Sustainable Cities and Society 67 (2021) 102702

Contents lists available at ScienceDirect

Sustainable Cities and Society

journal homepage: [Link]/locate/scs

Intelligent management of bike sharing in smart cities using machine

learning and Internet of Things
El Arbi Abdellaoui Alaoui a, b, *, Stephane Cedric Koumetio Tekouabou c, d
a
EIGSI-Casablanca, 282 Route of the Oasis, Casablanca, Morocco
b
My Ismail University, Morocco
c
Center of Urban Systems (CUS), Mohammed VI Polytechnic University (UM6P), Hay Moulay Rachid, 43150 Ben Guerir, Morocco
d
Laboratory LAROSERI, Department of Computer Science, Faculty of Sciences, B.P. 20, 24000 El Jadida, Morocco

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

Keywords: Global ecological requirements are pushing city actors to opt for ecological solutions at all levels, including urban
Bike-sharing system (BSS) mobility. More sustainable Bike-sharing systems (BSS) have become an indispensable part of the transport offer
Management by world’s major metropolis. Like any computerized service system, they generate voluminous and complex data
Prediction
that the use of which is essentially limited to the management and operation of the system. The movements made
Smart cities
Internet of Things (IoT)
by system users can provide valuable information on many aspects of urban life including the spatial and
Regression temporal dynamics of travel in the city, on the place of the bicycle among other modes of transport, or on the
Ensemble models distribution of territorial and social inequalities in geographical space. In this paper, we study the problem of
intelligent management of shared bicycle systems. Indeed, the management of these systems faces many opti­
mization problems in its procession. Thus, to improve the BSS user’s satisfaction, it’s useful to inform the actors/
users in this system about the state of bike sharing for a station. For this, we propose an approach that integrates
in these systems both the new IoT for smart city technologies and machine learning in order to facilitate the task
of management, availability and profitability. In addition, we propose an automatic management system capable
of predicting the number of bikes shared per hour, day or month by taking several dynamic parameters.
Simulation results carried out on real data from London’s bike sharing system demonstrate the effectiveness of
the proposed model.

1. Introduction docked BSS, dockless BSS are centered on bike with a lock that is usually
integrated into the frame without fixed station allowing the user the
Bike-sharing systems have become a key element in urban transport flexibility to pick up and drop off the bike wherever he wants (Chen, van
policies over the last decade, as evidenced by the recent explosion in the Lierop, & Ettema, 2020; Ji, Ma, He, Jin, & Yuan, 2020). In either case,
number of bicycles in circulation in the world’s major cities. By the user can pay a daily, weekly, monthly or annual subscription
providing affordable access to bicycles, they should actively participate available immediately at the station or mobile application. Payment can
in the development of alternatives to motorized vehicles for urban travel be made by direct debit or by card. Each journey entitles the user to a
and contribute to the reduction of air pollution (Cao & Shen, 2019; period of free or very low cost use. Generally this period varies between
Leister, Vairo, Sims, & Bopp, 2018), noise levels and congestion prob­ 30 minutes and 1 hour, beyond which the price increases significantly,
lems affecting the world’s major cities. One of the keys to the success of encouraging short trips. However, the number of journeys is unlimited
the bike sharing system is its ease of use. For the majority of systems, a during the subscription period, allowing users to make several journeys
user has the option of picking up or dropping off a bike at one of the in a row, provided that they can leave the bicycle at a station. Thus,
many stations throughout the city in a fully automated manner. They are actual metropolis bike sharing program (BSP) consists both of docked
known as “docked BSS”. In recent years, an alternative based on dockless and dockless BSS. Predicting the number of shared biked over a time in
BSS has emerged in several metropolises, although remaining less such a metropolis is a challenge for the BSP (Bikeshare, 2013; Garcia-­
widely used than the first one. Instead of a fixed station as in the case of Gutierrez, Romero-Torres, & Gaytan-Iniestra, 2014).

* Corresponding author.
E-mail address: abdellaoui.e@[Link] (E.A. Abdellaoui Alaoui).

[Link]
Received 12 April 2020; Received in revised form 14 November 2020; Accepted 29 December 2020
Available online 9 January 2021
2210-6707/© 2021 Elsevier Ltd. All rights reserved.
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

BSS are a new generation of traditional bike rentals where the whole • The global system uses ensemble methods to predict the number of
process of joining, renting and returning bikes has become automatic. bike sharing in smart city;
Thanks to these systems, the user can easily rent a bike from a particular • The optimization of different ensemble-based models for predicting
position and return to another position. Currently, there are about 500 number of bike sharing in smart city;
bike-sharing programs worldwide, with over 500,000 bikes shared.
Today, there is great interest in these systems because of their important The rest of this paper is organized as follows: In Sections 1 and 3.1 we
role in traffic, environmental and health issues (Cao & Shen, 2019). first present respectively previous work on bike sharing and bake
In addition to the interesting real-world applications of bike-sharing, sharing approach for smart cities. We then describe the proposed
the characteristics of the data generated by these systems make them method to predict the number of bike shared in Section 4 before
attractive for research in sustainable urban planning (Lu, An, Hsu, & analyzing the results obtained in Section 5. Finally, we discuss the re­
Zhu, 2019). In contrast to other transport services such as bus or metro, sults of our model and its limitations in Section 6 and conclude our study
travel time, departure and arrival positions are explicitly recorded in in Section 7.
these systems. This features transform the BSP into a virtual sensoring
network that can be used to detect mobility in the city using new 2. Related work
technologies such as the Internet of Things (IoT). Indeed, BSP and the
IoT are strongly linked. Currently, most BSP use 2Gs for networking, but In parallel with the explosion of bike sharing programs around the
these systems will gradually move to narrow-band networks of the world, research on bike sharing has seen a burgeoning interest from
Internet of Things (NB-IoT) in the future (Chen, Miao, Hao, & Hwang, transportation researchers and beyond. The synthesis of the literature on
2017). Several bike-sharing companies have started using the NB-IoT bike sharing systems by Fishman et al., first in 2013 (Fishman, Wash­
(Sinha, Wei, & Hwang, 2017). Recently a study has been conducted ington, & Haworth, 2013) and again in 2015 (Fishman, 2016), reveals
by Zguira, Rivano, and Meddeb (2018) to facilitate the collection of BSP both the wealth of scientific publications on the subject and the diversity
data through the IoT. They proposed an efficient routing protocol, of research topics covered. These studies thus examine the many ques­
IoB-DTN, based on data aggregation that applies the Delay Tolerant tions raised by this new mode of transport, whether they are:
Network (DTN) for IoT applications which collect data from sensor (i) technical: does the system work properly? is it optimal? can it be
networks based on urban bike sharing systems (Zguira et al., 2018). improved and how?,
Therefore, most significant events in the city should be detected by (ii) Social: who uses BSS and why? How to improve access to BSS?
monitoring this data using the services of a reliable network. (iii) in relation to transport: what is the impact of BSS systems on
BSP typically aim to reduce traffic congestion, noise and air pollution other modes of transport?
(Bullock, Brereton, & Bailey, 2017; Cao & Shen, 2019) by providing (iv) or dealing with economic aspects, such as what the activity of
free/affordable access to bicycles for short-distance travel in an urban bike sharing program tells us about the economic activity of a city.
area as opposed to motorized vehicles. The number of users on a given This enthusiasm can be explained by the need, both for political
day can vary considerably for these systems. The ability to forecast the decision-makers and the companies managing such systems, to know
number of hourly users can enable entities such as companies and and describe BSS systems in order to develop and support the use of bike
governments that oversee these systems to manage them more effi­ sharing in the city.
ciently and cost-effectively. Predicting shared bikes is also very impor­ In addition, unlike traditional transportation systems that require the
tant for urban planning and business model for companies investing in establishment of often cumbersome and costly infrastructure (like using
shared bike programs in urban areas (Bikeshare, 2013). Our goal is to sensors or field surveys in order to obtain partial system data), the
use and optimize machine learning based models that effectively predict technologies employed in the third-generation BSS currently in opera­
the total number of shared bikes that will be used during a given hour, a tion provide rapid access to digital data on multiple aspects of bike
given month, a given season, etc., using the information available on sharing. This data, provided by the operators of “JCDecaux developer”
that hour/day within the bike system, taking into account several factors platform Jcdecaux developer (2020) allows, for example, access to the
(Eren & Uz, 2019; Mattson & Godavarthy, 2017). Thus, it is useful to status of the stations of all the systems managed by cyclocity in real time.
inform the actors/users in this system about the bike sharing status for a It has fostered collaboration between specialists in socio-economic is­
station or around a user’s location. sues, who until then had been dealing with these problems, and re­
In this paper, we propose a combination of IoT and machine learning searchers in data analysis (signal processing, data mining, etc.), in order
techniques through regression ensemble methods to optimize the man­ to find answers to the problems mentioned above, in data sets that can
agement of self-service shared bike systems in smarts cities that we also be large and require sophisticated techniques.
want to be green and sustainable cities. This type of combination has
already shown good performance in the case of intelligent parking ga­ 2.1. Bike sharing as a mode of transport
rages (Tekouabou, Cherif, & Silkan, 2020). Indeed if the IoT allows to
collect a large variety of data from the different dynamic sources asso­ Beyond surveys carried out by institutional bodies (Transportation,
ciated to our BSS, after data cleaning and transformation, the robust 2013) or the operators themselves (Bikeshare, 2013; London, 2013)
ensemble methods could allow to predict the number of shared bikes per consisting of qualitative or quantitative descriptions of the use of the
hour, per day, per month, etc., thus allowing to optimize and automate system and its management, the place of bike sharing among urban
the efficient management of the system. We now have a volume of data travel modes has been the subject of many studies, seeking to under­
that allows us to study the external environment with precision. The city stand the reasons for the success of BSP. They relate to the analysis of a
is fully digitized and the amount of data produced in the cities will particular system, such as Montreal (Morency, Trepanier, & Godefroy,
continue to double every 2 years. These data are all integrated and 2011), Hangzhou (Shaheen, Zhang, Martin, & Guzman, 2011), Bangkok
processed automatically by the prediction algorithms. They enable (Ueasangkomsate, 2020), London (Noland & Ishaque, 2006) or Dublin
predictions to be made such as the fill rate of a bike share station, system (Tuama, 2015), or comparative studies between several systems in
or program; the time it takes to find a parking space on the road, public North America, Western Europe and Asia (Lohry, 2020; Midgley, 2009;
transit trips and even the comfort or stress felt by pedestrians. Parkes, Marsden, Shaheen, & Cohen, 2013; Shaheen, Guzman, & Zhang,
The key contributions of this work are: 2010).
The cohabitation of bike shared with other modes of transportation
• The integration of an IoT-based system in smart cities and particu­ and particularly the effects of BSP on the reduction of car use have been
larly in the case of bike sharing; scrutinised: urban agglomerations are seeking to promote cycling as part

