AI-Powered Urban Mobility Analysis for Advanced Traffic Flow

Forecasting

Sarah Di Grande

, Mariaelena Berlotti

and Salvatore Cavalieri

Department of Electrical Electronic and Computer Engineering, University of Catania, Viale A.Doria n.6, Catania, Italy

Keywords: Traffic Flow, Forecasting, Machine Learning, Clustering.

Abstract: Rapid global urbanization has resulted in burgeoning metropolitan populations, posing significant challenges

for managing transportation infrastructure. Despite various attempts to address these issues, persistent

challenges hinder urban growth. This study emphasizes the crucial need for effective traffic flow forecasting

in city traffic management systems, with Catania serving as a case study due to its notable traffic congestion.

Predicting traffic flow encounters obstacles, such as the cost and feasibility of deploying sensors across all

roads. To overcome this, the authors suggest an innovative two-level machine learning approach, involving

an unsupervised clustering model to extract patterns from extensive sensor-generated big data, followed by

supervised machine learning models forecasting traffic within individual clusters. Notably, this method allows

predictions for roads without sensor data by leveraging a small subset of alternative data sources.

1 INTRODUCTION

According to recent studies, more than half of the

population of the world currently resides in cities and,

in a few decades, this percentage is expected to rise

(ONU, 2019). This ever-increasing urban population

has led to an exponential rise in the number of

vehicles, putting transport systems under enormous

pressure and causing problems such as congestion

control, increased travel times, traffic, accidents, and

traffic law violations (Xu et al., 2020). Despite the

many attempts to mitigate these problems, traffic

congestion with its associated issues persists and

slows down the development of urban areas.

In recent years, the evolution of big data

technology has revolutionized problem-solving in

transportation (Abouaïssa et al., 2016). The field of

the Internet of Things (IoT) within Information and

Communication Technologies has gained

prominence thanks to possibility of creating a web of

interconnected devices accessible via the Internet.

This network facilitates easy data exchange through

various communication channels like Wi-Fi, RFID,

WSN, NFC, Bluetooth, and more (Swarnamugi and

Chinnaiyan, 2018). The proliferation of connected

https://orcid.org/0009-0008-8895-2175

https://orcid.org/0009-0007-6564-704X

https://orcid.org/0000-0001-9077-3688

devices in smart city setups contributes to an

exponential increase in collected data volumes

(Zantalis et al., 2019). The growth of computational

technologies coupled with the progressive

development of models for the analysis of the

abundant data, facilitates the development of

sophisticated algorithms crucial for traffic analysis

In the present paper, the authors propose a

machine learning approach to predict traffic flow

having input data available from sensors distributed

around the transportation system of an urban

scenario. This paper presents an extended version of

the work developed in (Berlotti et al., 2023). The

authors have enhanced the earlier research by

introducing a more intricate model, trained using one

year of data instead of the initial 3-month period. This

model is capable of detecting diverse patterns,

considering also variations across different months.

Furthermore, additional experiments were conducted

to test the models during holidays.

The paper is structured as follows: Section 2 will

provide an overview of the state of the art regarding

the paper subject; its content will be finalized to the

novelty of the proposed approach. In Section 3 the

authors provide a detailed explanation of the

Di Grande, S., Berlotti, M. and Cavalieri, S.

AI-Powered Urban Mobility Analysis for Advanced Trafﬁc Flow Forecasting.

DOI: 10.5220/0012625900003714

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2024), pages 57-64

ISBN: 978-989-758-702-3; ISSN: 2184-4968

proposed approach. Section 4 will display the

principal outcomes of the proposal testing. Finally, in

Section 5 concluding remarks will summarize the

contents of the paper.

2 RELATED WORKS

In this section, the authors provide an overview of the

methods for the forecasting of traffic flow existing in

the current literature to identify the differences with

respect to the approach presented in this study.

In the literature, current traffic flow prediction

methods are broadly categorized into three groups.

