Contents lists available at ScienceDirect

Transportation Research Interdisciplinary Perspectives

Study on facility location of air taxi skyports using a prescriptive

analytics approach

A R T I C L E I N F O
A B S T R A C T
Keywords:
Urban air mobility The current study identifies infrastructure locations for air taxi operations in New York City (NYC) using a
Electric vertical takeoff and landing prescriptive analytics approach. Based on the potential air taxi demand estimated by prior studies, 14 unique
Air taxi sites are identified by an unsupervised machine learning technique called clustering large applications (CLARA)
Facility location in Step-1. However, developing all the suggested stations simultaneously might be challenging for any business,
Unsupervised machine learning and integer
and therefore, an integer programming model is developed in Step-2 to propose opening these centers in multiple
programming approach
phases, considering parameters such as rental cost, daily number of trips, easy access to road facilities, and
Prescriptive analytics
employee salary, while maximizing the demand satisfaction. Sensitivity analysis is then conducted to identify the
impact of various market penetration strategies (such as aggressive, defensive and balanced), inflation in the
rental cost and employee salary, and the emergence of latent demand on the phased opening of the facilities. The
clustering algorithm results indicate seven air taxi facilities to be established in the Manhattan borough, four in
Queens, two in Brooklyn, and one in New Jersey, spread across three phases. Larger vertiports are suggested for
airports, such as John F Kennedy and LaGuardia international airports, since they serve approXimately 55% of
the overall demand. Results also showed that varying percentages of the budgeted rental cost and employee
salary have minimal effect on the total number of stations.

1. Introduction
provide a more efficient, faster, and cheaper way to commute across
Rapid urbanization and accelerated economic development have the city.
led to a considerable increase in automotive vehicles. As a result, traffic Air taxis are based on the electric Vertical Takeoff and Landing
congestion in urban areas has been exacerbated in recent times due to (eVTOL) concept that can provide urban air mobility (UAM) (Strau-
a failure to meet the demand requirements (Sun et al., 2019; binger et al., 2020; Rajendran and Srinivas, 2020). These battery-
Geronimo et al., 2021). This has a significant impact on the national powered vehicles can travel up to 100 miles and provide substantial
economy (Kan et al., 2019; Kohan and Ale, 2020; Rajendran et al., time savings compared to ground transportation (Holden and Goel,
2021) as well as on an individual’s health (Sanchez et al., 2020). A 2016; Rajendran et al., 2021). Several other logistics companies, such
recent study found that the total hours of delays due to traffic as Airbus, Boeing, and German Lilium, are in the process of designing,
congestion increased from 1.8 billion hours in 1982 to 8.8 billion hours developing, and testing this innovative service (Sun et al., 2018).
in 2017 in 494 urban cities in the United States (Schrank et al., 2019). Numerous studies have been conducted to analyze the impact of air taxis
Consequently, the congestion cost (quantified value for total fuel on the environment and other key operational parameters like vehicle
wasted and time delay) rose to $179 billion from $15 billion during the speed, cruise altitude, and travel time (Bacchini and Cestino, 2019;
same time period (Urban Mobility Report, 2019). Similarly, the cost of Enconniere et al., 2017; Rajendran and Pagel, 2020).
death due to air pollution from excess traffic was estimated to be Typically, the air taxi stations are classified as vertiports and
approXimately $500 million in 2010 (Sanchez et al., 2020). Therefore, vertistops in the literature (Rajendran and Shulman, 2020). A vertistop
to tackle this issue, various logistics and transportation companies is a facility that is solely used for customer pickup and dropoff,
across the globe have started to venture into emerging aviation whereas a vertiport can also be used for battery charging, maintenance,
technology called air taxi, which is expected to docking, and repair, along with passenger movement from one location
to another, similar to
vertistops (Holden and Goel, 2016). Previous literature has focused on
utilizing either clustering algorithms (He et al., 2019; Rajendran and multimodal warm start approach with k-means clustering algorithm to
Zack, 2019) or mathematical programming (He et al., 2018; Kadri et determine 21 potential air taxi stations in New York City (NYC). Demand
al., 2020) for infrastructure locations of emerging logistics. However, estimation for this soaring service was made by employing publicly
no research is available on air taxi location selection using a phased available regular taxi data and parameters provided by Holden and
approach (See Table 1). Goel (2016). They also examined the effect of various parameters, such
The current study focuses on developing tools to gain insights into as customer satisfaction and percentage time savings on prospective
strategic decision planning for the optimal location of air taxi stations sites. Similarly, Lim and Hwang (2019) suggested a k-means algorithm
in any metropolitan city in a two-step process. First, an unsupervised for selecting skyport stations based on three heavily used routes in the
ma- chine learning (clustering) algorithm is used to generate k number Seoul metro area. However, cluster centroids generated from k-means
of potential stations based on certain eligibility parameters defined in approach are highly dependent on the initial value selected, and it
the literature. Second, an integer programming model is developed produces a different result with every run (Zahra et al., 2015). Research
that proposes the opening of k facilities obtained from the first step in n utilizing a mathematical model to identify optimal air taxi hubs was
different phases to maximize the demand fulfillment in each phase conducted by Rath and Chow (2019), with the objective of minimizing
considering several constraints. the total travel cost.
Clustering algorithm, an unsupervised machine learning technique,
2. Literature review is also applied for strategic facility locations in other emerging tech-
nologies, such as charging stations for electric vehicles (Andrenacci
Recent studies have focused on developing concept vehicles and et al., 2016; Helmus et al., 2020; Marino and Marufuzzaman, 2020;
examined the feasibility of implementation of UAM by analyzing the Riemann et al., 2015; Zhang et al., 2019a,b) and drone delivery systems
performance parameters (Al Haddad et al., 2020; Johnson et al., 2018; (Ferrandez et al., 2016; Salama and Srinivas, 2020). Andrenacci et al.
Silva et al., 2018; Vascik and Hansman, 2018). The efficiency of three (2016) utilized a fuzzy k-means cluster technique to generate optimal
different VTOL types, electric quadrotor, hybrid side by side charging stations for electric vehicles in Rome. Similarly, Zhang et al.
helicopter, and turbo electric-powered tilt-wing, were discussed by (2019a,b) explored the amalgamation of barycentric methods with k-
Johnson et al. (2018). They considered several factors, such as aircraft means technique to reduce the impact of demand dispersion on site
optimization, drag minimization, propulsion efficiency, and selection. On the other hand, Helmus et al. (2020) applied Gaussian
operational effectiveness. Silva et al. (2018) extended this research by miXture and partitioning around medoids (PAM) in a two-phase clus-
investigating another type of vehicle called lift plus cruise. The authors tering approach. The first phase was used to identify 13 most common
concluded that lift plus cruise type is a superior alternative for turbo- charging durations, while nine different user types were established in
electric propulsion than tilt-wing. Several criteria impacting the second phase. Marino and Marufuzzaman (2020) investigated a two-
customers’ opinion, such as service reliability, percentage time savings, step approach of integrating k-means algorithm with a stochastic
and cost of availing the facility, were identified by Al Haddad et al. model to determine potential charging station candidates and
(2020) through an online survey of residents in Munich. Similarly, minimize total energy usage in Peru. They employed principal
Vascik and Hansman (2018) established three additional constraints in component analysis to reduce the dimensionality of the dataset to
aircraft noise, a control system for air traffic, and ground improve the algorithm’s performance. They observed an 18.24 %
infrastructure availability that potentially affects the market growth of reduction in energy costs due to the integration strategy.
UAM. This research is one of the first to explore and provide insights into
The literature on infrastructure location for air taxis is still in the the emerging transportation mode known as air taxis. The current
evolutionary stage. Multiple tools, such as clustering algorithms (Lim study utilizes an unsupervised machine learning technique called
and Hwang, 2019; Rajendran and Zack, 2019), mathematical models clustering large applications (CLARA) for generating optimal locations
(Rath and Chow, 2019), and simulation (Balac et al., 2019), have been based on the given dataset. Previous research suggests that CLARA has
utilized to estimate demand and determine optimal vertiport and ver- a signifi- cant advantage over other clustering algorithms such as k-
tistop locations. Rajendran and Zack (2019) proposed integrating a means by reducing the impact of outliers and being non-sensitive to
noise (Popat