2
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

of policies to control energy consumption and greenhouse gas emissions, temporal distribution of activities. Most of these studies highlight re­
mainly from motorised transport (DeMaio, 2009). Several studies have lationships between the use of bicycles, the time of day and a
shown the low impact of bike sharing on vehicle use (Bikeshare, 2013; geographical and socio-economic description of the city.
Midgley, 2011; Murphy, 2010; Yang, Haixiao, & Qing, 2020), through Several efforts have also been made, particularly by research groups
field surveys. The vast majority of bike shared user’s are those of soft from the exact sciences, to forecast demand in order to plan the system
transport modes (public transport, walking, etc.). The role of BSP in the and anticipate its development (where to locate stations in the city, how
promotion of cycling in cities has also been studied through the transi­ many locations per station, etc.), both through theoretical (Bortolussi &
tion between shared and personal bikes (Castillo-Manzano, Hillston, 2015; Chen & Sun, 2015; Dell’Olio, Ibeas, & Moura, 2020;
Castro-Nuno, & Lopez-Valpuesta, 2015) or factors influencing the García-Palomares, Gutiérrez, & Latorre, 2012; Krykewycz, Puchalsky,
practice of shared bikes (Kim, Shin, Im, & Park, 2020). The health Rocks, Bonnette, & Jaskiewicz, 2010; Lin & Yang, 2011; Saharidis,
benefits of cycling induced by the use of BSS have also been the subject Fragkogios, & Zygouri, 2020) and data-driven approaches as in Lyon
of several studies (Doorley, Pakrashi, & Ghosh, 2015; Fishman, Wash­ (Borgnat et al., 2011, 2013), in Hangzhou (Woodcock et al., 2014) or in
ington, & Haworth, 2012; Rojas-Rueda, de Nazelle, Tainio, & Nieu­ Montreal (Faghih-Imani, Eluru, El-Geneidy, Rabbat, & Haq, 2014). The
wenhuijsen, 2011; Woodcock, Tainio, Cheshire, O’Brien, & Goodman, latter problem was recently the subject of a competition held in
2014), and the London (Woodcock et al., 2014) example illustrates the conjunction with a conference on data mining, reflecting the popularity
overall positive impact of cycling in cities on certain categories of of this topic (Subramaniyan, 2020). Other issues related to the smooth
population. The characterization of users, and their motivations, are also day-to-day operation of the system (bike maintenance, station
at the origin of many studies, mainly conducted by sociological re­ balancing, etc.) were also studied, once again through the development
searchers. This work has been carried out either through individual of theoretical models (Nair & Miller-Hooks, 2011; Sayarshad, Tavassoli,
surveys, as in Montreal (Bachand-Marleau, Lee, & El-Geneidy, 2012; & Zhao, 2012; Schuijbroek, Hampshire, & Van Hoeve, 2017; Vogel,
Fuller et al., 2011), or data analysis methods, like in London (Ogilvie & 2020; Vogel, Saavedra, & Mattfeld, 2020; Waserhole, Jost, & Brauner,
Goodman, 2012). Topics include understanding the behaviors involved 2013) and the analysis of real data (Nair & Miller-Hooks, 2011;
in the relationship between users and bike-sharing (Bordagaray, Ibeas, & Schuijbroek et al., 2017; Vogel, Greiser, & Mattfeld, 2011).
dell’Olio, 2012), studying the social characteristics of BSS users such as Table 1 provides an overview of publications dealing with data
gender (Beecham & Wood, 2014) or social level (Hoe & Kaloustian, analysis for the study of bike sharing, classified by city and research
2014), and the factors that differentiate between regular cyclists with theme, for a selection of cities.
their own bikes and BSS users (Buck et al., 2013).

2.3. System prediction

2.2. Bike sharing data analysis
In the literature, we find some work based on machine learning
The analysis of BSP through data is also the source of many studies, mechanisms to determine the resources in the BSS. In Table 2 we have
often multi-disciplinary. Bike sharing is characterized by temporal and grouped together works dealing with resource prediction for Bike
spatial analyses, for example through the classification of bicycle sta­ sharing.
tions, as in Paris (Etienne & Latifa, 2014; Nair & Miller-Hooks, 2011;
Randriamanamihaga, Côme, Oukhellou, & Govaert, 2014), Barcelona 3. Bike sharing in smart city
(Froehlich, Neumann, & Oliver, 2008, 2020; Kaltenbrunner, Meza,
Grivolla, Codina, & Banchs, 2010) or London (Lathia, Ahmed, & Capra, 3.1. Concept of smart city
2012) to name the main ones. Comparative multi-city approaches have
also been implemented (Austwick, O’Brien, Strano, & Viana, 2013; Smart cities are a concept developed in recent years that aims to offer
Bouveyron, Côme, & Jacques, 2015; O’brien, Cheshire, & Batty, 2014; solutions to improve the management of cities, particularly through the
Sarkar, Lathia, & Mascolo, 2015), and reflect the strong similarity be­ use of new technologies (Bouzguenda, Alalouch, & Fava, 2019; Yigit­
tween all the systems, particularly with regard to the spatial and canlar et al., 2019). The urban space is seen as a complex superposition

Table 1
Summary of the related work of dealing with data analysis for the study of bike-sharing systems, classified by city and research topics, on a selection of the most studied
cities.
City Spatial and temporal Regulation Flow prediction Classification of stations Others
analysis

Barcelona Kaltenbrunner et al. Froehlich et al. (2008), Froehlich et al. (2020), Sarkar
(2010), O’brien et al. et al. (2015)
(2014)
Dublin Yoon, Pinelli, and
Calabrese (2012)
Hangzhou Xu, Ying, Wu, and Lin
(2013)
London O’brien et al. (2014) Bouveyron et al. (2015), Lathia et al. (2012), Sarkar Ogilvie and
et al. (2015), Tribastone, Clark, Gast, Gilmore, and Goodman (2012)
Reijsbergen (2014), Austwick et al. (2013)
Lyon Borgnat et al. (2011), Borgnat et al. (2011), Borgnat et al. (2013), Bouveyron et al. (2015) Jensen, Rouquier,
O’brien et al. (2014) Borgnat, Fleury, Robardet, Ovtracht, and
and Scherrer (2020) Robardet (2010)
Montreal O’brien et al. (2014) Faghih-Imani et al. (2014)
Paris Etienne and Latifa Nair, Miller-Hooks, Bouveyron et al. (2015), Etienne and Latifa (2014),
(2014), O’brien et al. Hampshire, and Bušić Randriamanamihaga et al. (2014)
(2014) (2013)
Vienna O’brien et al. (2014) Vogel et al. (2011) Vogel et al. (2011) Test
Washington O’brien et al. (2014) Schuijbroek et al. Austwick et al. (2013) Fanaee-T and Gama
(2017) (2014)

3
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Table 2
Summary of related work on usage prediction.
Ref City Timespan Algorithm Performance measures Best performances

