applied

sciences
Article
Prediction of Vacant Parking Spaces in Multiple Parking Lots:
A DWT-ConvGRU-BRC Model
Liangpeng Gao 1,2 , Wenli Fan 1 , Zhiyuan Hu 1 and Wenliang Jian 1,3, *

1 School of Transportation, Fujian University of Technology, Fuzhou 350108, China

2 School of Transportation, Southeast University, Nanjing 210036, China
3 College of Transportation Engineering, Tongji University, Shanghai 201804, China
* Correspondence: [email protected]

Abstract: For cities, the problem of “difficult parking and chaotic parking” increases carbon emissions
and reduces quality of life. Accurately and efficiently predicting the availability of vacant parking
spaces (VPSs) can help motorists reduce the time spent looking for a parking space and reduce
greenhouse gas pollution. This paper proposes a deep learning model called DWT-ConvGRU-BRC to
predict the future availability of VPSs in multiple parking lots. The model first uses a discrete wavelet
transform (DWT) to denoise the historical parking data and then extracts the temporal correlation
of the parking lots themselves and the spatial correlation between different parking lots using a
convolutional gated recurrent unit network (ConvGRU) while using a BN-ReLU-Conv (1 × 1) module
to further improve the propagation and reuse of features in the prediction process. In addition, the
model uses availability, temperature, humidity, wind speed, weekdays, and weekends as inputs to
improve the accuracy of the forecasts. The model performance is evaluated through a case study
of 11 parking lots in Santa Monica. The DWT-ConvGRU-BRC model outperforms the LSTM and
GRU baseline methods, with an average testing MAPE of 2.12% when predicting multiple parking
lot occupancies over the subsequent 60 min.

Keywords: parking prediction; deep learning; discrete wavelet transform; convolutional gated
recurrent unit network; multiple parking lots

Citation: Gao, L.; Fan, W.; Hu, Z.;

Jian, W. Prediction of Vacant Parking 1. Introduction
Spaces in Multiple Parking Lots: A With economic and population growth, motor vehicle ownership is growing rapidly,
DWT-ConvGRU-BRC Model. Appl. which has exacerbated the imbalance between the supply and demand of vacant parking
Sci. 2023, 13, 3791. https://doi.org/
spaces (VPSs) in cities. Drivers typically spend 3.5–14 min looking for a VPS and cruising to
10.3390/app13063791
find a VPS accounts for 8–74% of traffic [1]. Excessive time spent by drivers looking for VPSs
Academic Editor: Yu-Dong Zhang increases time costs, fuel consumption, and emissions and leads to traffic congestion [2].
Parking difficulties and disorderly parking problems are often affected by accessibility,
Received: 21 February 2023
parking prices, and the number of VPSs.
Revised: 14 March 2023
To address this problem, some parking management and inducement systems have
Accepted: 15 March 2023
been developed to provide real-time VPS information [3–6]. These systems typically collect
Published: 16 March 2023
real-time available parking data using cameras and sensors [7,8]. In addition, several
crowd-sensing-based schemes monitor the availability of street parking using mobile
communication devices and in-vehicle sensors [9,10]. However, these parking guidance
Copyright: © 2023 by the authors. systems cannot guarantee the real-time nature of a VPS. That is, when a driver arrives at the
Licensee MDPI, Basel, Switzerland. designated parking space, the parking space may already be occupied. Due to the high cost
This article is an open access article of sensor equipment and maintaining real-time parking information, Y et al. [11] and Baidu
distributed under the terms and Maps proposed a model named Du-parking to estimate the real-time parking availability
conditions of the Creative Commons of the whole city. Therefore, to enable car owners to purposefully find parking spaces, it is
Attribution (CC BY) license (https:// necessary to develop a parking guidance information system with predictive algorithms
creativecommons.org/licenses/by/
that can help drivers plan driving routes to find VPSs and reduce driving costs [12], which
4.0/).

Appl. Sci. 2023, 13, 3791. https://doi.org/10.3390/app13063791 https://www.mdpi.com/journal/applsci

Appl. Sci. 2023, 13, 3791 2 of 16

can assist traffic planning and management to reduce energy consumption and traffic
congestion [2].
In this paper, we propose a DWT-ConvGRU-BRC model. This model consists of a
discrete wavelet transform (DWT) [13], convolutional GRU networks (ConvGRUs) [14],
a two-layer linear network, and a composite function of three consecutive operations,
i.e., batch normalization (BN), rectified linear activation (ReLU), and a 1 × 1 convolution
(Conv), denoted BRC. First, we use the DWT to denoise the VPS data. Noise reduction
before forecasting can eliminate the volatility of the VPS data themselves. Then, we use a
deep learning-based prediction model that leverages ConvGRUs and a two-layer linear
network to incorporate the spatial–temporal features of multiple data sources acquired
in networks. Finally, the propagation and reuse of features in the prediction process are
further improved using the BRC composite function.
This paper contributes to the literature in the following ways:
• We propose a deep learning-based parking space prediction model from the perspec-
tive of multiple parking lots. The model considers the processing of parking noise data
as well as the spatial correlation of multiple parking lots and the temporal correlation
of the parking lots themselves and uses a variety of factors, including parking lot
occupancy, temperature, humidity, wind speed, weekdays, and holidays, to predict
the number of available VPSs.
• Our proposed DWT-ConvGRU-BRC model can simultaneously predict the number
of available parking spaces in multiple parking lots. Specifically, a ConvGRU is used
to capture the spatial–temporal features of multiple parking lots, a two-layer linear
layer is used to extract external influences, and BRC is used to further improve the
propagation and reuse of features in the prediction process.
• The performance of the method is evaluated with a case study in the Santa Mon-
ica area. According to the results, the model outperforms other baseline methods,
including LSTM, GRU, ConvGRU, and dConvLSTM-DCN models. Moreover, the
results prove the improvement in prediction accuracy from the DWT and the effec-
tiveness of incorporating weekday, vacation, and weather features into parking lot
occupancy predictions.
The rest of this paper is organized as follows: Section 2 summarizes the literature
review. Section 3 describes the detailed DWT-ConvGRU-BRC prediction model. Section 4
presents the results and analysis of the comparison experiments. Finally, we provide our
conclusions and discuss possible future work in Section 5.

2. Literature Review
Access to VPS data has become easier with breakthroughs in sensor technology. How-
ever, the VPS data obtained in practical applications are often subject to different degrees
of noise pollution. How to effectively process the data collected by sensors and improve
the accuracy of algorithms is a thorny problem that many existing prediction methods
still face. To solve this problem, wavelet analysis has been applied in some recent studies
and has proven to be effective. For example, Li et al. [15] used the wavelet function for
multiscale wavelet decomposition and reconstruction of VPS data using the hidden layer
function of a wavelet neural network to improve prediction accuracy. Ji et al. [16] proposed
a multistep prediction study of impacted parking spaces based on a WT in combination
with a multistep prediction strategy using threshold noise reduction to further improve the
prediction accuracy. Therefore, effective noise removal helps to improve the efficiency and
accuracy of prediction.
Predicting the occupancy rate of multiple parking lots is one of the necessary links
to solve the “difficult parking” problem. In recent years, VPS prediction has been di-
vided into two categories: one is based on a statistical prediction model, and the other is
based on machine learning (ML) and deep learning (DL). For statistical prediction models,
Caliskan et al. [17] combined continuous Markov and queuing theory models to predict
the occupancy status of parking lots in the destination area. On this basis, Xiao et al. [18]
Appl. Sci. 2023, 13, 3791 3 of 16