2
 parameters, such as cost of the trip, willingness to fly, and time saved. If
the passenger prefers to use the
and Panda, 2019). CLARA, an extension of k-medoids clustering tech-
nique, has been widely used in recent years with technological Rajendran and Zack, 2019) suggests certain unique constraints that are
evolution (Swarndeep Saket and Pandya, 2016). This observation is incorporated in the present study. First, a customer is eligible to utilize
supported by the findings in the current study. The algorithm air taxi services only if the estimated time saved is more than 40 %
proposed in the current analysis also performed significantly better for compared to that of ground transportation. Second, the first and last
similar demand satis- faction rates than the existing literature leg of the trip should not exceed one mile each. Third, the vertiports
(Rajendran and Zack, 2019) that utilized a k-means clustering and vertistops should be able to fulfill at least 70 % of the eligible
technique. passenger demand. Therefore, a clustering algorithm that enables the
The present research is also one of the first to propose an integer decision- makers to adjust the specific number of cluster centers is
programming (IP) model for developing multi-phase location selection used to pro- vide recommendations on air taxi station locations. The
decisions for efficient air taxi network operations. Optimization models value of the centroids is increased until all the constraints mentioned
are predominantly used if there are infinite solution alternatives. In above are satisfied. However, any logistic company would intend to
this problem under study, there are 14 site locations. In each phase, a minimize the number of facility locations, and thus, the objective of the
de- cision has to be made on whether each station has to be opened in algorithm is to minimize the number of stations while fulfilling the
that phase or not. Therefore, the number of solution alternatives is above conditions. The results from the first stage of the research are
extremely large to be handled by multiple criteria ranking algorithms, given as input for the mathematical model that is used to determine the
such as AHP. Therefore, an integer programming model has been specific locations to be opened in each phase while maximizing the
developed in this study. Several factors affecting the decision-making overall demand fulfillment.
policy, such as rental cost, average employee salary, road facilities, and
the number of trips per day, are considered in the model, with the 4.1. Clustering algorithm
objective of maxi- mizing the demand fulfillment rate. Furthermore,
the current study also explores the impact of various market We use an unsupervised learning technique in this research as it
penetration strategies, influence of inflation in the rental cost and automatically identifies inherent pattern/associations present in the
average employee salaries, as well as the emergence of latent demand data. Specifically, clustering algorithm, an unsupervised machine
on the location opening decisions. learning approach, is used to partition millions of pickup and dropoff
coordinates into non-overlapping groups so that it is convenient to
3. System overview determine the infrastructure operating facility. In particular, the parti-
tioning around medoids (PAM) method focuses on minimizing the
The sequence of events for a standard air taxi network operation is total dissimilarity between the data points and the nearest k
presented in Fig. 1 (Holden and Goel, 2016; Salama and Srinivas, representative object in the data, also known as medoid, making it more
2020; Rajendran and Shulman, 2020; Rajendran and Pagel, 2020; robust than the k-means technique. Equation (2) presents the objective
Rajendran and Sinha, 2020). Customers would use a ridesharing app function for PAM (Schubert and Rousseeuw, 2019). A typical PAM
for entering their pickup location and final destination, similar to the algorithm integrates two techniques called BUILD and SWAP for
existing on- demand cab booking system. The app would then compare initial cluster selection and generates locally optimum solutions (since
the travel cost for several available transit mediums, such as public the global optimum problem is NP-hard) by improving the clusters,
transportation (buses, subways, commuter trains, etc.), regular taxis, respectively (Schubert and Rousseeuw, 2019).
and air taxis, if the individual is eligible for availing of the services. ∑ → 2
Typically, passenger preference to use ATS would be based on P= | x — xc | →min.P (1)
→
x∑∈ m
ATS, they can elect to walk from the pickup point to the nearest station mink m d(x , m ) (2)
(first leg) and from the destination skyport to drop off location (last =1...K i k
i=1
leg). Logistic companies would also give the option to cover both
these in-
stances (i.e., first and last legs) through a car trip. For initial construction using the BUILD algorithm, a point that has
the smallest distance to all other points in the dataset is considered as the
4. Methodology first medoid. Subsequently, further k—1 points that reduce the value of
Equation (2) are considered as medoids. Each data point is assigned to
The current investigation integrates an iterative clustering the nearest specific cluster. The SWAP algorithm then considers all
algorithm with a mathematical model to identify potential vertiports possible exchanges between a non-medoid and medoid similar to a
and verti- stops for air taxis in a metropolitan city (as shown in Fig. 2). greedy steepest-descent method to improve the initial clustering. This
Previous literature on the strategic location of air taxis (Holden and is repeated until no further reduction of the objective function is
Goel, 2016; possible. However, this technique increases the run time of the
algorithm

