IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 2, June 2024, pp. 2323~2332

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i2.pp2323-2332  2323

Smart traffic forecasting: leveraging adaptive machine learning

and big data analytics for traffic flow prediction

Idriss Moumen, Jaafar Abouchabaka, Najat Rafalia

Department of Computer Science, Faculty of Sciences, Ibn Tofail University, Kenitra, Morocco

Article Info ABSTRACT

Article history: The issue of road traffic congestion has become increasingly apparent in
modern times. With the rise of urbanization, technological advancements, and
Received Jun 26, 2023 an increase in the number of vehicles on the road, almost all major cities are
Revised Sep 29, 2023 experiencing poor traffic environments and low road efficiency. To address
Accepted Dec 16, 2023 this problem, re-searchers have turned to diverse data resources and focused
on predicting traffic flow, a crucial issue in intelligent transportation systems
(ITS) that can help al-leviate congestion. By analyzing data from correlated
Keywords: roads and vehicles, such as speed, density, and flow rate, it is possible to
anticipate traffic congestion and patterns. This paper presents an adaptive
Big data analytics traffic system that utilizes supervised ma-chine learning and big data analytics
Deep learning to predict traffic flow. The system monitors and extracts relevant traffic flow
Intelligent transportation data, analyzes and processes the data, and stores it to enhance the model's
systems accuracy and effectiveness. A simulation was conducted by the authors to
Machine learning showcase the proposed solution. The outcomes of the study car-ry substantial
Neural networks implications for transportation systems, offering valuable insights for
Traffic flow prediction enhancing traffic flow management.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Idriss Moumen
Department of Computer Science, Faculty of Sciences, Ibn Tofail University
B.P 133, University Campus, Kenitra, Morocco
Email: [email protected]

1. INTRODUCTION
The surge in the number of vehicles on roads, which has resulted in significant traffic congestion [1],
is causing adverse environmental and economic impacts and reducing mobility [2]. To tackle these challenges,
experts are using intelligent transportation systems (ITS) to improve traffic management and enhance the
overall transportation experience [3], [4]. With the emergence of big data analytics and the proliferation of
wireless communication technologies, various sources can gather an extensive volume of real-time
transportation data. This data creates novel opportunities for traffic flow prediction [5], which is pivotal for
traffic management, route optimization, and other ITS applications. Using statistical and machine learning
(ML) techniques, predictive models can be created to detect patterns and make predictions about traffic flow.
Recently, deep learning (DL), which is a ML method, has piqued the attention of both academic and industrial
researchers. DL has been shown to be useful in a wide range of tasks, including classification, natural language
processing, reducing dimensionality, object detection [6], and motion modeling. This has been demonstrated
in numerous studies, such as [7]–[11]. DL algorithms use multi-layer or deep structures to find underlying
properties in data from the most basic to the most complex. Revealing substantial amounts of structure within
the data. Furthermore, due to their unique qualities, such as distributed storage and large parallel structure,
neural networks have become the target of substantial research by numerous experts and scholars. Numerous
studies have been conducted in this field, employing a variety of methods such as Kalman state space filtering

Journal homepage: http://ijai.iaescore.com

2324  ISSN: 2252-8938

models [12], support vector machine (SVM) models [13], neuro-fuzzy systems [14], autoregressive integrated
moving average models [15], radial basis function neural network models [16], [17], Type-2 fuzzy logic
approach [18], k-nearest neighbor (KNN) model [19], binary neural network models [20], [21], Bayesian
network models [22], [23], back propagation neural network models [24], [25]. Recently, researchers have
combined artificial neural network (ANN) with empirical mode decomposition and auto-regressive integrated
moving average (ARIMA) to increase forecasting accuracy [26]. ARIMA is frequently contrasted with hybrid
models like the long short-term memory (LSTM) [27]. However, but there have been comparisons between
ARIMA and Facebook Prophet [28], as well as between ARIMA, LSTM, and Facebook Prophet [29].
Weytjens et al. [30] compared multi-layer LSTM networks with ARIMA and Facebook Prophet for forecasting
cash flow or demand, while Abbasimehr et al. [31] used multi-layer LSTM networks to accomplish the same
thing. Because of their ease of use and minimal data requirements, seasonal auto-regressive integrated moving
average (SARIMA) models are common [32]. Over the years, the use of ML has become common among
researchers to predict traffic injurie several models like elman recurrent neural network (ERNN) [33], LSTM
[34]–[37], and extreme gradient boosting (XGBoost) [38] have been used and proven to increase the precision
of forecasts. In this study, three tree ML models-logistics regression (LR), linear regressor, and decision tree
(DT)-as well as two DL models-Facebook Prophet and LSTM-based on recurrent neural networks (RNNs)-are
compared for the task of predicting traffic flow at an intersection. The goal is to use these models to modernize
the traffic light system by improving traffic flow without having to completely alter the system, making its
implementation more practical. The experiments show that all models are effective at forecasting vehicle flow
and can be implemented in a smart traffic light system. The remaining parts of the essay are arranged as follows.
The various data-analytics-based strategies for predicting traffic flow are described in section 2 using their
respective methods. The experimental findings are discussed in section 3. Final observations are described in
section 4.