The first category comprises statistical methods based

on mathematical theory. For instance, (J. Liu and

Guan, 2004) proposed a History Average Model (HA

model) for static prediction in urban traffic control

systems. Instead, according to (Lin et al., 2009) the

Autoregressive Integrated Moving Average model

(ARIMA model) is suitable for predicting stable

traffic flow by considering the sequence as a random

time sequence (Zhou et al., 2020). The second

category involves machine learning (ML) techniques

such as Regression analysis (Zhou et al., 2020) and

Boosting algorithms (Y. Liu et al., 2020) like

LightGBM (Chen and Guestrin, 2016), and CatBoost

(Ke et al., 2017), often used to identify patterns within

historical data progression and to forecasting and

regression problems. Finally, the third category

encompasses Deep Learning (DL) techniques,

particularly neural networks like Back Propagation

(BP) (Vijayalakshmi et al., 2021) and Long Short-

Term Neural Network (LSTM) (Li et al., 2020).

Models such as ST-ResNet (Ma et al., 2015) and

spatiotemporal graph convolutional networks

(ASTGCN) (J. Zhang et al., 2017) use various

architectures to predict traffic flow by modeling

congested traffic and attention mechanisms.

The approach of the authors predominantly relies

on the CatBoost model, employed differently from

existing literature, as elaborated further in subsequent

sections. In the present study, the input data given to

the model are obtained from sensors installed on

urban roads. Clearly, installing the sensors on all the

roads of an urban scenario is not possible due to costs

and other practical reasons. To address the challenge

of expensive and limited sensor installations on every

road, the authors use the data collected by sensors

installed in a subset of roads to predict both traffic on

the same roads and on roads lacking of sensor data.

Most current approaches in literature tries to

address this challenge by examining spatio-temporal

characteristics between neighboring and distant

sensors to predict traffic flow in urban areas lacking

from data but similar to the ones of the collected data

(Guo et al., 2019). For example, (Y. Zhang et al.,

2023) introduced a spatio-temporal traffic flow

estimation model that utilizes data from multiple

locations within the network. The approach

incorporates various features beyond solely relying

on traffic flow data.

The approach of the authors revolves around

utilizing a two-level machine learning method using

only traffic flow data. An unsupervised clustering

model organizes sensor data into clusters, while a

supervised machine learning model predicts traffic

flow for each cluster. This approach involves

assigning roads to clusters using distance metrics,

enabling precise prediction by employing specific

forecasting models trained on comprehensive sensor

data. Section 3 will delve into a detailed description

of this proposed approach.

3 PROPOSED APPROACH

In this section, the authors will provide details

regarding the proposed approach. In the analysis real

data from a network of traffic sensors situated in

Catania, Italy, were utilized. The most important

problem today in the traffic flow of Catania is

congestion. Over time, the population of the city

expanses, forming a unified urban network that

extends beyond municipal boundaries resulting in

considerable traffic pressure on Catania, in daily

congestion in the central area and in amplified

environmental pollution levels.

This situation intensified substantially leading to

the critical need to effectively manage traffic flow in

Catania through the implementation of forecasting

methodologies.

As previously stated, the primary idea proposed

by the authors is to use a two-level machine learning

approach, combining unsupervised and supervised

models. First, an unsupervised model is utilized to

extract patterns from the traffic flow time series

collected from sensors, organizing them into multiple

clusters. Within each cluster, a supervised machine

learning model will be then developed to predict

traffic flow for each time series belonging to the same

cluster. Using distance metrics enables the allocation

of roads to clusters with minimal observations,

facilitating predictions. Once the relevant cluster for

a specific road segment is identified, a machine

learning model trained on the traffic flow of segments

within that cluster is employed to predict traffic in the

new segment. Essentially, distinct models are created

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

for each cluster, allowing the forecasting of traffic

flow for roads with limited observations sharing

similar patterns.