3
Fig. 1. Conceptual representation of the UAM network.

4
 Fig. 2. Proposed framework.

significantly, making it complicated to use for large data sets (Schubert Sets and indices
and Rousseeuw, 2019). To reduce this complexity, a technique known as
clustering large applications (CLARA), a variant of partitioning around i∈I Set of phases

medoids (PAM), is developed. j ∈ JDecision Set of site location

CLARA chooses a sample set from the complete data record and variables
then
utilizes PAM to produce optimal medoids for that representative data. In δij 1 if location j is opened in phase i
order to mitigate sampling bias, the process of sampling and clustering is
0 otherwise
repeated numerous times until the next medoid sets are generated as
final clustering (Kaufman and Rousseeuw, 2008). It has a faster execu-
tion time and requires less storage when compared to PAM (Schubert Input parameters
and Rousseeuw, 2019). A study by Wei et al. (2003) concluded that rj Rental cost for opening a site at location j

CLARA outperformed other algorithms in clustering quality along with pjtjsjfj Populatio/n density at location j
computational time and is less susceptible to data randomness and
Number of trips per day per 1000 customers at location j
cluster distinctness degree. Therefore, CLARA has been used to deter-
mine the optimal station locations for this study. Algorithm 1 presents Average employee salary at location j
the pseudocode used by CLARA for a data set D (Kaufman and Rous-
seeuw, 2008). Distance between an ATS station at location j and its nearest
roadway (road facility)
Algorithm 1. Pseudo Code for CLARA Algorithm. Dj Demand coverage at each location j
Ni Maximum number of sites to be opened at each phase i
1 Create number of samplings = s;
Ri Maximum allowable budgeted rental cost for each phase i
2 Repeat ‘s’ times Si Maximum allowable salary expenditure in phase i
3 Draw sample data set from D; Ti Minimum number of trips per day per 1000 population for
4 ∑mof medoids M = {m1, m2......mk};
Apply PAM to generate a set phase i
5 Compute total score TS = d(xi, mk);
Fi Minimum required coverage from the nearest road facility
for phase i
i=1
6 If
7 TS < minimum total score:
8 Return M; Objective function
9 End If;
10 End Repeat;
11 Assign all data points to M;
Max.
∑∑
δij × Dj (3)
4.2. Mathematical model i j

As indicated earlier, in this study, we use the integer programming

(IP) method for opening the facilities in a phased approach. IP models Constraints
have been widely adopted for site selection problems for other services,
such as drones (e.g., Chauhan et al., 2019; Kim et al., 2019) and ∑
electric bicycles (Zhang et al., 2019a,b; Chen et al., 2020). The δij = 1 ∀j (4)
locations sug- gested by the clustering algorithms are used as an input i

to the integer programming model, which is developed to determine ∑

the set of stations to be opened in different phases. The constraints are δij ≤ Ni ∀i (5)
j
selected based on several parameters affecting the strategic location
decisions as suggested in previous literature (Hawas et al., 2016, ∑
Tzeng et al., 2002, Tzeng et al., 2005). rj × δij ≤ Ri ∀i (6)
i
Notations.