2. MATERIALS AND METHODS

The data used in this research were obtained from different locations on the roads of England for
2020, which is a public dataset for traffic prediction derived from a variety of traffic sensors. Data on a 4
million rows and 37 columns with diverse data types such as strings, integers, and dates. This research uses
five models-linear regression, LR, DT, LSTM and Facebook Prophet models. Figure 1 shows the typical
architecture of the proposed model to predict the future traffic flow at an intersection.

Linear Regression

Logistic Regression
Model
Train Data
Collect Data
Decision Tree Model
Traffic Data Preprocessing
Test Data
LSTM Model

Facebook Prophet
Model

Validation model

Traffic light system

Figure 1. The flowchart presented illustrates the process of predicting traffic flow using AI-based models and
experimental methods

Int J Artif Intell, Vol. 13, No. 2, June 2024: 2323-2332

Int J Artif Intell ISSN: 2252-8938  2325

In order to analyze preprocessed data and evaluate the effectiveness of traffic flow prediction models,
three standard classification algorithms with Spark/MLlib implementations, namely LR and DT, were
employed for model training and evaluation. Prior to the training and evaluation process, the data underwent
preprocessing steps such as feature extraction, normalization, and principal component analysis. It is important
to note that the same preprocessing techniques were applied consistently to the data before training and
evaluating each algorithm.

2.1. Linear regression model

Linear regression is widely used in various applications due to its simplicity and interpretability. It is
especially useful when there is a linear relationship between the input and output variables. The model’s ability
to provide insights into the strength and direction of the relationship between variables makes it valuable in
fields like economics, finance, and social sciences. The linear regression has an equation of the (1):

𝑌 = 𝛽0 + 𝛽1 𝑋 + 𝜀 (1)

where 𝑌 is the response variable, 𝑋 is the predictor variable, 𝛽0 and 𝛽1 are the regression coefficients or
regression parameters, and 𝜀 is the error term. The regression coefficients 𝛽0 and 𝛽1 determine the intercept
and slope of the regression line, respectively. The error term 𝜀 accounts for the discrepancy between the
predicted data and the observed data, as it represents the unexplained variability in the response variable not
captured by the predictor variable. The predicted value form of the predicted data is (2):

𝑌̂ = 𝛽̂0 + 𝛽̂1 𝑋 + 𝜀 (2)

In regression analysis, the term 𝑌̂ represents the fitted or predicted value, while 𝛽̂ represents the
estimated regression coefficients. The fitted value is calculated based on the observed data used to derive the
estimates of the regression coefficients 𝛽̂ , corresponding to one of the 𝑛 observations in the dataset. On the
other hand, the predicted values are generated for any arbitrary set of predictor variable values different from
those present in the observed data. In essence, the fitted value is specific to the observed data points used during
model training, while predicted values can be generated for any new combination of predictor variables beyond
the scope of the training data.

2.2. Logistic regression model

LR is a classification technique that belongs to the group of linear regression methods. These methods
are mathematically formulated as convex optimization problems. The primary objective of LR is to identify a
set of weights that, when linearly combined, effectively forecast a dependent variable while minimizing the
discrepancy between the predicted and actual values. In formal terms, the optimization problem entails finding
the optimal vector of weights, denoted as 𝑤, which minimizes the loss function 𝐿. This problem is posed within
the context of 𝑛 training data feature vectors, represented as 𝑥𝑖 , with a length of 𝑑, along with their
corresponding labels, denoted as 𝑦𝑖 .
1
𝑚𝑖𝑛𝑤∈ℝ𝑑 𝑓(𝑥) = ∑𝑛𝑖=1 𝐿(𝑤, 𝑥𝑖 , 𝑦𝑖 ) (3)
𝑛

when training LR models, the employed loss function is the logistic loss function, which operates on a linear
combination of weights and features (the 𝑤 𝑇 𝑥 term). This loss function is specifically designed for LR and
facilitates the optimization process by quantifying the discrepancy between predicted values and actual labels
(𝑤 𝑇 𝑥 term).
𝑇
𝐿(𝑤, 𝑥, 𝑦) = log(1 + 𝑒 −𝑦𝑤 𝑥 ) (4)

The trained model utilizes a logistic sigmoid function to make predictions by transforming the linear
combination of features and weights. This sigmoid function is commonly used in binary classification tasks,
where the output is mapped to a probability score between 0 and 1. By applying the sigmoid function, the
model can convert the raw linear combination into a probability, allowing it to determine the likelihood of a
binary outcome, such as whether an event will occur or not.
1
𝑓(𝑤; 𝑥) = 𝑇 (5)
(1+𝑒 −𝑤 𝑥 )

