See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/276120258

A Real-Time Parking Prediction System for Smart Cities

Article  in  Journal of Intelligent Transportation Systems Technology Planning and Operations · April 2015
DOI: 10.1080/15472450.2015.1037955

CITATIONS READS
146 6,843

4 authors:

Eleni I Vlahogianni Konstantinos Kepaptsoglou

National Technical University of Athens National Technical University of Athens
192 PUBLICATIONS   5,586 CITATIONS    128 PUBLICATIONS   1,627 CITATIONS   

SEE PROFILE SEE PROFILE

Vassileios Tsetsos Matthew G. Karlaftis

National and Kapodistrian University of Athens National Technical University of Athens
55 PUBLICATIONS   649 CITATIONS    213 PUBLICATIONS   10,232 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Campus solar roads: Optimization of solar panel and electric charging station location for university bus route View project

2 be safe View project

All content following this page was uploaded by Eleni I Vlahogianni on 24 August 2015.

The user has requested enhancement of the downloaded file.

 A Real-Time Parking Prediction System for Smart Cities

Eleni I. Vlahogianni*, Ph.D., Konstantinos Kepaptsoglou, Ph.D., Vassileios Tsetsos, Ph.D.

and Matthew G. Karlaftis†, Ph.D.
*corresponding author
School of Civil Engineering
National Technical University of Athens
5 Iroon Polytechniou Str, Zografou Campus, 15773 Athens Greece
Tel: +30 210 772 1369; Fax: +30 210 772 1454
e-mail: [email protected]

ABSTRACT
A methodological framework for multiple steps ahead parking availability prediction is
presented. Two different types of predictions are provided: i. the probability a free space
to continue being free in subsequent time intervals, and ii. the short-term parking
occupancy prediction in selected regions of an urban road network. The available data
come from a wide network of on-street parking sensors in the “smart” city of Santander,
Spain. The sensor network is segmented in four different regions and, then, survival
and neural network models are developed for each region separately. Findings show
that the Weibull parametric models best describe the probability of a parking space to
continue to be free in the forthcoming time intervals. Moreover, simple genetically
optimized Multilayer Perceptrons accurately predict region parking occupancy rates up
to 30 minutes in the future by exploiting 1 minute data. Finally, the real time, web
based, implementation of the proposed parking prediction availability system is
presented.

Keywords: Parking occupancy, duration modeling, neural networks, parking sensors,

smart city, internet of things

1
1. INTRODUCTION
Parking availability prediction is among the most important factors affecting both
private car based trip decisions and traffic conditions in urban areas. Drivers’ decisions
are temporally dependent, implying that they are influenced by past experience, as well
as real-time (on road) perceptions. Parking is such a case where prior knowledge on
possible prevailing conditions (e.g. difficulty in finding a parking space, off-street
parking costs, and so on) affects drivers’ parking decisions. At the same time, vehicles
in search of free parking spaces negatively impact traffic conditions and the
environment. In this context, parking information provision is a research area of
particular interest, since modern communication technologies offer alternative ways of
delivering information to travelers in a timely and effective manner.
Various systems developed for providing parking information and guidance have been
proposed by researchers in the past (Oh et al. 2002, Liu et al. 2006, Caicedo et al. 2012,
Giuffrè et al. 2012); parking information is usually disseminated by variable message
signs or through the internet, cellular phones, PDA, and GIS technologies (Liu et al.
2006, Giuffrè et al. 2012). Teodorovic and Lucic (2006) argue that, although parking
guidance systems may not affect the occupancy rate or average parking duration,
drivers tend to greatly appreciate the information provided by such systems. This is
because such systems have certain positive effects on a city’s traffic operations, as well
as on the personal trips of road users. First, they significantly increase the probability
of finding free parking spaces and mitigate frustration of those drivers/visitors
unfamiliar with the city center. They are supposed to decrease queues in front of parking
garages and decrease total vehicle-miles travelled (particularly in the city center).
Additionally, they help road users to optimize their trips and, thus, improve vehicles’
energy consumption and decrease emissions.
The usefulness of such predictive parking information is straightforwardly understood.
If all drivers act without information and make “uninformed” choices, they will
probably resort to similar optimal decisions leading to induced long waiting times,
queues and increased parking circling. On the other hand, dissemination of accurate
and timely parking availability information may lead to improved driver decisions and
parking searching (Caicedo et al. 2006). From a transport operator perspective, accurate
parking availability predictions may lead to better management of the system,
congestion mitigation due to queue formation avoidance, and so on.
The objective of this paper is to exploit statistical and computational intelligence
methods for developing a methodology that can be used for multiple steps ahead on-
street parking prediction in “smart” urban areas. This work takes advantage of massive
real-time parking data availability, obtained by an extended parking sensor network
available in the “smart” city of Santander, Spain. Models are developed for predicting
expected parking occupancy along with the probability of finding free parking spaces.
Traditional survival analysis models, as well as neural network models are developed.
The methodology is evaluated and a real-time, web based system exploiting the
proposed prediction models, for the city of Santander, Spain, is presented.

