international journal of hydrogen energy 34 (2009) 6045–6064

Available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/he

Optimization of hydrogen stations in Florida using the

Flow-Refueling Location Model

Michael Kubya,*, Lee Linesb, Ronald Schultzc, Zhixiao Xiec, Jong-Geun Kima, Seow Limd
a
School of Geographical Sciences, Arizona State University, Tempe, AZ 85287-5302, USA
b
Department of Environmental Studies, Rollins College, 1000 Holt Ave., Box 2753, Winter Park, FL 32789-4499, USA
c
Department of Geosciences, Florida Atlantic University, Boca Raton, FL 33431, USA
d
Salt River Project, 1521 N. Project Drive, Tempe, AZ 85281-1298, USA

article info abstract

Article history: This paper develops and applies a model that locates hydrogen stations to refuel the
Received 26 September 2008 maximum volume of vehicle flows. Inputs to the model include a road network with
Received in revised form average speeds; the origin–destination flow volumes between each origin and destination;
6 May 2009 a maximum driving range between refueling stops; and the number of stations to build.
Accepted 10 May 2009 The Flow-Refueling Location Model maximizes the flow volumes that can be refueled,
Available online 24 June 2009 measured either in number of trips or vehicle-miles traveled. Geographic Information
Systems and heuristic algorithms are integrated in a spatial decision support system that
Keywords: researchers can use to develop data, enter assumptions, analyze scenarios, evaluate
Optimal tradeoffs, and map results. For the Florida Hydrogen Initiative, we used this model to
Location investigate strategies for rolling out an initial refueling infrastructure in Florida at two
Refuel different scales of analysis: metropolitan Orlando and statewide. By analyzing a variety of
Infrastructure scenarios at both scales of analysis, we identify a robust set of stations that perform well
Station under a variety of assumptions, and develop a strategy for phasing in clustered and con-
Model necting stations in several stages or tiers.
Intercepting ª 2009 International Association for Hydrogen Energy. Published by Elsevier Ltd. All rights
Capturing reserved.
Network

1. Introduction plan how to phase in a network of refueling stations for intra-

city trips in the Orlando metropolitan area and for statewide
Many studies have highlighted refueling infrastructure as one intercity travel across Florida.
of the most formidable barriers to the transition to This research was conducted for the Florida Hydrogen
a hydrogen-based road transportation system [1–4]. Given the Initiative, Inc. (FHI), a non-profit organization established to aid
high cost of building new fueling stations, it is essential to the development of a robust hydrogen industry in Florida. In
coordinate the locations of the initial stations in a network addition to optimizing a station network, our project for FHI
that facilitates the maximum utilization by consumers. analyzed the feasibility of a hydrogen rental-car business based
Numerous models have been proposed to optimize a network at the Orlando International Airport (OIA) [17,18]. The rental car
of fueling stations [5–16]. This paper uses one such model, the analysis is closely related to the station analysis because most
Flow-Refueling Location Model developed by Kuby and Lim, to car renters in Orlando visit only a small handful of destinations,

* Corresponding author. Tel.: þ1 480 965 7533; fax: þ1 480 965 8313.
E-mail address: [email protected] (M. Kuby).
0360-3199/$ – see front matter ª 2009 International Association for Hydrogen Energy. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijhydene.2009.05.050
6046 international journal of hydrogen energy 34 (2009) 6045–6064

Nomenclature MILP mixed-integer linear program

NREL National Renewable Energy Laboratory
CNG compressed natural gas
OD origin-destination
FCLM Flow-Capturing Location Model
OIA Orlando International Airport
FDOT Florida Department of Transportation
OR operations research
FHI Florida Hydrogen Initiative
SDSS spatial decision support system
FRLM Flow-Refueling Location Model
TAZ traffic analysis zone
GIS Geographic Information System
VMT vehicle-miles traveled

making it possible for a small number of refueling stations to Hydrogen Highway with 200 refueling stations spaced 20 miles
meet the needs of most rental car customers. apart by 2010 at a cost of up to $100 million [19]. Though Cal-
Section 2 of this paper reviews other efforts underway ifornia backed off from this ambitious goal, current plans for
worldwide to develop initial networksdconsisting of clusters Phase 1 call for 50–100 stations by 2010. The existing stations are
or corridorsdof hydrogen refueling stations. Section 3 reviews clustered mainly in the Los Angeles and San Francisco–Sacra-
the literature on approaches to planning and modeling the mento regions [22,23]. The idea is to provide a higher level of
best locations for a network of refueling stations. Section 4 service to a smaller but densely populated and polluted area,
introduces our approach using the Flow-Refueling Location thus maximizing the number of more-likely buyers. Stations
Model (FRLM) to optimize and compare various strategies for connecting these clusters along rural interstate highways are
developing an initial hydrogen-refueling infrastructure for targeted for Phase 2. The State of California allocated $6.5
Florida. The model uses operations research (OR) and million to develop stations in both 2005 and 2006.
Geographic Information System (GIS) techniques, along with New York’s H2-NET hopes to open 20 stations between
sophisticated and detailed transport databases, to locate New York City and Buffalo and add hydrogen-fueling capa-
a coordinated set of stations. The model maximizes the bility to 70 existing CNG stations by 2020. Illinois is planning
potential for consumers to refuel as many trips (or, alterna- a 2H2 Hydrogen Highway, and the Northern H Fuels Network
tively, vehicle-miles traveled) as possible. Section 5 describes in the Great Plains region of the US and Canada proposed 12
the data used. We then apply the model to a number of stations spaced 200 km apart by 2012.
scenarios for the statewide network (Section 6) and the At the national scale, the National Renewable Energy
Orlando area (Section 7). In Section 8, we synthesize the two Laboratory (NREL) analyzed the potential development of
sets of scenarios with findings from our companion study of a national backbone network to make long-distance, interstate
a hydrogen rental-car business in Florida to develop a strategy trips possible by 2020 [24]. They estimated that 284 stations are
for phasing in a network of refueling stations. This strategy needed, at a total construction cost of $837 million. Fourteen of
prioritizes and coordinates station development in Orlando these stations are suggested for Florida, including six at
and statewide into several tiers or stages. Section 9 offers existing CNG stations. At the same time, a team of researchers
some brief conclusions. from Oak Ridge National Laboratory and National Renewable
The emphasis in this paper is on developing a coordinated Energy Laboratory is recommending to ‘‘concentrate on
system of multiple stations, rather than choosing suitable establishing networks of fueling stations in a limited number
parcels of land for particular station sites. Site criteria have of urban centers during the transition period. Strategically
been studied elsewhere, and include accessibility, traffic, fleet placing stations in major urban centers will maximize
operations, safety, host partners’ experience with gaseous coverage and permit a cost-effective approach to providing the
fuels, hydrogen supply, adequate space, logistical factors, early infrastructure’’ [25, p. xii]. They target Los Angeles and
energy sources, and zoning [19]. New York for the earliest clusters. President Bush and the U.S.
Department of Energy set 2015 as the date for reaching
a variety of technical and commercial targets to allow the
2. Hydrogen refueling infrastructure private sector to make a commercialization decision [26].
developments Outside of the USA, the European Union’s Joint Technology
Initiative hopes to kick-start commercialization by 2015 [27].
Many regions in the US and around the world have begun The Scandinavian Hydrogen Highway Project consisting of
planning and deploying initial networks of hydrogen refueling HyNor (Norway), Hydrogen Link (Denmark) and HyFuture
stations in order to break the well-known chicken-and-egg (Sweden) will link Oslo with Copenhagen. In Canada,
cycle. As of June 2008, there were an estimated 172 stations a Hydrogen Highway will connect Vancouver to Whistler Ski
worldwide, with 37% in the US, 15% in Germany, and 13% in Resort for the 2010 Olympics.
Japan [20]. By the end of 2008, the number of stations world- There are currently two hydrogen-fueling stations in Florida,
wide was projected to reach 200 [21]. both in the Orlando area. The Boggy Creek Hydrogen Refueling
California has forged ahead of any other region in the world Station opened near the Orlando airport in May 2007 to refuel
in developing a refueling infrastructure. By mid-2008, 32 rental-car shuttle buses. The other station, a mobile refueling
hydrogen stations were built and operating in California, with station operated by Progress Energy near suburban Oviedo,
16 more planned [20], though not all are open to the public. opened in December 2007 but was scheduled for
Initially (in April 2004), California proposed a statewide decommissioning.
 international journal of hydrogen energy 34 (2009) 6045–6064 6047