5
∑
sj × δij ≤ Si ∀i (7)
i

6
∑
tj × δij ≥ Ti ∀i (8) literature (Rajendran and Zack, 2019) has been employed in the
i present investigation. However, during this estimation, their study
∑ worked on the assumption that there will be one passenger per ride,
fj × δij ≥ Fi ∀i (9) which is cir- cumvented in this study. As a result, about 6.5 million
i
estimated air taxi passenger records are utilized for the clustering
The objective is to maximize the demand fulfillment at each phase algorithm in the present study, whereas the previous investigation
(Equation (3)). The demand data at each location is generated based considered about 4 million data points. The mean demand in a five
on the total number of riders eligible to avail of the air taxi facility. The minute time interval was observed to be about 22 customers, while the
constraints utilized in the model are given by Equations (4) – (9). median, standard devia- tion and range was observed to be
Equation (4) ensures that a facility can be opened in only one phase. approXimately 18, 22 and 74 pas- sengers on average.
Equation (5) places a restriction on the number of establishments to be Fig. 3 depicts the customer demand distribution across the
deployed in each phase. Previous research on multi-period facility lo- different days of the week. The demand is the lowest during the
cations concluded that more stations should be opened during the earlier weekends, and it gradually increases over the weekdays. The demand is
periods to maximize the total flow (Chung and Kwon, 2015; Li et al., observed to be the highest on Thursday (21 % of the total demand),
2016). Therefore, in the present investigation, it is assumed that while the second highest is on Friday (19 % of the total). Fig. 4
Ni > Ni+1 > ⋯ > NI. Equation (6) describes that the total cost of renting illustrates the passenger demand distribution across the different
the site should be less than the budgeted rental cost for that specific months of the year. The de- mand is highest during the last quarter of
phase. Similarly, Equation (7) ensures that the average salary is less than the year while lowest during Q1. On a monthly basis, the maximum
the maximum allowable salary expenditure for that phase. On the demand is expected to take place during May, which is anticipated due
other hand, Equations (8) and (9) pose a restriction on the total to the increase in tourist visits during summer. Table 2 presents the
number of trips per day per 1000 population and accessibility of air descriptive statistics of the average rental cost, population coverage,
taxi stations to a major road facility. trips per day, road facility, and employment cost in each region.
Clearly, the operational cost (rental cost and employee salary)
4.3. Integer programming solution algorithm associated with Manhattan is the highest compared to the other
boroughs.
An exact method is considered for obtaining the solutions for Phase-2
of this study. The solution space for the model is the set of values of δij 5.2. Facility location decision using CLARA
that satisfies all the model constraints. It is important to note that all
the decision variables in the proposed IP model are binary, and thus For the base case, the proposed framework generates a total of 14
there is an exponential increase in solution space with the problem size sites across five boroughs in New York City (Manhattan, Queens,
(Srinivas and Rajendran, 2017). Branch and bound technique has been Brooklyn, Bronx, and Staten Island) when compared to Rajendran and
widely used to solve IP problems over the past several decades and Zack (2019), that proposed 21 locations. Seven sites (#2, #3, #5, #9,
have proven to be effective. Therefore, in this study, the branch and #11, #12 and #13) are suggested to be built in Manhattan while four
bound algorithm that is given in Algorithm 2 is used (adapted from locations (#1, #4, #7 and #8) are proposed in Queens. Similarly, the
Srinivas and Rajen- dran, 2019). number of stations identified in Brooklyn and New Jersey are two (#6
and #10) and one (#14), respectively. Table 3 presents the locations
Algorithm 2. Branch and Bound Algorithm for Solving Binary Integer
along with their site number. Python programming language is used
Programming Model (adapted from Srinivas and Rajendran, 2019).
for developing the CLARA technique in this study.
1
Initialize the proposed optimization model without binary restrictions (PR ) as a In order to evaluate the effectiveness of the CLARA technique, the
sub-problem,PS ←R Davies–Bouldin index (DBI) metric values obtained for both the tradi-
2
Set a very high current best objective function value,Z* ← + ∞
tional k-means approach and CLARA are compared. The DBI metric of k-
3
While NS ∕= ϕ do means for the considered dataset is 0.42, while for CLARA it is 0.31
4
Choose a sub-problem Pk ∈ PS (with a percentage deviation of nearly 35 % from the k-means
5
Obtain the optimal solution, sk and objective function value Zk by solving pk as a
linear program
6
if Zk > Z* then
7
Remove pk from the set of sub-problems,PS ←PS \pk
8
else
9
if Sk satisfies binary restrictions then
10
Updated current best solution and objective function: s* ←Sk and Z* ←Zk
11
else
12
Choose a variable xij in pk which does not satisfy binary restrictions
13
Create two sub-problems (p1 and p2 ) by adding an additional constraint which
k k
forces xij = 0 in
{ } { }
14
p1k and xij = 1 in p2k (i.e., p1k = pk ∪ xij = 0 and nk2 = nk ∪ xij = 1 )
S S
{ 1 2}
15 Update the list of sub-problems, P ←P ∪ kp , kp \
16 end if
{pk}
17 end if
18 end while
19 return optimal solution s* and optimal objective function Z*