2. PARKING MODELING: OBJECTIVES AND METHODS

Parking modeling has been a topic of interest since the 1970s and has been studied from
various angles including parking patterns, impacts on traffic, off-street parking
technologies, parking policy, choice and location models, as well as economic models
2
describing parking conditions (Van Der Goot 1982, Arnott and Rowse 1991, Wong et
al. 2000, Hensher and King 2001, Bonsall and Palmer 2004, Ibeas t al. 2011). An early
review by Young et al. (1991) reported three categories of parking related models:
driver behavior with respect to parking (parking choice), optimal positioning of parking
lots (parking allocation), and interaction of parking operations with other transportation
system elements and infrastructures (parking interaction).
Parking availability prediction, on the other hand, coupled with modern capabilities on
data collection and processing, along with ITS based exploitation of such information,
although indicated as a challenge since the early 1990s (Bonsall 1991), has been
introduced in recent years. David et al. (2000) proposed a model for event-oriented
forecasting of parking occupancy based on standardized daily distribution occupancy
rate curves and on-line data obtained from parking lots equipped with detectors. On-
street occupancy estimation in the absence of detectors was investigated by David and
Keller (2001). An event driven model was developed for that purpose using historical
socioeconomic and parking specific data; the model was successfully validated for the
city of Munich, Germany. Teodorovic and Lucic (2006) developed a system based on
fuzzy logic, simulation and optimization models, which decides whether to accept or
reject new parking requests in real-time, according to estimated availability for parking
lots. Martens and Benenson (2008) exploited agent-based modeling for representing
parking behavior by integrating the effect of considering real-time and expected parking
availability, prices, and parking enforcement. Caicedo (2009) developed a discrete
choice model for combining on-line and historical data for real-time, off street parking
availability prediction; this model was later used by Caicedo et al. (2012). Fubuyasi et
al. (2011) developed ParkPGH, a system for predicting parking availability in eight
Pittsburg parking facilities using historical and real-time data. A fuzzy logic model for
estimating the uncertainty of peak period parking availability in park-and-ride facilities
was proposed by Chen et al. (2013).
Modern technological advances have revolutionized the ways of monitoring and
recording transportation operations and data especially in the case of parking (Thornton
et al. 2014). These have, however, been exploited for parking prediction to a lesser
extent, and have mainly focused on off-street parking facilities.

2.1. Smart Cities and Parking Prediction Challenges

Cities are characterized as “smart” when their transportation and communication (ICT)
infrastructures along with their human and social capital investments cooperate and
actively support sustainable growth and high quality of life, through participatory action
and engagement, while preserving natural resources (Caragliu et al. 2009). Indeed, as
noted by Komninos (2009), innovation and use of ICT for improving capacity of
infrastructures are key elements of “smart” cities. A novel type of “smart” city
infrastructure, applicable to the transportation sector is the so-called Internet of Things
(IoT). IoT consists of a variety of devices or objects – such as Radio-Frequency
IDentification (RFID) tags, sensors, actuators, mobile phones, and so on – which,
through unique addressing schemes, are able to interact with each other and cooperate
with their neighbors to reach common goals (Atzori et al. 2010). By continuously
collecting, analyzing and redistributing transportation information, IoT networks can
offer valuable, real time information to both travelers and operators, and thus support
and improve the operations of ITS, traffic and public transportation systems. Although
one can trace several reasons that may prevent IoT to be fully developed in urban

3
environments, this unique technological paradigm is expected to substantially support
sustainable development of future smart cities (Vlacheas et al. 2013).
The SmartSantander project is such a case of an IoT architecture deployed in the city
of Santander, to achieve a massive deployment of sensors and network communications
in order to provide efficient and equitable transportation and other services to citizens
(http://www.smartsantander.eu/). The IoT network of the SmartSantander project
consists of (Gutiérrez et al. 2013):
 IoT nodes: these are responsible for sensing and collecting information from the
natural and socioeconomic environment and activities (temperature, CO, noise,
light, car presence etc.).
 Repeaters: These nodes are placed high above ground in street lights, semaphores,
information panels, etc, and behave as forwarding nodes of IoT information.
 Gateways: Both IoT nodes and repeaters are configured to send all the information
to gateways. Once information is received by gateways it can be either stored
locally or sent to central processing units through different interfaces (for
example, WiFi, GPRS/UMTS or Ethernet).
SmartSantander and similar testbeds fuse high resolution datasets stemming from both
static and mobile sensors. Such data refer to macroscopic traffic flow in road sections
of interest, detailed parking information in real-time in urban areas, as well as high
resolution transit information. In the case examined in this paper, having parking
sensors collecting real time information on their occupancy provides accurate high
resolution parking information. New technologies can replace common parking metrics
of average duration, turnover rate and occupancy in extended time windows of 1 to 3
hours (which are manually collected in small scale regions) with accurate, both
aggregate and disaggregate information, on occupancy and parking duration. In this
framework there are several questions that may arise; this paper focuses on the
following:
 Is it possible to predict parking occupancy using time series modeling
approaches based on data collected from an IoT network of sensors?
 How accurate are parking occupancy predictions produced by such models in
relation to the predictive horizon?
 What are the statistical properties of parking space duration and how can we
predict the probability of having free parking spaces in the area of interest?

3. MODELING PARKING AVAILABILITY BASED ON SENSOR DATA

3.1 Methodology
The architecture of the proposed parking occupancy prediction system is presented in
Figure 1. In each sliding time window T , parking efficiency is defined by the following
metrics:
 Duration ( Dt ) of free parking space: The average time duration that a slot is
free, over a certain time period
 Occupancy ( Ot ): The percentage of parking slots occupied during a predefined
time period.