Smart traffic forecasting: leveraging adaptive machine learning and big data analytics … (Idriss Moumen)
2326  ISSN: 2252-8938

1, 𝑓(𝑤; 𝑥) > 0.5

𝑐𝑙𝑎𝑠𝑠(𝑤; 𝑥) = { (6)
0, 𝑓(𝑤; 𝑥) ≤ 0.5

2.3. Decision trees model

Classification models, such as DTs, partition the solution space into binary classes through recursive
splitting. This process involves creating a large DT with branches based on a chosen metric. In this study,
entropy and Gini impurity were considered as metrics for determining the splitting criterion at each branch.
Entropy aims to maximize the information gain of the split, rapidly narrowing down the predicted state choices.
Formally, the split 𝑠 is selected at each tree node to divide the dataset 𝐷 of size 𝑁 into two subsets, 𝐷𝑙𝑒𝑓𝑡 and
𝐷𝑟𝑖𝑔ℎ𝑡 with sizes 𝑁𝑙𝑒𝑓𝑡 and 𝑁𝑟𝑖𝑔ℎ𝑡 The objective is to maximize the entropy 𝐸(𝑥) relative to the number of
discrete classes 𝐶 where 𝑓𝑖 represents the frequency of class 𝑖 at a node.

𝐸(𝑥) = ∑𝐶𝑖=1 −𝑓𝑖 log 𝑓𝑖 (7)

𝑁𝑙𝑒𝑓𝑡 𝑁𝑟𝑖𝑔ℎ𝑡
𝑎𝑟𝑔𝑠 max (𝐸(𝐷) − 𝐸(𝐷𝑙𝑒𝑓𝑡 , S) − 𝐸(𝐷𝑟𝑖𝑔ℎ𝑡 , S)) (8)
𝑁 𝑁

The Gini impurity is a metric used in DT algorithms to measure the degree of impurity or uncertainty
in a particular split. Unlike the information gain, which aims to find the most efficient split, the Gini impurity
seeks to minimize the chances of misclassification after the split. It computes the impurity index 𝐺(𝑥) based
on the probability of misclassifying a randomly chosen element from the data distribution at a specific node,
and then the DT algorithm selects the split that maximizes this Gini impurity value. The Gini impurity 𝐺(𝑥) is
computed as (9) and (10):

𝐺(𝑥) = ∑𝐶𝑖=1 −𝑓𝑖 log 𝑓𝑖 (9)

𝑁𝑙𝑒𝑓𝑡 𝑁𝑟𝑖𝑔ℎ𝑡
𝑎𝑟𝑔𝑠 max (𝐺(𝐷) − 𝐺(𝐷𝑙𝑒𝑓𝑡 , S) − 𝐺(𝐷𝑟𝑖𝑔ℎ𝑡 , S)) (10)
𝑁 𝑁

2.4. Long short-term memory model

LSTM, a prominent model within RNNs, is widely employed due to its capacity to retain information
from sequential data [39]. Typically, an LSTM configuration with a hidden layer comprising 4 neurons and a
single output producing 12 predicted values is utilized. Training occurs over 250 epochs, and it is imperative
that the training set remains unshuffled. Figure 2 offers an illustrative depiction of the LSTM predictor’s
implementation.

Figure 2. The LSTM cell consists of an input gate, an output gate and a forget gate

LSTM, is a type of RNN known for its ability to handle long-term dependencies in sequential data. It
comprises three crucial gates: the forget gate, the input gate, and the output gate. Each gate performs a specific
function in the processing of information within the network. The forget gate determines which information
from the previous time step should be discarded, the input gate decides what new information to incorporate,

Int J Artif Intell, Vol. 13, No. 2, June 2024: 2323-2332

Int J Artif Intell ISSN: 2252-8938  2327

and the output gate regulates the output state of the LSTM cell. These gates enable LSTMs to selectively retain
important information, learn relevant patterns, and update their internal states, making them particularly
effective in tasks involving sequential data, such as natural language processing and time series analysis. The
notations are as follows: 𝑥𝑡 is the input value of the current time step, 𝑓𝑡 is the forget gate of the current time
step, ℎ𝑡 and ℎ𝑡−1 is hidden states representing short-term memory for the current and previous time steps, 𝑐𝑡 is
cell states representing long-term memory for the current and previous time steps, 𝜎 is sigmoid activation
function, 𝑡𝑎𝑛ℎ is non-linear activation function allowing error learning in multiple neuron layers, 𝑖𝑡 is input
gate for the current time step, and 𝑜𝑡 is output gate for the current time step.
The forgetting gate plays a crucial role in RNNs and LSTM networks. Its main function is to evaluate
the importance of information stored in the intermediate and previous layers of the network. By using a
mathematical representation, the forgetting gate determines which information should be retained for the
current task and which should be discarded, thereby allowing the model to focus on the most relevant
information for making accurate predictions or solving specific problems. The forgetting gate can be
mathematically represented in (11):

