Journal of Transport Geography 18 (2010) 750–758

Contents lists available at ScienceDirect

Journal of Transport Geography

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

Comparing the p-median and flow-refueling models for locating

alternative-fuel stations
Christopher Upchurch a,*, Michael Kuby b,1
a
Department of Geography, University of South Carolina, 709 Bull St., Columbia, SC 29208, USA
b
School of Geographical Sciences and Urban Planning, Arizona State University, Tempe, AZ 85287, USA

a r t i c l e i n f o a b s t r a c t

Keywords: The p-median and flow-refueling models are two of the more popular models for optimal location of
Refueling facilities alternative-fuel stations. The p-median model, one of the most widely used location models of any kind,
Optimal location locates p facilities and allocates demand nodes to them to minimize total weighted distance traveled. In
Electric vehicle comparison, the flow-refueling location model (FRLM) is a path-based demand model that locates p sta-
Hydrogen
tions to maximize the number of trips on their shortest paths that can be refueled. For a path to be con-
Biofuel
Natural gas
sidered refuelable, one or more stations must be located on the path in a way that allows the round trip to
be completed without running out of fuel, given the vehicle driving range. In this paper, we analyze how
well the facilities located by each model perform on the other’s objective function on road networks in
Florida. While each objective function degrades somewhat when facilities are located by the other model,
the stations located by the flow-refueling model generally do better on the p-median objective than the
stations located by the p-median model do on the flow-refueling objective. This difference between
the two models is even more pronounced at the state scale than at the metropolitan scale. In addition,
the optimal locations for the FRLM tend to be more much more stable as p increases than those located
by the p-median model.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction
Over the next several decades, as countries transition from gas-
oline and diesel to alternative fuels, they will need to invest heavily
in new refueling station infrastructure. Many studies have empha-
sized the critical role that refueling stations play in facilitating the
development of alternative fuels (California Environmental Protec-
tion Agency, 2005; Greene et al., 2008; Huleatt-James, 2008;
Melaina, 2003; Melaina and Bremson, 2008; Ogden, 1999). The
which-comes-first ‘‘chicken-and-egg” dilemma involving alterna-
tive-fuel (alt-fuel) stations and vehicles is widely acknowledged
by researchers and industry representatives alike (Melendez,
2006; National Research Council, 2004; US Department of Energy,
2002). A common strategy for breaking this cycle and building to-
wards necessary economies of scale involves government require-
ments for alternative-fuel fleets for government agencies, utilities,
and other large organizations, plus government subsidies for de-
pot-based fuel stations. However, recent studies have highlighted
the difficulty of transitioning from fleets to consumers (Melendez,
2006). In the early stages of transition, when consumer demand for
* Corresponding author. Tel.: +1 803 777 5867; fax: +1 803 777 4972.
E-mail addresses: chris.upchurch@sc.edu (C. Upchurch), mikekuby@asu.edu (M.
Kuby).
1
Tel.: +1 480 730 8457.
fuel may not be high enough to support private stations without
subsidies, it is especially important that the initial networks of sta-
tions be located in a way that maximizes the potential for consum-
ers to adopt alt-fuel vehicles.
In the literature, several approaches have been used to locate
refueling stations optimally. One group of studies has employed
variants of the p-median model, perhaps the most widely used
model in the field of optimal facility location analysis. The
p-median is a location-allocation model that locates a given num-
ber p of facilities, and allocates demand nodes i to facilities j to
minimize the total distance traveled by consumers to facilities
(Hakimi, 1964; Revelle and Swain, 1970). For locating alterna-
tive-fuel stations, the p-median model has the appeal of locating
stations convenient to where people live. Several studies have
demonstrated empirically that consumers generally prefer to re-
fuel near their homes (Sperling and Kitamura, 1986; Kitamura
and Sperling, 1987). The p-median model was first applied to fuel
stations by Goodchild and Noronha (1987), who used it as one of
the objectives in a multiobjective programming model for ratio-
nalizing stations from existing gas station networks. For alterna-
tive fuels, it has been used in studies by Nicholas et al. (2004)
and Nicholas and Ogden (2006) and adopted for several major
studies of the transition to hydrogen by Oak Ridge National
Laboratory (Greene et al., 2008). Lin et al. (2008) developed what
they called the ‘‘fuel travel-back” approach that is structurally