4
The free space duration is an indication of how often a parking space becomes available,
whereas occupancy relates the parking accumulation, meaning the number of parked
vehicles in the study area at any specific time period, to the parking capacity.
The proposed methodology has two modules. The first module is a real-time time series
occupancy prediction scheme based on recurrent artificial neural networks. Simple, yet,
flexible memory mechanisms will be applied in order to replicate the temporal
dynamics of parking occupancy Ot . The model is presented with past information of
occupancy ( Ot  ,..., Ot ( m1) ) to predict occupancy one step ahead Ot . The
characteristics of the model’s memory (the time delay τ and the dimension m) will be
evaluated through a nonlinear analysis of the dynamics of parking occupancy. The use
of exogenous variables, such as the type of day, or day of week, will be evaluated
relative to the improvement on the accuracy of predictions they impose.
<Figure 1>
When considering short-term prediction systems that operate in real-time and in an
“intelligent” technology-based environment, the effectiveness depends, mostly, on
predicting traffic information in a timely manner (Smith and Oswald 2003,
(Vlahogianni et al. 2006). Real-time system effectiveness depends both on the results
and on the time in which these are produced (Shin and Ramanathan 1994). The
computational time for making a prediction mainly depends on the functional form of
the prediction system; empirical results show that data-driven prediction systems that
include recursive data-search algorithms exhibit ‘best’ prediction accuracy but need
extensive computational time for convergence at acceptable results (Smith and Oswald
2003). For this, the final structure of the prediction model is kept as simple and flexible
as possible. The trade-off between simplicity and efficiency is studied in a preliminary
stage of analysis using genetic algorithms to optimize the structural and learning
parameters of the different models.
The second is a static approach for estimating the probability of finding available
parking space with relation to a series of factors such as the type of day (weekday,
weekend), and the time period (peak, off-peak, morning evening). The modeling will
be based on survival analysis. Parametric hazard based modeling may be developed
under a variety of functional forms. The model will be also presented with the predicted
value of occupancy as provided by the first module. The output of the model (the
probability of finding an available parking space and the anticipated number of
available space) can be visually depicted using various graph methods (such as heat
maps), to provide useful information to users.

3.2 Neural Network Models for Time Series Prediction

Time series modeling is a popular approach for making predictions in transportation
problems (Karlaftis and Vlahogianni 2011). This approach is suitable for analyzing
parking occupancy due to the temporal structure of the performance measures of
parking systems. A common prediction strategy implemented in transportation
problems is based on the autoregressive moving average family of models. These
models are relatively straightforward mathematically and easy to produce; however,
they are severely constrained by stationarity and linearity, characteristics that most
frequently violated in real transportation time series. Treating non-stationarity and
nonlinearity may lead to a tedious process without achieving the desired levels of
accuracy in predictions and modeling reliability (Washington et al. 2010). Neural
5
Networks (NNs) for time series provides a good alternative, as they relax many of these
constraints, and also appear to provide short-term forecasting models that are more
adaptable to sudden shifts in the data (Vlahogianni 2009). A recent study on traffic time
series prediction has shown the structural equivalencies between nonlinear univariate
and multivariate ARIMA models with exogenous variables and dynamic forms of
Multilayer Perceptrons (MLPs) (Vlahogianni and Karlaftis 2014). The simplest of all
is the NAR(p) structure of order p. In general, The MLP presented with p lagged values
of parking occupancy Ok ( k=1,…,n) may act as a predictor of the form (Mandic and
Chambers 2001):
I  p 
Oˆ k  hˆ(Ok 1 ,..., Ok  p )  Wi f   wij Ok  j  i  (1)
i 1  j 1 
Where f    is a smooth monotonic function, Wi and wij are the weights of the
connections (synapses) - coefficients - estimated through learning (training) thereby
obtaining an estimate of the nonlinear approximation ĥ . The network converges to an
 O 
2
estimate of ĥ by minimizing the residuals k k  Oˆ k using a learning algorithm
(usually back-propagation).
For a time series consideration in a NN framework, the MLP should be modified to
account for the time sequence of events under study. This is usually accomplished by
adding memory structures in than MLP that retain the effect of past information to the
system and use it during learning. The memory is accomplished using local - at an intra-
neuron level - and global - between neurons of different layers - recurrent connections
in a neural network. Memory mechanisms may be of a simple tap delay form realizing
structures:
O(t )  O(t   ),..., O(t  (m  1) ) (2)

where τ is the delay and m the dimension of the horizon of past information introduced
to the model or of other more complex mathematical forms (Gamma memory and so
on). The networks with memory usually require cumbersome and slow learning
procedures that may not be always stable. To avoid this, static MLPs can be externally
modified to represent the temporal characteristics of transportation time series (e.g.
parking occupancy) in a manner resembling the common statistical prediction
techniques. The introduction of such data inputs in MLPs that are unchanged in their
internal structural logic may conceptually approximate very complex and multivariate
statistical structures with equal efficiency as classical MLPs (Vlahogianni et al. 2005).