𝑓𝑡 = 𝜎(𝑥𝑡 × 𝑊𝑓 + ℎ𝑡−1 × 𝑊𝑓 ) (11)

– Input gate: following the forgetting gate, the input gate updates and integrates data into the memory cell
using an activation function. The specific formula for the input gate is as (12):

𝑖𝑡 = 𝜎(𝑥𝑡 × 𝑊𝑖 + ℎ𝑡−1 × 𝑊𝑖 ) (12)

– Output gate: the output gate governs the model's output by incorporating the weight of the control state
𝑐𝑡 with the current LSTM hidden layer. The initial output is obtained through an activation function and
subsequently normalized using the tanh function. The expression for the output gate is as (13) and (14):

𝑜𝑡 = 𝜎(𝑥𝑡 × 𝑊𝑜 + ℎ𝑡−1 × 𝑊𝑜 ) (13)

ℎ𝑡 = 𝑜𝑡 × 𝑡𝑎𝑛ℎ(𝑐𝑡 ) (14)

2.5. Facebook Prophet model

The Prophet model, introduced by Facebook Inc. in 2017 [28], is an additive model specifically
designed for time series prediction. According to Google’s official presentation [40], it demonstrates superior
performance when applied to time series data with pronounced seasonal effects and an ample historical data
spanning multiple seasons. Prophet exhibits resilience in handling missing data, trend shifts, and outliers [41].
Since its release, the Prophet model has gained significant popularity in the field of time series analysis. It
decomposes the time series into three key components: the seasonal term 𝑆𝑡 , the trend term 𝑇𝑡 , and the residual
term 𝑅𝑡 are as (15):

𝑦𝑡 = 𝑆𝑡 + 𝑇𝑡 + 𝑅𝑡 (15)

The Prophet model goes beyond basic time series forecasting by incorporating the influence of
holidays, denoted as ℎ(𝑡), into its predictions. This integration allows the model to account for the significant
variations in data patterns that often occur during holidays. By considering the impact of holidays on the time
series, the Prophet model becomes more adaptable and accurate in capturing real-world scenarios, making it a
valuable tool for forecasting in diverse industries and applications.

𝑦𝑡 = 𝑔(𝑡) + 𝑠(𝑡) + ℎ(𝑡) + 𝜀𝑡 (16)

The model’s robustness and capacity to handle missing data and outliers make it stand out in the field
of data analysis. Its ability to fit a diverse range of data with reasonable accuracy further cements its popularity
among data analysts, especially when dealing with time series prediction tasks. With its versatility and reliable
performance, this model has become a preferred choice for tackling complex real-world datasets, enabling
analysts to make more informed decisions and predictions.

2.6. Model evaluation

After completing the training phase, it is crucial to evaluate the prediction model’s accuracy using
testing data. In comparing forecasting methods that share the same unit, the widely utilized metric is the root

Smart traffic forecasting: leveraging adaptive machine learning and big data analytics … (Idriss Moumen)
2328  ISSN: 2252-8938

mean square error (RMSE). The RMSE is calculated using a specific equation, which quantifies the differences
between the predicted values and the actual values in the testing dataset. This evaluation metric provides a
measure of the model’s predictive performance and allows for meaningful comparisons between different
forecasting approaches.

1
RMSE(𝑦, 𝑦̂ ) = √ ∑𝑛−1 ̂𝑖 )2
𝑖=0 (𝑦𝑖 − 𝑦 (17)
𝑛

Where 𝑛 is number of samples, 𝑦 is observed traffic flow, and 𝑦̂ is predicted traffic flow.

2.7. Data exploration

This analysis aids in identifying attribute pairs with high correlation, offering insights into the key
attributes for model construction. Figure 3 displays the correlation matrix, which showcases strong correlations
among attributes such as start_junction_road_name, end_junction_road_name, day, month, year, road_name,
red-time, and green-time. The correlation matrix allows us to discern the relationships between these attributes
and their potential impact on traffic patterns, enabling us to make informed decisions during the model
development process.