0966-6923/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jtrangeo.2010.06.015
 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758
similar to the p-median model but with nodes weighted by the
quantity of fuel consumed on segments ‘‘pointing to” demand
node i (instead of population) and with travel time between
demand nodes i and candidate facility locations j substituted
for distance. Essentially, this model uses vehicle-miles traveled
(VMT) data to minimize the total travel time for all fuel (in
gallon-minutes) to travel from where it is burned back to the
nearest station. The p-median model has the additional advan-
tage of having simple data requirements. Road network data
and population data are widely available in GIS format from
various sources (e.g., US Census Bureau, 2010), and inter-node
distances can be easily computed using a custom program, in this
case a script written in the Python programming language.
A second approach aims to locate stations on high-traffic routes.
In addition to the p-median objective they used, Goodchild and
Noronha (1987) employed a second objective that maximizes the
traffic flows on the roads passing by a station. Melendez and
Milbrandt (2005) considered only roads with at least 20,000 vehi-
cles per day in their GIS analysis of a national hydrogen station
network. This approach recognizes that many drivers refuel on
their way to somewhere else, and tries to maximize the passing
traffic. Nicholas (2010) operationalized a version of this criteria
as the total vehicle-kilometers traveled within an aggregated zone.
Bapna et al. (2002) introduced an objective that is a hybrid of the
first two types, which maximizes the population on covered links.
The potential problem with traffic-count methods, however, is
they count the same trips by the same drivers more than once if
the trip travels multiple links, even though drivers might refuel
only once. As a result, the traffic-count or VMT methods could
locate stations on several adjacent links of a high-volume freeway.
These methods are therefore probably best suited as a secondary
objective that competes with a primary objective that would
spread the stations around, as in Goodchild and Noronha (1987),
or as a ‘‘threshold”-type constraint guaranteeing a certain mini-
mum potential demand for each station.
A third general approach to locating refueling stations maxi-
mizes passing flows without double counting. This approach orig-
inated with Hodgson’s (1990) flow-capturing location model, later
termed the flow-intercepting location model (FILM) by Berman
et al. (1992). These models are classified as path-based or flow-
demand models. The basic units of demand in these models are
not points in space representing where people live (p-median
models), nor network links (traffic-count models), but flows on
paths across a network representing the routes people travel. The
basic objective of the FILM is to locate p facilities to maximize
the number of trips intercepted. A demand is considered captured
or intercepted if there is a facility anywhere along the path. The
standard FILM counts each flow intercepted only once, regardless
of how many stations are along its route.2 Behaviorally, the FILM
is well-suited for facilities at which consumers stop along their
way to somewhere else rather than making a special trip from home
and back. Given that drivers rarely make special-purpose trips from
home to stations and back solely to refuel their vehicles, it can be
argued that flow capturing provides a behaviorally realistic basis
for locating refueling stations.
There are two main problems, however, in applying the basic
FILM to locating refueling stations. First, the model requires a ma-
trix of traffic flows from origins to destinations, each of which must
then be assigned to a particular likely path through the network.
These ‘‘trip table” data are more challenging to work with than
population data, and are not always available for all regions and
2
Extensions of the FILM have considered multiple exposures, such as to billboards
or inspection stations, with diminishing benefits of subsequent exposures (Berman
et al., 1995a,b; Zeng et al., 2008, 2010a,b; Berman, 1997).
751
geographic scales. The second problem is that, for longer inter-city
trips, one station anywhere along the path may not be enough to
enable a vehicle with a limited driving range to complete the trip
without running out of fuel. This is especially a problem for bat-
tery-powered electric vehicles and hydrogen vehicles because of
the limited energy storage capabilities of these technologies. To
address this limitation, Kuby and Lim (2005) developed the flow-
refueling location model (FRLM). The FRLM counts a flow as
refueled only if a combination of stations exists on a path that can
successfully refuel the round trip between the origin and destina-
tion, given the assumed driving range of vehicles. Like the FILM,
the FRLM tries to maximize the number of trips that can poten-
tially be refueled by p stations. The FRLM has been applied to
real-world networks at both the metropolitan scale and state scale
in Florida (Kuby et al., 2009) and Arizona (Kuby et al., 2004), and
has been extended to stations with limited capacities (Upchurch
et al., 2009), locations along arcs (Kuby and Lim, 2007), and maxi-
mizing trip-miles instead of trips (Kuby et al., 2009). At the metro-
politan scale, if no round trips are longer than the assumed vehicle
driving range, the FRLM reduces to the FILM.
In this paper, we compare the two main approaches—the node-
based p-median model and the flow-based FRLM—in terms of how
well each one does in satisfying the other’s objective. Other
researchers have combined traditional point-based demands with
flow demands in several ways (Berman, 1997; Berman and Krass,
1998; Hodgson and Rosing, 1992), but using FILM rather than
FRLM, and not applied to refueling stations. Using the Orlando
metro area and statewide Florida networks developed in Kuby
et al. (2009), we locate p stations to maximize the flows refueled,
and calculate how well the solutions perform in minimizing the
total weighted distance traveled from population nodes to stations,
that is, the p-median objective. Then, we locate p stations to min-
imize the p-median objective and calculate how well the stations
so located would be able to refuel the trips in the trip table. Our
purpose here is not to assess which model more accurately repre-
sents typical consumer refueling behavior. Rather, assuming that
both models capture an important aspect of refueling behavior—
refueling near home (p-median) and refueling on the way
(FRLM)—our goal is to assess which model does better in satisfying
the other’s objective. We also investigate which model provides
more stable solutions in which locations that are optimal for
smaller numbers of stations remain optimal when networks are
expanded with additional stations.
2. Model descriptions
The p-median model minimizes the total distance between pop-
ulation and the closest facility. The formulation for the p-median
model is as follows:
XX
Min hi dij Y ij ð1Þ
i j
Subject to :
X
Y ij ¼ 1 8 i ð2Þ
j
X
Xj ¼ p ð3Þ
j
Y ij  X j 6 0 8 i; j ð4Þ
X j ¼ 0; 1 8 j ð5Þ
Y ij ¼ 0; 1 8 i; j ð6Þ
where Yij is 1 if customer i is served by facility j, 0 if not; Xj is 1 if a
facility is located at candidate site j, 0 if not; hi is demand at location
i; dij is distance from location i to location j; p is the number of facil-
ities to be located.
752 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758