5. Results

5.1. Data description

Since the air taxi operations are currently in the trial and testing
stage, the estimated air taxi NYC demand data used in previous Fig. 3. Estimated air taxi demand distribution across different days of

7
the week.

8
 the maximum, siX stations could be opened in the first phase, followed
by five and four in the second and third stages, respectively. The
impact of changing these values on the opening decisions is discussed
in the sensitivity analysis section. Based on the results, it is observed
that all suggested sites are located in different districts.
Fig. 5 shows the various infrastructure locations suggested in the
three phases. It is observed that approXimately half of the overall air
taxi demand is fulfilled by the siX stations in Phase #1. Whereas Ozone
Park (#7) caters to less than one percent of the demand, and thus, it is
rec- ommended that logistic companies develop one common vertiport
to serve the customers at Ozone park and its nearest proposed station
at JFK International Airport (#8). In Phase #1, the model indicates
three centers to be built in Manhattan, two in Queens and one in New
Jersey (#14). Phases two and three have a demand satisfaction rate of
almost 40 % and 10 %, respectively. Similar to Phase #1, the model
suggests three stations be developed in Manhattan in Phase #2 along
with fa- cilities for LaGuardia International airport (#1) and 84th
Street, Brooklyn (#10).
Fig. 4. Estimated air taxi demand distribution across different months of
the year. 5.4. Sensitivity analysis

algorithm), indicating that CLARA performs better for the estimated In this section, the impact of various input parameters on the
air taxi demand dataset. This was also observed in previous studies phased opening of each center is examined. As in the base case, the
comparing the two algorithms (Arora et al., 2016; Gupta and Panda, integer programming model determines which station has to be
2019). opened in each phase in the sensitivity analysis settings depending on
It is observed that the locations associated with the two airport sta- the model objective (maximizing the demand satisfaction) bounded by
tions in Queens (#8 and #1), John F. Kennedy (JFK) and LaGuardia the con- straints given in Equations (4) – (9). These constraints set the
International Airport (LAG), cater to over 55 % of the total customer upper bound on factors, such as rental cost and employee salary, while
demand. Surprisingly, demand from the Newark Liberty International imposing a lower bound on the daily number of trips.
Airport (#14) is found to be the least when compared to other sites.
Rajendran and Zack (2019) also reported similar findings. Therefore, it 5.4.1. Impact of market penetration strategies
is proposed to build a large vertiport at JFK and LAG for maximum As mentioned earlier, the number of stations in the base case de-
demand fulfillment. Following that, Washington Square Park (#11), W creases based on the sequence (6, 5, 3) for the three phases,
31st Street near Empire State Building (#12), and Time Square (#13) in respectively. The aim of this section is to analyze the impact of
Manhattan cater to about 30 % of the aggregate demand. However, the variation in the tuple on specific locations and the overall demand
algorithm does not suggest major locations, such as Yankee Stadium in realization. Multiple sce- narios considered in the present study with
Bronx and Columbia University in Upper Manhattan, which is regards to the change in the
counterintuitive.
Table 3
5.3. Phased approach for opening facilities using an integer programming Cluster centers suggested in the proposed study.
(IP) model Center Number Locations

1 LaGuardia International Airport

The input parameters required for the mathematical model 2 263 Nagle Avenue
(average rental cost, trips per day per 1000 population, average salary, 3 881–899 Amsterdam Avenue
and road facilities) are obtained from various official sources 4 36th Street, Astoria, Queens
5 854 Park Avenue, near Central Park
(Department of City Planning, NYU Furman Center, and United States
6 Prospect Park West, Brooklyn
Census Bureau). EXcel solver was used to obtain the results of the 7 Ozone Park, Queens
phased approach. 8 John F. Kennedy International Airport
Based on the results obtained from the clustering algorithm, all 14 9 200 Water Street, Lower Manhattan
sites can be expected to function within three phases for the base sce- 10 84th Street, Brooklyn
11 Washington Square Park
nario. EXisting literature on multi-period strategic location decisions
12 W 31Street near Empire State Building
suggests the deployment of more stations in the earlier phase to maxi- 13 Times Square
mize demand fulfillment than later stages (Chung and Kwon, 2015; Li 14 Newark Liberty International Airport
et al., 2016). Therefore, for the base case, a condition is imposed that at