The results outlined in this paper will underscore

that each model, having been trained on extensive

time series within the same cluster, can effectively

generate forecasts for similar but unseen series

requiring only a minimal number of observations

from these new series. For these roads lacking

sensors, since a small subset of traffic data is crucial

for the forecasting models, it can be obtained from

alternative sources such as Floating Car Data.

3.1 Data Acquisition

Data employed in the model refers to sensors-data

about the traffic flow of Catania city. Located in the

eastern part of Sicily, Italy, Catania has a population

of around 300,000 inhabitants across approximately

183 km2 (Medina-Salgado et al., 2022). The city is

part of a larger metropolitan area encompassing the

main municipality and 26 nearby urban centers.

Sensor-based date have been collected through 21

microwave traffic counters known as MOBILTRAF

300 by FAMAS (www.famassystem.it/it/prodotto/

mobiltraf-300), placed across the Catania urban area.

When any vehicle crosses the electromagnetic field

generated by two MobilTraf300 sensors, these units

capture different vehicle-related data, including the

date and time of passage, the travel direction, and the

specific transit lane. To access and retrieve these data,

FAMAS's traffic manager software, known as

MobilTraf MANAGER, was used.

Twelve traffic counters (TCs)—those that were

operational at the time of data download—were

chosen from among all the ones present. The period

under analysis spans from January 1, 2022, to

December 31, 2022 with data recorded at 5-minute

intervals.

Each traffic counter corresponds to a specific

road, showcasing different characteristics. In the

following section the authors will describe the steps

involved in the preprocessing.

3.2 Data Preprocessing

The roads analysed can be categorized as single-lane

roads, two-lane roads in the same direction, or two-

lane roads in the opposite direction. Based on the

characteristics of each road, distinct data

preprocessing steps were applied. In details, for roads

with two lanes in the same direction, the vehicle

counts from both lanes were summed up in a

consolidated time series representing the total vehicle

count for that road. Conversely, time series related to

roads with two lanes in opposite directions were

disaggregated into distinct time series to capture

information about vehicles traveling in separate

directions on the same road.

Post a pivot transformation, the final dataset was

composed of one column for Timestamp and

additional columns representing the total vehicle

count for each road and direction.

The next step was the data cleaning. Two types of

missing values were identified in the dataset: sensor

malfunctions, when a specific TC broke and failed to

retrieve traffic information, and outliers. Outliers in

the time series were detected using boxplots and

replaced with missing values.

To address missing observations, the technique

chosen involves filling in missing values using a

time-based averaging method. This function

calculates the mean of traffic values for the same

road, day of the week, and time within the same

month.

Lastly, aiming to train the machine learning

model with hourly data, data have been aggregated

per hour using the sum as the aggregation function,

resulting in the total vehicle count recorded for a

specific street per hour.

The final dataset comprised 15 columns and 8769

rows, encompassing all the hours of the day across

365 days, equating to one year of data.

3.3 Clustering

The paper aims to create a ML solution to accurately

predict traffic flows both on the roads with sensors

and on the ones where sensors are not installed. To do

this, the authors use a machine learning model trained

on a set of sensors sharing similar characteristics to

forecast traffic flow on a road with a very limited

number of observations possessing resemblances to

the sensor-equipped group. Consequently, a

clustering step is applied.

The study utilizes Time Series K-means

(TSkmeans), an adapted version of the traditional K-

means algorithm designed specifically for clustering

time series data (Huang et al., 2016). In contrast to

standard K-means, which focuses solely on data point

values, TSkmeans incorporates temporal

relationships, considering both values and their

temporal aspects in cluster formation. Notably,

TSkmeans employs the Dynamic Time Warping

(DTW) metric instead of the conventional Euclidean

distance for measuring similarity among temporal

sequences. The initial step in TSKmeans clustering

involves determining the appropriate number of

AI-Powered Urban Mobility Analysis for Advanced Trafﬁc Flow Forecasting

clusters (K), achieved through the use of a silhouette

score.