‘‘flow-intercepting’’ models [34–48]. In Flow-Capturing Loca-

3. Prior research on refueling infrastructure tion Models (FCLM), demand consists of paths through
modeling a network instead of points of origin for trips to the facility
and back. The FCLM locates facilities conveniently on the
In recent years, a large number of papers and reports have origin-to-destination routes that drivers use on trips they
addressed the infrastructure needs for the transition to already make. The basic model locates p facilities so as to
hydrogen energy. While numerous studies have examined the intercept as many trips as possible. These models have been
supply chain from production to delivery to fueling stations used for locating ‘‘discretionary’’ facilities, such as ATMs,
[15], or the total number of initial hydrogen stations needed convenience stores, and fast food, at which people stop on
[28,29], our focus here is on methods used to solve optimally their way to somewhere else rather than make a special trip
for the locations of a coordinated network of stations. We from home to facility and back. In our opinion, the FCLM
group these approaches into GIS models and operations provides a realistic behavioral basis for locating alternative-
research (OR) models. fuel stations because people tend to refuel on their way to
Melaina [30] proposed a GIS method that yields exact station somewhere else.
locations and station sizes based on the idea of condensing the Kuby and Lim [8,49] modified the basic FCLM for the
existing network of gasoline stations into clusters. The NREL purpose of locating hydrogen-refueling stations. The FCLM
study discussed above [24] used GIS to develop a national assumes that a single facility anywhere on a path is enough to
network of stations to enable long-distance trips on interstate capture that demand. For serving long trips, however, a single
highways. They selected heavily traveled interstates with over station may not be enough to refuel a trip from an origin to
20,000 vehicles per day, narrowed that set by choosing major a destination and back again because vehicles have a finite
north–south and east–west routes, and then placed stations driving range. US DOE has set a target of 300 miles for the
along those routes guided by other GIS data layers such as driving range of hydrogen vehicles, but the practical limit, or
hydrogen production plants, other alt-fuel stations, population, what drivers would be comfortable with, may be far less. For
and US highway intersections. Stations were placed no more reference, the 2009 Honda FCX Clarity has a driving range of
than 50 miles (80.5 km) apart in the east and in urban areas, and 240 miles. To deal with fuel limitations, Kuby and Lim adapted
no more than 100 miles (160.9 km) apart in the west. the FCLM by incorporating a driving range parameter. The
Although GIS is a powerful tool for integrating detailed resulting Flow-Refueling Location Model (FRLM) is a mixed-
spatial data layers, it is not ideal for ‘‘combinatorial optimi- integer linear programming problem (MILP) that maximizes
zation,’’ in which the model must choose a combination of the number of trips that can potentially be refueled with
locations from a large set of candidate sites. Given the astro- a given number of stations p:
nomical number of combinations that are possible for many
X
real-world problems, many researchers turn to operations max Z ¼ fq yq (1)
research (OR) techniques. One of the earliest OR papers for q˛Q

locations of gasoline stations developed a model to open and Subject to:

close gasoline stations to maximize a company’s market share
X
[5]. They based market share on two kinds of demand: home- bqh vh  yq ; c q˛Q (2)
to-facility trips, and traffic volumes on network links. Bapna h˛H

et al. [6] used multiobjective programming to locate refor-

mulated gasoline stations in India. One objective minimized ahk xk  vh ; c h˛H; k˛K (3)
the sum of travelers’ costs and station investment costs, while
the second maximized the population on enabled links. These X
xk ¼ p (4)
objectives are optimized subject to a constraint that it is k˛K
possible to travel from every node to every other node by at
xk ; vh ; yq ˛f0; 1g; c k; h; q (5)
least one route given the driving range of vehicles. The fuel-
travel-back model, which uses VMT data as the demand,
where:
locates stations to minimize the average refueling travel time
for a random motorist [12]. Bersani et al. [13] developed
q ¼ index of OD pairs (and, by implication, the shortest paths
a competitive multiobjective model for a petrol company to
for each pair)
minimize its investment cost for converting some existing
Q ¼ set of all OD pairs
stations to hydrogen while maximizing the demand that can
fq ¼ flow volume on the shortest path between OD pair q
be satisfied. In their model, demand is related to existing
(number of vehicle-trips per time period)
gasoline sales, storage capacity, and distances to competitors. 
1; if fq is captured
Several papers have used a variant of the p-median model yq ¼
0; otherwise
to locate hydrogen-refueling stations in various urban areas in
California [5,7,9,12,31]. The p-median modeldone of the most k ¼ potential facility location
widely used OR models for facility locationdlocates a given K ¼ set of all potential facility locations

number ( p) of stations so as to minimize the total distance 1; if a facility is located at k
xk ¼
from population nodes to their nearest open station [32,33]. 0; otherwise
Beginning in 1990, researchers began developing a new p ¼ the number of facilities to be located
approach to facility location, known as ‘‘flow-capturing’’ or h ¼ index of combinations of facilities
6048 international journal of hydrogen energy 34 (2009) 6045–6064

resulting Spatial Decision Support System (SDSS) performs

three primary tasks [17]:

1. Processing transport network data from GIS sources into

a format suitable for the solution algorithms.
Fig. 1 – Example showing it is not possible to complete
2. Solving the FRLM for a given driving range and number of
a round trip from A to B and back assuming a 100-mile
facilities.
driving range.
3. Displaying the outputs in map and graph form.

H ¼ set of all potential facility combinations The SDSS is implemented by extending the ArcGIS Desktop

1; if facility k is in combination h interface using the ESRI ArcObjects and Microsoft .NET tech-
ahk ¼
0; otherwise nologies. It provides an intuitive user interface for data input,
 data conversion, model execution options, and display results.
1; if facility combination h can refuel OD pair q
bqh ¼ For the preprocessing stage, the SDSS uses five main input
0; otherwise
 map data layers: a point layer of population centers, which are
1; if all facilities in combination h are open
vh ¼ used as the origins and destinations of trips; a point layer of
0; otherwise
road junctions; a point layer of candidate facility sites; a line
layer of road network arcs; and a line layer of shortest paths.
Constraint (2) prevents a trip on path q from being counted
Non-spatial input data to the SDSS includes the vehicle range,
as refueled (variable yq ¼ 1) unless a valid combination of
the number of facilities to be built, and a trip table for all OD
stations h is open (variable vh ¼ 1) that can refuel a vehicle on
pairs. The SDSS uses ESRI ArcMap and ArcCatalog to create
path q given the driving range specified (coefficient bqh ¼ 1).
GIS data in the ESRI personal geodatabase format. It generates
The driving range assumption is embedded in the bqh coeffi-
the shortest path between all OD pairs using the ArcGIS
cients. A subroutine developed by Kuby and Lim [8] uses the
Network Analyst tool. It also converts all GIS data, with real-
driving range parameter to determine whether a particular
world locations and shapes, into a simplified and connected
combination h of stations is capable of refueling a vehicle on
network of nodes, arcs, and paths for the solution algorithms,
the round trip on path q without running out of fuel. If so, the
which have no need for real-world location information.
coefficient bqh is set to 1. Constraint (3) ensures that a combi-
While MILP software can solve small versions of the FRLM
nation of stations h is considered open (vh ¼ 1) only if all of the
optimally, we developed two heuristic solution algorithms to
individual stations k in that combination are open (xk ¼ 1). The
solve larger, more realistic cases of the modeldalthough they
coefficients ahk indicate whether a station k is a member of
are not guaranteed to find the global optimal solution in every
station combination h. Constraint (4) sets the number of
case. For this paper, we used a greedy-adding-and-substitu-
stations to p, a user input. Finally, (5) stipulates that a station
tion algorithm [50,51]. A simple greedy algorithmdwithout
or a combination of stations is either open or not, and likewise
substitutiondwould add one facility at a time at the site that
that any origin-destination flow can be considered either
increases the objective function the most, that is, that refuels
refuelable or not. The objective function (1) maximizes the
the most flow volume above and beyond what the previous
flow volume (number of vehicle trips) that can be refueled.
facilities were able to refuel. The simple greedy algorithm,
Upchurch et al. [14] developed an alternative objective (6) that
however, has been shown to be suboptimal for the FRLM [8].
multiplies each trip q by its distance dq. This objective is
The greedy substitution algorithm, on the other hand, allows
measured in vehicle-miles traveled. It gives preference to
the model to swap unused candidate sites for chosen candi-
siting stations to refuel longer trips, which would consume
date sites at each iteration. So if, for instance, after choosing
more hydrogen than shorter trips:
the 1st to 8th sites, the 3rd site becomes partly redundant, the
X
max Z ¼ fq dq yq (6) substitution algorithm could swap another site for it, and
q˛Q possibly perform several additional substitutions, before
moving on to add the 9th site. The second solution algorithm
The vehicle driving range is related to, but not equivalent to,
in the SDSS is a genetic algorithm, but it was not used to
station spacing. For example, assume a driving range of 100
generate the results here. Both algorithms allow the user to set
miles and stations located 100 miles apart at nodes A and C
certain parameters:
(Fig. 1). A round trip between nodes A and C could be
completed without running out of fuel, as could the round trip
1. A driving range for vehicles (a distance).
between B and C. The round trip from A to B, however, would
2. The number of stations to locate ( p).
require 160 miles without refueling and thus could not be
3. The objective function to be maximized (either the number
completed.
of trips or vehicle-miles traveled that can potentially be
refueled).
4. Any stations the analyst wishes to force into the solution
4. A spatial decision support system for the (either pre-existing or desired).
flow refueling location model
After completing a model run, the solution algorithms
For the FHI, we integrated the FRLM into a GIS interface using output the results directly to text files, the Access database,
ArcGIS software and Microsoft Visual Studio.net. The and directly to the open ArcMap window. The selected
 international journal of hydrogen energy 34 (2009) 6045–6064 6049