Fanaee-T and Gama Washington 2 years Weka’s regressors -Relative Absolute Error -Root Relative Squared 29.98% 39.27%
(2014) DC Error
Yoon et al. (2012) Dublin 27 days Modified ARIMA Root Mean Square Error 5 min: 0.91 60 min:
3.47
Froehlich et al. (2020) Barcelona 13 weeks Bayesian network Relative Absolute Error 0.08
Borgnat et al. (2011) Lyon 2 years + 8 Linear Regression Mean Relative Error: 12%
months
Kaltenbrunner et al. Barcelona 7 weeks Auto-Regressive Moving Mean Absolute Error 1.39
(2010) Average

of infrastructures grouping together municipal services such as trans­ These data are very often found in the form of networks, energy or
port, water, energy and energy supply, or systems communications communication networks. Fig. 1 shows, for example, the main inter­
(Sharifi, 2020). Above these infrastructures, a new information layer is vened acts in the design of a smart city. The sensors also form networks
put in place, making it possible to obtain, with the help of sensors, and can exchange information according to their distance in order to
real-time information on many aspects of the systems in order to opti­ better control, for example, urban pollution (Sharifi, 2020). This
mise maintenance, inform citizens and, in a second stage, understand particular structure, found in many systems, is the subject of a new
the functioning of these systems in order to improve them. As Jeremy theory that has been developing in recent years: the science of complex
Rifkin explains (Rifkin, 2011), smart grids, or smart energy grids, pro­ networks.
pose a new paradigm in energy management, shifting from an approach
in which a company produces and delivers energy to consumers, to
3.2. Bike sharing program (BSP)
another model in which each household produces energy and distributes
it to the grid. The concept of Smart Cities extends this paradigm to the
A BSP or bike sharing, or self-service bike sharing consists of making
city level, offering inhabitants the means to become responsible players
bikes available to the public, free of charge or not. This type of mobility
in the management of the city. Due to the large volume of data gener­
service allows people to travel locally, mainly in urban areas, while
ated, their variety and speed of acquisition, issues related to smart cities
preserving the environment. This bike rental is a form of collaborative
are part of the datamass (Sharifi, 2020).
consumption and thus removes three obstacles to cycling: parking at
The first smart city projects still remain evasive about their aims,
home, theft and maintenance of one’s personal bike. While they are
contenting themselves with general ideas. Fig. 1 shows an example of a
equally good with and without docks, they have the particularity of
smart city. The vision of the smart city is broken down into four ideas
being centered on a bike.
(Sharifi, 2020):
3.2.1. Bike
• The taking into account environmental issues and clean energy
The concept of bike sharing is a neologism of shared bikes and cor­
constraints;
responds to a service that allows an individual to move around easily
• The networking of the actors among themselves;
without the need to have a personal bike. The bike is an energy-efficient,
• The transition from property to use;
safe, CO2-neutral and space-saving transportation mode (Cao & Shen,
• The integration of new technologies (like the IoT).
2019). It has a small ecological footprint (if used). In urban areas, for
short trips, it is a good alternative to the car. For longer trips or to get to
Research on smart cities requires a very broad set of skills given the
work, always in an urban environment, it is an excellent complement to
diversity of problems encountered in cities. It also requires the devel­
public transport, as it multiplies the area served. Although at the
opment of techniques for extracting knowledge adapted to urban data.
beginning of the 21st century, the majority of the BSS were docked,
nowadays, the BSP (see Fig. 2) are made up of both docked (Fig. 2a) and
dockless (Fig. 2b) BSS which have emerged recently in several cities
such as London, New York, San Francisco, Beijing, etc.

3.2.2. Docked BSS

A docked bike-sharing system consists of bikes, roads, and fixed
stations, as shown in Fig. 2a. Unlike dockless BSS, the station plays a
very important role in this type of BSS, which is still very successful
today and particularly in London city.

3.2.3. Station in docked BSS

The main characteristic of docked BSS is based on fixed stations
where users can pick up or drop off the shared bike on a self-service
basis. In such system, each station is composed of a transactional ter­
minal and several bike anchorage points. Each transactional terminal is
equipped with a radio identification reader (RFID) and sensors that
transmit information to the transactional terminal. However, since this
obligation to find a station is a constraint for users and also affects the
flexibility of the system, it has led to the emergence of dockless BSS as in
the case of the London BSP.

3.2.4. Dockless BSS

Fig. 1. Fundamental concept of the smart city. According to Sherriff, Adams, Blazejewski, Davies, and Kamerāde

4
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Fig. 2. Example of Bike Sharing Program (BSP) consisting of: (a) Docked BSS and (b) Dokless BSS.

(2020), the utility of a bike share service is dependent on accessibility 4. Methodology

and availability of the bikes. The location of bikes is therefore an
important consideration. Like docked BSS, a dockless BSS also makes In this section we propose an approach to predict the total number of
bikes available to the public, free of charge or not. It allows people to bike shares. Fig. 3 describes an overview of the proposed methodology.
travel locally, mainly in urban areas. Dockeless BSS consists of bikes and This approach is mainly based on the introduction of different steps of
roads without fixed parking stations as shown in Fig. 2b. The latter are the supervised regression process. It consists first of all of a new
also known as floating bike or fourth generation BSS. Thus, dockless bike enrichment method to improve the quality of the data and to take into
rental systems consist of a bike with a lock that is usually integrated into account other relevant factors (e.g. environmental factors). Secondly,
the frame and does not require a docking station (Ji et al., 2020; Lazarus, there is the implementation of machine learning algorithms to predict
Pourquier, Feng, Hammel, & Shaheen, 2020). the number of shared bikes. Thirdly, we adjust the parameters of each
model. Finally, we propose adequate evaluation tools based on regres­
3.3. Factors affecting bike sharing in smart sities sion performance metrics.

Many factors could affect bike sharing in smart cities and can be: 4.1. Smart system for Bike-sharing management
collective (economic, cultural and social), individual (physical condi­
tions, personal constraints, etc.) or meteorological factors. They play an An intelligent bike sharing system can be defined as a machine that
important role in the shared bike and should be considered when pro­ integrates a server (computer) connected to the Internet that can collect
moting paratransit cycling as a mode of travel. Several papers have been and analyse data and communicate with other systems. It is therefore a
published with great interest for the time of day, day of the week, month, classic intelligent system adapted to the context of bike sharing (see
etc. (Zhang, Wen, Qiu, Wang, & Abbas, 2019). However, as weather Section 3.2.2). Such a system must therefore effectively perform three
issues are frequently cited as a major impediment to use bike-sharing. functions: data collection, analysis and transmission. The intelligence
including weather condition in BSP designing is one of the issues of embedded in these systems is part of the development of the IoT where
this paper. almost everything can be provided in the form of unique identifiers, and
allows data to be transferred automatically over a network without
human-human or human-machine interaction. Nowadays, one of the

Fig. 3. Global ensemble-based system for real-time average number of bike-share prediction.

5
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

most important factors of intelligent systems is the inclusion of advanced characteristics to divide each node gives error rates that compare
software systems based on artificial intelligence that make it possible to favorably to other ensemble methods (Géron, 2019). The algorithm 1
anticipate a user’s needs in order to better satisfy them. This artificial describes the Random Forest principle (Breiman & Cutler, 2005; Cutler,
intelligence is based on the use of machine learning algorithms. In this Cutler, & Stevens, 2012).
part, we propose a process for building a machine learning model based
Algorithm 1. Random Forest Regressor (RFR) Breiman and Cutler
on ensemble methods to optimize the prediction of bike sharing in an
(2005), Cutler, Cutler, & Stevens(2012)
smart city. The proposed approach consists of three steps illustrated in
Fig. 4. Input •x the observation to predict;
• dn the observation;
• B the number of Trees;
4.2. Data collection and prepocessing
• m ∈ N the number of candidate variables to cut a node.
1∑B
Output h(x) = h(x, yk )
The first step to achieve best prediction according to flowchart 3 is B k=1
therefore the collecting and the aggregation of sufficient data, for 1: for k = 1 to B do
2: Draw a bootstrap sample in dn
solving the problem. The second step is the pre-processing of the
3: Construct a Classification And Regression Tree (CART) Razi and Athappilly
collected data, i.e. a reduction to just what we need, as well as their (2005) on this bootstrap sample, each cutoff is selected by minimizing the cost
cleaning, transformation and standardization. The goal of this step is to function of CART over a set of m randomly selected variables among the p. We
increase the accuracy of the model and to optimize the execution time note h(., k) the built tree.
4: end for
and learning as well as its size. More details about the dataset is in
Section 5.1.
where h(x, yk) is a collection of predictors per tree where yk is a string
4.3. Ensemble-based modeling of i.i.d. random variables. The random forest predictor is obtained by
aggregating this collection of trees.
The final goal of a machine learning work is to build a prediction
model. Indeed, one of the main challenges of machine learning is to 4.3.2. Bagging regressor (BR)
design efficient regression systems based on a set of representative ex­ Bagging (Breiman, 1996; Chen & Ren, 2009), meaning ”bootstrap
amples of a data population. Among the different approaches to address aggregating”, was one of the first methods proposed for the creation of a
this type of problem, combining a set of weak individual regressors to set of regressors. The Bagging technique redefines the learning set for
form a single regression system called the set of regressors has attracted each regressor. It consists in generating several versions of the same
growing interest from the scientific community. Recent work of research training database (called bootstraps) and using each of them to train a
has shown that certain principles of combining classifiers are particu­ basic estimator using the same training algorithm. The different boot­
larly effective, such as Boosting, Bagging, Random Subspaces, or more straps are built by random draws with delivery of examples of the
recently Random Forests. The efficiency of regressor combinations is original training data set. Thus, the outputs of the classifiers are then
mainly based on their ability to take advantage of the complementarities combined by majority vote or any other merger rule Breiman (1996).
of the individual regressors, with the aim of improving performance as
Algorithm 2. Bagging Regressor (BR) Breiman (1996)
much as possible by generalizing the set (Tay, Chui, Ong, & Ng, 2013).
The learning machine offers many different models to build a rela­ Input •x the observation to predict;
tionship between two magnitudes, the X input and the Y output, but the •dn the observation;
•B the number of Trees;
ideal is to build the most synthetic relationship possible. The simpler it
•m ∈ N the number of candidate variables to cut a node.
is, the easier it is to interpret. Ensemble methods are good at finding 1 ∑B
Output the estimator m(x) m
relationships between specific values of X and Y. Indeed, ensemble (or ̂ =
B k=1
̂ k (x)