Before clustering, since the time series considered

exhibit widely varying value ranges, data

normalization was needed. Normalizing the data

enables to establish a uniform baseline to prevent the

clustering algorithm from reacting to feature scales.

The normalization technique used in the analysis is

Min-Max scaling, which transforms the range of each

variable to a standardized 0-1 scale.

After clustering data, the next step for the analysis

involves the creation of the forecasting model; next

paragraph will describe more in detail this step.

3.4 Forecasting

The purpose of this step of the analysis is to find out

the most suitable machine learning algorithm for

traffic flow forecasting. All the machine learning

models proposed, were implemented using Darts,

Python library. (Time Series Made Easy in Python —

Darts Documentation, n.d.)

First, the dataset was divided into a training set

spanning from January 1, 2022, to December 16,

2022, and a test set spanning from December 17,

2022, to December 31, 2022.

Next, the Catboost algorithm was compared with

various machine learning algorithms, with default

hyperparameters.

The authors considered the following metrics to

evaluate models’ performances: mean absolute error

(MAE) (Prokhorenkova et al., 2019), symmetric

mean absolute percentage error (SMAPE), mean

squared error (MSE) (Dorogush et al., 2018) and the

root mean square error (RMSE). For each of these

metrics, lower values denote better model

performance. It is important to note that while

SMAPE is the main performance metric used to

choose the best model, other metrics like MAE, MSE,

and RMSE are also taken into consideration as

supporting indicators during the evaluation process.

According to all these metrics, Catboost emerged

as the best-performing algorithm and was considered

to proceed with the analysis.

Proposed by (Herzen et al., 2023), the CatBoost

algorithm is a Gradient Boosting Decision Tree

(GBDT) framework that merges weak learners as

symmetric decision tree, to generate a stronger

predictive model. Ensemble methods like CatBoost

process sequentially a series of simple decision trees,

trying to reduce the errors done in the models

previously trained for optimizing performances.

To test the approach multiple times, different

models were trained repeatedly, leaving out one

specific time series from the training data each time,

and then evaluating the model's performance based

on the omitted time series. The purpose of this

methodology is to assess robustness and

generalization capabilities of the approach training

models on various combinations of the available time

series data.

Different CatBoost models were created and

tested for different sets of hyperparameters, using

Optuna Python library (Optuna: A Hyperparameter

Optimization Framework — Optuna 3.5.0

Documentation, n.d.). A total of 100 trials were used

to create and compare 100 different models.

The chosen objective function to be optimized

for each training set was the validation loss, used to

quantify the performance of machine learning models

on a validation dataset during hyperparameter

optimization. The last 24 hours of the training set

were used for validating the model.

Walk-forward validation method was

implemented as a validation technique. This

validation method stands out from standard cross-

validation approaches by maintaining the temporal

order of data, making it particularly suitable for

capturing time-dependent patterns in time series data.

The validation process initiates with an initial

training period covering historical time series data

from January to December 16

. Subsequently, the

model undergoes iterative phases, where it is

retrained and makes predictions for upcoming time

steps within the sequence. The performance

assessment occurs continuously as predictions are

compared with actual values, mimicking the dynamic

nature of real-world scenarios. This regular retraining

process enables the model to adapt dynamically to

evolving data distributions or patterns over time,

thereby significantly enhancing its practical efficacy.

4 RESULTS

In this section, the authors will present the results and

discuss the outcomes obtained from the clustering and

forecasting phases.

4.1 Clustering

As a result of the clustering process, a specific

configuration emerged, yielding a silhouette score of

0.52. This outcome is favorable, indicating a

reasonably clear distinction between the clusters

formed. Experts in the field have also validated the

effectiveness of the clustering algorithm in grouping

roads that share similar characteristics.

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

The results of the clustering procedure are

visually depicted in Figure 1. Due to limitations in

space, the figure showcases data collected over just

one week, despite the algorithm utilizing an entire

year of data as input. Upon observation, it becomes

apparent that the first cluster comprises 5 time series,

the second cluster consists of 3 time series, and the

third cluster encompasses 5 time series. The fourth

cluster, comprising only one time series, was omitted

from the visual representation due to its singular

nature.