proposed a continuous-time Markov M\M\C\C model for predicting available parking

spaces. Caicedo et al. [19] proposed a real-time available dynamic algorithm using historical
information to predict the availability of each parking lot. In addition, Rajabioun et al. [20]
developed a vector spatiotemporal autoregressive model that can be used to predict the
availability of parking spaces at a driver’s estimated arrival time at both on-street and off-
street parking locations. Peng et al. [21] modelled the discrete occupancy rate of a parking
lot as a nonstationary Poisson process and proposed a cost-effective method for searching
for parking spaces. Abdeen et al. [22] proposed a smart parking algorithm that varied
the weights of five factors (availability, gate wait time, parking cost, traffic congestion,
and driving distance to the parking lot) to achieve balanced traffic allocation and parking
best use of the field. In fact, these statistical prediction models are highly dependent on
assumptions about the arrival and departure process and therefore have difficulty adapting
to the dramatic fluctuations in parking traffic flow.
For ML/DP prediction models, researchers have applied models such as regression
trees, support vector machines (SVMs), support vector regression (SVR), neural networks,
K-nearest neighbour (KNN), and random forests models to predict parking availability [23–27].
Hu et al. [28] combined support vector regression (SVR) and the fruit fly optimization
algorithm (FOA) to predict the number of vacant parking spaces. Fan et al. [29] optimized
a multi-step long short-term memory recurrent neural network (LSTM-NN) model with
a grid search method to predict the number of vacant parking spaces. Moreover, many
scholars have combined nonlinear system theory and optimization algorithms with neural
networks to improve prediction accuracy. For example, Vlahogianni et al. [30] used a
genetic algorithm-optimized multilayer perception (MLP) to predict the occupancy rate of a
regional parking lot over the subsequent 30 min. Camero et al. [31] used a genetic algorithm
(GA) combined with a recurrent neural network (RNN) to predict parking occupancy in
Birmingham. Zeng et al. [32] combined a wavelet transform (WT) with bi-directional LSTM
(Bi-LSTM) to further improve the prediction accuracy using threshold noise reduction.
In addition, some scholars have considered the influence of external factors, such
as weather and holidays, on VPS forecasting. Fokker et al. [33] explored the influence of
external factors such as weather on parking occupancy and found that external factors
improved the predictive performance by 8%. In Zhang’s [34] work, a PewLSTM was
proposed for predicting parking behaviour by combining the effects of weather and parking
periodicity. Zeng et al. [35] proposed a stacked gated recurrent unit (GRU)-LSTM model
that combined the efficiency of a GRU and the accuracy of LSTM and incorporated various
factors as inputs, such as weather, to predict the availability of parking spaces. ML/DL
methods can automatically learn from past samples to better describe complex nonlinear
problems. However, the ML/DP methods described above only consider the temporal
correlation of VPS data and fail to consider the spatial correlation of VPSs in multiple
parking lots.
Therefore, this paper proposes a model called DWT-ConvGRU-BRC to predict the
number of VPSs in multiple parking lots. Our model combines the advantages of wavelet
transform. While capturing the spatial–temporal correlation of multiple parking lot data,
it also takes external factors such as weather as input to improve the accuracy of the
model. Previous research related to our methodology includes DWT-Bi-LSTM [32] and
dConvLSTM-DCN [36].

3. Methodology
3.1. Data Description
To evaluate the performance of the proposed prediction model, we conducted a case
study in Santa Monica, CA, USA (longitude range: [−118.499378, −188.49361], latitude
range: [34.019575, 34.010806]) [37], which has 11 parking lots scattered over the road
network, as illustrated in Figure 1. The data were collected from 6 April 2021 to 13 May
2021. The number of VPSs was collected every 5 min, resulting in 10,944 pieces of historical
data per parking lot.
 To evaluate the performance of the proposed prediction model, we conducted a case
study in Santa Monica, CA, USA (longitude range: [−118.499378, −188.49361], latitude
range: [34.019575, 34.010806]) [37], which has 11 parking lots scattered over the road net-
work, as illustrated in Figure 1. The data were collected from 6 April 2021 to 13 May 2021.
Appl. Sci. 2023, 13, 3791 4 of 16
The number of VPSs was collected every 5 min, resulting in 10,944 pieces of historical data
per parking lot.

Figure1.1.Distribution
Figure Distributionofofparking
parkinglots
lotsininthe
theregion.
region.

Weuse
We useaa100 100mm×× 100 100 mm grid
grid to todivide
dividethe thetarget
targetarea
areaintointoHH×× W W grids
grids (Figure
(Figure 1).
1).
Eachparking
Each parkinglot lotininthe
theregion
regionisis distributed
distributed inin a different
a different grid,
grid, and and a grid
a grid without
without a park-
a parking
ingdistribution
lot lot distribution is considered
is considered to have to nohave no in
VPSs VPSs
thatingrid.
thatThen,
grid. theThen, the number
number of VPSsofinVPSs
the
in the
area at area
time at
t istime t is denoted
denoted as: as:
 
vt (v )
1,1(1,1) vt (v1,2(1,2)
) ·· · vt (v1,W ) )
(1,W

 (2,1t ) t ) t )
 vt (2,1) vt (2,2(2,2) · · · vt (2,W (2,W ) 
Vt = 
Vt   v. t v. t . .  v..t   (1)
= ..

.. . .  (1)
     
vt( H,1) vt ( H,2) · · · vt ( H,W ) 
( H ,1) ( H ,2) ( H ,W )
vt vt  vt 
where each element in the matrix, denoted as vt (h,w) , h ∈ [0, H], and w ∈ [0, W], is the
where each
number element
of VPSs in theingrid
the (h,
matrix, denoted
w). This area isasdivided , h∈
vt ( h , w ) into a [0,H]
total of, and w ∈ [0,
60 grids, withW]
H ,being
is the
10 and W being 6.
number of VPSs in the grid (h, w). This area is divided into a total of 60 grids, with H being
The 11 parking lots we selected in the grid area are the St1–St9 parking lots, Lot1
10 and W being 6.
parking lot, and Library parking lot. These parking lots are mainly distributed in recre-
The 11 parking lots we selected in the grid area are the St1–St9 parking lots, Lot1
ational, commercial, and residential areas. It is worth noting that there are similarities
parking lot, and Library parking lot. These parking lots are mainly distributed in recrea-
and differences in the evolution of the number of spaces in these parking lots. We can
tional, commercial, and residential areas. It is worth noting that there are similarities and
imagine that the closer the parking lot types and the closer the distance, the more they
differences in the evolution of the number of spaces in these parking lots. We can imagine
should have the characteristics of time–space correlation. We take the Lot1, St5, and St7
that the lots
parking closerasthe parkingto
examples lotmine
typesthe
andcharacteristics
the closer the distance, the parking
of different more they should
lots fromhave
the
perspective of spatiotemporal correlation, considering that the St5 and St7 parkinglots
the characteristics of time–space correlation. We take the Lot1, St5, and St7 parking as
lots
examples to mine the characteristics of different parking lots from the
represent commercial areas and are close to each other, and the Lot1 parking lot represents perspective of spa-
tiotemporal correlation,
entertainment areas. considering that the St5 and St7 parking lots represent commer-
cial Figures
areas and are close to
2 and 3 show the each other, and thecharacteristics
spatiotemporal Lot1 parking lot represents
of these entertainment
3 parking lots. The
areas.represents the time interval. The y-axis represents the change in VPSs, where a
x-axis
positive number represents the outflow of vehicles. The larger the number is, the greater
the number of VPSs. Figure 2 shows that the inflow on weekends is significantly higher
than that on weekdays during the almost full day in the Lot1 parking lot. The Lot1 parking
lot represents an entertainment area, which tends to be crowded on weekends. In contrast,
between 7 a.m. and 9 a.m., the inflow of the St5 and St7 parking lots is higher on weekdays
than on weekends, which may be influenced by the parking of mall workers. In addition,
we explore the impact of weather factors on the availability of VPSs.
 number
number ofof VPSs.
VPSs. Figure
Figure 22 shows
shows that
that the
the inflow
inflow on
on weekends
weekends is is significantly
significantly higher
higher than
than
that
that on
on weekdays
weekdays during
during the
the almost
almost full
full day
day in
in the
the Lot1
Lot1 parking
parking lot.
lot. The
The Lot1
Lot1 parking
parking lotlot
represents an entertainment area, which tends to be crowded on weekends.
represents an entertainment area, which tends to be crowded on weekends. In contrast, In contrast,
between
between 77 AM
AM and
and 99 AM,
AM, the
the inflow
inflow of
of the
the St5
St5 and
and St7
St7 parking
parking lots
lots is
is higher
higher on
on weekdays
weekdays
Appl. Sci. 2023, 13, 3791
than
than on weekends, which may be influenced by the parking of mall workers. In addition,
on weekends, which may be influenced by the parking of mall workers. In 5 of 16
addition,
we explore the impact of weather factors on the availability of
we explore the impact of weather factors on the availability of VPSs. VPSs.