9
 Fig. 5. Infrastructure locations to be opened in different phases.

number of centers in each phase are (i) aggressive market penetration, (ii)
defensive market penetration, and (iii) balanced market penetration. Phase #3 for other instances. Only Ozone Park (#7) and Washington
Under the aggressive market penetration, the exponential decrease (9, 3, Square Park (#11) are found to be in the same phase under the expo-
2) and linear decrease (7, 5, 2) is examined. In the case of defensive nential increase and exponential decrease. On the other hand, only si X
market penetration, an exponential increase (2, 3, 9) and linear centers remain in the same phase for the linear increase and linear
increase (2, 5, 7) scenarios are considered, while the balanced market decrease scenarios.
penetra- tion follows a stable phased opening approach (5, 5, 4). It is
observed that the highest demand satisfaction rate in Phase #1 5.4.2. Impact of rental cost (RC)
(approXimately 60 In this section, the impact of inflation in the rental cost (RC) on
%) is achieved for an exponential increase in sites. Both LAG (#1) and infrastructure opening decisions is observed. In the US, the average in-
JFK (#8) are recommended in the first phase of this scenario. Linear crease in rental cost each year is nearly 5 % (Arnold, 2022). Therefore, in
increase and linear decrease showcase the maximum percentage de- this analysis, the rental cost is increased in steps by 5 % while keeping
mand fulfillment for Phase #2 (around 70 %) and Phase #3 (around 65 the number of facilities at each stage constant for all settings. It is
%), respectively. In the balanced setting, demand satisfaction follows observed that a 5 % rise in cost from RC1 to RC2 leads to si X locations
an increasing trend, which is counterintuitive. being shifted from their respective initial phases. Both Central Park
In terms of locations, it is interesting to note that the model (#5) and JFK (#8) are recommended to change from Phase #1 to Phase
suggests LAG (#1) in Phase #2 in all scenarios except for an #3 in RC2, which results in an increase in the demand fulfillment for
exponential increase in the number of stations. Newark Liberty Phase #3 to approXimately 45 %. On the other hand, only 36th Street,
International Airport (#14) is reported in Phase #1 for every instance Astoria (#4), which has one of the lowest rental costs amongst all the
besides the exponential and linear increase settings. Another station other locations, switched from Phase #3 to Phase #1. Similarly, two
near Central Park (#5) is pro- posed in Phase #1 for all the cases in stations, Prospect Park (#6) and Washington Square Park (#11),
which the number of sites de- creases, while it is recommended for changed from Phase #3 to Phase #2. However, the demand satisfaction
higher stages in other situations, as shown in Table 4. Similarly, 881 for Phase #2 in RC2 remains similar to the base case scenario. It is
Amsterdam Avenue (#3) is indicated in Phase #1 for both linear and interesting to note that despite a linear increase in the cost, no further
exponential decrease, while it is desired in changes are obtained

Table 4
Variation in number of stations in each phase.
Center # Base case (6, 5, 3) EXponential decrease (9, 3, 2) Linear Decrease (7, 5, 2) EXponential increase (2, 3, 9) Linear increase (2, 5, 7) Balance (5, 5, 4)

1 2 2 2 1 2 2
2 1 3 1 3 3 1
3 2 1 1 3 3 3
4 3 1 3 2 3 3
5 1 1 1 3 2 3
6 3 1 1 3 2 2
7 1 2 1 2 1 1
8 1 3 2 1 3 3
9 1 1 2 3 2 2
10 2 1 2 3 2 2
11 3 2 1 2 3 2
12 2 1 2 3 1 1
13 2 1 3 3 3 1