Figure 1: Time series divided into clusters.

4.2 Models Comparison

Taking as reference clustering results, short-term

forecasting was implemented.

Initially, the authors compared the performances

of different machine learning models to identify the

best performing algorithm.

Two critical hyperparameters had to be set: the

input chunk length fixed at 168 hours (equivalent to 7

days), and the output chunk length was set to 24

hours. Essentially, this configuration means that the

model utilized data from the previous 7 days to

predict the forthcoming 24 hours. All the others

hyperparameters were set to the default values. The

results are documented in Table I.

Table 1: Performances comparison with default

hyperparameters.

Algorithm

Performance Metrics

SMAPE MAE MSE RMSE

LR 33.301 0.082 36.565 0.112

LSTM 199.406 0.571 125.61 0.596

CatBoost 32.985 0.078 28.441 0.013

DLinear 33.915 0.084 31.366 0.014

LightGBM 33.8 0.08 26.767 0.014

N-Hits 34.328 0.084 31.585 0.014

Transf 38.71 0.099 36.165 0.019

N-Beats 34.944 0.087 32.011 0.015

B-RNN 51.067 0.152 68.324 0.034

N-Linear 34.385 0.088 37.116 0.014

TCN 40.699 0.118 39.194 0.022

TiDe 35.631 0.095 49.398 0.015

As can be seen from Table 1, CatBoost emerged

as the best-performing algorithm, leading to its

selection for creating the final models.

Next step of the analysis was the optimization of

the CatBoost model. As explained in Section 3.4,

each time an optimized forecasting model is created

for a cluster, a time series belonging to that group is

excluded from the training to be used as a test.

It is important to note that upon examining the

resulting optimized models for the three clusters, it is

evident that the longest duration needed to generate

forecasts is 383 hours, equivalent to approximately 16

days, a minimal number of observations.

4.3 Forecasting

As previously said in Section 3.5, the CatBoost model

was tested on two weeks comprising data from 17th,

to 31

December, 2022. This choice was dictated by

the fact that in Italy last week of December is

Christmas week, during which traffic flow is different

from the normal.

Table 2 and Table 3 show the results obtained for

each sensor in the three clusters, for the roads

included in the training of the model.

AI-Powered Urban Mobility Analysis for Advanced Trafﬁc Flow Forecasting

Table 2: Performances test week from 17th to 23rd Dember

2022 for roads included in the training.

Sensor

Performance Metrics

MAE SMAPE MSE RMSE

MT10a 0.0519 16.7697

0.0044 0.0133

MT10

0.0483 9.9464 0.0047 0.0124

MT6

0.0568 18.5326 0.0061 0.0139

MT6

0.0451 14.1015 0.0040 0.0115

MT7

0.0481 14.4138 0.0042 0.0138

MT13a 0.1258 37.1682 0.0322 0.0416

MT13

0.1223 41.9823 0.0273 0.0425

MT17a 0.0963 31.4148 0.0163 0.0333

MT14a 0.0423 25.3374 0.0056 0.0134

MT14

0.0878 34.5072 0.0123 0.0265

MT18

0.0473 19.2189 0.0064 0.0194

MT9

0.0732 24.1288 0.0124 0.0218

MT9

0.0981 44.1747 0.0185 0.0343

Table 3: Performances test week from 24th to 31st

December 2022 for roads included in the training.