aggregation) methods for statistical learning algorithms are based on the 1: for k = 1 to B do
2: Draw a bootstrap sample dkn in dn
idea of combining the predictions of several predictors (or classifiers) for
3: Adjust the regressor on this bootstrap sample m(x).
better generalization and to compensate for the possible shortcomings of
̂
4: end for
individual predictors, such as support vector machine regressor (SVMR)
(Mavroforakis & Theodoridis, 2006) and linear regression (LR) (Coppi,
4.3.3. XGBoost regressor (XGBR)
D’Urso, Giordani, & Santoro, 2006). In general, there are two families of
In this section, we will present the XGBoost algorithm proposed by
such methods:
Chen and Guestrin. The name XGBoost is an abbreviation of ”eXtreme
Gradient Boosting”. This algorithm is a continuation of GBM (bradient
• Bagging methods (Chen & Ren, 2009) where the principle is to
boosting machines). The interest of XGBoost lies in the correction of a
average several predictions hoping for a better result following the
major problem of Gradient Boosting Tree, namely the fact of considering
reduction of variance of the mean estimator. We find in this family
the potential loss for each possible separation (split) at a given node to
the following algorithms: Random Forest Regressor (RFR), Bagging
create a new branch (Chen, 2020). Indeed, in the case where we have a
Regressor (BR).
lot of explanatory variables, there are a large number of possible splits,
• Boosting methods (Freund, Schapire, & Abe, 1999) where the pa­
which leads to high computation times. Fig. 5 shows the algorithm
rameters are iterative adapted to produce a better mix. We find in
process.
this family the following algorithms: XGBoost Regressor (XGBR),
As depicted in Fig. 5, at each iteration of the gradient boosting al­
AdaBoosting Regressor (ABR).
gorithm, the residual will be used to correct the previous predictor so
that the specified loss function can be optimized. As an improvement,
In the following we will explore each of these algorithm classes.
the regularization is added to the loss function to establish the objective
function in XGBoost measuring the performance of the model Θ Zhang
4.3.1. Random forest regressor
et al. (2018). Such performance Θ is given by Eq. (1) where: τ denotes
Random Forest are a combination of decision trees, where each tree
the trained parameters from the given data; Φ is the training loss func­
depends on the values of an independently sampled random vector with
tion, such as square loss or logistic loss, which measures how well the
the same distribution for all trees in the forest. The generalization error
model fits on training data and Ω is the regularization term Zhang et al.
of a forest of trees depends on the strength of the individual trees in the
(2018).
forest and the correlation between them. Using a random selection of

6
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Fig. 4. Smart system for bike-sharing management.

1 ∑m
Ω(fk ) = γT + λ ω2 (5)
2 i=1 j

From Eq. (5), γ denotes the complexity of each leaf, T denotes the
number of leaves in a tree, λ denotes a parameter to scale the penalty,
and ω is the vector of scores on the leaves. Then, the first-order along
with the second-order Taylor expansions are taken to the loss function in
XGBoost. More details about the process of XGBoost algorithm are given
by Chen (2020), Zhang et al. (2018).

4.3.4. AdaBoosting regressor (ABR)

Adaboost was able to solve several optimization problems that were
causing problems with the boosting algorithms. It takes up the general
principle of boosting by constructing a linear combination of hypothe­
Fig. 5. Algorithm process of XGBoost. ses, designed by iterative executing a weak leaner algorithm, on prob­
ability distributions calculated on the basis of the results of the previous
iterations (Freund & Schapire, 1997, Freund et al., 2020). This is to
Θ(τ) = Φ(τ) + Ω(τ) (1)
allow the learner to focus his or her attention on examples that are
The algorithm can independently determine the types of loss func­ difficult to learn (Mishra, Mishra, & Santra, 2017). Algorithm 3 de­
tions used for model evaluation. To reduce the risk of overfitting, an scribes the principle of the Adaboost algorithm.
additional adjustment term is added to the model. Knowing that the base
Algorithm 3. AdaBoosting Regressor (ABR) Freund and Schapire
models fm consist of decision trees, we can write the output of the
(1997), Freund et al. (2020)
ensemble model ̂ y i from a space W, which is voted or averaged over m
base trees by Eq. (2).
∑
m
1∑ M
yi =
̂ fm (xi ), fm ∈ W, (2) g0 (.) = argminc f (x) = L(yi , c)
i=1 n m=1

For n predictions, the objective function at the tth iteration can be

written by Eq. (3) below. Input •x the observation to predict;
•dn = (x1, y1), .. ., (xn, yn) the sample;
∑
n ∑
t
• h the weak rule;
Θ(t) = Φ(yi , ̂
yi) + Ω(fk ) (3)
•λ a regularization parameter such as 0 < λ ≤ 1.
i=1 k=1
•M the number of iterations.;
Output the estimator ̂g M (x)
y i of the
Given objective function of Eq. (3), the corresponding output ̂
(t)
1: Initialization:
ensemble model will be given by Eq. (4) as follows:
∑
m 1∑ M

(4)
1) g0 (.) = argminc f (x) = L(yi , c)
y (t)
̂ i = y (t−
fm (xi ) = ̂ i + ft (xi ) n m=1
i=1
2: for m = 1 to M do
The regularization term Ω(fk) for a decision tree is defined by Chen 3: ∂
(2020) as follows: Calculate the opposite of the gradient L(y, g) and evaluate it at points gm1 (xi ):
∂y

(continued on next page)

7
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

(continued )
[ ]
∂l(yi , gm (xi ))
Ui = − , i = 1, ..., n.
∂g(xi ) g(xi )=gm− 1 (xi )

4: Adjust the weak rule on the sample (x1, U1), .. ., (xn, Un), we notice hm the rule
thus defined.
5: Update: gm(x) = gm− 1(x) + λhm(x)
6: end for

5. Experiments and results analysis

5.1. Bike sharing dataset

The used dataset in our experiments comes from the repository of the
city of London in England. The data come from three different sources:
the TfL government website1 that shares London’s bike sharing data, the
weather site2 on which the meteo data for the same period were
Fig. 6. Parameters that influence the number of bikes share.
collected and finally the government site3 allowing to obtain working
days, holidays and weekends. Data were collected from 04/01/2015 to
03/01/2017 for these three sources. relevant learning algorithm for the task is a real challenge. These choices
These information are publicly available on the links associated with have a direct impact on the quality of the system.
each of the data sources. The data from the three sources was then
aggregated according to time, day and weather information and
5.2. Experimental protocol
analyzed in the 5.4.1 section. The type of the attributes as mentioned in
the third column of Table 3, corresponds to the type of original data
The aim of this section is to describe the experimental protocol to test
collected for each of these attributes. Thus, prior to this analysis, some
and evaluate the performances of the proposed models for the prediction
features have been transformed as follows4 :
number of bike shared per hour or day. The implementations have being
carried out on a windows platform under python 3.7 using principally
• Season: there are 4 seasons corresponding to: spring coded by (0),
sklearn library. The computer used is of the ′′ Asus′′ brand with the
summer coded by (1), fall coded by (2) and winter coded by (3).
following configuration: 8 GB of RAM, intel core i7 processor and an
• Working day: if the day is neither the weekend nor a public holiday,
NVIDIA Geforce 930M for graphic card. For the experiments, the dataset
the value is 1. Otherwise, 0;
has been divided into 5 folds and the tests were performed with cross
• Weather situation: we have eight cases; Clear or Few clouds is coded
validation using 4 folds for training and the one for testing.
by (1), Partly cloudy is coded by (2), Fragmented clouds is coded by
The performances of the involved algorithms have been compared in
(3), Fog more cloudy is coded by (4), Rain is coded by (7), Heavy rain
terms of r-square (R2), root mean log square error (RMLSE), mean ab­
or ice pallets plus thunderstorm is coded by (10), Snowfall is coded
solute error (MAE), mean square error (RMSE) and processing time.
by (26) and Icing fog is coded by (94).
These performance measure are the commonly used to evaluate de
regression models. They have been involved in many previous works
After codification, all features Vij were normalized according to the
associated to bike sharing including Tekouabou et al. (2020).
formula of Eq. (6) to minimize the impact of variable scales on predictor
The MAE defines by Eq. (9) consists of calculating the arithmetic
performance.
mean of the absolute values of the differences between the actually bike
Vij − minj (Vij ) shared rates and those predicted by the machine learning model
Vij ⟵ (6)
maxj (Vij ) − minj (Vij ) Tekouabou et al. (2020). Since the objective is to measure the error
produced by the BSS, the lower the MAE (or RMSE) obtained, the better
Table 6
the prediction quality, meaning that the predicted performance is closer
The relevance of the extracted characteristics and the choice of a
to the reference. Let Y = [y1, .. ., yN] the predicted bike sharing rates and
Y
̂ = [̂y 1 ; ...; ̂
y N ] the actual bike shared number for N instances with an
Table 3
Details about bike shared dataset. y , Eq. (7), (9), (8) are used to compute the RMLSE, MAE, R2
average of ̂
and RMSE respectively as follows:
N Attribute Description Type
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1 timestamp Timestamp field for grouping the data Numerical
∑N
i=1 (log(y) − log(̂ y ))2
2 cnt Count of a new bike shares Numerical RMLSE = (7)
N
3 t1 Real temperature in ◦ C Numerical
4 t2 Apparent temperature in ◦ C Numerical ∑N
5 hum Humidity in percentage (%) Numerical i=1 (y − ̂y )2
6 windspeed Twind speed in km/h Numerical MSE = (8)
N
7 weather_code Category of the weather Categorical
8 isholiday Boolean field - 1 holiday / 0 non holiday Boolean ∑N ⃒⃒ ⃒
y⃒
i=1 y − ̂
9 isweekend Boolean field - 1 if the day is weekend Boolean MAE = (9)
10 season Category field meteorological seasons Categorical N
∑N
y)
(y − ̂
R2 = 1 − ∑Ni=1 (10)
i=1 y
(̂ − y)
̂
1
Https://[Link]/ The choice of a single metric does not always allow the models to be
2
[Link] separated. If the RLMSE and the MSE are known to give the error
3
[Link] characterized by the variance and the mean between the predicted and
4
[Link] the actual, favouring the effects of large deviations, the absolute error