(a)
(a) (b)
(b) (c)
(c)
Figure
Figure 2.
2. Impact
Figure Impact
2. Impactof
of holiday
holiday events
of holidayevents (weekdays
(weekdaysor
(weekdays
events ororweekends)
weekends) on
weekends) on VPS
on VPS changes
VPS changes (7 AM
AM to
(7 a.m.
changes (7 to 777 PM).
to p.m.).(a)
PM). (a)
Lot1
Lot1 parking
(a)parking lot; (b)
lot; (b)
Lot1 parking St5
lot;St5 parking
(b)parking lot; (c)
lot;lot;
St5 parking St7
(c) (c) parking
St7St7
parking lot.
parkinglot.
lot.

(a)
(a) (b)
(b) (c)
(c)
Figure
Figure 3.
3. Impact
Impact of
ofofweather
weather on VPS changes (7 AM to
to777p.m.).
PM). (a) Lot1 parking lot; (b) St5 parking
Figure 3. Impact weatheron
onVPS
VPS changes (7a.m.
changes (7 AMto PM).(a)
(a)Lot1
Lot1parking
parkinglot;lot;
(b)(b)
St5St5 parking
parking
lot; (c)
lot;lot; St7
(c) (c) parking
St7St7
parking lot.
lot.
parking lot.

The
The number
The numberof
number ofofparking
parking occupancies
parking occupancies isis
occupancies different
isdifferent
differentforfor
for hazardous
hazardous
hazardous weather
weather
weather and
and normal
and normal normal
weather.
weather. It can be imagined that when encountering hazardous weather such as heavy
weather. It can
It canbe imagined
be imagined that
that when
when encountering
encountering hazardous
hazardous weather
weather such such
as as
heavy heavy
rain,
rain, heavy
heavy snow,
rain, heavy
snow, and
snow, andsmog,
and smog, people
people may
smog, people may
may reduce
reduce travel
travel in
reduce travel
in private
private cars,
in private
cars, so
so the
cars, so the
the number
number
number of
of
of
available parking spaces will increase. The results for the three representative parking lots
available
available parking
parking spaces
spaces will
will increase. The results for the three representative parking lots
are shown in Figure 3. For theincrease. The results
entertainment for the three
area represented representative
by Lot1 parking lots
and the commercial
are shown
areareas
shown in
in Figure
Figure 3.
3. For
For the
the entertainment
entertainment area
area represented
represented by
by Lot1
Lot1 and
and the
the commercial
commercial
represented by St5 and St7, a significant decrease in parking occupancy was observed
areas
areas represented
for represented
all hours of the by St5
by day and
St5 under St7,
and St7, a significant
a significant
hazardous weatherdecrease
decrease in parking
in parking
conditions. To assess occupancy
the impactwas
occupancy of ob-
was ob-
served
served for
for all
hazardous hours
hours of
allweather the
the day
day under
ofconditions under hazardous
hazardous
on parking demand, weather
weather conditions.
conditions.
we define weather to To
Tobeassess
assess the
the impact
considered impact
of
of hazardous
hazardous
hazardous if weather
weather conditions
conditions
one or more on
on parking
parking
of the following demand,
demand,are
conditions we define
wemet:
define weather
weather
(1) fog to
to be
or snow, (2)considered
be considered
wind
hazardous
hazardous
speed greaterif
if one
one or more
or 39
than more of
of the
km/h, the following
following conditions
(3) precipitation conditions are
are met:
intensity greater met:
than(1)
(1) fog or
or snow,
foginches
0.15 snow, (2)
(2) wind
per hour. wind
speed
All greater
other than
conditions 39 km/h,
are (3)
consideredprecipitation
normal intensity
weather greater
conditions. than
Similar
speed greater than 39 km/h, (3) precipitation intensity greater than 0.15 inches per hour. 0.15
to theinches
researchperof hour.
All
All other
Yang etconditions
other conditions are
al. [38] and Zhaoconsidered
are considered normal
normal weather
et al. [39], we conduct
weather conditions.
ablation experiments Similar
conditions. Similar to
in Section 4the
to the research
research of
to explore
of
the influence of external factors such as weather on parking prediction.
Yang
Yang etet al.
al. [38]
[38] and
and Zhao
Zhao et
et al.
al. [39],
[39], we
we conduct
conduct ablation
ablation experiments
experiments in
in Section
Section 44 to
to ex-
ex-
plore
plore the
the influence
influence of
of
3.2. Prediction Model external
external factors
factors such
such as
as weather
weather on
on parking
parking prediction.
prediction.
The DWT-ConvGRU-BRC model provided in this study consists of four components,
3.2. Prediction
3.2.namely,
Prediction Model
Modelcomponent, three ConvGRU components, the meta-info feature extraction
the DWT
component, and the BRC component (Figure 4). The first component is the DWT module,
which performs noise reduction on VPS data by means of the db3 wavelet basis function.
The second component is the ConvGRU module. A CNN can capture spatial correlation
well but not temporal correlation. A GRU and LSTM can both model temporal correlation
well, but a GRU maintains the prediction accuracy and reduces the running speed compared
to LSTM. Therefore, an integration of a CNN and GRU to form a three-layer ConvGRU
network can capture both temporal and spatial correlations. The third component is a
two-layer linear layer module that incorporates external factors such as temperature, wind
 tion. The second component is the ConvGRU module. A CNN can capture spatial cor
lation well but not temporal correlation. A GRU and LSTM can both model temporal c
relation well, but a GRU maintains the prediction accuracy and reduces the running spe
compared to LSTM. Therefore, an integration of a CNN and GRU to form a three-lay
Appl. Sci. 2023, 13, 3791 6 of 16
ConvGRU network can capture both temporal and spatial correlations. The third comp
nent is a two-layer linear layer module that incorporates external factors such as temp
ature,speed,
windhumidity,
speed, humidity,
weekdays, weekdays,
and vacationsandintovacations
the model into the model
to enhance to enhance
the accuracy of the
curacy of long-term
long-term forecasts.
forecasts. Finally,
Finally, feature feature
fusion fusion isusing
is performed performed
the BRC using
layer tothe BRC layer
obtain
obtain predictions via the sigmoid function.
predictions via the sigmoid function.

FigureFigure
4. DWT-ConvGRU-BRC
4. DWT-ConvGRU-BRC model.
model.