The objective function (1) minimizes the demand-weighted dis- To solve the FRLM, this study used a greedy substitution heuris-
tance between each demand node and the nearest facility. The first tic (Lim and Kuby, 2010). This heuristic includes two distinct algo-
constraint (2) ensures that each demand node will be allocated to rithms, run one after the other. The greedy algorithm places the
one and only one facility. The second constraint (3) sets the num- first facility at the candidate site where it can refuel the greatest
ber of facilities to open to exactly p. The next constraint (4) pre- amount of flow. Additional facilities are added one at a time until
vents demand from being allocated to candidate sites that do not the algorithm has located p facilities. Each additional facility is
have facilities. The last two constraints ensure that Xi and Yij have placed where it can refuel the maximum amount of additional
Boolean values of 0 or 1. flow. While greedy algorithms perform optimally on some location
The formulation of the FRLM is as follows: problems, this is not the case with the FRLM. The FRLM requires
X specific combinations of stations to refuel flows that are longer
Max fq Y q ð7Þ than the vehicle’s range. Because the algorithm adds stations one
q2Q at a time, a purely greedy heuristic will often miss these combina-
Subject to : tions of stations.
X To make it more likely that the heuristic will find these combi-
bqh v h P Y q 8q2Q ð8Þ
nations (and thus the optimal solution), this study uses a substitu-
h2H
tion algorithm similar to the one used for the p-median heuristic.
ahk X k P v h 8 h 2 H; k 2 K ð9Þ
X The substitution algorithm uses the output of each step of the gree-
Xk ¼ p ð10Þ dy algorithm as its input. It goes through each facility on the list
k and attempts to replace it with every possible candidate site. If a
X k 2 f0; 1g 8 k ð11Þ candidate site increases the amount of flow captured, it is substi-
0 6 Y q 6 1 8 q; 0 6 v h 6 1 8 h ð12Þ tuted for the original facility (if more than one candidate site
improves the output, the best site is substituted). After each substi-
where Yq is 1 if path q is refuelable, 0 if not; vh = 1 if all facilities in tution the process begins again and the algorithm attempts to
combination h are open, 0 if not; Xk is 1 if facility k is open, 0 if not; replace another facility with a better candidate site. Lim and Kuby
fq is the total volume on path q; bqh is 1 if combination h can refuel (2010) compared the best solutions from the greedy substitution
path q, 0 if not; ahk is 1 if combination h includes facility k, 0 if not.
The objective function (7) maximizes the amount of flow that
can be refueled. The first constraint (8) ensures that flows can only
be refueled if there is a combination of open facilities that can re-
fuel the entire round trip. In this mixed-integer linear program-
ming (MILP) formulation, the combinations of facilities that can
refuel each flow would have to be precalculated based on the dis-
tance between candidate sites along each path and the range of the
vehicle (see Kuby and Lim (2005) for the algorithm for generating
all feasible combinations). The second constraint (9) prevents a
combination from being used unless all facilities in that combina-
tion are open. The third constraint (10) requires exactly p facilities
to be opened. The fourth constraint (11) ensures that only whole
facilities will be located. The last constraint (12) prevents the pro-
portion of the flow refueled from exceeding 1.