8
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

may reflect the effect of accuracy in predicting the rate of bike sharing. remains high but gradually decreases, losing 2300 movements per hour
R2 will give himself the proportion of the actual rate that was correctly between 11 pm and 5 am.
predicted. The optimal model will result from the homogeneity between Fig. 8e shows the bike sharing rate obtained as a function of time for
all three metrics. the three patterns (workdays, weekend and holidays). For the first
pattern ”workdays”, we can notice that there are two peaks of bike-
5.3. Hyper-parameters tuning and optimization of the models sharing activities from 7:00 am to 9:00 am, then from 5:00 pm to 7:00
pm, i.e. during the periods of going to and from work. For weekend and
Optimal models for a specific task are obtained by hyper-parameters holiday reasons, we can notice that the activity of shared bikes is from
adjustments. For each algorithm, we performed several setting for a 10 am to 8 pm.
better prediction of the number of shared bikes. We would like to In order to clearly visualise the use of bikes at different times of the
compare the performance of the different learning models presented in day, the bike usage data for each period from 0 to 24 hours is extracted
Section 4.3 with the prediction of the number of shared bikes in the city as shown in 8 b. We notice at the beginning of the week the rate of bike
of London. For the bagging methods, the most important hyper- sharing is low, in the middle of the week the rate of bike sharing in­
parameters are the number of iterations, the bagging error (EOOB) and creases. However, during the weekend the number of rentals is much
the best weak learner. For the boosting method, we had to find the lower.
number of iterations, the learning rate and the best weak learner The use of bikes in different situations is more weather related, so in
(Tekouabou et al., 2020). Fig. 7 illustrates the tests that were performed order to show the impact of weather conditions on the bike usage data
to select these parameters using a simple decision tree as a weak learner. for each situation are extracted, as shown in Fig. 8d. This figure shows
Fig. 7a and b provide the optimal parameters for the ABR and XGBR that the number of bikes shared in better weather conditions is very high
boosting methods, respectively. From Fig. 7a, we can see that optimal compared to bad weather.
for ABR model has a learning rate of 1 while its number of estimators Fig. 8c shows the bike sharing rate for each season. It can be noted
must be greater than 40. From Fig. 7b, we notice that the optimal that the use of shared bikes is quite considerable during summer and fall.
learning rate and number of estimators for XGBR must be at least greater Fig. 8f shows the bike sharing rate obtained as a function of months
than 0.05 and 20, respectively. Fig. 7c shows that for the bagging models for the three reasons (workdays, weekends and holidays). For all three
(BR and RFR), the out-of-bag error (EOOB) is almost equal to 0 for 20 or reasons, we notice a considerable movement of bike sharing between the
more estimators. Knowing these best parameters allows us to experi­ months of April and October. For the rest of the year, the sharing rate is
ment and find the most efficient model for our prediction task. very low. This is true for all three patterns.

5.4.2. Predictive result analysis

5.4. Results analysis To show the reliability of our model, we compare for the different
algorithms, the validation and test performances by computing the MSE,
In this section we discuss the obtained results. We first present the MAE, RMLSE and R2 scores of each algorithm. The obtained perfor­
data analysis on the total number of bike shared on given BSS of London. mance values are grouped below in Table 4 and are also illustrated
We then analyse the predictive model results from our learning graphically.
algorithms. Fig. 9 illustrates the performance of the different algorithms used for
the R2 metric. From Fig. 9, we can see that the bagging methods RFR and
5.4.1. Data analysis BR give the best scores for both validation (0.9532, and 0.9531
Based on our database from the London BSS we described in Section respectively) and testing (0.9548 and 0.955 respectively)). They are
5.1, we can quickly create diagrams to determine how the number of followed by the boosting algorithms, in particular XGBR then ABR,
bike shared is affected by the available features. To do this, we now see which give respectively 0.9513 then 0.8264 for validation and 0.9513
some graphs that illustrate these different relations. then 0.8036 for the test. Finally individual algorithms SVMR then LR
According to Fig. 8a, we can segment the day time range into three give the lowest R2 score with respectively 0.7892 then 0.6538 for vali­
segments. Thus, after a first peak at the time of arrival at work (7–9 am), dation and 0.8115 then 0.6807 for the test.
the second peak is at lunchtime, between 12 and 1 pm (about 1500 The same trend is confirmed for the other three performance metrics
movements per hour). From 5 p.m. to 7 p.m., the number of bike rentals associated with the different calculated errors. Figs. 10–12 illustrate the
shared is at its highest with nearly 2800 movements per hour, then errors corresponding to the MAE, RMSLE and MSE measurements
declines steadily from 2800 to 500. In the evening, the use of the system

Fig. 7. Tuning and optimization of hyper-parameters of the proposed models.

9
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Fig. 8. Distribution of the average number of bike shared: (a) by hour of day; (b) by day of week; (c) by climatic season (d) according to the weather represented by
codes, (e) time of day grouped by workdays, weekends and holidays (f) by month-days grouped by workdays, weekends and holidays.

Table 4
Evaluation of BSS performance prediction with ensemble models in terms of MAE, MSE, RMLSE and R2.
Algorithm R2 MAE MSE RMSLE

Validation score Test score Validation score Test score Validation score Test score Validation score Test score

LR 0.6538 0.6807 0.1272 0.1241 0.0268 0.0253 0.0821 0.078

RFR 0.9532 0.9548 0.0406 0.0413 0.0036 0.0036 0.0319 0.03
XGBR 0.9513 0.9513 0.0417 0.0421 0.0037 0.0039 0.0348 0.031
SVMR 0.7893 0.8115 0.1057 0.1011 0.0163 0.0149 0.0619 0.0585
ABR 0.8264 0.8036 0.0934 0.101 0.0136 0.0156 0.0596 0.0608
BR 0.9531 0.955 0.0406 0.0412 0.0036 0.0036 0.032 0.0299

10
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Fig. 9. Comparison of the performances of the used regression algorithms ac­ Fig. 12. Comparison of the performances of the used regression algorithms
cording R2. according RMSLE.

respectively. From these figures whose values are summarized in

Table 4, we can notice that overall the bagging methods always give the
best scores and therefore the lowest errors both in validation and testing.
The values obtained for RF and BR are respectively 0.0406, 0.0036 and
0.0319 in validation then 0.0413, 0.0036 and 0.03 in test for MAE, MSE
and RMLSE. These best performances are followed by those of boosting
methods such as XGBR then ABR in order of performance. Among these
two algorithms based on boosting, XGBR is more efficient with respec­
tively 0.0417, 0.0037, 0.0348 in validation then 0.0421, 0.0039, 0.031
in test while ABR gave 0.0934, 0.0136, 0.0596 in validation then 0.101,
0.0156, 0.0608 in test and this, for the trio of metrics MAE, MSE and
RMLSE. Finally the individual algorithms that almost gave the lowest
score corresponding to the highest errors. For the metrics MAE, MSE and
RMLSE, the values obtained are respectively 0.1057, 0.0163, 0.0619 in
validation then 0.1011, 0.0149, 0.0585 in test for SVMR and finally the
lowest scores of 0.1272, 0.0268, 0.0821 in validation then 0.1241,
0.0253, 0.078 in test attributed to LR for the three metrics. We can see
that the bagging technique is more favourable on all metrics to build the
best models. The validation score values are roughly equal to the test
Fig. 10. Comparison of the performances of the used regression algorithms values, which shows that there is no over-fitting. However, for the in­
according MAE.
dividual models, there is some few under-fitting which would have been
corrected by the ensemble strategy to give the best scores by eliminating
bias and variance.
The conclusions made by performance analysis are also confirmed by
the distribution patterns of 200 randomly drawn instances represented
in Fig. 13. The points of the distribution correspond to the couple real
value of the bikes sharing in abscissa and the predictive values associ­
ated in ordinate and this for the six algorithms used. From Fig. 13, we
can clearly see that the distribution of RFR and BR is better. Then comes
XGBR, then ABR and finally SVMR and LR have the worst distributions.
The next section is devoted to the discussion of the results obtained.