facilities and the refuelable paths are displayed directly on the junctions. Once each network was built, we generated short-
map with real locations and shapes. These optimal facilities est paths minimizing travel time between each OD pair (e.g.,
and covered paths can be saved as separate layers for each Fig. 2). Our Florida-based researchers checked many paths one
scenario, making it easy to visually compare results for by one to see if they followed realistic routes that drivers
different scenarios. The percentage of trips or VMT covered by would take, and calibrated the speeds associated with
each solution is output to a table from which tradeoff graphs different classes of roads, added missing road links, or
such as Fig. 8 can be made. changed OD locations accordingly.
The Orlando study area includes all of Orange and Semi-
nole counties and the northwest region of Osceola county
5. Data including the cities of Kissimmee and Saint Cloud (Fig. 3). We
obtained GIS highway networks and maximum speed data
The three universities involved in this project worked closely from the Florida Department of Transportation (FDOT), and
together to build, simplify, check, and calibrate detailed and detailed street networks from ESRI, Inc. FDOT divides the
realistic GIS network databases for the Orlando and statewide Orlando study area into 358 traffic analysis zones (TAZs),
case studies (see [17] for details). For each case study, we which we aggregated to 102 larger zones. We obtained
obtained GIS road network databases and then corrected a matrix of weekday OD trip volumes from FDOT, which we
topology errors in the raw network and simplified the road aggregated into a 102  102 matrix, ignoring short intra-zonal
network by eliminating minor and duplicative roads while flows. Speeds on arterial streets were reduced from posted
retaining needed connectivity. Next we aggregated nearby speed limits by 15% to reflect slower driving conditions.
traffic analysis zones (TAZs) together, being careful to ensure For intercity travel, the study area included the entire state
contiguity and compactness, and selected a single origin– of Florida (Fig. 4). Using ESRI and FDOT road networks, we
destination (OD) point to represent each area, considering included all interstate highways, toll roads, and US highways,
the locations of major intersections and traffic generators as well as selected state highways important for intercity
(for other TAZ aggregation in flow-intercepting models, see trips. We reduced the speeds on all roads other than limited-
[46–48]). The OD points are a subset of the larger set of road access highways by 15%. For modeling intercity trips, the basic

Fig. 2 – Sample path from Tallahassee (Leon County) to Clearwater (103011).

6050 international journal of hydrogen energy 34 (2009) 6045–6064

Fig. 3 – Network for the Orlando case study. Orlando’s central business district is near the intersection of I-4 and the East–
West Expressway.

level of spatial aggregation was by county, with some excep- the generalized cost formula and the aggregated populations
tions. Large urban counties consisting of several distinct of i and j into (7). The resulting Tij values were then stan-
urban areas in the Miami, Orlando, and Tampa areas were dardized to a percentage of the statewide total and used as the
further subdivided into separate ODs, and some small rural intercity trip volumes fq. This study does not include out-of-
counties were aggregated together. In all, Florida was aggre- state flows to or from Florida.
gated into 74 ODs. For the base cases for the Orlando and statewide models,
The FDOT trip data used for the Orlando study does not the trip volume data consider trips made by all Orlando and
include statewide travel flows. We therefore used a spatial Florida residents. All consumers, however, are not equally
interaction or gravity model [52,53] to estimate intercity flows likely to purchase hydrogen vehicles in the early stages of
Tij based on zone populations Pi and Pj and a friction function commercialization. Therefore, we also ran the model with
FFij for the route from i to j: demand weighted by a hydrogen demand factor. Our method
was loosely based on a 2006 report by NREL [56]. We used their
Tij ¼ Pi  Pj  FFij (7) variables, relative weights, and 7-point scales whenever
possible, but adapted their method to estimate per capita
We used a friction function for intercity home-based social-
demand factors that could be used as multipliers on the fq
recreational and vacation trips from Michigan’s statewide
vehicle flow total. Based on Table 1, we calculated a weighted
travel forecasting model [54], cited by a Federal Highways
average hydrogen demand score on a scale from 1 to 7 for each
Administration guidebook [55] as an example of best practice
aggregated TAZ (see base layer in Figs. 3 and 4). Then, after
in statewide travel forecasting:
linear conversion to a multiplier between 0 and 1, we
computed a weighted average multiplier for each origin–
FFij ¼ 50  GC0:114
ij  e0:03GCij (8)
destination pair and multiplied the number of origin–desti-
where GCij is the generalized cost between zones i and j, nation trips by it. Given that ‘‘there is no single best data
calculated as 0.75  miles þ 0.5  minutes. To estimate trip classification method,’’ and that NREL modified the number of
volumes, we applied equations in (7) and (8) to each OD pair, classes and weights in a subsequent publication [10, p. 4] these
plugging the travel time and length of the shortest path into scenarios should be interpreted mainly as a way to assess the
 international journal of hydrogen energy 34 (2009) 6045–6064 6051

Fig. 4 – Network used for the statewide case study.

sensitivity of the optimal locations to assumptions about 50-mile spacing as a proxy for capturing shorter trips because
demand levels. the model uses data for both short and long intercity trips. If
The base case scenario for both statewide and Orlando case capturing short trips that would otherwise fall between
studies uses 100 miles as a reasonable or ‘‘safe’’ vehicle range. stations justifies spacing stations closer than 100 miles apart,
This number represents a distance drivers would feel the model will recognize that. We also ran scenarios assuming
comfortable traveling between stations on a round trip, rather a 50-mile and 75-mile range.
than the technological maximum range of the vehicle,
currently around 180–300 miles for today’s prototypes. The
NRC [2] assumes a 300-mile driving range for their infra- 6. Florida state-scale results
structure scenarios, and the US DOE has set 300 miles as the
goal for 2015 [26]. The original California Hydrogen Highway We ran a variety of scenarios locating stations optimally for
plan was to space stations every 20 miles along interstates, intercity trips within Florida. Scenarios varied by objective
although their current thinking is closer to 50 miles. The NREL function (trips, VMT, weighted trips, weighted VMT), driving
national analysis spaces stations 100 miles apart west of the range (100, 75, 50 miles), and whether or not stations were
Mississippi River but 50 miles apart in the eastern US and forced into the solution for ‘‘non-model-based’’ reasons. For
urban areas, because interstates in the east are ‘‘used exten- 5 stations, the model locates four in SE Florida and one in
sively for short trips’’ [24]. We adopt NREL’s 100-mile Tampa (Fig. 5). These five stations are capable of refueling
assumption as a reasonable driving range that allows for 62% of the estimated intercity trips in Florida. The large
driver error, suboptimal vehicle performance, improper filling, number of intercity trips in this area of 5.4 million people is
side trips, detours, and closed stations. With the FRLM, the result of many large population nodes with short
however, we do not need to assume NREL’s more conservative distances between them. Clustering stations in the greater
6052 international journal of hydrogen energy 34 (2009) 6045–6064

that are longer than 100 miles round trip. Outside of this
Table 1 – Hydrogen consumer demand scoring and
weighting system. cluster, the other station in Tampa facilitates round trips
between the area’s largest city and some large surrounding
Data layer Spatial Weight
cities such as St. Petersburg, Lakeland, Sarasota, and
units (%)
Clearwater.
Median household income Census tract 23 The optimal system of ten stations for maximizing
Percentage of households with 2 þ vehicles Census tract 23
unweighted trips includes two major clusters: a linear one in
Pct. of workers age 16þ who commute Census tract 18
SE Florida and another connecting the Tampa–St. Petersburg
more than 20 min
Percentage of people with bachelor’s Census tract 18 (2.6 million people) and Orlando (1.9 million) metropolitan
degrees areas. Each cluster now has five stations, and the stations that
Clean cities coalitions County 18 were optimal with only five stations remain optimal with 10
(Fig. 5). These 10 stations are able to refuel 77% of the estimated
intercity trips in Florida. With 15 stations, the model adds three
stations to the Orlando–Tampa cluster, one station to south-
Miami-Palm Beach area is therefore a smart strategy for east Florida, and one in Jacksonville (population 1.2 million).
serving many trips with a few stations. In addition, the The stations that were optimal within the first five and the first
linear arrangement of cities along the coast means that 10 remain optimal with 15, which is not always the case with
most intercity trips travel north and south on the main this kind of model, adding confidence to these priorities. As we
highways, and by stringing the stations along the coast, the saw before in the case of Tampa, the single station in down-
arrangement of these four stations allows trips to be made town Jacksonville potentially enables trips from central

Fig. 5 – Optimal 15 stations maximizing trips with 100-mile vehicle range.

 international journal of hydrogen energy 34 (2009) 6045–6064 6053

Jacksonville to the five surrounding counties. These 15 stations number of trips on each shortest path by its distance in order
refuel 84% of estimated intercity trips (Fig. 5). to prioritize serving longer trips that use more fuel. The results
A new strategy emerges when expanding the network to 20 are surprisingly similar (Fig. 6). Twelve of the locations that
stations (not shown). The model begins adding some con- were optimal for maximizing trips remain optimal for maxi-
necting/bridging stations, such as on I-95 at Daytona, Cocoa mizing VMT, while others shift only slightly. There continue
Beach, and Vero Beach, facilitating north–south trips among to be clusters of stations around Miami, Orlando, and Tampa,
nearly every pair of cities from Coral Gables in the south as well as a station in Jacksonville. The priority order,
almost to the Georgia border in the north. Despite enabling however, is different, with one of first five stations shifted to
trips up and down the east coast and further down the west Orlando to make the 170-mile Orlando–Tampa round trip
coast, these five stations only increase the percentage of possible. Despite the increased emphasis on serving longer
refuelable intercity trips from 84% to 89%, because people trips, however, there would still not be enough VMT served in
make fewer long-distance trips than short-distance trips. the Panhandle to justify placement of scarce refueling infra-
With 25 stations, the model adds stations as far north as structure there.
Gainesville (University of Florida main campus), and adds to Two examples highlight the differences in the trips and
the clusters in the Orlando, Tampa, and Miami areas (Fig. 6). VMT solutions. The VMT model locates a station on I-75
Note the lack of stations in the Florida Panhandle, and the fact crossing the Everglades, because of the higher weight placed
that some stations that were optimal with 5–15 stations are no on long-distance trips between the Miami and Tampa–St.
longer optimal with 25. Petersburg conurbations. To support this strategy, several
Next, we maximized vehicle-miles traveled (VMT) instead stations in the Miami area shift location slightly in order to be
of the number of trips. The VMT objective multiplies the at key junctions where trips funnel towards I-75 while still

Fig. 6 – Comparison of optimal networks of 25 stations maximizing trips and VMT, for a 100-mile vehicle range.
6054 international journal of hydrogen energy 34 (2009) 6045–6064

being able to serve short intercity trips within southeast to add fewer trips than the previous station, as the best loca-
Florida. The second notable change is that the station in tions are used up. Previous research [8,57] has shown,
Gainesville in the max-trips scenario shifts eastward onto US- however, that returns are not strictly diminishing, because
301, the fastest route between Jacksonville and Tampa. While sometimes it takes two additional stations to be able to refuel
Gainesville has a larger population than this node, a station in a high-volume but longer trip. This is visible in the curve for
Gainesville would serve primarily shorter intercity trips, maximizing trips assuming a 50-mile range, although it is also
whereas the station on US-301 facilitates fewer but longer possible that the concavity at 7–8 facilities could be examples
trips. of the substitution algorithm being unable to find the global
Fig. 7 presents results for 25 stations maximizing trips, optimal solution. Second, the curves for the shorter range are
assuming driving ranges of 50 miles compared with 100 miles. lower than the curves for the longer range, simply due to the
With a 50-mile range, the optimal network of 25 stations need for spacing stations closer together to serve the same
breaks up into two independent clusters in the greater Miami long-distance trips. Third, a given number of stations can
and Tampa–Orlando areas, with a single station in Jackson- potentially refuel a higher percentage of trips than of VMT
ville. With the shorter range, more stations are needed to because the latter consists more of long trips that require
provide connectivity in the two main clusters, leaving too few more stations to serve. Note that the percentage of trips or
stations to provide linkages between these clusters. VMT enabled is to some extent an artifact of the structure of
Fig. 8 shows tradeoff curves for both the max-trips and the network constructed by the analysts. With a more-
max-VMT objectives for driving ranges of 50, 75, and 100 detailed network and more origin and destination nodes, the
miles. First, as expected, all curves show a general pattern of same set of stations would likely intercept a smaller
diminishing marginal returns. Each subsequent station tends percentage of intercity routes.

Fig. 7 – Optimal 25 stations with 100-mile and 50-mile driving range assumptions.
 international journal of hydrogen energy 34 (2009) 6045–6064 6055

100% university adopters and the possible tourist rental-car

90% market in the Orlando area may more than make up for the
loss of 1% of intercity travelers. In any case, 17 stations in the
80%
maximizing trips scenario and 10 stations in the maximizing
70%
Percent Refueled

VMT scenario remain in exactly the same locations, showing

60% that these are robust locations regardless of whether the
50%
rental-car stations and university stations are forced into the
Max Trips (100 mi) top 25.
40% Max Trips (75 mi) Finally, we ran a number of infrastructure scenarios using
30% Max Trips (50 mi)
the weighted hydrogen consumer demand factors based on
Max VMT (100 mi)
20% NREL’s GIS model, with and without forcing the three rental
Max VMT (75 mi)
10% Max VMT (50 mi)
stations and two university stations into the solution, and
maximizing weighted trips or weighted VMT. All weighted
0%
0 5 10 15 20 25
demand scenarios assumed a 100-mile driving range. In the
weighted demand scenarios, the Miami-Palm Beach and
Number of Stations
Orlando metropolitan areas are predicted to have higher
Fig. 8 – Tradeoff between number of stations and percent of consumer demand than the Tampa and Jacksonville areas
demand able to be refueled for the statewide network. mainly because they have Clean Cities coalitions to facilitate
adoption, but also because of demographics and longer
commutes. As a result, there is slightly more clustering in
The station networks in the previous scenarios were opti- these two urban areas and earlier connecting stations
mized with no a priori assumptions about any particular station between them. Overall, however, many locations remain the
being part of the network. There may, however, be certain same, which helps lead us to a robust final set of priorities for
locations that decision-makers wish to include for tourism, the initial hydrogen-refueling infrastructure in Florida.
political, or scientific reasons, even though the model based on
intercity estimated trip demand does not select them. In
another part of the FHI study [18], we found that the bundles of 7. Orlando metropolitan-scale results
trips made by approximately 64% of car renters at Orlando
International Airport (OIA) could be serviced by three stations: The Orlando area was chosen as the study area for analysis at
at OIA (station already exists), downtown Orlando, and the the metropolitan scale because it was the focus of our feasi-
theme parks. Of these, only downtown Orlando was consis- bility study on hydrogen rental cars [18], and is home to
tently chosen by the model based on intercity trips. Many trips Florida’s first two hydrogen stations. We treat the station
to the theme parks and OIA are intra-city trips that are not located at OIA as existing in all scenarios, but the mobile
modeled in the statewide network. Yet these two stations play refueling unit in Oviedo, scheduled for decommissioning in
a key role in a possible hydrogen rental-car business model. late 2008, is not treated as existing in any scenario. The main
Likewise, no scenarios with 25 or fewer stations selected Tal- difference between analysis at the state and local scales is that
lahassee based on the intercity trips to, from, or through it, yet the range of the vehicle becomes a non-factor. Hardly any
a station in Tallahassee may be justified on other legitimate round trips in Orlando exceed the 100-mile safe vehicle range
grounds. As the site of Florida State University, a station in in length. Therefore, one station anywhere on a path can
Tallahassee may be important for scientific research and envi- refuel almost any OD pair.1
ronmental education. Similarly, a station near the state capital The base case for the Orlando network is maximizing trips
in Tallahassee could be important for demonstrating hydrogen based on the FDOT trip table, with only the airport location
technology and for fueling governmental fleets. Finally, locked into place. Fig. 9 shows the optimal locations for 25
a station in Tallahassee may be justified by the intra-city trips of stations, color-coded by priority. We generated these priori-
the highly educated workforce there. Similar considerations ties by first solving the model for five stations, then 10
point to a station in Gainesville near the University of Florida at stations, and so on up to 25 stations. Only two stations are
I-75. For this reason, we ran some scenarios in which these substituted for others during the greedy-substitution algo-
three rental-car stations and two university stations are forced rithm, and only one is not part of the final optimal solution for
into the solution, and the rest of the station network was opti- 25 stations (see Fig. 9 caption for explanation). The fact that
mized around these five fixed stations. these are the only two substitutions suggests that if stations
Forcing the placement of these five stations involves
a large sacrifice in terms of the number of intercity trips 1
For those few round trips longer than 100 miles, the path is
serveddbut only in the early stages of infrastructure devel- likely to follow a major highway, and a single station anywhere
opment. Alone, these five stations serve only 7% of intercity near the middle of the path could easily serve the trip by refueling
trips estimated by the gravity model, versus 65% for the five it in both directions. For a round trip of, say, 110 miles (the longest
in our Orlando network), the only way that a single station would
stations optimally located in southeast Florida and Tampa.
not be able to refuel the vehicle before it reached the 100-mile
However, adding 20 optimal stations in coordination with
mark is if the station were located within five miles of either the
the five fixed stations, it is possible to serve 94% of the esti- origin or destination. For applications such as these where the
mated intercity trips, only 1% less than if all 25 stations were vehicle range exceeds the longest round trip distance, the FRLM
optimized freely. The un-modeled benefits of early reduces to the simpler flow-capturing model.
6056 international journal of hydrogen energy 34 (2009) 6045–6064

Fig. 9 – Optimal 25 stations in the Orlando area maximizing trips with 100-mile vehicle range. The color shown for each
location indicates the highest priority level at which it was ever selected. The reason why six new stations are shown as
optimal for p [ 11–15 and p [ 16–20, and only four for p [ 21–25, is due to substitutions during the greedy adding-with-
substitution algorithm. Station A was optimal for 6–10 facilities, but shifts its location less than a mile south to B for 11–15
stations and remains there for 16–20 and 21–25 facilities. Meanwhile, Station C, at a freeway exit east of the theme parks, is
shown as first appearing in the optimal solution for 6–10 stations. It remains optimal for 11–15 stations, is replaced by
a different site for 16–20, and re-emerges as optimal for 21–25 stations. The station that replaced C in the 16–20 solution
remains optimal for 21–25.

are developed in this order, future stations are not likely to While these freeway sites are on high-traffic roads, they are
cannibalize the demand of earlier stations. After the existing not necessarily at the next highest traffic sites in the network.
station at OIA, the best location is in downtown Orlando, near What is more important is that they intercept high-traffic
the intersection of I-4 and the East–West Expressway, which flows that were not captured by the previously chosen
could intercept and potentially refuel 14% of the daily trips. stations. The freeway stations in the top 10 are located
Of the eight next best station locations, six are on major a substantial distance away from each other. A certain
freeways with high passing traffic volumes as well as high amount of cannibalization of demand is inevitable, but the
crossing traffic flows and/or originating and ending flows. FRLM tries to maximize the unique flow volume that can be
 international journal of hydrogen energy 34 (2009) 6045–6064 6057

refueled with each additional station, and does not consider arrive at a prioritized list of stations coordinated across the
any benefits of redundancy. The model also places two of the two scales of analysis. This approach recognizes that the
top 10 stations at heavy-traffic intersections of major arterial FRLM is capable of analyzing tens of thousands of origin–
streets. destination pairs and sorting through trillions of possible
As important as where the model locates stations is where station combinations in ways that are impossible without the
the model does not locate stations. In contrast with other benefits of OR and GIS. At the same time, we recognize the
methods such as the p-median model, the FRLM does not limitations of the model and the datasets, and the importance
necessarily try to spread the locations evenly around metro of factors that are not included in the models. In addition, no
Orlando to minimize the average distance from residential single scenario can determine the best station network. The
zones to their nearest stations. Many OD nodes remain far final factor dictating this approach is that the Orlando trip
from any of the top 10 sites, and large areas appear to be under- table was based on FDOT data while the statewide trip table
served; however, appearances can be deceiving. Many trips was based on gravity model estimates, which meant that the
may start from neighborhoods without a station but pass by Orlando and statewide refueling networks had to be analyzed
stations in other areas. In fact, 53% of trips pass through at separately and coordinated exogenously.
least one of the top 10 station sites. For this reason, we looked for station locations that
After these first 10 stations, the subsequent sets of five add consistently perform well across a variety of scenarios. The
13%, 9%, and 7% as diminishing returns set in. Not surpris- most robust locations can coordinate well with a variety of
ingly, fewer ‘‘freeway funnel’’ points remain to intercept large other stations in refueling short trips by themselves and
volumes of passing flows that were not already intercepted by longer trips in combination with other stations, regardless of
previous locations. Thus, the model eventually begins to vehicle range, demand weighting, and whether we are maxi-
adopt a different strategy, spreading stations 16–25 around mizing trips or VMT. In devising a strategy for the Orlando
Orlando in more suburban locations on arterial streets. area, we had to take into account that some stations would
There is empirical evidence that consumers prefer to refuel also serve the intercity and rental-car trips in addition to the
near their homes [31,58]. This raises the question whether the intra-Orlando trips. Thus, any station included in the final
FRLM or the p-median model is more behaviorally realistic. statewide network had to also be included in the final Orlando
When presented with a choice of many stations along their network at the same location. Finally, our team of four pro-
driving route, we agree that most drivers would prefer fessorsdwith expertise in transport modeling, geography,
a station near home, all else being equal. Stations near home and tourism and a combined 60 years living and driving in
are more familiar, and by refueling near home drivers can Floridadweighed the importance of various factors not
avoid getting off the freeway to refuel and then getting back included in the model to develop the location strategy pre-
on. Locating stations in each and every neighborhood, sented below.
however, is a luxury that may not be affordable when building Table 2 lists the higher-priority stations and their highest
the first set of hydrogen stations at $1–3 million apiece. For the rank in various scenarios. The full table, including every
initial refueling infrastructure, we argue that it is most station that was optimal in at least one scenario, can be found
important to locate on the routes of as many trips as possible. in [17]. Rows represent stations, columns represent model
With only a handful of stations, we believe that early adopters scenarios, and cells relate whether the station was optimal in
will be willing to refuel far from home as long as the station is the scenario, and if so, in which tier.2 Based on all the model
on their regular commuting or shopping route and does not results, we sort the stations into tiers of overall Top 5, Top 10,
require a detour. A station far from home but on the way may and so on through Top 25 as in Table 2. In ranking them, we
in fact be more convenient than a station that is closer but consider synergies, spacing, timing, and local intra-city
requires a special trip or detour to reach. Stations next to consumer demand. We do not attempt to rank stations within
highways with large passing flow volumes can enable refuel- each tier of five. For the statewide network of intercity trips,
ing by a larger share of potential adopters. we place greater emphasis on the scenarios that include the
We also ran scenarios maximizing VMT, using weighted three stations in Orlando needed for serving rental-car trips
demands based on the strength of consumer demand, and and the two flagship university stations. We also put more
forcing in the stations needed for a rental-car business. emphasis on the weighted demand scenarios and the 100-
Because we already treat the airport station as fixed, and the mile range scenarios. Finally, because our focus in the state-
downtown station is always the first station added, the station wide network is on facilitating long-distance trips, we place
at the center of the theme parks is the only one that needs to more emphasis on maximizing VMT.
be forced in to optimize the refueling network around the Our final priorities for 25 stations are quite close to the
three targeted rental-car stations. ‘‘Max VMT-Range 100-Weighted Demand-Five Required Sites’’
scenario in the last column of Table 2, with the substitution of
Kennedy Space Center for the Shady Hills station, some minor
8. Priorities for Infrastructure in Florida and relocations among nearby sites, and some shuffling of
Orlando stations among the tiers. These 25 higher-priority stations
generally perform very well across the board, coordinate well
In planning an initial hydrogen refueling infrastructure in
Florida, our team of researchers synthesized the results of 2
Some scenarios were solved separately for 5, 10, 15, 20, or 25
a large number of model scenarios for both the statewide and stations, which allowed us to prioritize the stations into Top 5,
Orlando networks using our judgment and local knowledge to Top 10, and so on. Others were solved only for p ¼ 10 and 25.
 6058
Table 2 – Statewide infrastructure strategy.a
Jct City Intersection Max Max Max Max Max Max Trips range Max trips range Max trips range 100 Max VMT range 100
trips VMT trips trips VMT 100 five required 100 weighted weighted demand five weighted demand five
range range range range range sites demand required sitesb required sites
100 100 50b 75c 75b

First tier – top 3 locations

190 Orlando Boggy Creek Rd. Existing Existing Existing
Int’l Airport
183 Orlando I-4 & East West Expwy Top 10 Top 5 Top 25 Top 10 Top 25 Required Rental Required Rental Station Required Rental Station

international journal of hydrogen energy 34 (2009) 6045–6064

Downtown Station
197 Orlando Epcot Ctr Dr. between I-4 & Required Rental Required Rental Station Required Rental Station
Theme World Dr. Station
Parks

Second tier – top 10 locations

242 Delray I-95 & W Atlantic Ave Top 5 Top 15 Top 25 Top 10 Top 15
Beach
256 Ft I-95 & Sunrise Blvd Top 5 Top 5 * Top 10 Top 10
Lauderdale
266 Miami I-75 & Palmetto Expwy Top 5 Top 10 Top 10 Top 10 Top 5 Top 25 Top 10
Lakes
107 Tampa I-4 & I-275 Top 5 Top 5 Top 25 Top 25 Top 10 Top 5 Top 25 Top 10
155 Gainesville I-75 & W Newberry Rd Top 25 Required Top 25 Required University Station Required University Station
University Station
163 Wildwood I-75 & Florida’s Tpke Top 15 Top 10 Top 25 Top 25 Top 20 Top 20 Top 25 Top 15
292 Tallahassee Tennessee St & Monroe St Required Univ./ Required Univ./Govt. Required Univ./Govt.
Govt. Station Station Station

Third tier – top 15 locations

109 Mango I-4 & I-75 Top 25 Top 25 Top 15 Top 25 Top 15
270 Miami I-95 & I-195 Top 5 Top 5 Top 25 Top 25 Top 10 Top 10
169 Sanford S Orlando Dr & Eastern Top 10 Top 15 * * Top 15 Top 10 Top 25 Top 20
Beltway
209 St. I-275 & 5th Ave N Top 10 Top 10 Top 25 Top 10 Top 25 Top 15 Top 10 Top 25 Top 15
Petersburg
Kennedy Visitor’s center, not included
Space in statewide network.
Center

Fourth tier – top 20 locations

93 Daytona I-95 & I-4 Top 20 Top 25 Top 25 Top 25 Top 25 Top 25 Top 25
Beach
217 Fort Pierce St Hwy 70 between I-95 and Top 25 Top 25 Top 15 Top 25
Florida’s Turnpike
82 Jacksonville I-95 & I-10 Top 15 Top 25 Top 25 Top 25 Top 15 Top 15 Top 25 Top 25
149 June Park I-95 & Coast Hwy Top 15 Top 20 Top 15 Top 25 Top 15
(near
Melbourne)
232 Palm Beach I-95 & PGA Blvd Top 10 Top 15 Top 25 Top 25
Gardens
 international journal of hydrogen energy 34 (2009) 6045–6064 6059

with each other for serving longer distance trips, and canni-

a Indicates that another location within the same city was in the Top 25 for the scenario in question. The other location(s) can be found further down in the table. Does not apply to large cities such as
balize each other’s demand as little as possible.
As can be seen in Fig. 10, we propose developing clusters of
stations and connecting stations along major interstates in
stages, so that as each tier is constructed, the clusters and the
* connections between them grow in a coordinated way. The
first tier of three stations consists of the airport, downtown,
Top 25

Top 20

Top 20
and theme park stations in Orlando needed for the hydrogen
rental-car business. Given Orlando’s head start in the
hydrogen industry with its existing station at the airport, we
see it as the key to getting hydrogen moving in Florida. The
second tier of seven stations rounds out the Top 10. It creates
a hydrogen corridor from Miami Lakes to Ft. Lauderdale to
Delray Beach, as well as a connected triangle between Tampa,
*

Orlando, and Gainesville. The third tier fleshes out the

Orlando, Tampa, and Miami clusters. The fourth tier
Top 25
Top 25

completes the network up I-95 from Palm Beach Gardens to

Jacksonville. The fifth tier extends the I-75 network north and
south of Tampa–St. Petersburg, as well as adding to the Miami
and Orlando clusters and shortening the distance between
stations on Florida’s Turnpike. Fig. 10 shows that the spacing
*

of the connecting stations between the clusters along I-95, I-4,

Top 15

and Florida’s Turnpike never exceed 100 miles.

We ranked the two university stations in the second tier.
The Gainesville station was actually optimal in a number of
scenarios, based on intercity trips to and from Gainesville and
along I-75. It becomes a higher priority considering the
hydrogen research program at the University of Florida, as
Top 25
Top 25
Top 25

Top 20

well as university vehicle fleets and the greater likelihood of

staff and students as early adopters. Similar arguments apply
to Florida State University in Tallahassee, minus the through-
Top 25

Top 25

traffic benefits but plus the benefits to state government fleets

*
*

and for purposes of political demonstrations and leadership. A

c Stations in this column are either Top 10 or Top 25. Other p values were not analyzed.
b Stations in these columns are not ranked. Only one scenario (for p ¼ 5) was analyzed.

number of other universities around the US have hydrogen

Top 25

Top 25
Top 25

stations (e.g., Penn State, UC-Davis, UCLA, UC-Irvine, UC-

Riverside, and CSU-LA).
*

Miami, Orlando, and Jacksonville, where different locations may be far apart.

Kennedy Space Center is placed in the third tier, even

Top 25
Top 25
Top 25

though we assume the station would be at the NASA Visitor

Center. To be conservative, we assumed it is too far out of the
way to refuel trips along I-95. Nevertheless, we include it in
Top 25
*

the third tier because of its importance to the proposed rental-

car business, its attractiveness as a tourist destination to
technophiles, its use of hydrogen for rockets, and the avail-
Florida’s Tpke & County Line Top 15

Top 15
*

ability of hydrogen for a station. Daytona is another location

of particular interest because of its car-racing heritage.
Although it was rarely chosen in the Top 20, it was consis-
Florida’s Tpke and Orange
201 Kissimmee Osceola Parkway between

tently in the Top 25 and we moved it up to the 4th tier in order

to phase in the I-95 connecting stations as a group. About 5%
of car renters included it in their bundle of trips with the
I-75 & Duncan Rd
I-75 & St Hwy 52
I-75 & St Hwy 72

Orlando theme parks and/or downtown Orlando, and it is also

Blossom Trail

a popular weekend destination from Orlando.

In prioritizing stations for Orlando intra-city traffic flows,
Fifth tier – top 25 locations

we consider scenarios using the likely-consumer-demand

Rd

multipliers to be the most important. Secondly, we consider

maximizing trips more important than maximizing VMT for
263 Miramar

81 Sarasota

the urban network. In the urban network, in contrast to the

50 Solana
98 Pasco

statewide network, it is actually easier to refuel the longer

trips, because the long trips are more likely to use freeways
and more likely to encounter at least one station somewhere
along the route. Our emphasis, therefore, is on refueling as
6060 international journal of hydrogen energy 34 (2009) 6045–6064

Fig. 10 – Statewide refueling infrastructure strategy.

many trips as possible, both short and long. An initial network The three stations needed for the rental business are joined in
of 11 hydrogen-refueling stations prioritized into two tiers the Top 5 by stations at funnel points of the network that
could refuel about 54% of weighted trips (Fig. 11; Table 3). capture many trips not otherwise refueled by the downtown,
Development in coordinated stages is less important than in airport, and theme-park stations. We propose placing the
the statewide network because distances within the metro theme park station near I-4 so it would be accessible to
area are not long enough to require multiple refuelings on any through traffic for the statewide network. Detailed explana-
given trip. The stations are grouped into tiers mainly accord- tions for the other high-priority stations and a full list of all
ing to their potential to add to the total trips that can be stations that were optimal in any run can be found in [17].
refueled. Five of the stations are also important in the state-
wide network.
Compared with the statewide network, there is more 9. Conclusions and future research
consistency across Orlando scenarios, leading to a robust plan
with reduced uncertainty. Eleven stations are included Having already summarized the station location priorities for
because there are 10 that capture substantial amounts of the Florida statewide network and Orlando metropolitan
traffic and/or play key roles in the rental-car business model, travel, a number of general policy conclusions can be drawn
and an 11th station is needed at a key highway intersection for from these analyses that may be useful for other regions in
the statewide network. Eight of the 11 stations are consis- their transition to hydrogen vehicles. First, although the FRLM
tently in the Top 10 in every scenario and are robust choices. does not assume either a clustering or bridging strategy, the
 international journal of hydrogen energy 34 (2009) 6045–6064 6061

Fig. 11 – Orlando refueling infrastructure strategy.

results appear to suggest a strategy somewhat similar to critical component of the strategy considering the distance,
California, beginning with clusters in the major cities and later trip volumes, and network structure of the region in question.
building bridging stations to facilitate trips between urban It is important to place bridging stations where they can also
regions. serve crossing traffic flows and local traffic flows.
The initial clusters in and near large metropolitan areas Careful thought and planning must be given to the spacing
can refuel the high trip volumes between heavily populated of bridging stations. Spacing stations too far apart could lead
nodes that are close together. Clustering also enables the to emergency situations and stranding of vehicles that could
stations to work together to refuel medium-length trips that compromise safety and generate bad publicity. On the other
require multiple stations along the travel route. hand, spacing stations too closely sacrifices coverage of trips
There is no consistent formula or guideline to determine elsewhere and could lead to duplication and underutilization.
when bridging stations should be added to connect the clus- Our choice of a 100-mile driving range is consistent with
ters. Given the high cost of building hydrogen stations, it is NREL’s assumptions, and should be adequate for dealing with
important to use a model such as the FRLM to determine this detours, getting lost, incomplete refueling, and station
6062 international journal of hydrogen energy 34 (2009) 6045–6064

Table 3 – Orlando infrastructure strategy.

Jct Intersection Place Max Max Trips Max Trips Max VMT
Trips Airport þ Disney Airport þ Disney Airport þ Disney
Airport weighted demand weighted demand

First tier – top 5

29 Aloma Ave & Semoran Blvd East of Winter Top 5 Top 5 Top 5 Top 10
Park
144 Epcot Ctr Dr between I-4 & World Theme Parks Top 20 Required Rental Required Rental Station Required Rental Station
Dr Station
176 I-4 & W Maitland Blvd Maitland Top 5 Top 5 Top 5 Top 10
202 I-4 & East-West Expwy Downtown Top 5 Top 5 Top 5 Top 5
305 S Access Rd & Airport Blvd Orlando Existing Existing Existing Existing
International
Airport

Second tier – top 11 stations

66 Orange Blossom Trail & N John Northwest Top 10 Top 10 Top 10 Top 10
Young Pky Orlando
126 S Orange Blossom Trail & Sand Florida Mall Top 15 Top 15 Top 15 Top 20
Lake Rd
148 S Orlando Dr between Cent Sanford - Lake Top 20 Top 20 Top 15 Top 25
Florida Greenway & Lake Mary Mary
Blvd
168 I-4 & W Lake Mary Blvd Heathrow Top 10 Top 10 Top 10 Top 5
180 Cent Florida Greenway & E Union Park Top 10 Top 10 Top 10/25 Top 10/15/25
Colonial Dr
210 Osceola Pky between Florida’s Kissimmee Top 10 Top 10 Top 10 Top 20
Tpke & Orange Blossom Trail

closures. A smaller driving range, such as 50 miles, would only truly be near a small fraction of residents’ homes, but can
lower the percentage of intercity trips and VMT that can be be directly en route for over half of all trips.
refueled by a given number of stations, perhaps unneces- Not all residents are equally likely to purchase hydrogen
sarily. It would also further cluster the optimal 25 stations and vehicles when they become available. NREL has estimated
reduce coverage for other areas. Whatever the driving range, it geographic differences in consumer demand, and market
should be enforced as a maximum spacing only. Strict regular analysis may reveal certain neighborhoods as ‘‘hot’’ markets
spacing will waste scarce resources. for alternative-fuel vehicles. Regardless of whether a model
To get the most benefit out of the initial infrastructure uses flow demands like the FRLM or point demands like the p-
rollout, it is important to locate stations at funnel points on median, some scenarios should be conducted using weighted
the road network through which many trips pass, from many demands reflecting consumer preferences.
origins to many destinations. This is best achieved by locating Finally, we recognize that factors not included in the
many of the first stations on major freeways with high model’s data set should be taken into consideration. In the
volumes of passing traffic, where they intersect with other case of Florida, unique synergistic opportunities exist with
freeways or major arterials with high volumes of crossing, a possible OIA rental-car business, Florida’s flagship univer-
originating, or ending traffic. It is equally important that these sities, Kennedy Space Center and Disney, and the state capital
funnel points duplicate or cannibalize each other as little as in Tallahassee. Other regions will offer their own unique
possible. Thus, stations on major freeways should not be too opportunities that are not reflected in their trip tables, and
close together. At the intra-metropolitan scale, the highest thus it is important that the decision support system be able to
priority locations will not necessarily be spread evenly across solve a variety of scenarios quickly and display the results
the network to minimize average distance from where people visually. Tradeoff curves based on model results can help
live to their nearest stations. As more stations are added, decision-makers visualize the diminishing marginal returns
however, optimal locations are increasingly spread around to typical of these kinds of problems, but the number of stations
smaller and smaller funnel points in suburban areas, and the to build is ultimately a political decision.
overall network gradually begins to resemble an evenly spread Additional research could improve on these results in
distribution that would minimize average distance to stations a number of ways. To begin with, a more detailed analysis of
but favor freeway locations. the Miami and Tampa metropolitan areas, similar to the
While other empirical research has shown that consumers Orlando network analysis, should be performed. In the
tend to refuel their conventional vehicles near their homes, statewide network, the Miami-Palm Beach area is represented
we caution against concluding that the initial set of hydrogen by nine OD nodes, while the Tampa area stretching from
stations should be located according to that principle. We Sarasota to Pasco County to Lakeland is represented by eight
believe it is more important to locate the early set of stations nodes, compared with over 100 nodes in the Orlando intra-city
along the routes people travel rather than locate them near network. Modeling these two metropolitan areas at the same
their homes, especially because the first 10 or so stations can level of detail as Orlando was beyond the limited scope of this
 international journal of hydrogen energy 34 (2009) 6045–6064 6063

project. Future research could also incorporate out-of-state [7] Nicholas MA, Handy SL, Sperling D. Using Geographic
trips, which were omitted from these data, limiting the traffic Information Systems to evaluate siting and networks of
in the Panhandle. hydrogen stations. Transp Res Rec 1880;2004:126–34.
[8] Kuby MJ, Lim S. The flow-refueling location problem for
We are working on several enhancements to the FRLM
alternative-fuel vehicles. Socio Econ Plann Sci 2005;39:
model. First, the model could allow drivers to detour from 125–45.
their shortest paths up to some limit or with decreasing [9] Nicholas MA, Ogden J. Detailed analysis of urban station
willingness, as it has been incorporated in flow-intercepting siting for California hydrogen highway network. Transp Res
models [37]. Second, models need to account for the different Rec 1983;2006:121–8.
fixed cost of different sites, based on land prices and co- [10] Melendez M, Milbrandt A. Regional consumer hydrogen
demand and optimal hydrogen refueling station siting.
location possibilities with fuel depots, hydrogen production
Golden, CO: National Renewable Energy Laboratory; 2008.
facilities, and other alt-fuel stations [24]. Third, the FRLM and
Report No.: NREL/TP-540-42224.
the p-median models could be combined into a multiobjective [11] Melaina M, Bremson J. Refueling availability for alternative
model to identify solutions in which stations can refuel many fuel vehicle markets: sufficient urban station coverage.
paths and minimize distance from stations to residential areas Energy Policy 2008;36(8):3223–31.
[36]. Fourth, we need to develop consistent trip tables to [12] Lin Z, Ogden J, Fan Y, Chen C. The fuel-travel-back approach
optimize local-scale and state-scale networks simultaneously. to hydrogen station siting. Int J Hydrogen Energy 2008;33(12):
3096–101.
[13] Bersani C, Minciardi R, Sacile R, Trasforini E. Network
planning of fuelling service stations in a near-term
competitive scenario of the hydrogen economy. Socio Econ
Acknowledgements Plann Sci 2009;43(1):55–71.
[14] Upchurch C, Kuby MJ, Lim S. A capacitated model for location
The Florida Hydrogen Initiative (FHI Agreement No. 2005-01) of alternative-fuel stations. Geog Anal 2009;41:127–48.
[15] Conzelmann G: Cirillo R, Mintz M, Macal C, North M. In:
and the US Department of Energy (Grant Award No. DE-FC36-
Modeling the evolution of a new hydrogen infrastructure
04G014225) provided funding for this project. Prior to that, NSF using a complex adaptive using a complex adaptive systems
supported the initial development of the FRLM under NSF approach. Proceedings of the National Hydrogen Association,
Grant No. 0214630. Dash Ltd. provided Xpress-MP software Los Angeles, CA; April 26–30 2004.
under their Academic Partnership Program. We wish to thank [16] Ogden JM. Developing an infrastructure for hydrogen
Steve Adams, Ed Levine, and Pam Portwood, who have served vehicles: a southern California case study. Int J Hydrogen
Energy 1999;24(8):709–30.
as Executive Directors of FHI, as well as John Masiello of
[17] Lines L, Kuby MJ, Schultz R, Xie Z. Hydrogen refueling
Progress Energy in Orlando; Herman Everett of NASA; Ray
infrastructure and rental car strategies for
Hobbs of Arizona Public Service; Gene Nemanich, former commercialization of hydrogen in Florida, Final report to
president of the National Hydrogen Association; Yongqiang Florida Hydrogen Initiative. US Dept of Energy; June 30, 2007.
Wu of Florida DOT; Roberto Miquel of Cambridge Systematics FHI Agreement No. 2005-01, Grant Award No. DE-FC36-
in Tallahassee; Dennis Hooker of MetroPlan Orlando; Margo 04G014225.
Melendez and Anelia Milbrandt of the National Renewable [18] Lines L, Kuby MJ, Schultz R, Clancy J, Xie Z. A rental car
strategy for commercialization of hydrogen in Florida. Int
Energy Laboratory; Karen Faussett of the Michigan DOT; Nicole
J Hydrogen Energy 2008;33(20):5312–25.
Barber of Florida DEP; M. John Hodgson and Weiping Zeng of
[19] California hydrogen blueprint plan, vols. 1 and 2. California
University of Alberta; and Max Wyman, of Terragenesis, Inc. Environmental Protection Agency. Available from: www.
hydrogenhighway.ca.gov/plan/plan.htm; May 2005 [Last
accessed March 31, 2007].
[20] Worldwide hydrogen fueling stations. Fuel Cells. Available
references from: www.fuelcells.org/info/charts/h2fuelingstations.pdf;
2000 [Last accessed July 24, 2008].
[21] Huleatt-James N. Hydrogen infrastructure survey. 2008;
[1] US Dept of Energy. National hydrogen energy roadmap. Available from: www.fuelcelltoday.com/media/pdf/surveys/
Washington, DC: US Dept of Energy; 2002. 2008-Infrastructure-Free.pdf [Last accessed July 24, 2008].
[2] National Research Council. The hydrogen economy: [22] California Fuel Cell Partnership. Driving for the future,
opportunities, costs, barriers, and R&D needs. Washington, Available from: www.fuelcellpartnership.org/ [Last accessed
DC: The National Academies Press; 2004. April 30, 2004].
[3] Melendez M. Transitioning to a hydrogen future: learning [23] California Fuel Cell Partnership. 2004–2007 plans: the next four
from the alternative fuels experience. Golden, CO: National years. Available from: www.fuelcellpartnership.org/fact_
Renewable Energy Laboratory; 2006. Technical Report No. sheets/factsheet_04-07.html [Last accessed March 31, 2007].
NREL/TP-540-39423. [24] Melendez M, Milbrandt A. Analysis of the hydrogen
[4] Ogden JM. Prospects for building a hydrogen energy infrastructure needed to enable commercial introduction of
infrastructure. Ann Rev Energy Environ 1999;24:227–79. hydrogen-fueled vehicles. National Renewable Energy
[5] Goodchild MF, Noronha VT. Location–allocation and Laboratory; 2005. Report No. CP-540-37903.
impulsive shopping: the case of gasoline retailing. In: Ghosh A, [25] Greene DL, Leiby PN, James B, Perez J, Melendez M,
Rushton G, editors. Spatial analysis and location–allocation Milbrandt A, et al. Analysis of the transition to hydrogen
models. New York: van Nostrand Reinhold; 1987. p. 121–36. fuel cell vehicles and the potential hydrogen energy
[6] Bapna R, Thakur LS, Nair SK. Infrastructure development for infrastructure requirements. Oak Ridge, TN: Oak Ridge
conversion to environmentally friendly fuel. Eur J Oper Res National Laboratory; March 2008. Report No. ORNL/TM-
2002;142(3):480–96. 2008/30.
6064 international journal of hydrogen energy 34 (2009) 6045–6064

[26] Hydrogen, fuel cells & infrastructure technologies program: [41] Hodgson MJ, Rosing KE, Zhang J. Locating vehicle inspection
multi-year research, development and demonstration plan. stations to protect a transportation network. Geogr Anal
US Dept of Energy, Office of Energy Efficiency and Renewable 1996;28:299–314.
Energy; 2007. [42] Berman O. Deterministic flow-demand location problems.
[27] Adamson KA, Crawley G. Fuel cell today, 2007 automotive J Oper Res Soc 1997;48(1):75–81.
infrastructure survey. Hertfordshire, UK: Royston. Available [43] Berman O, Krass D. Flow intercepting spatial interaction
from: www.fuelcelltoday.com/FuelCellToday/FCTFiles/ model: a new approach to optimal location of competitive
FCTArticleFiles/Article_1187_Automotive%20Infrastructure. facilities. Location Sci 1998;6(1–4):41–65.
pdf; 2007. [44] Wu TH, Lin JN. Solving the competitive discretionary service
[28] Melaina MW. Initiating hydrogen infrastructures: facility location problem. Eur J Oper Res 2003;144(2):366–78.
preliminary analysis of a sufficient number of initial [45] Yang J, Zhang M, Chen X. Class of the flow capturing location-
hydrogen stations in the US. Int J Hydrogen Energy 2003; allocation model with service radius. Xitong Gongcheng Lilun
28(7):743–55. yu Shijian/Syst Eng Theory Practice 2006;26(1):117–22.
[29] Melaina MW. Turn of the century refueling: a review of [46] Zeng W. GIS based facility location planning with different
innovations in early gasoline refueling methods and types of consumers. J Dissert 2007;1(2).
analogies for hydrogen. Energy Policy 2007;35(10):4919–34. [47] Zeng W, Hodgson MJ, Castillo I. The pickup problem:
[30] Melaina MW. Estimating relative station sizes in early consumers’ locational preferences in flow interception.
hydrogen station networks. In: Proceedings of the National Geogr Anal 2008;41:3–22.
Hydrogen Association, Washington, D.C.; 2005. [48] Zeng W, Hodgson MJ, Castillo I. A generalized model for
[31] Nicholas MA, Handy S, Sperling D. Hydrogen refueling locating facilities on a network with flow-based demand.
network analysis using Geographic Information Systems. In: Netw Spat Econ; doi: 10.1007/s11067-008-9073-8.
Proceedings of the National Hydrogen Association, Los [49] Kuby M, Lim S. Location of alternative-fuel stations using the
Angeles, CA; April 26–30 2004. Flow-Refueling Location Model and dispersion of candidate
[32] Hakimi SL. Optimum locations of switching centres and the sites on arcs. Netw Spat Econ 2007;7(2):129–52.
absolute centres and medians of a graph. Oper Res 1964;12: [50] Daskin MS. Network and discrete location analysis. New
450–9. York: John Wiley and Sons; 1995.
[33] Revelle CS, Swain R. Central facilities location. Geogr Anal [51] Lim S, Kuby M. Heuristic algorithms for siting alternative-
1970;2:30–42. fuel stations using the Flow-Refueling Location Model,
[34] Hodgson MJ. A flow capturing location allocation model. Unpublished results (under review); 2008.
Geogr Anal 1990;22:270–9. [52] Black WR. Transportation: A geographical analysis. New
[35] Berman O, Larson RC, Fouska N. Optimal location of York: Guilford; 2003.
discretionary service facilities. Transp Sci 1992;26:201–11. [53] Hanson S, Giuliano G. In: The geography of urban
[36] Hodgson MJ, Rosing KE. A network location–allocation model transportation. 3rd ed. New York: Guilford; 2004.
trading off flow capturing and p-median objectives. Ann [54] Statewide travel demand model: update and calibration,
Oper Res 1992;40(1-4):247–60. phase II. Ann Arbor, State of Michigan: KJS Associates; 1996.
[37] Berman O, Hodgson MJ, Krass D. Flow intercepting models. [55] Center for Urban Transportation Studies, University of
In: Drezner Z, editor. Facility location: a survey of Wisconsin–Milwaukee. Guidebook on statewide travel
applications and methods. New York: Springer; 1995. forecasting. Washington, DC: Federal Highways
p. 389–426. Administration, US Dept of Transportation; 1999.
[38] Berman O, Krass D, Chen WX. Locating flow-intercepting [56] Melendez M, Milbrandt A. Geographically based hydrogen
facilities: new approaches and results. Ann Oper Res 1995;60: consumer demand and infrastructure analysis. Golden, CO:
121–43. National Renewable Energy Laboratory; 2006. Report No.:
[39] Hodgson MJ, Rosing KE, Leontien A, Storrier G. Applying the NREL/TP-540-40373.
flow-capturing location–allocation model to an authentic [57] Kuby MJ, Lim S, Wang K. A model for optimal location of
network: Edmonton, Canada. Eur J Oper Res 1996;90(3): hydrogen refueling stations: an Arizona case study. In:
427–43. Proceedings of the National Hydrogen Association, Los
[40] Averbakh I, Berman O. Locating flow-capturing units on Angeles; April 26–30, 2004.
a network with multi-counting and diminishing returns to [58] Kitamura R, Sperling D. Refueling behavior of automobile
scale. Eur J Oper Res 1996;91(3):495–506. drivers. Transp Res A Pol 1987;21A(3):235–45.