 under lags 1-5,
correspondingly.

FIGURE 6. MAE outcome of AHOA-HDLTM methodology under all lags

FIGURE 5. RMSE outcome of AHOA-HDLTM system under all lags

Fig. 6 demonstrates the MAE outcome of the AHOA-
HDLTM method with recent approaches. The outcomes FIGURE 7. EC outcome of AHOA-HDLTM approach under all lags
signify that the PSOLS-SVM and FFOLS-SVM models have Table 4 and Fig. 8 represent the running time (RUNT)
shown ineffective performance with maximal MAE values. inspection of the AHOA-HDLTM technique with recent
Also, the hybrid LSSVM algorithm has attained a somewhat models. The results indicate that the PSOLS-SVM and
decreased MAE value whereas the AST2-FPOHDBN and FFOLS-SVM models have shown ineffective performance
IAROEL-TFMS systems have gained considerable MAE with maximum RUNT values. At the same time, the hybrid
values. But, the AHOA-HDLTM technique reports effective LSSVM model has attained a slightly decreased RUNT value
outcomes with lesser MAE values of 13.777, 11.186, 15.693, whereas the AST2-FPOHDBN and IAROEL-TFMS models
20.880, and 17.531 under lags 1-5, correspondingly. have obtained considerable RUNT values. Nevertheless, the
In Fig. 7, the EC results of the AHOA-HDLTM technique AHOA-HDLTM technique reports effectual outcomes with
are compared with recent models. The outcome highlighted minimal RUNT values of 1.06s, 1.18s, 2.61s, 2.14s, and 3.06s
that the AHOA-HDLTM technique offers improved under lags 1-5, correspondingly. These results confirmed the
performance with increased EC values under all lags. On lag enhanced predictive results of the AHOA-HDLTM method.
1, the AHOA-HDLTM technique provides a higher EC of
99.42 whereas the PSOLS-SVM, FFOLS-SVM, Hybrid LS-
SVM, AST2-FPOHDBN, and IAROEL-TFMS models
accomplish lower EC of 95.85%, 95.96%, 98.08%, 98.51%,
and 99.24%, respectively. Additionally, on lag 5, the AHOA-
HDLTM method offers superior EC of 99.71 whereas the
PSO- LSSVM, FFOLS-SVM, Hybrid LS-SVM, AST2-