Figure 3. Correlation matrix as heat map shows good correclation between day, month, year, roadname, start
junction road name, end junction road name, red time, green time

2.8. Data preprocessing

It is crucial to identify the types of data we'll be working with. In some cases, data might contain
irrelevant or absent components. To address this, we perform data cleaning, which encompasses handling
missing or noisy data, among other tasks. In this paper, we employed two data cleaning techniques: i) the
removal of outliers and redundant values (as we discovered several of these in the dataset), and ii) the
imputation of missing values (various methods exist for this purpose, but we opted to manually fill in missing
values by using the attribute’s mean or the most probable value). The primary goal of this approach is to
transform the data into formats that are appropriate for the data exploration process. This operation involves
the following techniques: Normalization: scaling data values within a specified range such as (0.0 to 1.0 or
-1.0 to 1.0). This is done for categorical columns to transform them into numerical values for processing in ML
algorithms; Attribute selection: in this strategy, new attributes are constructed from a determined set of
attributes to simplify the data exploration process. New columns are created based on a combination of other
attributes to obtain transformations such as normalization, standardization, scaling, and pivoting. Binning can
also be applied (based on the number of values, and treating missing values as a separate group). Data
replacement techniques can also be used, such as splitting, merging, and slicing.

Int J Artif Intell, Vol. 13, No. 2, June 2024: 2323-2332

Int J Artif Intell ISSN: 2252-8938  2329

3. RESULTS AND DISCUSSION

In order to maintain a systematic approach to the development process, the findings are elucidated in
discrete sub-chapters. ML models, specifically LR, linear regression, and DT, were utilized in this study, in
addition to time series models including LSTM and Facebook Prophet. PySpark was employed as the
framework for implementing these models.

3.1. Machine learning models

Table 1 presents the performance metrics of various ML models. Three different algorithms, namely
LR, linear regression, and DT, were evaluated using PySpark. For LR, the RMSE was 214.40, representing the
average difference between predicted and actual values. The R2 value, indicating the proportion of variance
explained by the independent variables, was 0.907. In contrast, linear regression achieved an RMSE of 140.126,
indicating a lower average difference compared to LR. The R2 value was higher at 0.923, indicating a better
fit to the data. The DT model yielded a higher RMSE of 306.145, indicating a relatively larger average
difference compared to logistic and linear regression. The R2 value for DT was 0.877, slightly lower than the
other models. Overall, linear regression exhibited superior performance with the lowest RMSE and highest R2
value among the evaluated models, highlighting it is predictive power in this context.

Table 1. Comparison of ML models performance metrics

Model RMSE R2
Logistic regression 214.40 0.907
Linear regression 140.126 0.923
Decison tree 306.145 0.877

3.2. Time series models

In this study, various time series models were employed to analyze and forecast events based on
historical data. Common types of time series models include ARIMA, smoothing-based models, and moving
average models. It is essential to consider that different models can yield different results when applied to the
same dataset, emphasizing the importance of selecting the most suitable model for a specific time series
analysis. In our study, we utilized LSTM and Facebook Prophet as specialized time series models for the
purpose of analysis and forecasting. The results obtained from applying the LSTM model on our dataset
demonstrated its effectiveness, as the predicted values of the number of vehicles closely resembled the actual
values. Figure 4 depicts a plot that compares the actual data with the predictions made by the LSTM model on
the testing dataset, providing visual evidence of the model’s accuracy.
The LSTM neural network was implemented in Python, and prior to training the model, the data was
normalized to enhance efficiency. Figure 5 provides a visual representation of the loss function performance
during the training phase. The chart demonstrates a consistent decrease in the loss values for the training set,
indicating that the LSTM model was effectively trained without overfitting. Unlike other models, the LSTM
predictor takes into account recent values, reducing the influence of seasonality and incorporating the current
trend. Additionally, the Prophet model excels in modeling as an additive system and displays proficiency in
identifying and showcasing seasonality patterns.
Traffic flow

Loss

Time Epoch

Figure 4. A plot between actual and prediction data Figure 5. Loss function performance during the
on testing data for number of vehicles with LSTM training phase
model

Smart traffic forecasting: leveraging adaptive machine learning and big data analytics … (Idriss Moumen)
2330  ISSN: 2252-8938

Figure 6 displays the time series graph generated using the Facebook Prophet model. The graph
illustrates a clear upward trend in the data, indicating an overall increase in values over time. Additionally,
there is a possibility of a slight curvature in the data, as the rate of increase appears to be accelerating. In such
cases, the quadratic model becomes a suitable choice for capturing the underlying pattern. It is worth noting
that the time series data does not exhibit a distinct upward or downward trend in general. The higher average
consumption observed in previous years can be attributed to the lack of recent data, during which road
occupancy was high. Therefore, when comparing year-by-year data, the road occupancy should remain
relatively stable.

ds

Figure 6. The time series graph realized with the Facebook prophet model shows a clear upward trend