3.3 Hazard-Based Free Parking Duration Modeling

Hazard-based duration modeling deals with the statistical representation of time to
event data, a very frequent form of data in transportation problems; typical example of
such transportation data is the time to clear and incident (Chung 2010, Vlahogianni and
Karlaftis 2013), the time until the end of congested phenomena (Stathopoulos and
Karlaftis 2002, Vlahogianni et al. 2011), the time until the end of transit vehicle repair
(Karlaftis 2011), the time to an activity (Zhong and Hunt 2010, Habib 2012), the time
to household evacuation under the emergence of a physical disaster (Hasan et al. 2013),
the time for a pedestrian to cross a signalized intersection approach (Tiwari et al. 2007),
the time to vehicle transaction (Rashidi et al. 2011), the time to complete an overtake
(Vlahogianni 2013) and so on. Extensive review on the hazard based transportation
6
modeling applications may be found in Hensher and Mannering (1994) and Bhat
(2000). Moreover, methodological, computational and estimation issues in duration
modeling with focus on transportation problems may be found in Washington et al.
(2010).
Parametric hazard-based modeling is based on two concepts: the survival function and
the hazard function. Let T be a non-negative random variable representing the time a
vehicle occupy a space, the survival function is defined as the probability that T is of
length at least t (i.e. a parking space is occupied at least t min) and is given by:
S (t )  P (T  t )  1  F (t ), 0  t , (3)
where F (t ) is the cumulative probability. The survival function S (t ) can have a variety
of shapes following certain restrictions; it is bounded by 0 and 1 and, as T cannot be
negative, S(0) = 1. Moreover, the larger the t, S never increases (and usually decreases).
For continuous survival data, the hazard function specifies the instantaneous failure rate
at T=t conditional upon survival to time t and defined as:
Pr(t  T  t | T  t ) f (t )
h(t )  lim  , (4)
t 0 t S (t )

where f (t ) is the probability density function:

dF (t ) dS (t )
f (t )   (5)
dt dt
By combining (2) and (3) we get:
d
h(t )   log S (t ) (6)
dt
The hazard function h(t) is always nonnegative and, unlike survival functions, has no
upper bound. Parametric hazard based models may have a range of different functional
forms aiming at modeling different distributional characteristics. The hazard rate
h(t X) with covariates is given by:

h(t X)  h0 (t ) exp(βX) (7)

where h0 (t ) is the baseline hazard rate, β is a vector of estimated parameters and X is
a vector of covariates.

4. IMPLEMENTATION AND RESULTS

4.1 Dataset
Parking information comes from a network of 400 sensors located in the area of
Santander (Figure 2). Parking sensors based on ferromagnetic technology, buried under
the asphalt, have been installed at the main free parking areas of the city center, in order
to detect parking sites availability in these zones. The parking sensors are deployed in
a regulated zone (working days from 10:00 to 14:00 and from 16:00 to 20:00) there is
a 2 hour time limit in the parking duration with a fare of 0.12 €/min. The above
restrictions do not apply for citizens living in the study area.

7
The information from the parking sensors is wirelessly transmitted to the IoT
corresponding repeaters that are installed in high-rise places (e.g. street lights,
semaphores information panels etc). The communication between IoT sensors and
repeaters is done through 802.15.4 protocol. All information is directed (through
802.15.4 protocol) and stored to the getaways (more details on the infrastructure
description may be found in
http://www.smartsantander.eu/index.php/testbeds/item/132-santander-summary).
The available dataset consists of time series of the state (free/occupied) of each sensor
every 1 minute for the months April to September 2013. With this information, it is
possible to straightforwardly calculate parking metrics such as parking accumulation,
occupancy, duration and so on. In this study we focus on: i. parking occupancy (%) of
a specific area, which is the percentage of parking spaces that are occupied by vehicles
within a time interval; ii. the turnover rate (veh/space/t), which is equal to the number
of vehicles per parking space in a time interval; and, iii. the duration that a parking
space is not occupied by vehicles.
<Figure 2>
For efficiently analyzing the parking characteristics, the entire area covered by the
parking sensors is further segmented into four (4) regions; this segmentation is
empirically done based on drivers’ feasible cruising paths, while searching for available
parking spaces. The specific regions are depicted in Figure 2. The four regions are of
different shape. Region 1 controls an area of approximately 1,800m2, Region 2 an area
of 12,000 m2. Regions 3 and 4 are 30,000 m2 and 76,000 m2 respectively. The proposed
segmentation is coarse enough to reflect the process of search for free parking space in
the study area and is dependent on the existing road network. Evidently, a more
exhaustive segmentation may be proposed (e.g. more than 4 regions), which may result
in more detailed information on the predictive parking occupancy patterns in the study
area.
Figure 3 and Figure 4 show the time series of parking occupancy (%) and the turnover
rate (veh/space/h) per region for a typical week. Further statistical testing shows that
there exist differences in the mean of occupancy and turnover rate between weekdays
and weekends for all regions. Moreover, high variability in average parking occupancy
for Region I in the morning and afternoon is observed; setting 85% as the critical
occupancy, we may distinguish between high occupancy and low occupancy periods
within a day. High/low occupancy periods for the other parking regions have an
occupancy threshold of 90%.
<Figure 3>
<Figure 4>
The duration of occupied parking spaces for all regions is best described by a Weibull
distribution; the survival curve has the following form: S (t )  exp(t p ) , with
parameters λ and p to be estimated. Figure 5 and Figure 6 show the survival
probabilities for the duration of occupied parking spaces per region. Higher
probabilities after 60 minutes on average are expected in Region I and II when
compared to those of Regions III and IV, where the probability of a space being
occupied after 60 minutes drops below 0.5.
<Figure 5>
<Figure 6>
8
4.2 Parking Occupancy Prediction
A unique MLP for each parking region is developed in order to predict the overall
occupancy (%) of parking spaces using past information on its evolution. The input
space is introduced with information on occupancy with a look back time window
ranging from 5 to 10 minutes. For the prediction system to be useful, multiple steps
ahead predictions are required. Here, the prediction horizon of the developed models is
extended from 1 step to 30 steps ahead (that is, from 1 to 30 minutes ahead).
Following previous research on short term prediction using neural networks in
transportation problems (a detailed description is given in Vlahogianni et al. 2005), the
neural networks are optimized with respect to their structural and learning properties
using genetic algorithms. Training is conducted using a simple back propagation
algorithm with genetically optimized learning rate and momentum. The structure of the
hidden layer (number of hidden units) is also optimized using genetic algorithms. The
model’s look back time window is also genetically optimized. Based on the above
description, a population of 100 generations is evaluated. Each generation involved the
development and testing of 50 chromosomes that include the number of hidden units h
(structure), the learning rate η and the momentum μ (learning), as well as the time
dependent inputs (parameters τ and m of Equation 2). A roulette selection mechanism
is employed. The crossover probability is set to 0.9 and the mutation probability to 0.02.
The fitness function is the mean square error in the cross-validation set.
The available data is divided into the training (60%), cross-validation (20%) and test
set (20%) used for training the models, evaluating the training of the models, and testing
the generalization power of the models, respectively. The genetic optimization showed
that an MLP of 8 hidden layers and a look back time window of 5 minutes in the past
may be efficiently used to predict parking occupancy (%) up to 30 steps in the future
with high accuracy.
Prediction results (test set) are seen in Table 1 with respect to the following forecasting
metrics:
 yˆ i   yi 
 Mean Absolute Error (MAE): i