1
14 1 1 1 3 3 1

1
in the other scenarios. Table 5 presents the suggested phases for all the
sites for various rental cost cases considered in the present study. Table 6
Variation in each phase for changes in number of trips per day per 1000
5.4.3. Impact of number of trips (TR) customers.
As described by Hawas et al. (2016) and Meyer (2001), the number Center TR1 (20 % TR2 (10 % TR3 TR4 (10 % TR5 (20 %
of trips per day per 1000 customers (TR) depicts the transportation Number decrease) decrease) (Base increase) increase)
case)
supply. In this section, TR is varied from TR1 to TR5 in steps of 10 %
for all three phases simultaneously, with the values in TR3 considered in 1 2 2 2 2 1
2 1 1 1 1 1
the base case. TR1 and TR2 investigate the effect of a decrease in 3 2 1 2 1 3
demand in the future whereas, TR4 and TR5 analyze the impact of the 4 3 3 3 3 3
emergence of latent demand for the ATS. As expected, a rise in the 5 3 3 1 2 2
initial number of available air taxi trips per day led to higher overall 6 3 1 3 1 3
7 1 3 1 2 2
demand satisfaction in Phase #1, supporting the findings in previous
8 1 1 1 1 1
literature. It is observed that the locations having the highest trips per 9 2 2 1 3 2
day per 1000 customers are selected in Phase #1 for TR5. Therefore, 10 1 1 2 1 2
LAG (#1) and W 31st Street (#12) are recommended to be developed 11 1 1 3 1 1
12 2 2 2 2 1
in Phase #1 for the TR5 setting, while the model proposes both these
13 1 2 2 2 1
locations to be built in Phase #2 for all other cases. These two sites 14 2 2 1 3 2
contribute heavily to increasing the customer fulfillment rate for Phase
#1 from the base case. On the other hand, a decrease in the number of
trips has a minimal impact on the demand satisfaction rate. While in the results for AS3, AS4, and AS5, with Phase #2 having the highest
most stations are pro- posed in a different phase for the various customer satisfaction rate (approXimately 70 %) when compared to the
settings, the overall demand percentage is almost the same for TR1, other two stages. Another critical pattern observed is that 881 Amster-
TR2, as well as the base case (TR3). While Times Square (#13) is dam Avenue (#3) and Times Square (#13) are recommended in Phase
expected to be in Phase #1 for TR5 because of a high number of daily #2 for AS1 and in Phase #1 and Phase #3 for all other settings.
trips, it is also proposed at the same stage for TR1. Interestingly, the
model recommends the location in Phase #2 for all other settings. 5.4.5. Impact of road facility (RF)
Another key observation is that only three locations, 263 Nagle Road facility (RF) is described as the feature required for the utili-
Avenue (#2), 36th Street, Astoria (#4), and JFK (#8), remain in the zation of alternative vehicles (Tzeng et al., 2005). In this section, it is
same phase for all five settings. This indicates that TR is an important varied at 10 % intervals, and the impact of reducing (RF1 and RF2) and
criterion for making air taxi infrastructure de- cisions. Table 6 shows improving (RF4 and RF5) road facilities are investigated in
the various vertiport and their corresponding comparison to the base case (RF3). As expected, creating more
phases for different TR settings. accessible access to the vertistops and vertiports led to an increase in
the demand satisfac- tion rate in Phase #1. This is because better
5.4.4. Impact of average salary (AS) accessibility leads to certain important sites such as Washington
Based on the existing trend, it is expected that the average Square Park (#11), W 31st Street (#12), and Times Square (#13) to be
employee salary (AS) will increase by about 3.5 % each year (Bremen, recommended in Phase #1 for RF4 and RF5. All three stations are
2022). Therefore, in this section, AS is increased linearly from the base proposed in later stages for all other scenarios.
case (AS1) by 5 % until 20 % (AS5) to study the impact of Interestingly, the model suggested locations having the two highest
inflation in employee salary on phased opening. Only five locations demand in JFK (#8) and LAG (#1) in Phase #3 for RF2 and RF5,
remain in the same phase for all the scenarios in this analysis, out of respectively. Potential reasoning for this could be that both JFK and LAG
which four are in Phase #1 (263 Nagle Avenue (#2), Central Park (#5), are situated on the outskirts of Queens, and any change in road infra-
Ozone Park (#7), and Newark Liberty International Airport (#14)). structure might have an impact on facilitating customer access. It is
Another interesting result to note is that despite this consistency, observed that only 263 Nagle Avenue (#2), 36th Street Astoria (#4),
the demand fulfillment in Phase #1 reduces drastically with a rise in and Ozone Park (#7) remain in the same stage for all five cases. Table 8
employee wages. This is probably because the model prefers cheaper showcases the different phases recommended for each station.
locations at the initial stages. The phases suggested for different loca-
tions are presented in Table 7. It is worth noting that there is no
change

1
Table 8 developed to recommend specific stations to be built in multiple
Variation in each phase for changes in road facility. stages. This would give logistic companies an opportunity to analyze
Center RF1 (20 % RF2 (10 % RF3 RF4 (10 % RF5 (20 % the market further and improve the existing services. The constraints
Number decrease) decrease) (Base increase) increase) chosen in the model were based on key factors impacting the strategic
case) location de- cisions in previous studies. Finally, sensitivity analyses
1 2 2 2 2 3 were conducted to study the influence of these constraints on the
2 1 1 1 1 1 centers and their cor- responding phases.
3 2 1 2 3 2
For a given demand fulfillment rate, the present study reported fewer
4 3 3 3 3 3
5 1 2 1 2 3 potential locations than the existing literature. The base case assumed
6 3 1 3 3 2 siX, five, and three sites to be built respectively in order to maximize
7 1 1 1 1 1 the demand fulfillment. ApproXimately 50 % of the overall demand
8 1 3 1 1 1
was satisfied in Phase #1. Whereas Phase #2 and #3 had close to 41 %
9 1 1 1 2 2
10 2 2 2 1 2
and 9
11 3 3 3 1 1 % demand completion rate. It was observed that out of the 14 stations,
12 2 2 2 2 1 the two major airports in the city (JFK and LAG) catered to over half
13 2 2 2 1 1 the potential demand. However, only JFK was proposed to be
14 1 1 1 2 2
developed in Phase #1, while LAG was suggested in Phase #2 for the
base setting. Parameters, such as the number of stations to be
5.5. Summary of model results across the different analysis recommended in each phase, the daily number of trips, and road
facilities, had a substantial impact on the demand fulfillment for each
Fig. 6 shows the heat map illustrating the frequency of selecting lo- stage. The results obtained in the present research performed better
cations in each phase, considering all the analyses conducted in this than the prior study on air taxi facility location (Rajendran and Zack,
study. Among the suggested locations, three sites (location # 2,7 and 2019). It is expected that the air taxis would improve citywide
14) are opened in Phase-1 for more than 50 % of the settings. These transportation, and the current research could provide a foundation for
locations are distrusted among different geographic locations logistic companies that might be inter- ested in venturing into the
(Manhattan, New- ark International Airport, and Queens). While sites eVTOL market.
#1,10 and 12 (LaGuardia International Airport, Brooklyn and
Manhattan) are being recommended to be opened in Phase-2 in a 6.1. Paper contributions
majority of the settings, site #4 (a small vertistop in Queens) is the
only location that is suggested to be established in Phase-3 for over 50 To the best of the authors’ knowledge, this paper is the first to pro-
% of the settings. pose a two-step procedure to not only determine the infrastructure lo-
cations for efficient air taxi operation but also provide recommendations
6. Conclusions on opening these sites in a phased manner. For a similar demand
fulfillment rate, the proposed technique exhibited superior performance
The current research focuses on proposing optimal infrastructure than previous literature (Rajendran and Zack, 2019). Furthermore, the
locations for air taxi network operation in urban cities using a pre- current study also explores the impact of various market penetration
scriptive analytics approach. While New York City was chosen as a case strategies, influence of inflation in the rental cost and average
study in the present investigation, the proposed approach can be easily employee salaries, as well as the emergence of latent demand on the
replicated for any other major city. The estimated air taxi demand data location opening decisions. The developed tool can be leveraged by
had about 6.5 million data records. Fourteen ideal operating sites for logistic companies for establishing emerging air taxi services in any
air taxi services using the Clustering large applications (CLARA) metropol- itan city.
technique have been identified. Next, an integer programming
model was