6. Discussion of the results

Smart cities offer a wide variety of ways for citizens to get around
depending to their needs. City managers and transportation service
providers need to meet both the mobility needs of users, the return on
investment and the sustainability (Banerjee, Kabir, Khadem, & Chavis,
2020; Sherriff et al., 2020). These requirements have fostered the
emergence of new urban mobility offers such as self-service shared bikes
which can be used several times as a sustainable modes of transport
Fig. 11. Comparison of the performances of the used regression algorithms responding favourably to the mobility needs of the inhabitants. In urban
according MSE.
planning of bike sharing program, the number of bikes shared is not
almost known and changes dynamically under the influence of several

11
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Fig. 13. Comparison of the regression distribution look for 200 first instances by the different algorithms.

factors including spatial distribution of bikes, time, weather conditions, recycling of used bicycles preserve the environment (Bikeshare, 2013;
seasons, etc (Ji et al., 2020). The duration of bike use is often unexpected Chen et al., 2020; Sherriff et al., 2020). Machine learning based models
in advance and the number of bikes shared also is often variable. can be applied in real life to manage and plan bike-sharing travels for
Moreover, all these parameters depend on several factors (Bikeshare, users over a period of time, as the system can plan the availability of
2013; Sherriff et al., 2020). Collecting the data associated to all these bikes for all users who request a bike-sharing travel. As illustrated in
factors through the IoT and analyzing them using machine learning al­ Fig. 3, the training of predictive systems is essentially based on the
gorithms would help to make BSS more cost-effective and accessible to extraction of characteristics associated with a training algorithm. Data
users. In order to offer more flexibility in service to the user, shared bikes pre-processing steps have made the data more suitable for efficient
programs can now involves docked and dockless BSS as in London. In training of predictive regression models. Several training algorithms
this paper, we aim to design a predictive model for self-service bike were tested including SVMR, LR, BR, RFR, XGBR and ABR. The simu­
sharing program planning and case of study in London city. As stated by lation results show that the RFR, BR, and XGBR achieved the best per­
Chen et al. (2020), self-service bike sharing programs certainly have formances taking into account factors that affect bike usage in BSS. In
advantages in terms of flexibility, convenience, accessibility and avail­ addition, most literature work does not take into account these factors
ability for the user. Even though the majority of BSS in London are (especially the meterological factors) that impact bike use in a BSS.
docked, there is also another alternative that is increasingly being (Borgnat et al., 2011; Fanaee-T & Gama, 2014; Kaltenbrunner et al.,
studied, which is dockless BSS. Indeed, dockless has advantages in terms 2010; Lu et al., 2019; Yoon et al., 2012; Zhang et al., 2019). Our
of flexibility, convenience, accessibility and availability for the user. approach has been evaluated by several evaluation metrics (R2, MAE,
However, one of the biggest problems of shared bikes is theft and MSE and RMSLE) unlike previous works (Borgnat et al., 2011; Fanaee-T
vandalism, which are more increased in self-service shared bike sharing & Gama, 2014; Kaltenbrunner et al., 2010; Lu et al., 2019; Yoon et al.,
systems (Chen et al., 2020; Floret, 2014; Zhang, Zhang, Duan, & Bryde, 2012; Zhang et al., 2019). Considering factors such as weather in the
2015). In addition, the business model and operational challenges prediction process guarantees a dynamic aspect of the model we have
associated to bike sharing may generate other more urgent problems, built compared to other works (Caggiani, Camporeale, Ottomanelli, &
such as vandalism and overproduction of bicycles, causing damage to Szeto, 2018; Raviv & Kolka, 2013).
the environment and threatening the sustainability of cities and society The deployment of the proposed model can be done in real time in
(Chen et al., 2020). Forecasting the number of bike shared is therefore order to automate and make reliable urgent decisions on the spatial and
very important both for the business model, the sustainability of BSS and temporal distribution of bike shared in self-service (Banerjee et al.,
therefore of cities. Strategically, planning the number of shared bikes 2020; Boufidis, Nikiforiadis, Chrysostomou, & Aifadopoulou, 2020;
allows to: (1) refine the spatio-temporal distribution of bicycles in the Conrow, Murray, & Fischer, 2018; Winslow & Mont, 2019). This is
city and therefore automate the display of the number of bicycles important to already improve the quality of BSS provided services by
available in real time for users as well as for city controllers and pro­ allowing automatically both users and providers to have an idea about
moters (Banerjee et al., 2020; Bikeshare, 2013). (2) For business models, the number of bicycles available in real time or when needed. Including
this will allow potential future investors to benchmark and evaluate the predictive analysis will thus allow for more autonomous and intelligent
profitability in the short, medium and even long term knowing that the management of the BSS. The stakes for investors is to have a clear idea of
turnover is both associated with the number of shared bicycles and the the profitability of BSS and to be accompanied in making strategic de­
use time that we explore in our model (Bikeshare, 2013; Sherriff et al., cisions. We complete by a real and concrete observation linked to the
2020). (3) The “Sustainability” already due to the alternative of shared fact that today the emergence of BSS depends on the conviction of users
bicycles which is more ecological compared to other modes of urban to use shared bikes and thus attract investors. This is a major challenge
mobility. The integration of the prediction model not only allows to for several countries. In this case, we show that the implementation of an
automate and improve the quality of service offered to users. It also to automatic predictive model will improve the service and thus attract the
make the system more intelligent by removing the need for physical maximum number of users while reassuring investors. The machine
resources to attract more users and thus better contribute to the sus­ learning based forecasting system allows an automatic, dynamic and
tainability of the city. For a sustainable city and society, urban mobility intelligent parametrization of the BSS. Its stakes for modern BSS are
is one of the important factors (Sherriff et al., 2020; Sun & Zhai, 2020). therefore strategic and multidimensional.
The smart mobility solution such as proposed BSP should confirm the The proposed system could thus serve as an effective tool for auto­
three pillars of sustainability including financial sustainability, social mating and improving the quality of service providing a dynamic BSS
sustainability, and environmental sustainability. Financial sustainability management. However, given all the challenges of management of bike-
can be achieved by providing a low-cost mobility service while sharing systems for the companies and governments that oversee these
improving the accuracy. Social sustainability can be obtained by the systems, the user decision model and user waiting time could be
customer satisfaction of the shared bike services close to the users and considered in future work. This will lead our future research to delve
finely tuned to their needs. Environmental sustainability can be ach­ deeper into the BSS most compromising between sustainability and user-
ieved by controlling the carbon and other toxic gas emissions by the centric service in the business or strategic model.
healthcare equipment. For example, the IoT, mobility sensors, spatial
distribution of bikes, station equipment’s if they exist and especially the