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

TABLE IV [3] Okic, A., Zanzi, L., Sciancalepore, V., Redondi, A. and Costa-Pérez,
RUNT OUTCOME OF AHOA-HDLTM APPROACH WITH RECENT MODELS X., 2021, May. π-ROAD: A learn-as-you-go framework for on-
UNDER ALL LAGS demand emergency slices in V2X scenarios. In IEEE INFOCOM
Running Time (sec) 2021-IEEE Conference on Computer Communications (pp. 1-10).
Hybrid AST2FP IEEE.
No. of PSOLS- FFOLS- IAROEL- AHOA- [4] Qi, W., Landfeldt, B., Song, Q., Guo, L. and Jamalipour, A., 2020.
LS- OH-
Lags SVM SVM TFMS HDLTM Traffic differentiated clustering routing in DSRC and C-V2X hybrid
SVM DBN
Lag = 1 11.45 14.38 9.98 3.99 2.26 1.06 vehicular networks. IEEE Transactions on Vehicular
Lag = 2 11.78 14.61 10.16 3.75 2.75 1.18 Technology, 69(7), pp.7723-7734.
Lag = 3 11.79 15.21 10.19 5.27 4.22 2.61 [5] Backfrieder, C., Ostermayer, G. and Mecklenbräuker, C.F., 2016.
Lag = 4 12.29 15.41 10.20 5.21 3.81 2.14 Increased traffic flow through node-based bottleneck prediction and
Lag = 5 12.36 15.43 10.62 5.20 4.42 3.06 V2X communication. IEEE Transactions on Intelligent
Transportation Systems, 18(2), pp.349-363.
[6] Song, R., Lyu, L., Jiang, W., Festag, A. and Knoll, A., 2023. V2X-
Boosted Federated Learning for Cooperative Intelligent
Transportation Systems with Contextual Client Selection. arXiv
preprint arXiv:2305.11654.
[7] Luo, L., Sheng, L., Yu, H. and Sun, G., 2021. Intersection-based V2X
routing via reinforcement learning in vehicular ad hoc networks. IEEE
Transactions on Intelligent Transportation Systems, 23(6), pp.5446-
5459.
[8] Xu, T., Guan, J., Hao, Y. and Sun, D., 2021, May. Real-time traffic
state predictor based on dynamic traveller behaviour. In Proceedings
of the Institution of Civil Engineers-Transport (pp. 1-11). Thomas
Telford Ltd.
[9] Yuan, X., Chen, J., Yang, J., Zhang, N., Yang, T., Han, T. and
Taherkordi, A., 2022. Fedstn: Graph representation driven federated
learning for edge computing enabled urban traffic flow
prediction. IEEE Transactions on Intelligent Transportation Systems.
[10] Zadobrischi, E., Cosovanu, L.M. and Dimian, M., 2020. Traffic flow
density model and dynamic traffic congestion model simulation based
on practice case with vehicle network and system traffic intelligent
communication. Symmetry, 12(7), p.1172.
FIGURE 8. RUNT outcome of AHOA-HDLTM approach under all lags [11] Abdullah, S.M., Periyasamy, M., Kamaludeen, N.A., Towfek, S.K.,
Marappan, R., Kidambi Raju, S., Alharbi, A.H. and Khafaga, D.S.,
V. CONCLUSION 2023. Optimizing Traffic Flow in Smart Cities: Soft GRU-Based
Recurrent Neural Networks for Enhanced Congestion Prediction
In this manuscript, a new AHOA-HDLTM method was Using Deep Learning. Sustainability, 15(7), p.5949.
developed in the ITS environment. The main purpose of the [12] Zheng, Y., Li, X., Xu, L. and Wen, N., 2022. A deep learning–based
AHOA-HDLTM system is to forecast the level of traffic flow approach for moving vehicle counting and short-term traffic
in the ITS, enabling proper traffic management in smart cities. prediction from video images. Frontiers in Environmental
Science, 10, p.905443.
To accomplish this, the AHOA-HDLTM approach includes [13] Pamuła, T. and Żochowska, R., 2023. Estimation and prediction of the
data preprocessing, ISSA-based feature selection, HELM- OD matrix in uncongested urban road network based on traffic flows
based classification, and AHOA-based parameter tuning. For using deep learning. Engineering Applications of Artificial
Intelligence, 117, p.105550.
traffic flow monitoring, the AHOA-HDLTM technique uses [14] Yao, B., Ma, A., Feng, R., Shen, X., Zhang, M. and Yao, Y., 2022. A
the HELM model, tailored for traffic data analysis, and deep learning framework about traffic flow forecasting for urban
extracts intricate features and patterns. At last, the predictive traffic emission monitoring system. Frontiers in public health, 9,
p.804298.
outcome of the HELM technique can be better by using an [15] Nigam, A. and Srivastava, S., 2023. Hybrid deep learning models for
AHOA-based hyperparameter selection process. The traffic stream variables prediction during rainfall. Multimodal
enhanced performance of the AHOA-HDLTM method can be Transportation, 2(1), p.100052.
[16] Qi, X., Mei, G., Tu, J., Xi, N. and Piccialli, F., 2022. A deep learning
assured using a detailed simulation analysis of diverse traffic
approach for long-term traffic flow prediction with multifactor fusion
data. The extensive results highlighted the better predictive using spatiotemporal graph convolutional network. IEEE
outcome of the AHOA-HDLTM system with other methods. Transactions on Intelligent Transportation Systems.
While the article primarily focuses on traffic flow prediction [17] Yang, H., Du, L., Zhang, G. and Ma, T., 2023. A traffic flow
dependency and dynamics based deep learning aided approach for
through the AHOA-HDLTM approach, there exists a critical network-wide traffic speed propagation prediction. Transportation
avenue for further investigation into the intricate research part B: methodological, 167, pp.99-117.
interconnection between traffic flow and traffic state. [18] Bharti, Redhu, P. and Kumar, K., 2023. Short-term traffic flow
prediction based on optimized deep learning neural network: PSO-Bi-
LSTM. Physica A: Statistical Mechanics and its Applications, 625,
REFERENCES p.129001.
[1] Jiang, L., Molnár, T.G. and Orosz, G., 2021. On the deployment of [19] Chen, C., Liu, B., Wan, S., Qiao, P. and Pei, Q., 2020. An edge traffic
V2X roadside units for traffic prediction. Transportation Research flow detection scheme based on deep learning in an intelligent
Part C: Emerging Technologies, 129, p.103238. transportation system. IEEE Transactions on Intelligent
[2] Shi, X., Liu, H., Wang, M., Li, X., Ciuffo, B., Work, D. and Kan, D., Transportation Systems, 22(3), pp.1840-1852.
2022, July. Inconsistency of AV Impacts on Traffic Flow: Predictions [20] Yu, K., Lin, L., Alazab, M., Tan, L. and Gu, B., 2020. Deep learning-
in Literature. In Automated Road Transportation Symposium (pp. based traffic safety solution for a mixture of autonomous and manual
165-173). Cham: Springer Nature Switzerland. vehicles in a 5G-enabled intelligent transportation system. IEEE

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3349032

transactions on intelligent transportation systems, 22(7), pp.4337-

4347.
[21] Liu, C. and Ke, L., 2023. Cloud assisted Internet of things intelligent
transportation system and the traffic control system in the smart
city. Journal of Control and Decision, 10(2), pp.174-187.
[22] Chen, J., Zhang, Y., Teng, S., Chen, Y., Zhang, H. and Wang, F.Y.,
2023. ACP-Based Energy-Efficient Schemes for Sustainable
Intelligent Transportation Systems. IEEE Transactions on Intelligent
Vehicles.
[23] Sulaiman, M.H. and Mustaffa, Z., 2023. An Application of Improved
Salp Swarm Algorithm for Optimal Power Flow Solution Considering
Stochastic Solar Power Generation. e-Prime-Advances in Electrical
Engineering, Electronics and Energy, p.100195.
[24] Du, F., Zhang, J., Ji, N., Shi, G. and Zhang, C., 2018. An effective
hierarchical extreme learning machine based multimodal fusion
framework. Neurocomputing, 322, pp.141-150.
[25] Selvaraj, A., Sundaram, V.R.S. and Mahdal, M., 2023. An Adaptive
Sleep Apnea Detection Model using Multi Cascaded Atrous based
Deep Learning Schemes with Hybrid Artificial Humming Bird Pity
Beetle Algorithm. IEEE Access.
[26] Ragab, M., Abdushkour, H.A., Maghrabi, L., Alsalman, D., Fayoumi,
A.G. and AL-Ghamdi, A.A.M., 2023. Improved Artificial Rabbits
Optimization with Ensemble Learning-Based Traffic Flow
Monitoring on Intelligent Transportation
System. Sustainability, 15(16), p.12601.

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4