1
Fig. 6. Heat map of different sites selected across the three phases.

1
6.2. Managerial implications performance of the facility.

While the present study focuses on integrating the clustering algo-

rithm and integer programming model for strategic infrastructure
location decisions in New York City, the methodology and several
related implications can be replicated for other major cities, as well.
Based on the results obtained, the following managerial recommenda-
tions are proposed:
From the results of the clustering algorithm, it is observe that there
is a potential to build 14 stations in NYC with seven facilities proposed
in the Manhattan borough.
Based on the taxi data, larger vertiports are suggested for JFK (#8)
and LAG (#1) since they serve approXimately 55 % of the overall
demand.
It is not essential to develop a center at Ozone Park (#7) in Phase
#1 as it caters to less than one percent of customers. Interested users
can leverage the nearest station at JFK.
Similarly, only a negligible proportion of consumers are expected to
avail the air taxi facilities from Newark Liberty International Airport
(#14). Therefore, it is recommended not to provide initial services until
latent demand arises.
It is noticed that market demand in Brooklyn is fairly low despite
having a few popular tourist destinations such as Prospect Park and
Ozone Park. Therefore, small vertistops are suggested to be
constructed for these sites.
It was observed that CLARA performed considerably better than the
k-means clustering algorithm for a large volume of data.
The total number of air taxi facilities to be developed in each phase
has a significant impact on the overall demand fulfillment for all
stages. The logistic companies can prefer to set up fewer
infrastructure lo- cations during the initial phase to observe
customer response to the product before investing heavily in other
locations. In such a situation, an exponential increase in the number of
stations and balanced scenarios would give a reasonably high demand
satisfaction rate for Phase #1
while minimizing incurred cost.
Results indicate that varying percentages of the budgeted rental
cost and employee salary have minimal effect on the total number of
stations.
A linear increase in the number of trips per day per 1000 customers
favors the locations having higher demand during the initial phases.
While it was expected that improving road facilities would expand
the demand satisfaction rate in Phase #1, it is interesting to note that
no substantial changes are viewed for decreasing road facilities when
compared to the base case.
Among the suggested locations, three sites in Manhattan, Newark
International Airport, and Queens are opened in Phase-1 for more than
50 % of the settings.
Sites in LaGuardia International Airport, Brooklyn and Manhattan,
are being recommended to be opened in Phase-2 in majority of the
settings.

6.3. Limitations and future works

While the current investigation is based on providing strategic

location selection decisions for urban air mobility, several operational
and tactical level insights have not been considered. Future work could
involve researching optimal scheduling and effective routing of air taxi
vehicles to reduce operating costs. This would also enable the logistic
companies to balance the capacity during the peak demand and off-
peak hours. In the present study, sensitivity analysis model parameters
are estimated solely based on the literature. Future work could
consider deriving these values using closed-form expressions.
Additionally, complementary market studies could be performed in the
future to analyze the latent demand and its impact on the suggested
infrastructure locations. Similarly, several factors such as air taxi
vehicle capacity and cost of the ride might affect public acceptance and
10
CRediT authorship contribution statement

Akhouri Amitanand Sinha: Methodology, Formal analysis, Vali-

dation, Investigation, Visualization. Suchithra Rajendran: Conceptu-
alization, Methodology, Resources, Supervision, Project
administration.

Declaration of Competing Interest

The authors declare that they have no known competing financial

interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

No data was used for the research described in the article.

10
11