12
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

7. Conclusion Chen, Q., & Sun, T. (2015). A model for the layout of bike stations in public bike-sharing
systems. Journal of Advanced Transportation, 49, 884–900.
Chen, M., Miao, Y., Hao, Y., & Hwang, K. (2017). Narrow band internet of things. IEEE
The new technologies offer by the smart cities and the IoT make daily Access, 5, 20557–20577.
life more comfortable as was presented at the beginning of this paper. Chen, Z., van Lierop, D., & Ettema, D. (2020). Dockless bike-sharing systems: what are
Thanks to these technologies we can retrieve data from several con­ the implications? Transport Reviews, 1–21.
Conrow, L., Murray, A. T., & Fischer, H. A. (2018). An optimization approach for
nected sources. These data allow us to better understand how this new equitable bicycle share station siting. Journal of Transport Geography, 69, 163–170.
form of transportation is used by city inhabitants. Thus, on the basis of Coppi, R., D’Urso, P., Giordani, P., & Santoro, A. (2006). Least squares estimation of a
real data from three sources of London BSS, we have set out two main linear regression model with lr fuzzy response. Computational Statistics & Data
Analysis, 51, 267–286.
objectives in this paper. The first was to study different aspects of the Cutler, A., Cutler, D. R., & Stevens, J. R. (2012). Random forests. in: Ensemble machine
shared bicycle system, through the analysis of travel data taking into learning, 157–175.
account several factors. The second was to develop a prediction Dell’Olio, L., Ibeas, A., & Moura, J. L. Implementing bike-sharing systems, in:
Proceedings of the Institution of Civil Engineers-Municipal Engineer, volume 164,
approach based on machine learning to help companies and stake­ Thomas Telford Ltd, 89-101.
holders in the shared bike sector to manage the system in an efficient DeMaio, P. (2009). Bike-sharing: History, impacts, models of provision, and future.
and cost-effective way. The simulation results show that the overall Journal of Public Transportation, 12, 3.
Doorley, R., Pakrashi, V., & Ghosh, B. (2015). Quantifying the health impacts of active
methods are more effective than the individual methods. The approach travel: assessment of methodologies. Transport Reviews, 35, 559–582.
also provides appropriate recommendations for those responsible for the Eren, E., & Uz, V. E. (2019). A review on bike-sharing: The factors affecting bike-sharing
system. demand (p. 101882). Sustainable Cities and Society.
Etienne, C., & Latifa, O. (2014). Model-based count series clustering for bike sharing
system usage mining: a case study with the vélib’system of paris. ACM Transactions
on Intelligent Systems and Technology (TIST), 5, 1–21.
Declaration of Competing Interest Faghih-Imani, A., Eluru, N., El-Geneidy, A. M., Rabbat, M., & Haq, U. (2014). How land-
use and urban form impact bicycle flows: evidence from the bicycle-sharing system
The authors report no declarations of interest. (bixi) in montreal. Journal of Transport Geography, 41, 306–314.
Fanaee-T, H., & Gama, J. (2014). Event labeling combining ensemble detectors and
background knowledge. Progress in Artificial Intelligence, 2, 113–127.
References Fishman, E., Washington, S., & Haworth, N. L. (2012). An evaluation framework for
assessing the impact of public bicycle share schemes.
Fishman, E., Washington, S., & Haworth, N. (2013). Bike share: a synthesis of the
Austwick, M. Z., O’Brien, O., Strano, E., & Viana, M. (2013). The structure of spatial
literature. Transport Reviews, 33, 148–165.
networks and communities in bicycle sharing systems. PLOS ONE, 8.
Fishman, E. (2016). Bikeshare: A review of recent literature. Transport Reviews, 36,
Bachand-Marleau, J., Lee, B. H., & El-Geneidy, A. M. (2012). Better understanding of
92–113.
factors influencing likelihood of using shared bicycle systems and frequency of use.
Floret, M. (2014). Bike-sharing and the promotion of urban sustainable mobility, Ph.D. thesis,
Transportation Research Record, 2314, 66–71.
uniwien.
Banerjee, S., Kabir, M. M., Khadem, N. K., & Chavis, C. (2020). Optimal locations for
Freund, Y., & Schapire, R. E. (1997). A decision-theoretic generalization of on-line
bikeshare stations: A new gis based spatial approach. Transportation Research
learning and an application to boosting. Journal of Computer and System Sciences, 55,
Interdisciplinary Perspectives, 100101.
119–139.
Beecham, R., & Wood, J. (2014). Exploring gendered cycling behaviours within a large-
Freund, Y., Schapire, R. E., et al. Experiments with a new boosting algorithm, in: icml,
scale behavioural data-set. Transportation Planning and Technology, 37, 83–97.
volume 96, Citeseer 148-156.
Bikeshare, C. (2013). Capital bikeshare member survey report. Washington, DC.
Freund, Y., Schapire, R., & Abe, N. (1999). A short introduction to boosting. Journal-
Bordagaray, M., Ibeas, A., & dell’Olio, L. (2012). Modeling user perception of public
Japanese Society For Artificial Intelligence, 14, 1612.
bicycle services. Procedia-Social and Behavioral Sciences, 54, 1308–1316.
Froehlich, J., Neumann, J., & Oliver, N. (2008). Measuring the pulse of the city through
Borgnat, P., Abry, P., Flandrin, P., Robardet, C., Rouquier, J.-B., & Fleury, E. (2011).
shared bicycle programs. Proc. of UrbanSense, 08, 16–20.
Shared bicycles in a city: A signal processing and data analysis perspective. Advances
Froehlich, J. E., Neumann, J., & Oliver, N. Sensing and predicting the pulse of the city
in Complex Systems, 14, 415–438.
through shared bicycling, in: Twenty-First International Joint Conference on
Borgnat, P., Robardet, C., Abry, P., Flandrin, P., Rouquier, J.-B., & Tremblay, N. (2013).
Artificial Intelligence.
A dynamical network view of lyon’s vélo’v shared bicycle system. in: Dynamics On
Fuller, D., Gauvin, L., Kestens, Y., Daniel, M., Fournier, M., Morency, P., & Drouin, L.
and Of Complex Networks, Volume 2, 267–284.
(2011). Use of a new public bicycle share program in montreal, canada. American
Borgnat, P., Fleury, E., Robardet, C., & Scherrer, A. Spatial analysis of dynamic
Journal of Preventive Medicine, 41, 80–83.
movements of vélo’v, lyon’s shared bicycle program.
Géron, A. (2019). Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow:
Bortolussi, L., & Hillston, J. (2015). Efficient checking of individual rewards properties in
Concepts, Tools, and Techniques to Build Intelligent Systems. O’Reilly Media.
markov population models. arXiv preprint arXiv:1509.08561.
García-Palomares, J. C., Gutiérrez, J., & Latorre, M. (2012). Optimizing the location of
Boufidis, N., Nikiforiadis, A., Chrysostomou, K., & Aifadopoulou, G. (2020).
stations in bike-sharing programs: A gis approach. Applied Geography, 35, 235–246.
Development of a station-level demand prediction and visualization tool to support
Garcia-Gutierrez, J., Romero-Torres, J., & Gaytan-Iniestra, J. (2014). Dimensioning of a
bike-sharing systems’ operators. Transportation Research Procedia, 47, 51–58.
bike sharing system (bss): a study case in Nezahualcoyotl, Mexico. Procedia-Social
Bouveyron, C., Côme, E., Jacques, J., et al. (2015). The discriminative functional mixture
and Behavioral Sciences, 162, 253–262.
model for a comparative analysis of bike sharing systems. The Annals of Applied
Hoe, N., & Kaloustian, T. (2014). Bike sharing in low-income communities: An analysis of
Statistics, 9, 1726–1760.
focus groups findings fall 2014.
Bouzguenda, I., Alalouch, C., & Fava, N. (2019). Towards smart sustainable cities: A
Jcdecaux developer https.//[Link]/.
review of the role digital citizen participation could play in advancing social
Jensen, P., Rouquier, J.-B., Ovtracht, N., & Robardet, C. (2010). Characterizing the speed
sustainability. Sustainable Cities and Society, 101627.
and paths of shared bicycle use in lyon. Transportation Research part D: Transport and
Breiman, L., & Cutler, A. (2005). Random forests. berkeley.
Environment, 15, 522–524.
Breiman, L. (1996). Bagging predictors. Machine learning, 24, 123–140.
Ji, Y., Ma, X., He, M., Jin, Y., & Yuan, Y. (2020). Comparison of usage regularity and its
Buck, D., Buehler, R., Happ, P., Rawls, B., Chung, P., & Borecki, N. (2013). Are bikeshare
determinants between docked and dockless bike-sharing systems: A case study in
users different from regular cyclists? a first look at short-term users, annual
Nanjing, China. Journal of Cleaner Production, 255, 120110.
members, and area cyclists in the Washington, DC, region. Transportation research
Kaltenbrunner, A., Meza, R., Grivolla, J., Codina, J., & Banchs, R. (2010). Urban cycles
record, 2387, 112–119.
and mobility patterns: Exploring and predicting trends in a bicycle-based public
Bullock, C., Brereton, F., & Bailey, S. (2017). The economic contribution of public bike-
transport system. Pervasive and Mobile Computing, 6, 455–466.
share to the sustainability and efficient functioning of cities. Sustainable cities and
Kim, D., Shin, H., Im, H., & Park, J. Factors influencing travel behaviors in bikesharing,
society, 28, 76–87.
in: Transportation Research Board 91st Annual Meeting.
Caggiani, L., Camporeale, R., Ottomanelli, M., & Szeto, W. Y. (2018). A modeling
Krykewycz, G. R., Puchalsky, C. M., Rocks, J., Bonnette, B., & Jaskiewicz, F. (2010).
framework for the dynamic management of free-floating bike-sharing systems.
Defining a primary market and estimating demand for major bicycle-sharing
Transportation Research Part C: Emerging Technologies, 87, 159–182.
program in Philadelphia, Pennsylvania. Transportation Research Record, 2143,
Cao, Y., & Shen, D. (2019). Contribution of shared bikes to carbon dioxide emission
117–124.
reduction and the economy in Beijing. Sustainable Cities and Society, 51, 101749.
Lathia, N., Ahmed, S., & Capra, L. (2012). Measuring the impact of opening the London
Castillo-Manzano, J. I., Castro-Nuno, M., & Lopez-Valpuesta, L. (2015). Analyzing the
shared bicycle scheme to casual users. Transportation Research part C: Emerging
transition from a public bicycle system to bicycle ownership: A complex relationship.
Technologies, 22, 88–102.
Transportation Research Part D: Transport and Environment, 38, 15–26.
Lazarus, J., Pourquier, J. C., Feng, F., Hammel, H., & Shaheen, S. (2020). Micromobility
Chen, T. & Guestrin, C. Xgboost: A scalable tree boosting system, in: Proceedings of the
evolution and expansion: Understanding how docked and dockless bikesharing
22nd acm sigkdd international conference on knowledge discovery and data mining
models complement and compete—a case study of San Francisco. Journal of
785-794.
Transport Geography, 84, 102620.
Chen, T., & Ren, J. (2009). Bagging for Gaussian process regression. Neurocomputing, 72,
1605–1610.