Sensor

Performance metrics

MAE SMAPE MSE RMSE

MT10a 0.0705 20.5463 0.0093 0.0342

MT10

0.0568 13.6693 0.0081 0.0287

MT6

0.0654 20.0405 0.0088 0.0345

MT6

0.0555 17.7132 0.0064 0.0306

MT7

0.0704 26.9439 0.0114 0.0438

MT13a 0.0980 27.7650 0.0144 0.0503

MT13

0.1015 32.6417 0.0187 0.0544

MT17a 0.1116 38.1443 0.0178 0.0500

MT14a 0.0531 36.9480 0.0082 0.0246

MT14

0.0764 30.1027 0.0112 0.0337

MT18

0.0682 32.9566 0.0128 0.0365

MT9

0.0811 36.6931 0.0168 0.0399

MT9

0.0865 52.2013 0.0164 0.0376

The next step of the analysis was to test optimized

models each time on the excluded time series.

Table 4 and 5 report the average performance

metrics computed each time a time series was

excluded from the three clusters, for the two test

weeks going from 17

to 31

December, 2022.

Table 4: Average performances test week from 17th to 23rd

December 2022 for roads excluded in the training.

Performance Metrics

MAE SMAPE MSE RMSE

1 0.0557 17.6324 0.0059 0.0169

2 0.1209 36.8653 0.0259 0. 0388

3 0.0740 31.6346 0.0116 0.2333

Table 5: Average performances test week from 24th to 31st

December 2022 for roads excluded in the training.

Performance Metrics

MAE SMAPE MSE RMSE

1 0. 0673 21.6607 0.0093 0.0298

2 0. 1110 34.3332 0.0190 0.0560

3 0. 0732 39.9600 0.0131 0.3142

Figures 2-4 display the true traffic values versus

the traffic flow predicted by CatBoost models during

the two test weeks from 17

to 31

December 2022,

on sensors that were excluded from the training

dataset.

Figure 2: Test excluded sensor cluster 1.

Figure 3: Test excluded sensor cluster 2.

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

Figure 4: Test excluded sensor cluster 3.

As can be seen, the optimized CatBoost models

tested for each cluster show relatively low SMAPE

values as concerns the tested week from 17

to 23

December, 2022. Instead, for the second tested week

which is the Christmas week, SMAPE tends to

increase suggesting that the predictions of the models

have some degree of error.

Results obtained depend on a model that has been

trained for one year. Thus, the model has never seen

during the training traffic flow patterns generated in

every street during the Christmas week. Moreover,

the only variable considered is the traffic flows.

Knowing such a limited time range, the results must

be considered impressive.

5 CONCLUSIONS

In this study, the authors propose a solution for a

traffic flow prediction both for roads where sensors

data are available and roads lacking from data for cost

and practicality reasons of sensors’ deployment on

every road. The authors address it with a novel two-

level machine learning approach, involving clustering

and forecasting models. The city of reference is

Catania, because of its complex transportation

network.

Using TSKMeans algorithm, time series were

divided into different clusters, highlighting not only

roads with similar patterns but also roads with similar

physical characteristics, as confirmed by domain

experts.

The forecasting process, where a distinct model

was generated for each cluster, yielded outstanding

outcomes when applied to the time series used in

training, employing the CatBoost algorithm.

Moreover, a tailored parameter optimization process

for each cluster facilitated the customized

configuration of hyperparameters.

Finally, this approach enables predictions for

roads lacking sensor data by utilizing a really small

subset of these new data, needing in input ranges

between 199 and 383 hours.

Future works plan to repeat this study with a

greater time range of data, to make the CatBoost

model more accurate in making predictions in the

presence of traffic flow patterns different from

normal, as it could happen during Christmas week.

Moreover, the authors plan to increase the number of

sensors considered in the analysis. Furthermore, data

from different sources (e.g. weather data, road

conditions as traffic jams and road works) will be

collected and given to the model to improve

forecasting.

ACKNOWLEDGEMENTS

The results presented in this paper have been

achieved under the research project of the Spoke 9

“Digital Society & Smart Cities” inside the “National

Centre for HPC, Big Data and Quantum Computing”

(Code CN00000013, CUP E63C22001000006). This

research project is currently running and is funded

under PNRR M4C2 Line 1.4, by Italian Ministry for

Research.