3.2.1. Discrete Wavelet Transform (DWT) Denoising

3.2.1. Discrete Wavelet
Time series Transform
data obtained (DWT)applications
in practical Denoisingare often contaminated by various
Time
forms series data of
and degrees obtained
noise. The indiscrete
practical applications
wavelet transform are often
is very contaminated
appropriate for noiseby vario
filtering, which makes it a good choice for time series data processing [40]. When time
formsseries
anddatadegrees of noise. The discrete wavelet transform is very appropriate for no
are decomposed by a DWT, the original signal is separated into approximate
filtering, whichand
coefficients makes
detailitcoefficients
a good choice for time
at different series
resolution dataThe
levels. processing
information [40]. When tim
of the
seriesoriginal
data are decomposed
signal is retained inby theawavelet
DWT, coefficients,
the original andsignal is separated
a perfect reconstruction into of approxim
the
coefficients
original and
data detail
can be coefficients
performed from at these
different resolution
coefficients. levels.
However, someTheofinformation
the detail of t
coefficients that represent the detailed motion in the data can be identified as noise. These
original signal is retained in the wavelet coefficients, and a perfect reconstruction of t
coefficients can then be set to zero prior to the DWT reconstruction process to filter out the
original
noise data
fromcanthe be performed
original from
time series, andthese coefficients.
reconstruction However,
involves some of
reconstructing thethe
timedetail co
ficients that
series fromrepresent the detailed
every component motion
except the noise. In inother
the words,
data can a DWT be isidentified as noise.
a discretization of The
coefficients can then be set to zero prior to the DWT reconstruction process to filter o
the scales and translations of the fundamental wavelet. A DWT can be defined as:
the noise from the original time series, and m
Z
reconstruction involves reconstructing the tim
Wϕ f (m, n) = 2− 2 f (t)ψ(2−m t − n)dt (2)
series from every component except the noise. R In other words, a DWT is a discretizati
of thewhere
scales and translations of the fundamental wavelet.
ψ is the complex conjugate of ψ, formula ψ satisfies
R +∞A DWT can be defined as:
ψ(t)dt = 0, and m and n
−∞
are integers. m
Appropriate wavelet basis functions are− very important − tomextract the features of park-
ing data. Kaplun= et al. [41]W ϕ f ( mthe
selected 2 2 ∫R wavelet
, n )appropriate f (t )ψ (2 t − n ) dt based on entropy
basis function
estimation in the matching pursuit algorithm. Bhavsar et al. [42] chose the appropriate
where ψ basis
wavelet is thefunction
complex conjugatethe
by calculating ψ
of magnitude
, formula ψ
of the mutual satisfies
information.+∞
∫−∞ ψIn(this
t ) dt =0,a
paper, we measure the dependence between two variables by calculating the normalized
m and n are integers.
 it is easy to choose a suitable wavelet basis function. The NMI is

2( H ( X ) − H ( X | Y ))
NMI ( X , Y ) =
Appl. Sci. 2023, 13, 3791
H ( X ) + H7(ofY16)

H (X ) H (Y )
mutual information (NMI), which places the mutual information between [0, 1], and it is
where
easy to choose a suitableand represent
wavelet basis function. The NMI isthe entropy
defined as: of variables
tively. H ( X | Y ) isNthe conditional
MI ( X, Y) =
2( H ( X ) −entropy
H ( X |Y )) for X given(3)Y .
H ( X ) + H (Y )
In this study, the basisfunctions of the compared wavelets are
where H ( X ) and H (Y ) represent the entropy of variables X and Y, respectively. H ( X |Y ) is
letconditional
the (sym3),entropy
and forcoiflet
X given(coif3).
Y. It can be seen from Figure 5 that d
relative to other
In this study, wavelet functions.
the basisfunctions After
of the compared experimental
wavelets are Daubechies compariso
(db3),
symlet (sym3), and coiflet (coif3). It can be seen from Figure 5 that db3 has the highest NMI
is selected
relative for wavelet
to other wavelet functions. decomposition of the experimental
After experimental comparison, the db3 wavelet basistime s
isdecomposition levels is 3,ofwhich
selected for wavelet decomposition can remove
the experimental time series,the
and noise while
the number of ma
decomposition levels is 3, which can remove the noise while maintaining the fluctuation
characteristics
characteristics of series
of the time the time
data as series data as much as possible.
much as possible.

Figure 5. Wavelet selection based on normalized mutual information.

Figure 5. Wavelet selection based on normalized mutual information.

We take the St7 parking lot as an example to show the specific process of the DWT

We take thesequence
St7 parking lotthree
as times,
an example tolow-frequency
show the spec
component. The original VPS data and its decomposition components are shown in Figure 6.
Red is the high-frequency decomposed and blue is the
component.
sequence. Thewas
The denoising original
performedVPS
usingdata and its
the threshold decomposition
method, compon
and then the denoised

6. Reddata.is the high-frequency sequence of thedecomposed three times,

time series was reconstructed. Figure 7 shows the comparison between the original and
denoised We can see that the overall regularity denoised data has increased,
quency
and the trendsequence. The denoising
tends to be smoother, which is morewas performed
suitable using the
for model construction. thresho
As in
Equation (1), the denoised data are Xt .
denoised time series was reconstructed. Figure 7 shows the comp
inal and denoised data. We can see that the overall regularity o
increased, and the trend tends to be smoother, which is more suit
tion. As in Equation (1), the denoised data are X t .
Appl.Sci.
Appl.
Appl. Sci.2023,
Sci. 2023, 13,x3791
2023,13,
13, xFOR
FORPEER
PEERREVIEW
REVIEW 888of
of17
of 17
16

Figure6.
Figure
Figure 6.db3
6. db3wavelet
db3 waveletdecomposition
wavelet decomposition diagram.
decompositiondiagram.
diagram.

Figure 7.
Figure7. Original
7.Original and
Originaland denoised
anddenoised time
denoisedtime series.
timeseries.
series.
Figure
3.2.2. Convolutional Gated Recurrent Unit (ConvGRU)
3.2.2. Convolutional
3.2.2. Convolutional Gated
Gated Recurrent
Recurrent Unit
Unit (ConvGRU)
(ConvGRU)
A key difference between a ConvGRU and GRU is that the former uses a convolution
operator rather than a fully concatenated operator. Therefore, a ConvGRU can better cap-
ture spatial–temporal correlations. Figure 8 shows the internal structure of the ConvGRU
 ture spatial–temporal correlations. Figure 8 shows the internal structure of the ConvGRU
cell, where CAT and UCAT denote the concatenation and splitting operations, respec-
tively. The detailed information flow of a ConvGRU is shown in the following equations:
=z σ (W * x + W * h ) (4)
t xz t hz t −1
Appl. Sci. 2023, 13, 3791 9 of 16

=r σ (W * x + W * h ) (5)
t xr t hr t −1

cell, where CAT and UCAT denote the concatenation and splitting operations, respectively.
∧ of a ConvGRU is shown in the following equations:
The detailed information flow
= h tanh(W * x + r  (W * h )) (6)
t xh t t hh t −1
zt = σ (Wxz ∗ xt + Whz ∗ ht−1 ) (4)

∧ Whr ∗ ht−1 )
rt = σ(Wxr ∗ xt + (5)
h = (1 − z )  h + z  h (7)
∧ t t t t t −1
ht = tanh(Wxh ∗ xt + rt (Whh ∗ ht−1 )) (6)
where zt denotes the update gate, rt denotes
∧
∧ the reset gate, and ht denotes the can-
h t = (1 − z t ) h t + z t h t −1 (7)
didate hidden state. xt denotes the input of the current step and ht and ht −1 denote
∧
the hidden
where state ofthe
zt denotes theupdate
currentgate,
and previous
rt denotessteps, gate, and x
respectively.
the reset and ht −the
htt denotes 1 are the in-
candidate
hidden state. xt denotes the input of the current step and ht and ht−1 denote the hid-
put and output vectors of the current time point, respectively, and W.z , W.r , and W.h
den state of the current and previous steps, respectively. xt and ht−1 are the input and
are the convolutional
output vectors of the kernels
current for each
time gate.
point, In addition,and
respectively, , W.rtanh
σW.zand , andare
W.hthe
aresigmoid
the con-
*
volutional kernels for each gate. In addition, σ and tanh are the sigmoid and hyperbolic
and hyperbolic tangent activation functions, respectively, is the convolution operation,
tangent activation functions, respectively, ∗ is the convolution operation, and denotes
and  denotes
the Hadamard the Hadamard product.
product.

Figure 8. Internal structure of the ConvGRU cell.