N
  yˆ  yi  
2
i 
 Root mean squared error (RMSE): i

N
1 yˆ  yi 
 Mean absolute percentage error (MAPE):  i 
N i yi 

  yˆ  yi     yˆ  yi  
2 2
i  i 
 Root relative squared error (RRSE): i i
N N
where yi  and yˆi  is the actual and predicted value with i  1,.., N , I  is the time step,
N is the number of samples, yi  is the last known value relative to the prediction step.
Results seem encouraging particularly for longer prediction horizons. For example, the
models are able to predict each region’s parking occupancy 15 minutes into the future
with less than 3.6% MAPE.
<Table 1>

9
Parking occupancy predictability is different between the four Regions. It is to note that
MLPs manage to produce predictions up to 30 minutes ahead with acceptable levels of
accuracy. Although this may hint a rather smooth temporal evolution of 1 minute
parking occupancy data, the efficiency of these models may prove to be critical in
delivering useful information services to travelers in urban areas.
Interestingly, all models perform well and produce predictions that are more accurate
than a naïve prediction (e.g. taking the last known value relative to the step of
prediction) as seen from the RRSE. The larger the predictive horizon the more accurate
the MLP becomes relative to a naïve prediction, an indication of the robustness of the
NNs in dealing with problems of ranging levels of complexity. This is a strong
indication that the proposed models may provide solid predictions that may enable
potential predictive information dissemination in real-time using web-based tools, cell
phone or other media.

4.3 Survival Analysis of Free Parking Spaces

Hazard based free parking space duration models are developed for each parking
region. Based on preliminary analysis, a Weibull survival function is used. Several
forms of Weibull regression models are tested with independent variables being the
TypeDay (weekend/weekday), the Period of day (high /low occupancy period) and the
Weekday (1 to 7, from Monday to Sunday). Results are seen in Table 2.
<Table 2>
As can be observed, different independent variables are significant in each parking
region when modeling free parking space duration. In Region I, free space durations
are dependent on the type of day (weekend or weekdays), whereas in Region II and III,
free space durations depend on the period of the day. Finally, in Region IV, free space
durations change with regard to the daily evolution of parking demand. Interesting
results arise from observing the survival probability curves (Figure 7) that show
whether the phenomenon will continue to be observed in the forthcoming intervals,
given it has lasted up to time t. For example, in Region I there is a probability of a
parking space continuing to be free after 5 minutes of 0.76 for weekdays; this increases
to 0.83 for weekends. The same probability for Region II spaces is 0.77 for high
occupancy periods and increases to 0.79 in low occupancy periods. In Region III the
probability of a space to be free after 5 minutes is 0.69 and 0.73 for high and low
occupancy periods respectively. Finally, in Region IV the differences of the same
probability are small but significant; the probability ranges from 0.63 in Mondays and
0.69 in Sundays.
<Figure 7>
The above survival curves may be used as a predictive tool for parking availability.
Interestingly, the estimations provided by the survival model may act complementary
to the parking occupancy predictions that require relatively extensive and continuous
data streams. This may significantly improve the real-time operability of an intelligent
parking prediction framework.