4. CONCLUSION
The escalating number of vehicles on roads has led to significant traffic congestion, causing
detrimental environmental and economic consequences while hampering mobility. To address these
challenges, experts have turned to ITS as a means to enhance traffic management and improve the overall
transportation experience. The advent of big data analytics and the proliferation of wireless communication
technologies have facilitated the collection of extensive real-time transportation data, opening up new
opportunities for traffic flow prediction. Statistical and ML techniques have been leveraged to develop
predictive models capable of detecting patterns and making accurate predictions about traffic flow. DL, a
powerful ML method, has garnered significant attention from both academia and industry for it is remarkable
capabilities in various tasks, including object detection, motion modeling, and natural language processing.
Researchers have explored diverse approaches, such as Kalman state space filtering models, SVM models,
neuro-fuzzy systems, and neural network models, to predict traffic flow. Recently, combining ANN with
empirical mode decomposition and ARIMA has shown promising results in increasing forecasting accuracy.
Comparisons have been made between different models, including ARIMA, LSTM, and Facebook Prophet,
highlighting the strengths and limitations of each. In this study, five models-LR, linear regressor, DT, Facebook
Prophet, and LSTM-were evaluated for the task of predicting traffic flow at an intersection, aiming to enhance
traffic light systems without significant system overhauls. The experimental results demonstrate the
effectiveness of all models in forecasting vehicle flow and their potential implementation in smart traffic light
systems.

REFERENCES
[1] K. Saito et al., “A regulatory circuit for piwi by the large Maf gene traffic jam in Drosophila,” Nature, vol. 461, no. 7268, pp. 1296–
1299, Oct. 2009, doi: 10.1038/nature08501.
[2] P. Anttila, J.-P. Tuovinen, and J. V. Niemi, “Primary NO2 emissions and their role in the development of NO2 concentrations in a
traffic environment,” Atmospheric Environment, vol. 45, no. 4, pp. 986–992, Feb. 2011, doi: 10.1016/j.atmosenv.2010.10.050.
[3] R. Sennett, The uses of disorder: Personal identity and city life. Brooklyn, New York: Verso, 2021.
[4] F. Chuang-Lin and W. De-Li, “Comprehensive measures and improvement of Chinese urbanization development quality,”
Geographical Research, vol. 30, no. 11, pp. 1931–1946, 2011.
[5] I. Moumen, J. Abouchabaka, and N. Rafalia, “Adaptive traffic lights based on traffic flow prediction using machine learning