Figure 8. Internal structure of the ConvGRU cell.
Specifically, given Xt (the denoised data) and ht−1 as inputs at time step t, the unit
firstSpecifically,
obtains the output X update
given of tthe(the denoised data)reset
gate and ht −1 as
andgates, inputs at time
respectively, step
using t, the unit
Equations (4)
∧
first
andobtains theusing
(5). Then, output of the update
Equation (6), thegate and reset
temporal gates,
hidden respectively,
state using Equations
h can be calculated, which
t ∧
considers
(4) and (5).both theusing
Then, Xt and the(6),
inputEquation hidden state ht−1hidden
the temporal stateby h
produced the
t previous ConvGRU
can be calculated,
operator. The final hidden state ht of the unit is produced by a linear combination of
which considers both the ∧ input X t and the hidden state ht −1 produced by the previous
temporal hidden state ht and previous hidden state ht−1 using Equation (7). We denote the
output of this part as follows:

Oc = h t ⊕ · · · ⊕ h t − 5 ( m − 2 ) ⊕ h t − 5 ( m −1 ) (8)

where m denotes the number of data divided into 5 min intervals and ⊕ is the concatena-
tion operation.

3.2.3. External Factor Extraction and Feature Learning

Parking spaces have obvious periodic characteristics. We analysed the impact of
external factors, such as weekdays, non-weekdays, hazardous weather, and normal weather,
Appl. Sci. 2023, 13, 3791 10 of 16

on the availability of parking spaces. External factors can affect parking events and are
an important part of the model. A two-layer linear layer was designed to consider the
impact of external factors on VPSs. We record the data collected from [43] on temperature,
wind speed, humidity, etc., and weekdays and non-weekdays as inputs to this section as n.
Bringing n into Equation (9), we obtain Ol ∈ R HW ×1 :

Ol = σ (wn,2 σ (wn,1 n + bn,1 ) + bn,2 ) (9)

where σ is the activation function and wn,i , bn,i , i = 1, 2 are the weights and deviations of
the i-th linear function. For feature learning, we convert the resulting output Ol ∈ R HW ×1
to Ol ∈ R H ×W via the Reshape function.
The outputs of the three ConvGRUs and the additional factor extraction component
are concatenated together, denoted as Oi = Oc ⊕ Ol , and fed into the BRC layer. The BRC
layer is a composite function of three consecutive operations, i.e., batch normalization (BN),
rectified linear activation (ReLU), and a 1 × 1 convolution (Conv). We use the BRC layer to
implement feature reuse and propagation. Finally, the prediction is obtained by applying
the sigmoid function.

4. Experimental Results
4.1. Experimental Setup and Evaluation Indicators
In our experiments, we select 60% of the data as the training set, 20% as the validation
set, and the rest as the test set. We normalize the denoised data using Equation (10) and
then slice it into the model using single-step moving window data of length 10. For training,
the gradient descent optimization algorithm is the Adam [44] algorithm, the learning rate
is 0.01, the loss function is the MAE, the epoch size is 32, the batch size is 32, and the
number of ConvGRU layers (k) is 3. The numbers of VPSs after 5, 15, 30, 45, and 60 min
are predicted accordingly. To avoid contingency, each prediction task was independently
repeated 30 times, and the mean values were taken as the results. The DWT-ConvGRU-
BRC model is implemented using PyTorch version 1.11.0, and the experimental equipment
includes a 12th Gen lntel(R) Core(TM) i5-12600KF processor and an NVIDIA GeForce GTX
30600Ti GPU with 16 GB memory.

x − min
x∗ = (10)
max − min
where max and min represent the maximum and minimum values of the sample data,
respectively, and max–min represents the range.
We use the mean absolute error (MAE), mean absolute percentage error (MAPE), and
root mean square error (RMSE) to measure the accuracy of the predicted values. All three
evaluation metrics have a range of [0, +∞) and are equal to 0 when the predicted value
exactly matches the true value, i.e., a perfect model; the larger the error is, the larger the
value. The definitions are shown in Equations (11)–(13).

1 n ∧
n i∑
MAE = | yi − yi | (11)
=1

∧
100% n |yi − yi |
n i∑
MAPE = (12)
=1
yi
s
1 n ∧ 2
RMSE = ∑
n i =1
( yi − yi ) (13)

∧
where yi denotes the actual VPSs, yi denotes the predicted VPSs, and n denotes the time step.
Appl. Sci. 2023, 13, 3791 11 of 16

4.2. Results and Analysis

Fifty experiments were carried out, and the mean was selected as the result to improve
the statistical significance of the difference in precision. Table 1 shows the detailed RMSE,
MAE, and MAPE data for our DWT-ConvGRU-BRC model. Figure 9 shows the effect
of the DWT-ConvGRU-BRC model in the Lot1, St5, and St7 parking lots for 5, 15, 30, 45,
and 60 min VPS predictions, where the x-axis represents the time interval, and the y-axis
represents the number of VPSs. Our model is robust in predicting the availability of VPSs
in multiple parking lots.

Table 1. Performance evaluation in terms of the MAE, RMSE, and MAPE.

DWT-ConvGRU-BRC DWT-ConvGRU-BRC
without External Factors with External Factors
5 min 15 min 30 min 45 min 60 min 5 min 15 min 30 min 45 min 60 min
RMSE St1 2.87 4.59 5.55 8.10 9.48 2.93 3.76 5.81 7.21 9.03
St2 5.72 4.56 7.26 8.88 10.98 3.22 4.31 6.70 9.29 10.26
St3 3.31 4.14 6.13 8.42 10.78 3.21 3.80 6.20 8.06 9.44
St4 3.74 4.97 8.52 9.75 13.90 3.65 4.74 7.46 10.43 11.80
St5 4.16 4.97 8.43 12.57 14.38 3.77 5.60 8.92 12.08 14.64
St6 5.20 6.17 9.47 12.12 15.00 4.86 6.07 9.79 11.98 14.83
St7 5.20 6.99 12.14 16.90 21.40 5.07 6.49 10.44 15.33 19.35
St8 6.14 7.21 12.18 16.60 22.18 5.42 6.97 11.14 16.38 20.74
St9 2.63 3.32 4.03 4.67 5.40 1.98 3.09 4.06 4.87 5.49
Lot1 9.21 12.27 18.84 34.60 41.72 9.19 12.01 19.54 27.38 34.50
Library 2.48 2.28 3.34 3.65 4.49 2.32 2.28 3.73 4.58 5.23
MAE St1 2.01 2.91 3.85 5.64 6.48 2.08 2.58 4.00 5.13 6.19
St2 3.22 4.31 6.70 9.29 10.26 2.29 3.12 4.91 6.51 7.44
St3 2.28 2.89 4.34 5.88 7.27 2.25 2.65 4.44 5.59 6.76
St4 2.67 3.54 5.88 6.77 9.89 2.56 3.40 5.33 7.24 8.17
St5 3.11 3.51 5.85 8.71 10.01 2.62 4.01 6.11 8.03 10.20
St6 3.77 4.36 6.88 8.77 11.43 3.43 4.36 6.99 8.80 11.08
St7 3.66 4.99 8.42 11.54 14.03 3.60 4.65 7.41 10.58 12.89
St8 4.09 4.93 8.35 11.10 14.26 3.80 4.97 7.79 11.38 13.78
St9 1.81 2.38 2.86 3.32 3.83 1.42 2.30 2.86 3.47 4.06
Lot1 5.57 7.77 11.75 21.60 25.53 5.67 7.68 12.67 18.22 23.18
Library 1.59 1.43 2.18 2.40 3.04 1.64 1.56 2.30 3.00 3.19
MAPE St1 0.82 1.24 1.56 2.30 2.64 0.84 1.07 1.65 2.07 2.51
(%) St2 0.76 0.63 1.01 1.25 1.52 0.44 0.60 0.94 1.25 1.43
St3 1.04 1.37 2.03 2.75 3.37 1.01 1.20 2.00 2.55 3.08
St4 0.56 0.74 1.26 1.46 2.12 0.54 0.71 1.13 1.56 1.76
St5 0.66 0.78 1.30 1.90 2.20 0.58 0.88 1.34 1.79 2.21
St6 0.70 0.84 1.33 1.68 2.18 0.64 0.82 1.32 1.66 2.11
St7 0.79 1.10 1.83 2.57 2.96 0.78 0.97 1.56 2.26 2.75
St8 0.62 0.75 1.27 1.69 2.19 0.57 0.74 1.17 1.67 2.08
St9 0.74 0.96 1.16 1.34 1.55 0.57 0.92 1.16 1.40 1.63
Lot1 0.68 0.93 1.60 3.49 4.14 0.69 0.95 1.64 2.32 3.07
Library 0.32 0.28 0.44 0.48 0.60 0.33 0.31 0.46 0.60 0.64