5. REAL WORLD APPLICATION

The above models have been implemented in the MITOS system (Multi-Input
TranspOrt planning System). MITOS is a system developed within SmartSantander
10
project (http://mitos.smartsantander.eu), in the city of Santander, Spain. MITOS goal is
to deliver novel Intelligent Transportation Services (ITS) to smart city citizens and
capture perceived user experience. The fact that an IoT (Internet of Things)
infrastructure is available, provides new ways of quantifying the effect that ITS may
have on the daily transportation of the commuters and the environment. The MITOS
platform provides an integrated multi-modal transportation guide, which allows citizens
and visitors to optimally choose their trips and get accurate information and guidance
before and during their trip. The MITOS platform includes the following services:
 Stop/Line/POI Survey and Search for public transport,
 Route Guide,
 Parking monitoring and short-term parking occupancy prediction.
The architecture of the MITOS platform is seen in Figure 8. The services and
applications deployed and demonstrated in the context of MITOS include a Web portal
for transportation information, which is used as a city guide for citizens and tourists,
and a mobile application for advanced transportation services. The last provides all type
of information required by commuters and enables intuitive ways for searching routes
and geo-referenced information based on the personal preferences and abilities of each
user.
<Figure 8>
An added value of the MITOS platform comes from the exploitation of various diverse
sensor sources such as:
 Participatory sensors: users will act as sensors that provide relevant traffic and
travel information in the form of free text or predefined messages (e.g. “heavy
traffic”, “too much noise”, etc.) and/or image.
 Environmental data sensors, measuring noise, temperature and CO/CO2
emissions.
 On-street parking space occupancy sensors.
 Traffic occupancy sensors
 On-vehicle devices (GPS) installed in buses.
The implementation of a parking prediction service involves retrieving parking sensor
data every minute (through the SmartSantander framework APIs) and storing them in
a relational database (MITOS DB). Next, the algorithms described earlier are executed,
and a continuous stream of predictions is generated. Those predicted values are stored
back in the database, so that related user requests on parking availability can be served
as fast as possible. Following user requests for parking occupancy, data are retrieved
and visualized on a digital Web map (Figure 9) and on a mobile phone (Figure 10). The
visualization involves data easily perceived by the end users, such as estimated
occupancy and estimated number of free parking slots in the short-term future (5, 10
and 15 minutes ahead).
<Figure 9>
<Figure 10>

6. CONCLUSIONS
In this paper we exploited high resolution parking occupancy data and developed and
tested a system for short and longer term parking availability prediction in urban areas.
11
The system encompasses two modules; the first introduces neural networks for the
prediction of the time series of parking occupancy in different regions of an urban
network, while the second applies survival analysis to predict the probability that a
parking space will be free in the following time intervals. Findings show that the
duration of free parking space follows a Weibull distribution. Moreover, the neural
networks adequately captured the temporal evolution of parking occupancy and may
accurately predict occupancy up to half an hour ahead. The proposed model, as it stands,
is flexible and of reduced complexity aiming to be implemented online. This is shown
by its incorporation within an innovative routing service (MITOS) for Santander, Spain,
using real time data from the city’s “smart” infrastructure.
From a methodological perspective, the proposed approach is tested on limited data that
may not claim to be representative of the monthly variations in parking demand. The
evolution of parking occupancy may significantly differ between summer and winter
months and may have other cyclicity features that were not addressed. Thus, a more
comprehensive dataset of historical parking occupancy and duration information may
improve forecasting. Towards the treatment of seasonality features, retraining strategies
based on the daily or monthly seasonality should be also considered. The treatment of
seasonal features may be done by introducing specific dummy variables to NNs, or by
using more complex NN structures, such as the temporal NNs that include memory
mechanisms. Nevertheless, these networks should be treated with caution due to the
difficulties in training and convergence.
Moreover, a critical limitation of the present approach is the lack of traffic data that
would have provided a more consistent formulation of the parking prediction problem
to the evolution of traffic demand. Correlations with traffic’s spatio-temporal demand
are expected to improve the predictability of parking occupancy. Nevertheless, these
recommended directions for improving forecasts may result to more complex and
difficult to train networks. Thus, a thorough consideration should be given to issues
related to computational efficiency during model development. Further, traffic spatio-
temporal patterns should be also jointly considered with the extent and coverage of the
free parking space prediction models. Evidently, area wide and street wide prediction
models may serve different purposes in relation to the type and characteristics of the
ITS application.
From a conceptual perspective, the proposed approach can provide the anticipated free
parking slots, as well as the probability of having free parking spaces in the specific
area. In the absence of real time data, the second information that is based on historical
data may be found extremely useful. The proposed approach may be extended to
include parking fares (where they apply), as well as underground parking stations near
each zone to enhance the explanatory power of the model and improve predictions.

ACKNOLEDGEMENTS
This work is part of the MITOS experiment, incorporated in the SmartSantader project
and funded by the European Union Seventh Framework Programme (FP7).