Data analyzed in the paper were provided by the

“ITS Laboratory” of the Department of Civil and

Architecture Engineering of Catania. Specifically, the

laboratory created within the project RE.S.E.T, is a

traffic monitoring, estimation and short-term

forecasting system, equipped with radar sensor and a

central control station for traffic data elaborations.

REFERENCES

Abouaïssa, H., Fliess, M., and Join, C. (2016). On short-

term traffic flow forecasting and its reliability. IFAC-

PapersOnLine, 49(12), 111–116. https://doi.org/10.101

6/j.ifacol.2016.07.559

Berlotti, M., Di Grande, S., Cavalieri, S., Torrisi, V., and

Inturri, G. (2023). Proposal of an AI based approach for

Urban Traffic Prediction from Mobility Data. IEEE

International Conference on Big Data (Big Data),

Sorrento, Italy. In press.

Chen, T., and Guestrin, C. (2016). XGBoost: A Scalable

Tree Boosting System. Proceedings of the 22nd ACM

SIGKDD International Conference on Knowledge

Discovery and Data Mining, 785–794. https://doi.org/

10.1145/2939672.2939785

Dorogush, A. V., Ershov, V., and Gulin, A. (2018).

CatBoost: Gradient boosting with categorical features

support (arXiv:1810.11363). arXiv. https://doi.org/

10.48550/arXiv.1810.11363

AI-Powered Urban Mobility Analysis for Advanced Trafﬁc Flow Forecasting

Guo, S., Lin, Y., Feng, N., Song, C., and Wan, H. (2019).

Attention Based Spatial-Temporal Graph

Convolutional Networks for Traffic Flow Forecasting.

Proceedings of the AAAI Conference on Artificial

Intelligence, 33(01), Article 01. https://doi.org/10.16

09/aaai.v33i01.3301922

Herzen, J., Lässig, F., Piazzetta, S. G., Neuer, T., Tafti, L.,

Raille, G., Van Pottelbergh, T., Pasieka, M., Skrodzki,

A., Huguenin, N., Dumonal, M., Kościsz, J., Bader, D.,

Gusset, F., Benheddi, M., Williamson, C., Kosinski,

M., Petrik, M., and Grosch, G. (2023). Darts: User-

friendly modern machine learning for time series. The

Journal of Machine Learning Research, 23(1),

124:5442-124:5447.

Huang, X., Ye, Y., Xiong, L., Lau, R. Y. K., Jiang, N., and

Wang, S. (2016). Time series k-means: A new k-means

type smooth subspace clustering for time series data.

Information Sciences, 367–368, 1–13. https://doi.org/

10.1016/j.ins.2016.05.040

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W.,

Ye, Q., and Liu, T.-Y. (2017). LightGBM: A Highly

Efficient Gradient Boosting Decision Tree. Advances

in Neural Information Processing Systems, 30.

https://papers.nips.cc/paper_files/paper/2017/hash/644

9f44a102fde848669bdd9eb6b76fa-Abstract.html

Li, Y., Huang, J., and Chen, H. (2020). Time Series

Prediction of Wireless Network Traffic Flow Based on

Wavelet Analysis and BP Neural Network. Journal of

Physics: Conference Series, 1533(3), 032098.

https://doi.org/10.1088/1742-6596/1533/3/032098

Lin, S.-L., Huang, H.-Q., Zhu, D.-Q., and Wang, T.-Z.

(2009). The application of space-time ARIMA model

on traffic flow forecasting. 2009 International

Conference on Machine Learning and Cybernetics, 6,

3408–3412. https://doi.org/10.1109/ICMLC.2009.5212

785

Liu, J., and Guan, W. (2004). A Summary of Traffic Flow

Forecasting Methods. Journal of Highway and

Transportation Research and Development, 21(3),

Article 3.