13
E.A. Abdellaoui Alaoui and S.C. Koumetio Tekouabou Sustainable Cities and Society 67 (2021) 102702

Leister, E. H., Vairo, N., Sims, D., & Bopp, M. (2018). Understanding bike share reach, Shaheen, S. A., Guzman, S., & Zhang, H. (2010). Bikesharing in Europe, the Americas,
use, access and function: An exploratory study. Sustainable Cities and Society, 43, and Asia: Past, present, and future. Transportation Research Record, 2143, 159–167.
191–196. Shaheen, S. A., Zhang, H., Martin, E., & Guzman, S. (2011). China’s hangzhou public
Lin, J.-R., & Yang, T.-H. (2011). Strategic design of public bicycle sharing systems with bicycle: understanding early adoption and behavioral response to bikesharing.
service level constraints. Transportation Research part E: Logistics and Transportation Transportation Research Record, 2247, 33–41.
Review, 47, 284–294. Sharifi, A. (2020). A typology of smart city assessment tools and indicator sets.
Lohry, G. F. & Yiu, A. Bikeshare in China as a public service: Comparing government-run Sustainable Cities and Society, 53, 101936.
and public-private partnership operation models, in: Natural Resources Forum, Sherriff, G., Adams, M., Blazejewski, L., Davies, N., & Kamerāde, D. (2020). From mobike
volume 39, Wiley Online Library, 41-52. to no bike in greater Manchester: Using the capabilities approach to explore Europe’s
London, T. (2013). Travel in London. London, UK: Transport London. first wave of dockless bike share. Journal of Transport Geography, 86, 102744.
Lu, M., An, K., Hsu, S.-C., & Zhu, R. (2019). Considering user behavior in free-floating Sinha, R. S., Wei, Y., & Hwang, S.-H. (2017). A survey on lpwa technology: Lora and nb-
bike sharing system design: A data-informed spatial agent-based model. Sustainable iot. Ict Express, 3, 14–21.
Cities and Society, 49, 101567. Subramaniyan, A. B. & Pan, R. Model reuse with bike rental station data, in: DC@ PKDD/
Mattson, J., & Godavarthy, R. (2017). Bike share in fargo, north dakota: Keys to success ECML.
and factors affecting ridership. Sustainable Cities and Society, 34, 174–182. Sun, C., & Zhai, Z. (2020). The efficacy of social distance and ventilation effectiveness in
Mavroforakis, M. E., & Theodoridis, S. (2006). A geometric approach to support vector preventing covid-19 transmission. Sustainable Cities and Society, 62, 102390.
machine (svm) classification. IEEE Transactions on Neural Networks, 17, 671–682. Tay, W.-L., Chui, C.-K., Ong, S.-H., & Ng, A. C.-M. (2013). Ensemble-based regression
Midgley, P. (2009). The role of smart bike-sharing systems in urban mobility. Journeys, 2, analysis of multimodal medical data for osteopenia diagnosis. Expert Systems with
23–31. Applications, 40, 811–819.
Midgley, P. (2011). Bicycle-sharing schemes: enhancing sustainable mobility in urban Tekouabou, S. C. K., Cherif, W., Silkan, H., et al. (2020). Improving parking availability
areas, United Nations. Department of Economic and Social Affairs, 8, 1–12. prediction in smart cities with iot and ensemble-based model. Journal of King Saud
Mishra, S., Mishra, D., & Santra, G. H. (2017). Adaptive boosting of weak regressors for University-Computer and Information Sciences.
forecasting of crop production considering climatic variability: An empirical Transportation, I. for N. Development Policy (New York, A. Gauthier, The bike-share
assessment. Journal of King Saud University-Computer and Information Sciences. planning guide, ITDP Institute for Planning & Development Policy, 2013.
Morency, C., Trepanier, M., & Godefroy, F. (2011). Insight into Montreal’s Bikesharing Tribastone, M., Clark, A., Gast, N., Gilmore, S., & Reijsbergen, D. (2014). Data validation
System, Technical Report. and requirements for case studies. QUANTICOL Deliverable, 5.
Murphy, H. (2010). Dublin bikes: An investigation in the context of multimodal transport. Tuama, D.Ó. (2015). Ripples through the city: Understanding the processes set in motion
Dublin: MSc Sustainable Development, Dublin Institute of Technology. through embedding a public bike sharing scheme in a city. Research in Transportation
Nair, R., & Miller-Hooks, E. (2011). Fleet management for vehicle sharing operations. Business & Management, 15, 15–27.
Transportation Science, 45, 524–540. Ueasangkomsate, P. Efficiency management of public bike-sharing system in Bangkok.
Nair, R., Miller-Hooks, E., Hampshire, R. C., & Bušić, A. (2013). Large-scale vehicle Vogel, P. & Mattfeld, D. C. Modeling of repositioning activities in bike-sharing systems,
sharing systems: analysis of vélib’. International Journal of Sustainable Transportation, in: World conference on transport research (WCTR).
7, 85–106. Vogel, P., Saavedra, B. A. N., & Mattfeld, D. C. A hybrid metaheuristic to solve the
Noland, R. B., & Ishaque, M. M. (2006). Smart bicycles in an urban area: Evaluation of a resource allocation problem in bike sharing systems, in: International Workshop on
pilot scheme in London. Journal of Public Transportation, 9, 5. Hybrid Metaheuristics, Springer, 16-29.
O’brien, O., Cheshire, J., & Batty, M. (2014). Mining bicycle sharing data for generating Vogel, P., Greiser, T., & Mattfeld, D. C. (2011). Understanding bike-sharing systems using
insights into sustainable transport systems. Journal of Transport Geography, 34, data mining: Exploring activity patterns. Procedia-Social and Behavioral Sciences, 20,
262–273. 514–523.
Ogilvie, F., & Goodman, A. (2012). Inequalities in usage of a public bicycle sharing Waserhole, A., Jost, V., & Brauner, N. (2013). Pricing techniques for self regulation in
scheme: socio-demographic predictors of uptake and usage of the London (uk) cycle vehicle sharing systems. Electronic Notes in Discrete Mathematics, 41, 149–156.
hire scheme. Preventive Medicine, 55, 40–45. Winslow, J., & Mont, O. (2019). Bicycle sharing: Sustainable value creation and
Parkes, S. D., Marsden, G., Shaheen, S. A., & Cohen, A. P. (2013). Understanding the institutionalisation strategies in Barcelona. Sustainability, 11, 728.
diffusion of public bikesharing systems: evidence from europe and north america. Woodcock, J., Tainio, M., Cheshire, J., O’Brien, O., & Goodman, A. (2014). Health effects
Journal of Transport Geography, 31, 94–103. of the London bicycle sharing system: health impact modelling study. BMJ, 348,
Randriamanamihaga, A. N., Côme, E., Oukhellou, L., & Govaert, G. (2014). Clustering g425.
the vélib? dynamic origin/destination flows using a family of poisson mixture Xu, H., Ying, J., Wu, H., & Lin, F. (2013). Public bicycle traffic flow prediction based on a
models. Neurocomputing, 141, 124–138. hybrid model. Applied Mathematics & Information Sciences, 7, 667.
Raviv, T., & Kolka, O. (2013). Optimal inventory management of a bike-sharing station. Yang, T., Haixiao, P., & Qing, S. Bike-sharing systems in beijing, shanghai and hangzhou
IIE Transactions, 45, 1077–1093. and their impact on travel behaviour, in: Transportation Research Board Annual
Razi, M. A., & Athappilly, K. (2005). A comparative predictive analysis of neural Meeting.
networks (nns), nonlinear regression and classification and regression tree (cart) Yigitcanlar, T., Kamruzzaman, M., Foth, M., Sabatini-Marques, J., da Costa, E., &
models. Expert Systems with Applications, 29, 65–74. Ioppolo, G. (2019). Can cities become smart without being sustainable? a systematic
Rifkin, J. (2011). The third industrial revolution: how lateral power is transforming energy, review of the literature. Sustainable Cities and Society, 45, 348–365.
the economy, and the world. Macmillan. Yoon, J. W., Pinelli, F., & Calabrese, F. (2012). Cityride: a predictive bike sharing journey
Rojas-Rueda, D., de Nazelle, A., Tainio, M., & Nieuwenhuijsen, M. J. (2011). The health advisor. in: 2012 IEEE 13th international conference on mobile data management, IEEE,
risks and benefits of cycling in urban environments compared with car use: health 306–311.
impact assessment study. BMJ, 343, d4521. Zguira, Y., Rivano, H., & Meddeb, A. (2018). Internet of bikes: A dtn protocol with data
Saharidis, G., Fragkogios, A., & Zygouri, E. A multi-periodic optimization modeling aggregation for urban data collection. Sensors, 18, 2819.
approach for the establishment of a bike sharing network: a case study of the city of Zhang, L., Zhang, J., Duan, Z.-y., & Bryde, D. (2015). Sustainable bike-sharing systems:
athens, in: Proceedings of the International MultiConference of Engineers and characteristics and commonalities across cases in urban china. Journal of Cleaner
Computer Scientists, volume 2, 1226-1231. Production, 97, 124–133.
Sarkar, A., Lathia, N., & Mascolo, C. (2015). Comparing cities’ cycling patterns using Zhang, D., Qian, L., Mao, B., Huang, C., Huang, B., & Si, Y. (2018). A data-driven design
online shared bicycle maps. Transportation, 42, 541–559. for fault detection of wind turbines using random forests and xgboost. IEEE Access, 6,
Sayarshad, H., Tavassoli, S., & Zhao, F. (2012). A multi-periodic optimization 21020–21031.
formulation for bike planning and bike utilization. Applied Mathematical Modelling, Zhang, Y., Wen, H., Qiu, F., Wang, Z., & Abbas, H. (2019). Ibike: Intelligent public
36, 4944–4951. bicycle services assisted by data analytics. Future Generation Computer Systems, 95,
Schuijbroek, J., Hampshire, R. C., & Van Hoeve, W.-J. (2017). Inventory rebalancing and 187–197.
vehicle routing in bike sharing systems. European Journal of Operational Research,
257, 992–1004.

14