REFERENCES
Arnott, R. and Rowse, J. (1991). Modeling parking. Journal of Urban Economics, 45(1),
97–124.
12
Atzori, L., Iera, A., and Morabito, F., (2010). The Internet of Things: A survey,
Computer Networks, 54(15), 2787-2805.
Bhat, C.R. (2000). Duration modeling. In: D.A. Hensher and K.J. Button, eds.
Handbook of transport modeling. London: Elsevier Science.
Bonsall, PW and Palmer, I. (2004). Modelling drivers’ car parking behaviour using data
from a travel choice simulator, Transportation Research Part C: Emerging
Technologies, 12(5), 321-347.
Bonsall. P.W. (1991). The changing face of parking-related data collection and
analysis: The role of new technologies. Transportation 18(1), 83-106.
Caicedo, F. (2009). The use of space availability information in “PARC” systems to
reduce search times in parking facilities. Transportation Research Part C: Emerging
Technologies, 17(1), 6–68.
Caicedo, F., Blazquez, C., and Miranda, P. (2012). Prediction of parking space
availability in real time. Expert Systems with Applications, 39(8), 7281-7290.
Caicedo, F., Robuste, F., Lopez-Pita, A. (2006). Parking management and modeling of
car park patron behavior in underground facilities. Transportation Research Record,
1956, 60–67.
Caragliu, A; Del Bo, C. and Nijkamp, P. (2009). Smart cities in Europe. Serie Research
Memoranda 0048 (VU University Amsterdam, Faculty of Economics, Business
Administration and Econometrics).
Chen, Z., Xia, J., and Irawan, B., (2013). Development of Fuzzy Logic Forecast Models
for Location-Based Parking Finding Services, Mathematical Problems in Engineering,
vol. 2013, Article ID 473471, 6 pages.
Chung, Y. (2010). Development of an accident duration prediction model on the Korean
Freeway Systems, Accident Analysis and Prevention, 42(2), 282–289.
David, A., & Keller, H. (2001). Event-Driven Modelling Of On-Street Parking
Probability. Proceedings of the 8th World Congress on Intelligent Transport Systems,
Sydney, New South Wales, Australia.
David, A., Overkamp, K., and Scheuerer, W. (2000). Event-oriented forecast of the
occupancy rate of parking spaces as part of a parking information service. Proceedings
of the 7th World Congress on Intelligent Systems, Turin, Italy, Nov 6th-9th.
Fabusuyi, T., Hampshire, R.C., Hill, V., Sasanuma, K. A (2011). Predictive Model and
Evaluation Framework for Smart Parking: The Case of ParkPGH. Proceedings of the
18th ITS World Congress, Orlando, Oct 16th-20th.
Giuffrè, T., Siniscalchi, M.S., Tesoriere, G. (2012). A Novel Architecture of Parking
Management for Smart Cities, Proceedia - Social and Behavioral Sciences, 53(3), 16-
28.
Gutiérrez, V., Galache, J.A., Sánchez, L., Muñoz, L., Hernández, J.M., Fernandes, J.,
Presser, M., (2013). SmartSantander: Internet of Things Research and Innovation
through Citizen Participation, Future Internet Assembly - Dublin (Ireland), Lecture
Notes in Computer Science, 7858, 173-186.
Habib, K. M. N. (2012). Modeling commuting mode choice jointly with work start time
and work duration, Transportation Research Part A: Policy and Practice, 46(1), 33-47.

13
Hasan, S., Mesa-Arango, R. Ukkusuri, S. (2013). A random-parameter hazard-based
model to understand household evacuation timing behavior, Transportation Research
Part C: Emerging Technologies, 27(1), 108–116.
Hensher, D. A. and Mannering, F.L. (1994). Hazard-based duration models and their
application to transport analysis, Transport Reviews: A Transnational Transdisciplinary
Journal, 14(1), 63-82.
Hensher, D.A., King, J., 2001. Parking demand and responsiveness to supply, pricing
and location in the Sydney central business district. Transportation Research Part A:
Policy and Practice 35, 177–196.

Ibeas, A., Cordera, R., dell'Olio, L., & Moura, J. L. (2011). Modelling demand in
restricted parking zones. Transportation Research Part a-Policy and Practice, 45(6),
485–498.
Karlaftis, M. G. and Vlahogianni, E. I. (2011). Statistics versus Neural Networks in
Transportation Research: Differences, Similarities and Some Insights, Transportation
Research Part C: Emerging Technologies, 19(3), 387-399.
Karlaftis, M. G.. (2011). Modeling transit vehicle repair duration and active service
time, Transportation Planning and Technology, 34(5), 433-442.
Komninos, N. (2009). Intelligent cities: towards interactive and global innovation
environments. International Journal of Innovation and Regional Development, 1(4),
337–355.
Liu, Q., Lu, H., Zou, B., and Li, Q. (2006). Design and development of parking
guidance information systems based on Web and GIS technology. In Proceedings of
the 6th international conference on ITS telecommunications, Chengdu, China.
Mandic D. P., and Chambers J. A., (2001). Recurrent Neural Networks for Prediction:
Learning Algorithms, Architectures, and Stability, Wiley Series in Adaptive and
Learning Systems for Signal Processing, Communications, and Control, ISBN
0471495174 (Print), John Wiley & Sons.
Martens, K., and Benenson, I. (2008). Evaluating urban parking policies with agent-
based model of driver parking behavior. Transportation Research Record, 2046, 37–44.
Oh, B., Lee, S., Kim, M., and Yang, Y. (2002). Spatial applications using 4S technology
for mobile environment. In Proceedings of the geoscience and remote sensing
symposium, Toronto, Canada.
Rashidi, T. H., Mohammadian A., Koppelman, F. S. (2011). Modeling
interdependencies between vehicle transaction, residential relocation and job change,
Transportation, 38(6), 909-932.
Shin K., and Ramanathan P. (1994). Real-time computing: A new discipline of
computer science and engineering. Proceedings of the IEEE, 82, 6-24.
Smith B. L., Oswald, R. K. (2003). Meeting Real-Time Requirements with Imprecise
Computations: A Case Study in Traffic Flow Forecasting. Computer Aided Civil and
Infrastructure Engineering, 18(3), 201-213.
Stathopoulos, A., and Karlaftis, M. G. (2002). Modeling duration of urban traffic
congestion. J. Transp. Eng., 128(6), 587–590.

14
Teodorovic, D., and Lucic, P. (2006). Intelligent parking systems. European Journal of
Operational Research, 175(3), 1666-1681.
Thornton, D. A., Redmill, K, and Coifman, B. (2014). Automated parking surveys from
a LIDAR equipped vehicle, Transportation Research Part C: Emerging Technologies,
39, 23-35.
Tiwari G., Bangdiwala, S., Saraswat, A. and Gaurav, S. (2007). Survival analysis:
Pedestrian risk exposure at signalized intersections Transportation Research Part F, 10,
(1), 77–89.
Van Der Goot, D., 1982. A model to describe the choice of parking places.
Transportation Research Part A: General 16, 109–115.