Liu, Y., Liu, C., and Zheng, Z. (2020). Traffic Congestion

and Duration Prediction Model Based on Regression

Analysis and Survival Analysis. Open Journal of

Business and Management, 8(2), Article 2.

https://doi.org/10.4236/ojbm.2020.82059

Ma, X., Tao, Z., Wang, Y., Yu, H., and Wang, Y. (2015).

Long short-term memory neural network for traffic

speed prediction using remote microwave sensor data.

Transportation Research Part C: Emerging

Technologies, 54, 187–197. https://doi.org/10.1016/

j.trc.2015.03.014

Majumdar, S., Subhani, M. M., Roullier, B., Anjum, A., and

Zhu, R. (2021). Congestion prediction for smart

sustainable cities using IoT and machine learning

approaches. Sustainable Cities and Society, 64, 102500.

https://doi.org/10.1016/j.scs.2020.102500

Medina-Salgado, B., Sánchez-DelaCruz, E., Pozos-Parra,

P., and Sierra, J. E. (2022). Urban traffic flow

prediction techniques: A review. Sustainable

Computing: Informatics and Systems, 35, 100739.

https://doi.org/10.1016/j.suscom.2022.100739

ONU. (2019). World Population Prospects 2019:

Highlights. United Nations Department for Economic

and Social Affairs, New York (US).

Optuna: A hyperparameter optimization framework—

Optuna 3.5.0 documentation. (n.d.). Retrieved 12

December 2023, from https://optuna.readthedocs.io/en/

stable/index.html

Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A.

V., and Gulin, A. (2019). CatBoost: Unbiased boosting

with categorical features (arXiv:1706.09516). arXiv.

https://doi.org/10.48550/arXiv.1706.09516

Swarnamugi, M., and Chinnaiyan, R. (2018). IoT Hybrid

Computing Model for Intelligent Transportation

System (ITS). 2018 Second International Conference

on Computing Methodologies and Communication

(ICCMC), 802–806. https://doi.org/10.1109/ICCMC.2

018.8487843

Time Series Made Easy in Python—Darts documentation.

(n.d.). Retrieved 16 October 2023, from

https://unit8co.github.io/darts/

Vijayalakshmi, B., Ramar, K., Jhanjhi, Nz., Verma, S.,

Kaliappan, M., Vijayalakshmi, K., Vimal, S., Kavita,

and Ghosh, U. (2021). An attention-based deep learning

model for traffic flow prediction using spatiotemporal

features towards sustainable smart city. International

Journal of Communication Systems, 34(3), e4609.

https://doi.org/10.1002/dac.4609

Xu, T., Han, G., Qi, X., Du, J., Lin, C., and Shu, L. (2020).

A Hybrid Machine Learning Model for Demand

Prediction of Edge-Computing-Based Bike-Sharing

System Using Internet of Things. IEEE Internet of

Things Journal, 7(8), 7345–7356. https://doi.org/

10.1109/JIOT.2020.2983089

Zantalis, F., Koulouras, G., Karabetsos, S., and Kandris, D.

(2019). A Review of Machine Learning and IoT in

Smart Transportation. Future Internet, 11(4), 94.

https://doi.org/10.3390/fi11040094

Zhang, J., Zheng, Y., and Qi, D. (2017). Deep Spatio-

Temporal Residual Networks for Citywide Crowd

Flows Prediction (arXiv:1610.00081). arXiv.

https://doi.org/10.48550/arXiv.1610.00081

Zhang, Y., Cheng, Q., Liu, Y., and Liu, Z. (2023). Full-

scale spatio-temporal traffic flow estimation for city-

wide networks: A transfer learning based approach.

Transportmetrica B: Transport Dynamics, 11(1), 869–

895. https://doi.org/10.1080/21680566.2022.2143453

Zhou, J., Yang, Q., Zhang, X., Han, C., and Sun, L. (2020).

Traffic Prediction Method for GEO Satellites

Combining ARIMA Model and Grey Model. Journal of

Shanghai Jiaotong University (Science), 25(1), 65–69.

https://doi.org/10.1007/s12204-019-2152-9

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems