Investigating Prediction Models for Vehicle Demand in a Service Industry
Ahmed Alzaidi, Siddhartha Shakya and Himadri Khargharia
EBTIC, Khalifa University, Abu Dhabi, U.A.E.
Keywords:
Machine Learning, Demand Forecasting, Resource Management.
Abstract:
Demand prediction is an important part of resource management. Higher forecasting accuracy leads to bet-
ter decision taking capabilities, especially in a competitive service-based business such as telecommunication
services. In this paper, a telecommunication service provider’s data on the use of vehicles by their employees
is analyzed and used to forecast the vehicle booking demand for the future at different geographical locations.
We implement multiple forecasting models and investigate the effect on forecasting accuracy of two predic-
tion strategies, namely the Direct multi-step forecasting strategy (DMS) and the Rolling mechanism strategy
(RMS). Moreover, the effect of different external inputs such as temperatures and holidays were tested. The
results show that both DMS and RMS can be used to forecast vehicle demand, with the highest improvement
in forecasting achieved through the addition of the holiday input, particularly by using the RMS strategy in
the majority of the cases.
1 INTRODUCTION
Flawless operation with constrained supply requires
an effective resource management strategy. The con-
tinuum in the supply can be ensured through the ac-
curate prediction of the demand with the exploita-
tion of historical data. Businesses such as telecom-
munication, utilities, retail, hotels, etc, recognize the
importance of accurate forecasting for demand espe-
cially when the supply is constrained. Indeed, oper-
ational efficiency and sustainable revenue growth in
businesses, with limited supply, are sensitive to poor
demand predictions (Azadeh et al., 2015). Hence,
corporations are keen to exploit machine learning and
other artificial intelligence methods for producing a
more accurate forecast. Higher operational efficiency
improves the quality of service leading to higher lev-
els of customer satisfaction.
However, in the service industry, maintaining ser-
vice standards is coupled with the availability of re-
strained resources to meet the demand. Based on the
types of industries, there could be many different re-
sources involved in providing services, such as vehi-
cles, specialized technical equipment, hardware loads
for when the device can sustain a certain limit of loads
(Herrer
´
ıa-Alonso et al., 2021), rooms or beds for in-
dividuals in hotels (Lee, 2018) or hospitals (Deschep-
per et al., 2021; Goic et al., 2021), etc. Generally,
the demand is predicted after analysis of the histori-
cal demand data besides other correlated data such as
weather, seasonality, geography, etc, to enhance the
forecasting accuracy and thus better manage the avail-
able resources. Different strategies such as Rolling
mechanism forecasting strategy (RMS) (Mu et al.,
2019), Direct multi-step forecasting strategy (DMS)
(Shi and Yeung, 2018) can be adopted by different
forecasting techniques for improving the accuracy.
Businesses such as utility companies, telecommu-
nication service providers, and car sharing companies
use vehicles to provide services to their consumers.
They normally own a fleet of vehicles. An employee
can request a vehicle for a certain hour or day and se-
lect a pick up location from the available locations.
This creates a record of the historical usage data in
different parking hubs and keeps track of the vehicles.
This historical usage data allows visualization of the
demand at respective parking hubs and can also be ex-
ploited to forecast the future demand for the vehicles
(Liu et al., 2021a; M
¨
uller and Bogenberger, 2015; Yu
et al., 2020), thus ensuring their availability upon the
requested booking date and enhancing operational ef-
ficiency.
In this work, we focus on the data provided by
our partner telecommunication service provider that
keeps a fleet of vehicles at different parking hubs. The
engineers can book the vehicles on a daily basis to
perform the tasks allocated to them. The choice of
parking hub for booking may depend on the starting
Alzaidi, A., Shakya, S. and Khargharia, H.
Investigating Prediction Models for Vehicle Demand in a Service Industry.
DOI: 10.5220/0011527400003332
In Proceedings of the 14th International Joint Conference on Computational Intelligence (IJCCI 2022), pages 359-366
ISBN: 978-989-758-611-8; ISSN: 2184-3236
Copyright © 2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
359
location of the engineer as well as the locations of the
tasks that they are performing. Hence, the demand for
vehicles can be different for different dates even for
the same parking hub. We investigate and evaluate
different forecasting strategies and forecasting mod-
els with the aim of accurately predicting the vehicle
booking demand for each parking hub. The ultimate
goal is to build a tool that can be used to manage re-
sources on a daily basis. We perform experiment with
the historical data provided by our partner telecom
and analyze the results to identify the best performing
method. Real-world data on the historical bookings
were combined with additional data involving official
holiday and temperature data to improve the accuracy.
We show the effectiveness of the tested methods and
provide a detail analysis of the results.
The rest of the paper is organized as follows. Sec-
tion 2 reviews related works, particularly focusing on
works, where the RMS and DMS were used and how
other factors were incorporated into the forecast. Sec-
tion 3 explains the data and the methods used. Section
4, presents the experiments and interprets the results,
and finally, section 5 provides a conclusion to the find-
ings.
2 BACKGROUND
There are many use cases where DMS and RMS were
used for forecasting. For example, DMS was ex-
ploited to predict energy prices and wind power with a
radial-basis functional network (Khalid, 2019). It was
found to outperform the recursive forecasting strat-
egy when used with a random forest algorithm for
wind speed prediction (Vassallo et al., 2020). An-
other study implemented the DMS with XGBoost al-
gorithm to predict the state of charge and terminal
voltage in lithium-ion batteries that are exposed to dif-
ferent loads (Dineva et al., 2021). The traffic speed
was predicted using DMS with an ensemble model
(Feng et al., 2021).
Many forecasting scenarios depend on the short
window from the recent past. Short-term predictions
of water levels in different water channels in a river
located in China were tested via the implementation
of an RMS with Long short-term memory (LSTM)
algorithm (Liu et al., 2021b). In (Yuan et al., 2021),
authors created an ensemble model composed of dif-
ferent models that used LSTM with RMS to forecast
the intensity of typhoons. A comparative study is
done in (Yuan et al., 2019) that found Convolutional
Long Short-Term Memory with Ensemble Empiri-
cal Mode Decomposition (EEMD-ConvLSTM) with
RMS to be more robust than the one-step forecasting
models exploiting Global Forecast System, Medium
Range Forecast (MRF), Model Ensemble Members
(ENSM) in forecasting North Atlantic Oscillation in-
dex. While (Du et al., 2016) uses RMS forecasting
with different algorithms to predict wind speed. The
demand was predicted with LSTM through the RMS
forecasting approach (Wang et al., 2021a).
In (Surakhi et al., 2021) the time lag influence on
the accuracy of the predictions is investigated and it is
found that the selection of time lags affect the accu-
racy significantly due to correlation strength between
the selected time lag values. The historical values
(time lags values) of electricity consumption alone
were found to be able to eliminate the need for ad-
ditional inputs such as weather data as they already
captured the effect of weather data, and emphasized
the need to select the optimal number of lagged val-
ues via genetic algorithms(Bouktif et al., 2018). Both
(Bakker et al., 2014; Wang et al., 2021b) found that
using weather input improved the forecasting accu-
racy. The weather, holiday, and accident were used in
the form of one-hot encoding as additional input to the
traffic forecasting models and were found to enhance
the predictions (Sun et al., 2020).
3 METHODOLOGY
Historical vehicle booking records of 91 days for 10
parking locations (hereafter referred to as parking 1,
parking 2, etc) were acquired. The data represent the
number of vehicles booked for the selected date at
each station. It is a continuous univariate time se-
ries data for the vehicle booking at different parking
hubs. As an example, bookings for parking hubs 3,
5, and 9 are shown in Figure. 1, which shows that
there is a shared pattern in booking demand except
for the period between the 16
th
of July and the 30
th
of
July where the booking requests dropped abnormally
and then raised sharply afterward. This pattern is ob-
served in most of the parking hubs data with the dif-
ference being the total booking request. Comparing
the different parking hubs, we found that the booking
request is lowest in parking hub 5 during all the peri-
ods. On the other hand, we found that parking hubs 8
and 9 are alternate as the most booked parking hub.
The weather plays an important role in everyday
transportation decisions and might affect the book-
ing of vehicles at different parking hubs. The fore-
casted temperature for the period between 15
th
May
and 17
th
August 2021 was collected (Figure. 2). It can
be observed that the temperature gradually increased
throughout the data.
The publicly announced holidays are shown in ta-
NCTA 2022 - 14th International Conference on Neural Computation Theory and Applications
360
Figure 1: Vehicle bookings by date for parking hubs 3, 5
and 9.
Figure 2: Temperature forecast data in Fahrenheit.
Table 1: Public holiday dates and the corresponding day of
the week.
Date Day of the week
19 July 2021 Monday
20 July 2021 Tuesday
21 July 2021 Wednesday
22 July 2021 Thursday
12 August 2021 Thursday
ble 1. The public holidays apply to all the sectors,
thus influencing the booking levels. The holiday dates
in the tables match the sudden changes in the trends
within the booking shown in Figure. 1 which indicate
a correlation between the holiday and vehicle book-
ings.
3.1 Rolling Mechanism Forecasting
(RMS)
Figure 3: The working concept of RMS.
RMS is a forecasting strategy that utilizes the recent
trends in data to forecast future values which are re-
ferred to as forecasting horizons (Mu et al., 2019). All
the values in the forecasting horizons are predicted us-
ing the same model that is fitted to take a fixed rolling
window of inputs when no external factors are consid-
ered. Figure 3 demonstrates the concept of the rolling
mechanism. When the first period in the horizon is
predicted, the fixed window of input is coming from
real data, yet when the second period is predicted the
first-period prediction is added as the most recent in-
put to the rolling window, and the input furthest from
the value to be predicted is removed. This new input
window is used to predict the second period. When
the third period is predicted the second-period predic-
tion is used as the most recent input and the value
furthest from the third period is removed. To explain
the concept mathematically, let B be the set of all the
vehicle bookings in a parking hub. Then we assume
B
rms
B to represent a subset of bookings considered
for RMS, such that
B
rms
= {b
t+hi
|1 i n, b
t+hi
B} (1)
where b
t+hi
represents the booking entry from
set B, n represents the size of the rolling window, h
represents the forecasting time step, t represents the
time step for the first forecasted period, then
b
t+h
= F (B
rms
) (2)
where, b
t+h
is the foretasted booking for the time
step h and F is the forecasting model. This forecast-
ing strategy has the advantage of adapting to changes
in data and the training phase of the model assigns
importance to the input values based on their proxim-
ity from the predicted horizon, thus adapting to data
trends. Additionally, the method is computationally
less expensive as one model is needed for predicting
all values in the horizon H. However, because the pre-
vious prediction is used in the forecast of other hori-
zons, the error is accumulated as the prediction steps
increase.
3.2 Direct Multi Step Forecasting
Strategy (DMS)
Figure 4: The working concept of DMS.
In DMS, each step ahead is predicted independently
via a different model. Each period is predicted via
Investigating Prediction Models for Vehicle Demand in a Service Industry
361
a separate model which is trained and validated us-
ing a modified data set specific to that perdition pe-
riod, yet all the models use the exact same input to
predict their corresponding periods (Shi and Yeung,
2018). The concept of DMS forecasting is illustrated
in Figure 4. To explain the DMS mathematically, lets
assume B
dms
B to represent a subset of bookings
considered for DMS, such that
B
dms
= {b
ti
|1 i n, b
ti
B} (3)
where b
ti
represent the booking entry from set B
and n represent the size of the input window, t repre-
sents the time step of the first forecasted period, then
b
t+h
= θ
h
(B
dms
) (4)
where,b
t+h
is the foretasted booking for the time
step h and θ
h
is the model specific for predicting the
booking b
t+h
at the forecasted time step h. The DMS
does not cause error accumulation as predicted hori-
zons are not used as input to predict further horizons
(Dossani, 2022). This advantage in DMS forecasting
made it interesting to test for booking demand fore-
casting. However, as all the models use the same in-
puts, the DMS does not capture the relationship be-
tween the different time steps of the modeled case
(Taieb et al., 2012). Furthermore, creating an indi-
vidual model for each horizon is a computationally
expensive process that becomes a liability when the
number of the forecasted horizon is large.
3.3 Machine Learning Algorithms
In this paper, we implement different machine
learning algorithms with multiple parameter set-
tings to forecast the booking demand. They in-
cludes K-nearest neighbour regressor (KNN), deci-
sion tree (DT), ridge regression (RR), Lasso Re-
gression (Lasso), liner regression (LR), random for-
est (RF), neural network (NN), stochastic gradi-
ent descent (SGD), support vector regressor (SVR)
(James et al., 2021), Gradient Boosting (Friedman,
2001) (GBoost), extrem gradiant boosing (Chen and
Guestrin, 2016) (XGBoost), light gradient boosting
machine (LGBM) (Ke et al., 2017), ExtraTreesRe-
gressor (Geurts et al., 2006), MLPRegressor (Hornik
et al., 1989) and ElasticNet (Zou and Hastie, 2003).
Due to limited space, we do not go into details of
these algorithms. Interested readers are reffered to
(James et al., 2021).
3.4 Model Formulation
It can be observed in Figure 1 that each parking hub
experience different booking level. Hence, we model
Table 2: Input Forms.
SL No Input Form Name D W T H
1 D7W0 7 0 - -
2 D14W0 14 0 - -
3 D6W3 6 3 - -
4 D7W0T 7 0 * -
5 D14W0T 14 0 * -
6 D6W3T 6 3 * -
7 D7W0T1 7 0 1 -
8 D14W0T1 14 0 1 -
9 D6W3T1 6 3 1 -
10 D7W0T3 7 0 3 -
11 D14W0T3 14 0 3 -
12 D6W3T3 6 3 3 -
13 D7W0H 7 0 - *
14 D14W0H 14 0 - *
15 D6W3H 6 3 - *
16 D7W0TH 7 0 * *
17 D14W0TH 14 0 * *
18 D6W3TH 6 3 1 *
19 D7W0T1H 7 0 1 *
20 D14W0T1H 14 0 1 *
21 D6W3T1H 6 3 1 *
22 D7W0T3H 7 0 3 *
23 D14W0T3H 14 0 3 *
24 D6W3T3H 6 3 3 *
each parking hub separately, i.e, for each parking hub,
we built 2 models (RMS and DMS), for each configu-
ration of input features, to test the effect of the differ-
ent model configurations on the forecasting accuracy.
In total, 24 feature configurations including different
external features and with different lag setups were
tested for each of the two models. Data for each of
these configurations were fitted on multiple machine
learning algorithms as listed in the previous section,
and for multiple different parameter settings of these
algorithms. The result for the model with the best
accuracy was taken as the result for that input config-
uration.
3.5 Data Preparation
In RMS, the target needs to be the booking period
directly after the last booking period input. On the
other hand, the DMS requires a specific data set for
each predicted period, such that each period model
is trained to predict multiple periods ahead with the
same input used in the other models. Fortunately,
the RMS input data can be modified to fit the DMS
requirement via shifting the target, thus training the
models to predict different steps ahead using the same
input.
NCTA 2022 - 14th International Conference on Neural Computation Theory and Applications
362
As shown in Table 2, 24 data inputs were gen-
erated for each of the parking hubs, termed input
forms, that represent different combinations of day
lags, week lags, temperature lags, and holidays. The
number after the D indicates the day lag, W indicates
the week lag, T indicates temperature input lag and H
indicates holiday inputs. If there was no number after
the T, this means that the temperature for the target
day was added to the inputs. Also ’-’ in the column
T and H represent the corresponding parameter was
not used, and ’*’ represent only the value for the tar-
get day was used in the input. Additionally, the day
of the week of the target date was used as an extra in-
put feature, encoded with one-hot encoding, creating
7 more binary inputs.
4 EXPERIMENTS AND RESULTS
In this section, the accuracy of RMS is compared to
DMS with different input forms and external input
combinations as shown in Table 2. Weighted abso-
lute percentage error (WAPE) was used as a measure
of accuracy. The data for the last 7 days were used
as a test set to calculate the accuracy. The model that
achieved the best WAPE accuracy was selected and
reported back with the results.
The KN, XGBoost , LGBM and RF all used
the default hyper parameters.The RF was tested
further through 13 different combination of max-
imum tree depth and number of estimator of
[(5,10),(5,15),(5,50),(7,80),(7,100),(7,120),(9,10),
(9,150), (11,10),(11,15),(11,100),(11,500),(13,700)]
respectively. DT algorithms was tested using 6
different maximum tree depths of 5, 7, 10, 15
and 20. The gradient boosting was tested with 11
different combination of hyper parameters, out of
which 7 combination only used number of estimators
and maximum depth with the respective values of
[(500,11),(500,3),(500,5),(100,11),(100,12),(100,13),
(100,14)]. The other combinations added the sub
sample parameter. The hyper parameter com-
bination in the respective order of number of
estimator , maximum depth and sub sample was
[(25,5,1),(30,5,1),(40,5,1,),(100,11,1)].
To make the results more comparable and readable
we present the model’s result in the form of accuracy
% as 1 WAPE, for all the different cases tested in
both the RMS and DMS. The result of demand pre-
diction across the 10 parking hubs for the lagged input
forms D7W0, D14W0, and D6W3 with and without
the cases of the temperature input that is added as the
temperature of the target day (T), lagged temperature
of 1 day with the target day (T1) and lagged temper-
ature of 3 days with the target day (T3) are presented
on tables 3, 4, 5 and 6. The tables report the perfor-
mance results of the models in both the RMS and the
DMS for all the forms. The models in tables 3 and
4 did not use a holiday input and the variation in ac-
curacy is presented in a boxplot (Figure 5) comparing
the RMS versus the DMS approach in all the data in-
puts forms. Similarly, tables 5 and 6 present the per-
formance results for the models that used the Holiday
input while Figure 6 present a comparative visualiza-
tion of the model’s performance in both the RMS and
DMS cases for the different combinations of inputs.
4.1 Models Performance
Table 3: 1-WAPE results for the different parking hub mod-
els using DMS and RMS without a holiday input and using
the basic forms and the basic forms with temperature inputs.
RMS DMS
Form D7W0 D14W0 D6W3 D7W0 D14W0 D6W3
Parking hub1 86.95 88.35 86.9 92.09 90.47 89.46
Parking hub2 85.58 84.6 90.43 84.06 91.82 91.92
Parking hub3 88.12 88.38 85.87 89.65 88 72.19
Parking hub4 75.94 83.53 73.33 76.11 80.65 81.73
Parking hub5 77.15 76.3 80.91 77.04 86.29 69.95
Parking hub6 83.72 83.68 84.04 83.84 77.53 75.53
Parking hub7 90.77 93 92.29 92.52 83.17 86.96
Parking hub8 89.44 87.94 89.57 87.67 85.1 84.95
Parking hub9 84.48 86.88 86.95 85.84 88.78 85.72
Parking hub10 82 84.66 83.7 80.44 88.96 83.44
Average 84.41 85.73 85.4 84.93 86.08 82.18
Stdev 4.67 4.16 5.15 5.49 4.28 6.97
Form D7W0T D14W0T D6W3T D7W0T D14W0T D6W3T
Parking hub1 85.84 88.22 86.35 94.46 87.45 85.38
Parking hub2 86.36 83.89 87.57 83.11 87.29 93.36
Parking hub3 86.17 87.94 85.87 85.98 85.18 79.25
Parking hub4 78.95 80.72 75.34 72.73 76.2 84.7
Parking hub5 81.71 77.18 82.11 75.7 85.77 77.82
Parking hub6 82.32 81.47 83.54 77.01 75.89 74.22
Parking hub7 89.62 91.56 91.72 86.31 82.63 88.66
Parking hub8 89.17 89.01 90.18 88.34 88.21 83.76
Parking hub9 85.16 86.3 86.66 87.86 83.42 84.87
Parking hub10 82.27 84.31 84.39 80.37 91.74 78.98
Average 84.76 85.06 85.37 83.19 84.38 83.1
Stdev 3.23 4.16 4.32 6.36 4.82 5.35
We can notice that, RMS results are more consistent
in comparison to the DMS forecasting when the holi-
day parameter was not used (Table 3 and 4). The dif-
ferent data input forms did not affect the RSM perfor-
mance as much as the DMS and the temperature input
did not significantly contribute to the improvement of
results. For the RMS, the D7W0 input accuracy im-
proved slightly when the temperature input of the tar-
get day or the temperature input of 1-day lag is added
to the RMS models while the opposite is observed for
input forms D14W0 and D6W3 for the prior condi-
tions. When a temperature lag of 3 was used, the
performance recorded was the lowest for all the input
forms. In the DMS case, the input forms D7W0 and
D14W0 experienced lower accuracy as more temper-
ature inputs were added. The only exception was for
Investigating Prediction Models for Vehicle Demand in a Service Industry
363
Table 4: 1-WAPE results for the different parking hub mod-
els using DMS and RMS without a holiday input and using
the basic forms with a temperature lag input of 1 and a tem-
perature lag input of 3.
RMS DMS
Form D7W0T1 D14W0T1 D6W3T1 D7W0T1 D14W0T1 D6W3T1
Parking hub1 86.27 88.99 87.24 90.35 89.77 86.31
Parking hub2 86.9 83.07 87.38 84.82 88.23 92.29
Parking hub3 86.97 87.13 85.87 85.74 85.52 77.06
Parking hub4 79.32 78.43 74.09 69.37 73.42 80.33
Parking hub5 82.93 77.59 83.56 73.97 76.44 72.09
Parking hub6 83.87 81.5 82.48 77.78 75.82 75.26
Parking hub7 92.69 90.6 91.49 86.12 80.76 79.24
Parking hub8 89.75 87.69 89.32 89.24 89.39 85.37
Parking hub9 85.44 86.29 88.77 87.21 81.32 84.46
Parking hub10 84 84.35 82.79 80.99 92.72 79.24
Average 85.81 84.56 85.3 82.56 83.34 81.16
Stdev 3.51 4.16 4.67 6.54 6.39 5.67
Form D7W0T3 D14W0T3 D6W3T3 D7W0T3 D14W0T3 D6W3T3
Parking hub1 85.27 86.95 85.94 86.61 83.53 85.24
Parking hub2 84.96 83.73 86.88 80.81 91.46 88.93
Parking hub3 86.1 88.39 85.87 86.34 86.88 83.7
Parking hub4 71.43 79.85 77.49 71.98 69.36 76.42
Parking hub5 81 75.56 81.1 70.13 67.61 78.71
Parking hub6 82.43 82.31 82.93 84.96 78.03 73.48
Parking hub7 90 89.13 90.07 81.91 80.8 78.22
Parking hub8 89.31 87.29 88.28 84.62 85.86 83.91
Parking hub9 84.76 86.41 86.02 83.51 76.62 81.78
Parking hub10 83.23 83.18 81 80.7 88.68 69.47
Average 83.85 84.28 84.56 81.16 80.88 79.99
Stdev 4.91 4.03 3.64 5.43 7.59 5.57
D6W3 with a target day temperature input (D6W3T),
the accuracy raised by around 1%.
From a different view, the parking hub level re-
sults reveal some interesting behavior. Parking hub 1
achieved the best accuracy with the form D7W0 us-
ing the DMS with an accuracy difference of 5% to
the D7W0 with RMS and 3.5% difference with the
best RMS basic form D14W0 for station 1. Addition-
ally, the temperature of the target day increased the
accuracy to around 94.5% which is a further improve-
ment. Parking hub 2 produced the best accuracy with
the RMS when D6W3 form was used (5% better than
D7W0), yet achieved 1% better performance when
D14W0 and D6W3 were used by the DMS. In the
case of the input form D7W0 in parking hubs 4 and
5, the accuracy was raised by around 3% by adding
the temperature of the target day. In parking hub 10,
the best demand prediction was achieved using the
D14W0 form with the temperature of the target day,
with a significant difference in the accuracy of 5%.
These cases indicate the DMS can make a significant
impact on the accuracy based on the used data.
Figure 5 shows the distribution of model perfor-
mance over 10 different parking hubs in all the cases
of tables 3 and 4. The interquartile range for the RMS
models is smaller in all the different cases, indicat-
ing more stability in the performance when compared
to the RMS. However, in some cases, and depending
on the data, it was observed that the DMS provides
some advantage. The maximum of the DMS models
for D7W0, D7W0T, D14W0T, D14W0T1, D6W3T,
and D6W3T1 is higher than the RMS models. The
Figure 5: Boxplot for accuracy of the models that did not
considered a holiday input in both the DMS and RMS ap-
proaches.
main issue with the DMS in the case study is the vari-
ance. The standard deviation of the performance pa-
rameter is higher for the DMS models than the RMS
models and this is clearly observed via the Figure 5
and supported by the tables 3 and 4.
In general, the RMS performance was better than
the DMS. However, on the level of the parking hub,
some DMS models improved the predictions signif-
icantly. Moreover, the temperature input did not
increase the accuracy significantly except for some
parking hubs which indicates that only some parking
hub booking demands are temperature dependent.
4.2 Model Performance with Additional
Holiday Input
Table 5: 1-WAPE results for the different parking hub mod-
els using DMS and RMS with a holiday input and using the
basic forms and the basic forms with temperature inputs.
RMS DMS
Form D7W0 D14W0 D6W3 D7W0 D14W0 D6W3
Parking hub1 90.62 93.01 94.13 91.72 92.17 92.73
Parking hub2 91.39 90.34 90.77 87.68 93.5 92.4
Parking hub3 86.1 84.54 85.87 83.57 78.76 77.18
Parking hub4 87.24 89.85 73.3 92.08 88.05 84.79
Parking hub5 88.47 88.87 89.71 87.31 88.14 87.35
Parking hub6 92.85 90.49 90.77 91.04 88 88.54
Parking hub7 91.7 93.08 89.61 90.24 95.34 92.84
Parking hub8 95.32 97.02 96.36 94.61 97.24 95.48
Parking hub9 91.85 90.31 90.04 90.38 92.39 91.35
Parking hub10 92.91 94.57 91.7 93.52 95.79 96.01
Average 90.84 91.21 89.23 90.22 90.94 89.87
Stdev 2.68 3.25 5.93 3.1 5.16 5.38
Form D7W0T D14W0T D6W3T D7W0T D14W0T D6W3T
Parking hub1 92.26 93.18 92.38 90.45 90.45 92.42
Parking hub2 91.29 90.35 90.18 85.56 85.56 92.97
Parking hub3 86.17 84.54 85.87 83.57 83.57 70.15
Parking hub4 85.12 89.18 74.09 88.74 88.74 82.55
Parking hub5 90.85 88.05 90.9 90.72 90.72 88.58
Parking hub6 92.66 92 93.84 89.97 89.97 87.06
Parking hub7 89.16 91.19 89.01 88.3 88.3 95.84
Parking hub8 96.15 97.16 95.79 94.99 94.99 91.85
Parking hub9 90.66 90.95 91.8 91.4 91.4 91.03
Parking hub10 90.15 92.58 91.73 91.11 91.11 90.43
Average 90.45 90.92 89.56 89.48 89.48 88.29
Stdev 3 3.17 5.75 3.03 3.03 6.96
NCTA 2022 - 14th International Conference on Neural Computation Theory and Applications
364
Table 6: 1-WAPE results for the different parking hub mod-
els using DMS and RMS with a holiday input and using the
basic forms with a temperature lag input of 1 and a temper-
ature lag input of 3.
RMS DMS
Form D7W0T1 D14W0T1 D6W3T1 D7W0T1 D14W0T1 D6W3T1
Parking hub1 91.34 93.36 91.62 91.13 94.11 93.94
Parking hub2 92.36 89.22 91.23 88.64 90.02 95.65
Parking hub3 86.17 84.54 85.87 83.57 81.96 72.02
Parking hub4 85.84 89.49 74.09 90.17 89.27 78.25
Parking hub5 93.9 87.26 90.85 85.69 90.31 87.86
Parking hub6 96.38 91.97 93.82 89.83 93.77 87.45
Parking hub7 89.9 87.69 87.08 89.22 94.24 95.04
Parking hub8 96.8 97.03 96.55 95.64 95.95 95.37
Parking hub9 91.94 90.74 91.95 89.59 90.45 90.19
Parking hub10 92.6 92.39 91.35 94.56 92.84 91.96
Average 91.72 90.37 89.44 89.8 91.29 88.77
Stdev 3.5 3.36 5.86 3.41 3.76 7.5
Form D7W0T3 D14W0T3 D6W3T3 D7W0T3 D14W0T3 D6W3T3
Parking hub1 93.97 92.05 92.53 93.99 93.05 96.62
Parking hub2 92.47 91.1 93.49 85.89 87.99 87.41
Parking hub3 86.1 84.54 85.87 84.83 79.68 71.6
Parking hub4 89.35 89.9 72.86 91.62 88.79 80.67
Parking hub5 90.16 87.7 91.46 81.54 87.81 88.88
Parking hub6 94.76 92.04 92.63 90.93 92.24 83.55
Parking hub7 91.43 88.44 86.05 80.65 86.22 92.19
Parking hub8 96.08 96.19 96.68 96.39 95.21 93.91
Parking hub9 94.72 93.87 93.55 91.36 90.78 93
Parking hub10 92.56 92.35 91.41 90.79 86.47 86.55
Average 92.16 90.82 89.65 88.8 88.82 87.44
Stdev 2.85 3.15 6.42 5.01 4.15 7.03
The addition of a categorical parameter representing
holiday (i.e, whether the target day is a holiday or a
normal day) made a significant improvement to all the
results. Each model’s accuracy was increased by no
less than 5% in comparison to the no holiday input
case. These can be observed in the Tables 5 and 6.
Figure 6: Boxplot for accuracy of the models that consid-
ered a holiday input in both the DMS and RMS approaches.
Also, Figure 6 indicates a huge improvement in
the performance of both the DMS and RMS, in com-
parison to the one in Figure 5. The boxplot shrink-
age and the abrupt improvement in accuracy ampli-
fied the importance of the holiday input for the fore-
cast. Additionally, the variance between the perfor-
mance of the different models decreased in all the
cases. An interesting observation can be seen in the
D6W3 form variants. There is an outlier in each in-
put form that has an accuracy lower than the major-
ity of the other models. When comparing the D6W3
in Figure 5 and Figure 6, the input form is causing
more outlier with lower accuracy than the input form
D14W0 and D7W0.
The analysis of the results indicates that the RMS
with a holiday input had the best performance. Al-
though in some case the DMS was observed to be
better than the RMS, the trade-off between the com-
putational cost and the performance make the RMS
the better strategy as a whole. The possible reason
behind the weakness of the DMS is the unsuitability
of the model and also the lack of previous day signal,
thus not capturing the trend. Moreover, the DMS fa-
vored the simplest data input forms (e.g. D7W0 with
and without a holiday) and was able to outperform the
RMS in only few of the cases when a long sequence of
day lags (D14W0) was provided, indicating the need
for a large number of correlated input to produce su-
perior results for specific parking hubs.
5 CONCLUSION
In this paper, we make use of a telecommunication
service provider’s data on the use of the fleet of vehi-
cles by its employees to analyze and forecast the ve-
hicle booking demand for the future. For that, a com-
parison of the accuracy results using the DMS against
the RMS is done. It was observed that the RMS per-
formance was superior to the DMS in the majority
of the cases and with a significantly lower computa-
tional cost. The holiday inputs were found to improve
the prediction quality by about 5% for both the RMS
and the DMS methods. The results suggest that, for
our problem, the RMS forecasting method is better
suited in comparison to the DMS. Some outliers were
seen in the accuracy of the results such as in park-
ing hub 5, and without any external inputs, where the
DMS showed improved accuracy when compared to
the RMS.
The tested models were built into a tool, which
trains the data with all models and automatically
keeps the one with the highest accuracy for each park-
ing hub. The built tool is being trialed by our partner
telecom and encouraging feedback is being received
on the effect of the better forecast have on the resource
management task.
REFERENCES
Azadeh, S. S., Marcotte, P., and Savard, G. (2015). A non-
parametric approach to demand forecasting in revenue
management. Computers & Operations Research,
63:23–31.
Bakker, M., Van Duist, H., Van Schagen, K., Vreeburg, J.,
and Rietveld, L. (2014). Improving the performance
of water demand forecasting models by using weather
input. Procedia Engineering, 70:93–102.
Investigating Prediction Models for Vehicle Demand in a Service Industry
365
Bouktif, S., Fiaz, A., Ouni, A., and Serhani, M. A.
(2018). Optimal deep learning lstm model for elec-
tric load forecasting using feature selection and ge-
netic algorithm: Comparison with machine learning
approaches. Energies, 11(7):1636.
Chen, T. and Guestrin, C. (2016). Xgboost: A scalable
tree boosting system. In Proceedings of the 22nd
ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining, KDD ’16, page
785–794, New York, NY, USA. Association for Com-
puting Machinery.
Deschepper, M., Eeckloo, K., Malfait, S., Benoit, D., Cal-
lens, S., and Vansteelandt, S. (2021). Prediction of
hospital bed capacity during the covid- 19 pandemic.
BMC health services research, 21(1):1–10.
Dineva, A., Csom
´
os, B., Sz, S. K., and Vajda, I. (2021).
Investigation of the performance of direct forecasting
strategy using machine learning in state-of-charge pre-
diction of li-ion batteries exposed to dynamic loads.
Journal of Energy Storage, 36:102351.
Dossani, A. (2022). Inference in direct multi-step and long
horizon forecasting regressions. Available at SSRN
4020488.
Du, P., Jin, Y., and Zhang, K. (2016). A hybrid multi-step
rolling forecasting model based on ssa and simulated
annealing—adaptive particle swarm optimization for
wind speed. Sustainability, 8(8):754.
Feng, B., Xu, J., Zhang, Y., and Lin, Y. (2021). Multi-
step traffic speed prediction based on ensemble learn-
ing on an urban road network. Applied Sciences,
11(10):4423.
Friedman, J. H. (2001). Greedy function approximation: a
gradient boosting machine. Annals of statistics, pages
1189–1232.
Geurts, P., Ernst, D., and Wehenkel, L. (2006). Extremely
randomized trees. Machine learning, 63(1):3–42.
Goic, M., Bozanic-Leal, M. S., Badal, M., and Basso, L. J.
(2021). Covid-19: Short-term forecast of icu beds in
times of crisis. Plos one, 16(1):e0245272.
Herrer
´
ıa-Alonso, S., Su
´
arez-Gonz
´
alez, A., Rodr
´
ıguez-
P
´
erez, M., Rodr
´
ıguez-Rubio, R. F., and L
´
opez-
Garc
´
ıa, C. (2021). Efficient wind speed forecast-
ing for resource-constrained sensor devices. Sensors,
21(3):983.
Hornik, K., Stinchcombe, M., and White, H. (1989). Multi-
layer feedforward networks are universal approxima-
tors. Neural networks, 2(5):359–366.
James, G., Witten, D., Hastie, T., and Tibshirani, R. (2021).
Statistical Learning, pages 15–57. Springer US, New
York, NY.
Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W.,
Ye, Q., and Liu, T.-Y. (2017). Lightgbm: A highly
efficient gradient boosting decision tree. Advances in
neural information processing systems, 30.
Khalid, M. (2019). Wind power economic dispatch–impact
of radial basis functional networks and battery energy
storage. IEEE Access, 7:36819–36832.
Lee, M. (2018). Modeling and forecasting hotel room de-
mand based on advance booking information. Tourism
Management, 66:62–71.
Liu, D., Lu, J., and Ma, W. (2021a). Real-time return de-
mand prediction based on multisource data of one-
way carsharing systems. Journal of Advanced Trans-
portation, 2021.
Liu, Y., Wang, H., Feng, W., and Huang, H. (2021b). Short
term real-time rolling forecast of urban river water
levels based on lstm: A case study in fuzhou city,
china. International Journal of Environmental Re-
search and Public Health, 18(17):9287.
Mu, B., Peng, C., Yuan, S., and Chen, L. (2019). Enso
forecasting over multiple time horizons using convl-
stm network and rolling mechanism. In 2019 inter-
national joint conference on neural networks (ijcnn),
pages 1–8. IEEE.
M
¨
uller, J. and Bogenberger, K. (2015). Time series analysis
of booking data of a free-floating carsharing system
in berlin. Transportation Research Procedia, 10:345–
354.
Shi, X. and Yeung, D.-Y. (2018). Machine learning for spa-
tiotemporal sequence forecasting: A survey. arXiv
preprint arXiv:1808.06865.
Sun, S., Wu, H., and Xiang, L. (2020). City-wide traf-
fic flow forecasting using a deep convolutional neural
network. Sensors, 20(2):421.
Surakhi, O., Zaidan, M. A., Fung, P. L., Hossein Motlagh,
N., Serhan, S., AlKhanafseh, M., Ghoniem, R. M.,
and Hussein, T. (2021). Time-lag selection for time-
series forecasting using neural network and heuristic
algorithm. Electronics, 10(20):2518.
Taieb, S. B., Bontempi, G., Atiya, A. F., and Sorjamaa, A.
(2012). A review and comparison of strategies for
multi-step ahead time series forecasting based on the
nn5 forecasting competition. Expert systems with ap-
plications, 39(8):7067–7083.
Vassallo, D., Krishnamurthy, R., Sherman, T., and Fer-
nando, H. J. (2020). Analysis of random forest mod-
eling strategies for multi-step wind speed forecasting.
Energies, 13(20):5488.
Wang, C.-C., Chien, C.-H., and Trappey, A. J. (2021a). On
the application of arima and lstm to predict order de-
mand based on short lead time and on-time delivery
requirements. Processes, 9(7):1157.
Wang, Y., Zhang, N., and Chen, X. (2021b). A short-term
residential load forecasting model based on lstm re-
current neural network considering weather features.
Energies, 14(10):2737.
Yu, D., Li, Z., Zhong, Q., Ai, Y., and Chen, W. (2020). De-
mand management of station-based car sharing sys-
tem based on deep learning forecasting. Journal of
Advanced Transportation, 2020.
Yuan, S., Luo, X., Mu, B., Li, J., and Dai, G. (2019). Pre-
diction of north atlantic oscillation index with convo-
lutional lstm based on ensemble empirical mode de-
composition. Atmosphere, 10(5):252.
Yuan, S., Wang, C., Mu, B., Zhou, F., and Duan, W. (2021).
Typhoon intensity forecasting based on lstm using the
rolling forecast method. Algorithms, 14(3):83.
Zou, H. and Hastie, T. (2003). Regression shrinkage and
selection via the elastic net, with applications to mi-
croarrays. JR Stat Soc Ser B, 67:301–20.
NCTA 2022 - 14th International Conference on Neural Computation Theory and Applications
366