Int J Artif Intell, Vol. 13, No. 2, June 2024: 2323-2332

Int J Artif Intell ISSN: 2252-8938  2331

models,” International Journal of Electrical and Computer Engineering (IJECE), vol. 13, no. 5, pp. 5813–5823, Oct. 2023, doi:
10.11591/ijece.v13i5.pp5813-5823.
[6] A. Garcia-Garcia, S. Orts-Escolano, S. Oprea, V. Villena-Martinez, and J. Garcia-Rodriguez, “A review on deep learning techniques
applied to semantic segmentation,” Computing Research Repository arXiv preprint, pp. 1-23, 2017.
[7] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Science, vol. 313, no. 5786,
pp. 504–507, Jul. 2006, doi: 10.1126/science.1127647.
[8] R. Collobert and J. Weston, “A unified architecture for natural language processing: Deep neural networks with multitask learningA
unified architecture for natural language processing: Deep neural networks with multitask learning,” in Proceedings of the 25th
International Conference on Machine Learning, pp. 160-167, 2008.
[9] I. J. Goodfellow, Y. Bulatov, J. Ibarz, S. Arnoud, and V. Shet, “Multi-digit number recognition from street view imagery using deep
convolutional neural networks,” International Conference on Learning Representations (ICLR), pp. 1-12, 2014.
[10] B. Huval, A. Coates, and A. Ng, “Deep learning for class-generic object detection,” arXiv- Computer Science, pp. 1-3, 2013.
[11] H.-C. Shin, M. R. Orton, D. J. Collins, S. J. Doran, and M. O. Leach, “Stacked autoencoders for unsupervised feature learning and
multiple organ detection in a pilot study using 4D patient data,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 35, no. 8, pp. 1930–1943, Aug. 2013, doi: 10.1109/TPAMI.2012.277.
[12] B. Ait-El-Fquih and I. Hoteit, “Fast Kalman-like filtering for large-dimensional linear and Gaussian state-space models,” IEEE
Transactions on Signal Processing, vol. 63, no. 21, pp. 5853–5867, Nov. 2015, doi: 10.1109/TSP.2015.2468674.
[13] Y. Zhang and Y. Liu, “Data imputation using least squares support vector machines in urban arterial streets,” IEEE Signal
Processing Letters, vol. 16, no. 5, pp. 414–417, May 2009, doi: 10.1109/LSP.2009.2016451.
[14] J. Perez, A. Gajate, V. Milanes, E. Onieva, and M. Santos, “Design and implementation of a neuro-fuzzy system for longitudinal
control of autonomous vehicles,” in International Conference on Fuzzy Systems, Jul. 2010, pp. 1–6, doi:
10.1109/FUZZY.2010.5584208.
[15] A. Guin, “Travel time prediction using a seasonal autoregressive integrated moving average time series model,” in 2006 IEEE
Intelligent Transportation Systems Conference, 2006, pp. 493–498, doi: 10.1109/ITSC.2006.1706789.
[16] H. Gao, J. Zhao, and L. Jia, “Short-term traffic flow forecasting model of Elman neural network based on dissimilation particle
swarm optimization,” in 2008 IEEE International Conference on Networking, Sensing and Control, Apr. 2008, pp. 1305–1309, doi:
10.1109/ICNSC.2008.4525419.
[17] L. Li, W.-H. Lin, and H. Liu, “Type-2 fuzzy logic approach for short-term traffic forecasting,” IEE Proceedings-Intelligent
Transport Systems, vol. 153, no. 1, pp. 33–40, 2006, doi: 10.1049/ip-its:20055009.
[18] A. X. Wang, S. S. Chukova, and B. P. Nguyen, “Ensemble k-nearest neighbors based on centroid displacement,” Information
Sciences, vol. 629, pp. 313–323, Jun. 2023, doi: 10.1016/j.ins.2023.02.004.
[19] X. Xu, X. Jin, D. Xiao, C. Ma, and S. C. Wong, “A hybrid autoregressive fractionally integrated moving average and nonlinear
autoregressive neural network model for short-term traffic flow prediction,” Journal of Intelligent Transportation Systems, vol. 27,
no. 1, pp. 1–18, Jan. 2023, doi: 10.1080/15472450.2021.1977639.
[20] L. Kerkelä, K. Seunarine, R. N. Henriques, J. D. Clayden, and C. A. Clark, “Improved reproducibility of diffusion kurtosis imaging
using regularized non-linear optimization informed by artificial neural networks,”arXiv-Physics, pp. 1-16, 2022.
[21] Y. Li, W. Zhao, and H. Fan, “A spatio-temporal graph neural network approach for traffic flow prediction,” Mathematics, vol. 10,
no. 10, pp. 1-14, May 2022, doi: 10.3390/math10101754.
[22] Y. Yang, S. Rasouli, and F. Liao, “Effects of life events and attitudes on vehicle transactions: A dynamic Bayesian network
approach,” Transportation Research Part C: Emerging Technologies, vol. 147, pp. 1-27, Feb. 2023, doi: 10.1016/j.trc.2022.103988.
[23] M. Gui, A. Pahwa, and S. Das, “Bayesian network model with Monte Carlo simulations for analysis of animal-related outages in
overhead distribution systems,” IEEE Transactions on Power Systems, vol. 26, no. 3, pp. 1618–1624, Aug. 2011, doi:
10.1109/TPWRS.2010.2101619.
[24] N. Brahimi, H. Zhang, L. Dai, and J. Zhang, “Modelling on car-sharing serial prediction based on machine learning and deep
learning,” Complexity, vol. 2022, pp. 1–20, Jan. 2022, doi: 10.1155/2022/8843000.
[25] C. Li, Y. Xie, H. Zhang, and X. Yan, “Dynamic division about traffic control sub-area based on back propagation neural network,”
in 2010 Second International Conference on Intelligent Human-Machine Systems and Cybernetics, Aug. 2010, pp. 22–25, doi:
10.1109/IHMSC.2010.104.
[26] Ü. Ç. Büyükşahin and Ş. Ertekin, “Improving forecasting accuracy of time series data using a new ARIMA-ANN hybrid method
and empirical mode decomposition,” Neurocomputing, vol. 361, pp. 151–163, Oct. 2019, doi: 10.1016/j.neucom.2019.05.099.
[27] W. Chen, H. Xu, L. Jia, and Y. Gao, “Machine learning model for Bitcoin exchange rate prediction using economic and technology
determinants,” International Journal of Forecasting, vol. 37, no. 1, pp. 28–43, Jan. 2021, doi: 10.1016/j.ijforecast.2020.02.008.
[28] S. J. Taylor and B. Letham, “Forecasting at scale,” Peerj Preprints, vol. 5, pp. 1–25, 2017, doi: 10.1080/00031305.2017.1380080.
[29] N. K. Chikkakrishna, C. Hardik, K. Deepika, and N. Sparsha, “Short-term traffic prediction using sarima and FbPROPHET,” in 2019
IEEE 16th India Council International Conference (INDICON), Dec. 2019, pp. 1–4, doi: 10.1109/INDICON47234.2019.9028937.
[30] H. Weytjens, E. Lohmann, and M. Kleinsteuber, “Cash flow prediction: MLP and LSTM compared to ARIMA and Prophet,”
Electronic Commerce Research, vol. 21, no. 2, pp. 371–391, Jun. 2021, doi: 10.1007/s10660-019-09362-7.
[31] H. Abbasimehr, M. Shabani, and M. Yousefi, “An optimized model using LSTM network for demand forecasting,” Computers and
Industrial Engineering, vol. 143, May 2020, doi: 10.1016/j.cie.2020.106435.
[32] S. Siami-Namini, N. Tavakoli, and A. Siami Namin, “A comparison of ARIMA and LSTM in forecasting time series,” in 2018 17th
IEEE International Conference on Machine Learning and Applications (ICMLA), Dec. 2018, pp. 1394–1401, doi:
10.1109/ICMLA.2018.00227.
[33] K. Mehmood, H. T. Ul Hassan, A. Raza, A. Altalbe, and H. Farooq, “Optimal power generation in energy-deficient scenarios using
bagging ensembles,” IEEE Access, vol. 7, pp. 155917–155929, 2019, doi: 10.1109/ACCESS.2019.2946640.
[34] C. Fan, K. Matkovic, and H. Hauser, “Sketch-based fast and accurate querying of time series using parameter-sharing LSTM
network,” IEEE Transactions on Visualization and Computer Graphics, vol. 27, no. 12, pp. 4495–4506, Dec. 2021, doi:
10.1109/TVCG.2020.3002950.
[35] M. Qiao, S. Yan, X. Tang, and C. Xu, “Deep convolutional and LSTM recurrent neural networks for rolling bearing fault diagnosis
under strong noises and variable loads,” IEEE Access, vol. 8, pp. 66257–66269, 2020, doi: 10.1109/ACCESS.2020.2985617.
[36] Y. Cui, H. Xu, J. Wu, Y. Sun, and J. Zhao, “Automatic vehicle tracking with roadside LiDAR data for the connected-vehicles
system,” IEEE Intelligent Systems, vol. 34, no. 3, pp. 44–51, May 2019, doi: 10.1109/MIS.2019.2918115.
[37] M. F. Tahir, C. Haoyong, K. Mehmood, N. A. Larik, A. Khan, and M. S. Javed, “Short term load forecasting using bootstrap
aggregating based ensemble artificial neural network,” Recent Advances in Electrical and Electronic Engineering (Formerly Recent
Patents on Electrical and Electronic Engineering), vol. 13, no. 7, pp. 980–992, Nov. 2020, doi:
10.2174/2213111607666191111095329.
Smart traffic forecasting: leveraging adaptive machine learning and big data analytics … (Idriss Moumen)
2332  ISSN: 2252-8938