Vlacheas, P.; Giaffreda, R.; Stavroulaki, V.; Kelaidonis, D.; Foteinos, V.; Poulios, G.;
Demestichas, P.; Somov, A.; Biswas, A.R.; Moessner, K., (2013). Enabling smart cities
through a cognitive management framework for the internet of things. IEEE
Communications Magazine, IEEE , 51(6), 102-111.
Vlahogianni E. I., Karlaftis, M. G., Golias, J. C. and Kourbelis, N. D. (2006). Pattern-
based Short-term Urban Traffic Predictor. Proceedings of the 9th International IEEE
Conference on Intelligent Transportation Systems, September, Torondo, Canada.
Vlahogianni, E. I. (2009). Enhancing Predictions in Signalized Arterials with
Information on Short-Term Traffic Flow Dynamics, Journal of Intelligent
Transportation Systems, 13(2), 73-84.
Vlahogianni, E. I. and Karlaftis, M.G. (2013). Fuzzy-Entropy Neural Network Freeway
Incident Duration Modeling with Single and Competing Uncertainties, submitted to
Computer Aided Civil and Infrastructure Engineering, 28(6), 420-433.
Vlahogianni, E. I., and Karlaftis, M G. (2014). Testing And Comparing Neural Network
and Statistical Approaches for Predicting Transportation Time Series, Transportation
Research Record, In press.
Vlahogianni, E. I., Karlaftis, M. G., and Golias, J. C. (2005). Optimized and meta-
optimized neural networks for short-term traffic flow prediction: A genetic
approach. Transportation Research Part C: Emerging Technologies, 13(3), 211-234.
Vlahogianni, E.I. (2013). Modeling of the Overtaking Duration in Two Lane Highways,
Transportation Research Part F: Traffic Psychology and Behavior, 20, 135-146.
Vlahogianni, E.I., Karlaftis, M.G., and Kepaptsoglou, K. (2011). Nonlinear
Autoregressive Conditional Duration Models For Traffic Congestion Estimation,
Journal of Probability and Statistics, Article ID 798953.
Washington S., Karlaftis M.G. and Mannering F.L. (2010). Statistical and econometric
methods for transportation data analysis, 2nd Edition. Chapman & Hall/CRC Press,
Boca Raton, FL.
Wong, S.C., Tong, C.O., Lam, W.C.H., Fung, R.Y.C., 2000. Development of parking
demand models in Hong Kong. Journal of Urban Planning and Development 126, 55–
74.
Young, W., Thompson, R.G., and Taylor, M.A.P., (1991). A review of urban parking
models. Transport Reviews, 11(1), 63–84.

15
Zhong, M. and Hunt, J.D. (2010). Exploring best-fit hazard functions and lifetime
regression models for urban weekend activities: Case study, Journal of Transportation
Engineering, 136(3), 255-266.

16
Table 1: Results for different prediction horizons (prediction step τ equals to 1
minute) for each parking region (test set).
Prediction Horizon
Region 1 1 min 5 min 15 min 30 min
MAE 0.007 0.023 0.030 0.032
RMSE 0.015 0.032 0.043 0.048
MAPE 0.869 2.713 3.564 3.913
RRSE 85.065 88.954 71.682 59.250
Region 2
MAE 0.006 0.019 0.024 0.028
RMSE 0.011 0.024 0.031 0.036
MAPE 0.749 2.192 2.824 3.262
RRSE 87.223 92.559 80.614 74.776
Region 3
MAE 0.004 0.012 0.014 0.015
RMSE 0.007 0.015 0.017 0.019
MAPE 0.412 1.213 1.415 1.576
RRSE 86.008 88.895 78.422 75.549
Region 4
MAE 0.004 0.010 0.012 0.014
RMSE 0.006 0.012 0.015 0.017
MAPE 0.436 1.049 1.302 1.480
RRSE 86.245 89.879 79.914 77.151

17
Table 2: Survival analysis results.
βi
Variable
Region I Region II Region III Region IV
Constant -2.98 -2.90 -2.65 -2.06
TypeDay 0.43 - - -
Period - 0.13 0.15 -
Weekday - - - -0.04
Log likelihood 770.86 -1764.64 -1932.55 -4367.82
LR chi2(1) 10.15 9.78 8.28 14.36
Prob > chi2 0.000 0.002 0.002 0.0002

18
Figure 1: Functional architecture of the proposed parking occupancy prediction
scheme.

19
Figure 2: Parking sensors location in the area of Santander (map accessed at:
http://mitos.ploigos.gr/ ).

20
Figure 3: Hourly evolution of parking occupancy (%) and turnover rate
(veh/space/h) for Regions I and II for a typical week.

21
Figure 4: Hourly evolution of parking occupancy (%) and turnover rate
(veh/space/h) for Regions III and IV for a typical week.

22
Figure 5: Survival probabilities for the duration of occupied parking spaces for
Regions I and II.

23
Figure 6: Survival probabilities for the duration of occupied parking spaces for
Regions III and IV.

24
 Region 1

Region 2

Region 3

Region 4

Figure 7: Survival curves based on the models developed for free space durations
in the four regions under study.

25
Figure 8: Architecture of MITOS platform.

26
Figure 9: Screenshot of Web parking prediction implementation.

27
 Figure 10: Screenshot of mobile parking prediction implementation.

28

View publication stats