From Section 3, we conclude that parking occupancy is affected by external factors

such as weather. This is consistent with the conclusions drawn in [33–35,38,39]. Therefore,
we performed ablation experiments in Table 1 to compare the effect of the presence or
absence of external factors on VPS prediction. We find that the inclusion of external factors
such as temperature, wind speed, weekdays, and weekends was beneficial in improving
the accuracy of the forecasts. For the RMSE, MAE, and MAPE, the forecasting models
with external factors outperformed those without external factors in 40, 35, and 39 of the
55 forecasting tasks, respectively. This shows that considering external factors can improve
the accuracy of VPS prediction, so it can be said that external factors have a significant
 where yi denotes the actual VPSs, yi denotes the predicted VPSs, and n denotes the
∧

time step.

Appl. Sci. 2023, 13, 3791 4.2. Results and Analysis 12 of 16

Fifty experiments were carried out, and the mean was selected as the result to im-
prove the statistical significance of the difference in precision. Table 1 shows the detailed
RMSE,
impact MAE, and MAPE
on parking data for
prediction. our DWT-ConvGRU-BRC
Figure 10 specifically shows the model. Figure 9effect
prediction shows the
of our
effect of the
proposed DWT-ConvGRU-BRC
DWT-ConvGRU-BRC model
model forin theparking
each Lot1, St5,
lot.and St7be
It can parking lots the
seen from for figure
5, 15,
30,
that45, and
our 60 min
model hasVPS
goodpredictions,
timeliness where the x-axis
in predicting the represents
number of the time
VPSs interval, parking
in multiple and the
y-axis represents the number of VPSs. Our model is robust in predicting theof
lots, and it is relatively robust to changes in the spatial–temporal correlation availability
VPSs and
of
canVPSs in multiple
accurately parking
achieve lots.
predictions.

(a) (b) (c)

Appl. Sci. 2023, 13, x FOR PEER REVIEW
Figure 9. Comparison between the DWT-ConvGRU-BRC prediction and actual values at 5, 13
15,of30,
17
Figure 9. Comparison between the DWT-ConvGRU-BRC prediction and actual values at 5, 15, 30, 45,
45, and 60 min. (a) St5 parking lot; (b) St7 parking lot; (c) Lot1 parking lot.
and 60 min. (a) St5 parking lot; (b) St7 parking lot; (c) Lot1 parking lot.

Table 1. Performance evaluation in terms of the MAE, RMSE, and MAPE.

DWT-ConvGRU-BRC DWT-ConvGRU-BRC
without External Factors with External Factors
5 min 15 min 30 min 45 min 60 min 5 min 15 min 30 min 45 min 60 min
RMSE St1 2.87 4.59 5.55 8.10 9.48 2.93 3.76 5.81 7.21 9.03
St2 5.72 4.56 7.26 8.88 10.98 3.22 4.31 6.70 9.29 10.26
St3 3.31 4.14 6.13 8.42 10.78 3.21 3.80 6.20 8.06 9.44
St4 3.74 4.97 8.52 9.75 13.90 3.65 4.74 7.46 10.43 11.80
St5 4.16 4.97 8.43 12.57 14.38 3.77 5.60 8.92 12.08 14.64
(a) (b) (c)
Figure 10.
Figure 10. DWT-ConvGRU-BRC model evaluation
DWT-ConvGRU-BRC model evaluation metrics
metrics for
for each
each parking
parking lot.
lot. (a)
(a) RMSE;
RMSE; (b)
(b) MAE;
MAE;
(c) MAPE.
(c) MAPE.

As some
As some models
models cancan only
only predict
predict for
for aa single
single parking
parking lot,
lot, we
we compared
compared the the effective-
effective-
ness of these models in predicting the number of VPSs at 5, 15, 30, 45, and 60 min,min,
ness of these models in predicting the number of VPSs at 5, 15, 30, 45, and 60 using
using the
the St7 parking lot as an example. The LSTM and GRU models can effectively
St7 parking lot as an example. The LSTM and GRU models can effectively extract temporal extract tem-
poral information
information from nonlinear
from nonlinear timedata,
time series series
butdata, but to
they fail they fail tothe
consider consider the spatial
spatial correlation
correlation
between between
parking lotsparking
within alots within
region. Thea region.
ConvGRU Themodel
ConvGRU model outperforms
outperforms the ConvLSTM the
ConvLSTM model in terms of running speed while capturing spatial–temporal
model in terms of running speed while capturing spatial–temporal correlations. After correla-
tions. After
wavelet wavelet
noise noisethe
reduction, reduction,
forecaststhe forecastssignificantly,
improved improved significantly,
and externaland external
factors, suchfac-
as
tors, such as weather, improved the accuracy of the long-term forecasts. As
weather, improved the accuracy of the long-term forecasts. As illustrated in Table 2, the illustrated in
Table 2, the
proposed proposed DWT-ConvGRU-BRC
DWT-ConvGRU-BRC model is superior
model is significantly significantly
to thesuperior
benchmarkto the bench-
methods.
mark methods.
Table 2. Comparison of operating results.
Table 2. Comparison of operating results.
Model Indicator 5 min 15 min 30 min 45 min 60 min
Model Indicator 5 min 15 min 30 min 45 min 60 min
RMSE 14.86 17.14 20.84 29.24 34.45
LSTM MAE RMSE
8.54 14.86
11.07 17.14
14.92 20.8426.7129.24 34.45
33.12
LSTM MAPE(%) MAE
1.32 8.54
1.72 11.07
2.31 14.92 4.1526.71 33.12
5.17
RMSE MAPE(%)
10.41 1.32
13.86 1.72
18.75 2.31 23.424.15 5.17
29.54
GRU MAE 7.80 11.40 15.15
RMSE 10.41 13.86 18.7519.2523.42 21.74
29.54
MAPE(%) 1.57 2.50 2.93 4.36 5.07
GRU MAE 7.80 11.40 15.15 19.25 21.74
MAPE(%) 1.57 2.50 2.93 4.36 5.07
RMSE 7.14 10.46 14.41 17.83 21.73
ConvGRU MAE 4.74 7.25 10.01 12.31 14.59
MAPE(%) 1.01 1.51 2.08 2.62 3.16
RMSE 7.70 10.55 14.28 16.87 20.01
Appl. Sci. 2023, 13, 3791 13 of 16

Table 2. Cont.

Model Indicator 5 min 15 min 30 min 45 min 60 min

RMSE 7.14 10.46 14.41 17.83 21.73
ConvGRU MAE 4.74 7.25 10.01 12.31 14.59
MAPE(%) 1.01 1.51 2.08 2.62 3.16
RMSE 7.70 10.55 14.28 16.87 20.01
ConvLSTM MAE 5.15 7.30 9.91 11.45 13.14
MAPE(%) 1.09 1.54 2.07 2.35 2.73
RMSE 7.23 10.75 15.10 17.88 19.43
dConvLSTM-DCN MAE 4.98 7.36 10.18 11.77 13.02
MAPE(%) 1.04 1.53 2.11 2.47 2.68
RMSE 5.2 6.99 12.14 16.9 21.4
DWT-ConvGRU MAE 3.66 4.99 8.42 11.54 14.03
MAPE(%) 0.79 1.1 1.83 2.57 2.96
RMSE 5.07 6.49 10.44 15.33 19.35
DWT-ConvGRU-BRC MAE 3.6 4.65 7.41 10.58 12.89
MAPE(%) 0.78 0.97 1.56 2.26 2.75

To illustrate the ability of the proposed model to predict the actual number of VPSs, in
Table 3, we present a detailed comparison between the actual and predicted number of St7
parking lot VPSs (from 10:00 to 11:00, 6 May 2021) output by the proposed DWT-ConvGRU-
BRC model. We also calculate the MAE, MAPE, and RMSE values for the time period. The
output values of the DWT-ConvGRU-BRC model are very close to the real values.

Table 3. Comparisons between the real and predicted numbers of VPSs at the St7 parking lot.

Time Real Predicted Values

Point Value 5 min 15 min 30 min 45 min 60 min
6 May 2021 10:00 652 647 643 654 674 669
6 May 2021 10:05 652 645 639 643 667 661
6 May 2021 10:10 644 646 636 632 659 655
6 May 2021 10:15 641 642 633 620 648 645
6 May 2021 10:20 643 641 632 622 637 636
6 May 2021 10:25 639 641 629 617 625 626
6 May 2021 10:30 638 642 625 614 611 615
6 May 2021 10:35 628 632 624 615 597 604
6 May 2021 10:40 615 623 624 606 593 591
6 May 2021 10:45 609 613 616 605 587 578
6 May 2021 10:50 603 600 606 605 582 564
6 May 2021 10:55 594 594 596 605 582 557
6 May 2021 11:00 584 584 585 591 575 554
MAE 3.23 7.54 12.08 17.15 20.69
MAPE(%) 0.51 1.19 1.91 2.74 3.36
RMSE 4.01 8.45 14.17 18.69 23.46

In addition, model performance evaluation should consider both prediction accuracy

and time consumption. Table 4 compares the running times per round of the LSTM, GRU,
ConvGRU, ConvLSTM, dConvLSTM-DN, DWT-ConvGRU, and DWT-ConvGRU-BRC
models. The DWT-ConvGRU-BRC model considers the effects of factors such as wavelet
noise reduction and external factors, so it is slightly inferior to the models that do not
consider these factors in terms of running speed, but compared to the dConvLSTM-DN
model proposed in [36], our model shows a significant improvement in running speed. In
conclusion, our proposed model not only improves the effectiveness of predictions but also
improves the running speed.
Appl. Sci. 2023, 13, 3791 14 of 16

Table 4. GPU runtimes of different prediction methods.

Model LSTM GRU ConvLSTM ConvGRU

Runtime (s/epoch) 3.34 2.46 17.21 15.07
Model dConvLSTM-DCN ConvGRU-BRC DWT-ConvGRU DWT-ConvGRU-BRC
Runtime (s/epoch) 21.82 15.51 16.04 16.28

5. Conclusions
This paper proposes a deep learning model for occupancy prediction of multiple
parking lots. The model incorporates DWT, ConvGRU, and BRC modules and has the
flexibility to take multiple spatial–temporal structured data sources as inputs. The perfor-
mance of the model is evaluated using a case study from 11 public parking lots in Santa
Monica, California, USA, in which VPS data, weather data, and weekday and weekend
data are used. The experimental results show that our model can achieve considerably
high accuracy with MAPEs of less than 2% for short-term predictions and less than 4% for
long-term predictions. The DWT-ConvGRU-BRC model significantly outperforms the base-
line LSTM and GRU methods. In general, we found that noise reduction of VPS data using
a DWT can improve prediction accuracy and that combining weather information and
weekday and weekend information can improve the performance of long-term predictions
of parking occupancy.
Prediction of available parking spaces is an integral part of parking guidance informa-
tion systems. Available parking space predictions can improve the effectiveness of parking
guidance system information, which can help drivers plan driving routes and find vacant
parking spaces. Furthermore, if we have a reliable parking prediction algorithm, we can
apply dynamic parking pricing to control the parking demand of each parking lot, thereby
assisting traffic planning and management and reducing energy consumption and traffic
congestion. In future work, we will concentrate on further improving the adaptability by
considering other external influences, such as POI information, traffic incident data, traffic
flow data, etc. At the same time, future research will also consider how the running time of
the model can be optimized while ensuring prediction accuracy.

Author Contributions: Methodology, W.J.; Formal analysis, Z.H.; Resources, W.J.; Writing—original
draft, W.F.; Writing—review & editing, L.G. All authors have read and agreed to the published
version of the manuscript.
Funding: This research was funded by Fujian Natural Science Foundation (Project No.: 2020J05194,
2021J05226, and 2022J01938).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Shoup, D.C. Cruising for parking. Transp. Policy 2006, 13, 479–486. [CrossRef]
2. Caicedo, F. Real-time parking information management to reduce search time, vehicle displacement and emissions. Transp. Res.
Part D Transp. Environ. 2010, 15, 228–234. [CrossRef]
3. Khanna, A.; Anand, R. IoT based smart parking system. In Proceedings of the IEEE 2016 International Conference on Internet of
Things and Applications (IOTA), Pune, India, 22–24 January 2016; pp. 266–270. [CrossRef]
4. Wang, H.; He, W. A reservation-based smart parking system. In Proceedings of the 2011 IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS), Shanghai, China, 5–10 April 2011; pp. 690–695. [CrossRef]
5. Yan, G.; Yang, W.; Rawat, D.B.; Olariu, S. SmartParking: A Secure and Intelligent Parking System. IEEE Intell. Transp. Syst. Mag.
2011, 3, 18–30. [CrossRef]
Appl. Sci. 2023, 13, 3791 15 of 16