[38] J. Luo, Z. Zhang, Y. Fu, and F. Rao, “Time series prediction of COVID-19 transmission in America using LSTM and XGBoost
algorithms,” Results in Physics, vol. 27, pp. 1-9, Aug. 2021, doi: 10.1016/j.rinp.2021.104462.
[39] R. Yu et al., “LSTM-EFG for wind power forecasting based on sequential correlation features,” Future Generation Computer
Systems, vol. 93, pp. 33–42, Apr. 2019, doi: 10.1016/j.future.2018.09.054.
[40] C. Xie et al., “Trend analysis and forecast of daily reported incidence of hand, foot and mouth disease in Hubei, China by Prophet
model,” Scientific Reports, vol. 11, no. 1, pp. 1-8, Jan. 2021, doi: 10.1038/s41598-021-81100-2.
[41] B. Rostami-Tabar and J. F. Rendon-Sanchez, “Forecasting COVID-19 daily cases using phone call data,” Applied Soft Computing,
vol. 100, pp. 1-11, Mar. 2021, doi: 10.1016/j.asoc.2020.106932.

BIOGRAPHIES OF AUTHORS

Idriss Moumen born in 1990 in Kenitra. He received his Master’s degree in

computer science, Computer Sciences Researches from Ibn Tofail University, Kenitra, Morocco.
He is a Ph.D. student in Computer Research Laboratory (LaRI) at Ibn Tofail. His research
interests include big data, data mining, ML, DL, and distributed computing. He can be contacted
at email: [email protected].

Jaafar Abouchabaka was born in Guersif, Morocco, 1968. He has obtained two
doctorates in Computer Sciences applied to mathematics from Mohammed V University, Rabat,
Morocco. Currently, he is a professor at Department of computer Sciences, Ibn Tofail University,
Kenitra, Morocco. His research interests are in concurrent and parallel programming, distributed
systems, multi agent systems, genetics algorithms, big data, and cloud computing. She can be
contacted at email: [email protected].

Najat Rafalia was born in Kenitra, Morocco, 1968. She has obtained three doctorates
in Computer Sciences from Mohammed V University, Rabat, Morocco by collaboration with
ENSEEIHT, Toulouse, France, and Ibn Tofail University, Kenitra, Morocco. Currently, she is a
professor at Department of Computer Sciences, Ibn Tofail University, Kenitra, Morocco. Her
research interests are in distributed systems, multi-agent systems, concurrent and parallel
programming, communication, security, big data, and cloud computing. He can be contacted at
email: [email protected].

Int J Artif Intell, Vol. 13, No. 2, June 2024: 2323-2332