6. Idris, M.; Leng, Y.; Tamil, E.; Noor, N.; Razak, Z. Car Park System: A Review of Smart Parking System and its Technology.
Inf. Technol. J. 2009, 8, 101–113. [CrossRef]
7. Parking Sensor Data Guide—SFMTA. Available online: https://www.sfmta.com/sites/default/files/reports-and-documents/20
18/08/sfpark_dataguide_parkingsensordata.pdf (accessed on 15 May 2022).
8. On-Street Parking Fee Collection Goes Digital in Cen-Tral Beijing. Available online: https://www.chinadaily.com.cn/a/201907/0
2/WS5d1af0b9a3105895c2e7b2bc.html (accessed on 16 May 2022).
9. Xu, B.; Wolfson, O.; Yang, J.; Stenneth, L.; Yu, P.S.; Nelson, P.C. Real-time street parking availability estimation. In Proceedings of
the 2013 IEEE 14th International Conference on Mobile Data Management, Milan, Italy, 3–6 June 2013; pp. 16–25. [CrossRef]
10. Bock, F.; Di Martino, S.; Origlia, A. Smart Parking: Using a Crowd of Taxis to Sense On-Street Parking Space Availability.
IEEE Trans. Intell. Transp. Syst. 2019, 21, 496–508. [CrossRef]
11. Rong, Y.; Xu, Z.; Yan, R.; Ma, X. Du-parking: Spatio-temporal big data tells you realtime parking availability. In Proceedings of
the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, London, UK, 19–23 August 2018;
pp. 646–654. [CrossRef]
12. Kotb, A.O.; Shen, Y.-C.; Zhu, X.; Huang, Y. iParker—A New Smart Car-Parking System Based on Dynamic Resource Allocation
and Pricing. IEEE Trans. Intell. Transp. Syst. 2016, 17, 2637–2647. [CrossRef]
13. Shensa, M. The discrete wavelet transform: Wedding the a trous and Mallat algorithms. IEEE Trans. Signal Process. 1992, 40,
2464–2482. [CrossRef]
14. Shi, X.; Gao, Z.; Lausen, L.; Wang, H. Deep learning for precipitation nowcasting: A benchmark and a new model. Adv. Neural Inf.
Process. Syst. 2017, 30, 5622–5632. [CrossRef]
15. Li, T.T.; Yang, Y.; Yu, P.; Zhao, H. Optimize the multi-step prediction of parking berths in the wavelet neural network. Chongqing
Norm. Univ. 2018, 35, 113–120.
16. Ji, Y.-J.; Gao, L.-P.; Chen, X.-S.; Guo, W.-H. Strategies for multi-step-ahead available parking spaces forecasting based on wavelet
transform. J. Central South Univ. 2017, 24, 1503–1512. [CrossRef]
17. Caliskan, M.; Barthels, A.; Scheuermann, B.; Mauve, M. Predicting parking lot occupancy in vehicular ad hoc networks.
In Proceedings of the 2007 IEEE 65th Vehicular Technology Conference-VTC2007-Spring, Dublin, Ireland, 22–25 April 2007;
pp. 277–281. [CrossRef]
18. Xiao, J.; Lou, Y.; Frisby, J. How likely am I to find parking?—A practical model-based framework for predicting parking availability.
Transp. Res. Part B Methodol. 2018, 112, 19–39. [CrossRef]
19. Caicedo, F.; Blazquez, C.A.; Miranda, P.A. Prediction of parking space availability in real time. Expert Syst. Appl. 2012, 39,
7281–7290. [CrossRef]
20. Rajabioun, T.; Ioannou, P.A. On-Street and Off-Street Parking Availability Prediction Using Multivariate Spatiotemporal Models.
IEEE Trans. Intell. Transp. Syst. 2015, 16, 2913–2924. [CrossRef]
21. Peng, L.; Li, H. Searching parking spaces in urban environments based on non-stationary Poisson process analysis. In Proceedings
of the 2016 IEEE 19th International Conference on Intelligent Transportation Systems (ITSC), Rio de Janeiro, Brazil, 1–4 November
2016; pp. 1951–1956. [CrossRef]
22. Abdeen, M.; Nemer, I.; Sheltami, T. A Balanced Algorithm for In-City Parking Allocation: A Case Study of Al Madinah City.
Sensors 2021, 21, 3148. [CrossRef] [PubMed]
23. Zheng, Y.; Rajasegarar, S.; Leckie, C. Parking availability prediction for sensor-enabled car parks in smart cities. In Proceedings of
the 2015 IEEE Tenth International Conference on Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP),
Singapore, 7–9 April 2015; pp. 1–6. [CrossRef]
24. Awan, F.M.; Saleem, Y.; Minerva, R.; Crespi, N. A Comparative Analysis of Machine/Deep Learning Models for Parking Space
Availability Prediction. Sensors 2020, 20, 322. [CrossRef]
25. Pflügler, C.; Köhn, T.; Schreieck, M.; Wiesche, M.; Krcmar, H. Predicting the availability of parking spaces with publicly available
data. Informatik 2016, 2, 361–374.
26. Chen, X. Parking Occupancy Prediction and Pattern Analysis; Technical Report CS229-2014; Department Computer Science, Stanford
University: Stanford, CA, USA, 2014.
27. Chen, B.; Pinelli, F.; Sinn, M.; Botea, A.; Calabrese, F. Uncertainty in urban mobility: Predicting waiting times for shared bicycles
and parking lots. In Proceedings of the 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013),
The Hague, The Netherlands, 6–9 October 2013; pp. 53–58. [CrossRef]
28. Fan, J.; Hu, Q.; Tang, Z. Predicting vacant parking space availability: An SVR method with fruit fly optimisation. IET Intell.
Transp. Syst. 2018, 12, 1414–1420. [CrossRef]
29. Fan, J.; Hu, Q.; Xu, Y.; Tang, Z. Predicting Vacant Parking Space Availability: A Long Short-Term Memory Approach. IEEE Intell.
Transp. Syst. Mag. 2020, 14, 129–143. [CrossRef]
30. Vlahogianni, E.I.; Kepaptsoglou, K.; Tsetsos, V.; Karlaftis, M.G. A Real-Time Parking Prediction System for Smart Cities. J. Intell.
Transp. Syst. 2015, 20, 192–204. [CrossRef]
31. Camero, A.; Toutouh, J.; Stolfi, D.H.; Alba, E. Evolutionary deep learning for car park occupancy prediction in smart cities.
In Proceedings of the International Conference on Learning and Intelligent Optimization, Chania, Greece, 27–31 May 2019;
Springer: Cham, Switzerland, 2019; pp. 386–401. [CrossRef]
Appl. Sci. 2023, 13, 3791 16 of 16

32. Zeng, C.; Ma, C.; Wang, K.; Cui, Z. Predicting vacant parking space availability: A DWT-Bi-LSTM model. Phys. A Stat. Mech.
Appl. 2022, 599, 127498. [CrossRef]
33. Fokker, E.S.; Koch, T.; van Leeuwen, M.; Dugundji, E.R. Short-Term Forecasting of Off-Street Parking Occupancy. Transp. Res. Rec.
J. Transp. Res. Board 2021, 2676, 637–654. [CrossRef]
34. Zhang, F.; Liu, Y.; Feng, N.; Yang, C.; Zhai, J.; Zhang, S.; He, B.; Lin, J.; Zhang, X.; Du, X. Periodic Weather-Aware LSTM With
Event Mechanism for Parking Behavior Prediction. IEEE Trans. Knowl. Data Eng. 2021, 34, 5896–5909. [CrossRef]
35. Zeng, C.; Ma, C.; Wang, K.; Cui, Z. Parking Occupancy Prediction Method Based on Multi Factors and Stacked GRU-LSTM.
IEEE Access 2022, 10, 47361–47370. [CrossRef]
36. Feng, Y.; Xu, Y.; Hu, Q.; Krishnamoorthy, S.; Tang, Z. Predicting vacant parking space availability zone-wisely: A hybrid deep
learning approach. Complex Intell. Syst. 2022, 8, 4145–4161. [CrossRef]
37. Santa Monica Open Data Portal. Available online: https://data.smgov.net/Transportation/Parking-Lot-Counts/ng8m-khuz
(accessed on 13 May 2022).
38. Yang, S.; Ma, W.; Pi, X.; Qian, S. A deep learning approach to real-time parking occupancy prediction in transportation networks
incorporating multiple spatio-temporal data sources. Transp. Res. Part C Emerg. Technol. 2019, 107, 248–265. [CrossRef]
39. Zhao, D.; Ju, C.; Zhu, G.; Ning, J.; Luo, D.; Zhang, D.; Ma, H. MePark: Using Meters as Sensors for Citywide On-Street Parking
Availability Prediction. IEEE Trans. Intell. Transp. Syst. 2021, 23, 7244–7257. [CrossRef]
40. Patil, R. Noise Reduction using Wavelet Transform and Singular Vector Decomposition. Procedia Comput. Sci. 2015, 54, 849–853.
[CrossRef]
41. Kaplun, D.; Voznesenskiy, A.; Romanov, S.; Nepomuceno, E.; Butusov, D. Optimal Estimation of Wavelet Decomposition Level
for a Matching Pursuit Algorithm. Entropy 2019, 21, 843. [CrossRef]
42. Bhavsar, K.; Vakharia, V.; Chaudhari, R.; Vora, J.; Pimenov, D.Y.; Giasin, K. A Comparative Study to Predict Bearing Degradation
Using Discrete Wavelet Transform (DWT), Tabular Generative Adversarial Networks (TGAN) and Machine Learning Models.
Machines 2022, 10, 176. [CrossRef]
43. Weather Data Website. Available online: https://www.wunderground.com (accessed on 20 May 2022).
44. Kingma, D.P.; Ba, J. Adam: A method for stochastic optimization. arXiv 2014, arXiv:1412.6980. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.