3. Heuristics
Fig. 1. Performance of the FRLM and p-median facility locations on the FRLM
The networks used in this study are reasonably large (the Orlan- objective for the Orlando network.
do area network has 287 nodes, while the Florida network has 293
nodes). This is too large to solve using the MILP formulations given
above due to the computing time and memory that would be
needed to compute and store all possible combinations of stations
that can refuel the longer paths, and then solve the subsequent
MILP (Lim and Kuby, 2010). To provide a quicker solution, this
study relies on heuristics to solve both the p-median problem
and FRLM.
The p-median problem was solved using a substitution algo-
rithm outlined by Daskin (1995). This heuristic is seeded with an
initial list of p facilities. The algorithm attempts to replace each
facility on the list with every possible candidate. If one of these
replacements results in a better performance on the objective func-
tion, the replacement facility is substituted for the original one and
the process begins anew. This continues until the algorithm finds a
list of facilities that cannot be improved by replacing any facility on
the list with any other candidate site. The substitution (a.k.a. ex-
change) algorithm widely attributed to Tietz and Bart (1968) is
one of the most widely used methods for solving the p-median
problem, and was described by Densham and Rushton (1992) as Fig. 2. Performance of the FRLM and p-median facility locations on the p-median
‘‘more robust than competing algorithms.” objective for the Orlando network.
 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758 753

algorithm for the FRLM with the global optimum solution from
mixed-integer programming on a test network. Using a driving
range longer than any of the arc lengths (which is typical of
real-world problems), the greedy substitution algorithm found
the optimal solution for 22 of the 25 p values, and the solutions
were never more than 0.52% below optimal.
4. Orlando results
The Orlando network dataset was developed by Kuby et al.
(2009). It includes highways and major surface streets in Orange
and Seminole counties, and the northwest portion of Osceola
County. The area is broken down into 102 demand zones, each of

Downtown
D
Do
ownto
ow
wnt
wn
nto
nto
tow
tow
own
own

Legend
FRLM stations

p-median stations

Candidate Sites
(population)
475 - 6768
6768 - 12961
12961 - 18641

18641 - 24506

24506 - 37073

Streets
(flow)
0 - 7228

7229 - 18362

18363 - 34798

34799 - 68458

68459 - 131917

Fig. 3. Map of the Orlando network with p = 5 solutions.

754 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758
which is represented by a node at a carefully selected intersection
within the zone. These demand nodes serve as the population cen-
ters in the p-median model and as the origins and destinations of
trips in the FRLM. Population data for the p-median model are
based on census data, while OD trip volume data are from the Flor-
ida Department of Transportation (FDOT). Travel times are based
on posted speed limits, with the speed reduced 15% on arterial
streets to reflect more realistic driving conditions. The vehicle
range for the FRLM station solution and objective is assumed to
be 100 miles. Because this is longer than the longest origin–desti-
nation distance in the Orlando area, this effectively turns the FRLM
into a flow-intercepting location model.
We used the p-median and FRLM heuristics to locate 1–25 alt-
fuel stations on the network. Each set of alt-fuel station locations
was evaluated using both the p-median and FRLM objective func-
tions, showing how well stations located using the FRLM perform
on the p-median objective and vice versa.
As expected, the FRLM stations perform better on the flow-refu-
eling objective than the p-median stations do (Fig. 1). The average
flow captured by the stations located by the FRLM is 39% higher
(better) than the flow captured by the stations located by the p-
median model. This represents a fairly substantial difference, par-
ticularly during the early phases of a transition to alternative fuels.
The p-median stations outperform the FRLM stations on the p-
median objective. The difference, however, is not as great as on
the flow-refueling objective (Fig. 2). The average demand-
weighted distance for the p-median stations is 26% lower (better)
than the FRLM stations. The proportional differences between the
two objectives are smallest for lower numbers of stations and grow
as the number of stations increases.
As seen in Fig. 3, when the p-median and FRLM are asked to lo-
cate five stations in Orlando, there are no common locations. Both
algorithms disperse their stations around the urban area, but the
FRLM locates its stations exclusively on high-volume highways,
often at major intersections. The p-median model locates some of
its stations in similar locations, but others are found in much less
traveled areas. Because the p-median model is attracted to popula-
tion, rather than traffic flow, hitting these high-volume intersec-
tions is a matter of happenstance. For instance, the FRLM locates
the first station in downtown Orlando, where two major highways
meet and a huge number of trips can be intercepted. The p-median
model does not locate any stations in the downtown area because
of the lack of residential population in and around the central busi-
ness district.
One important difference between the p-median solutions and
the FRLM solutions is the stability of the solutions as the number
of stations changes. The FRLM solutions are quite stable. It is very
rare for a station that was part of the solution for a given number of
stations to drop out of the solution for a higher number of stations.
Stations are somewhat ‘‘riveted” to the particular locations that
heavy traffic flows funnel through, and only move around to avoid
duplication and cannibalization of flows captured. The entire series
of solutions from 1 station to 25 stations only involves the selec-
tion of 29 distinct candidate sites. In contrast, the p-median solu-
tions are far less stable. Stations regularly drop out of the
solution and reappear. It is not hard to visualize why. With one sta-
tion, the p-median would tend to place it in the population-
weighted center of the network. With two stations, the center
would no longer be a good location, as one would want a station
in each half of the network. As a result, in our analysis, the p-med-
ian solutions for 1–25 stations involve a total of 60 different candi-
date sites. Because alternative-fuel stations are likely to be
completed over a moderately long time span, the performance of
a partially completed set of stations may be as important as the
performance of the completed system. The stability of the FRLM
solutions potentially makes it much easier to design a system that
can be built progressively and remain close to optimal even when
not all facilities are complete.
5. Florida results
The Florida network covers the entire state and includes all
interstate and US highways, toll roads, and selected state high-
ways. Travel times are again based on the posted speed limits
and are reduced by 15% on non-limited access highways. The state
is divided into 74 demand nodes (all counties or groups of coun-
ties) in order to model inter-city flows. Origin–destination flow
data, however, are not available at the state level, so flows are esti-
mated using a gravity model based on the Michigan statewide tra-
vel demand study (Kuby et al., 2009; KJS Associates, 1996). The
vehicle driving range is set at 100 miles.
The flow-refueling objective displays a very large difference be-
tween the FRLM and p-median solutions (Fig. 4). The p-median sta-
tions refuel an average of 79% less flow than the FRLM stations.
Unlike the Orlando area results, the p-median solutions do not dis-
play much of a trend towards increasing performance on the FRLM
objective as the number of stations increases. This abysmal perfor-
mance at the state scale stands in stark contrast to the perfor-
mance of the p-median solutions on the FRLM objective at the
metropolitan area scale in Orlando. For several values of p, the per-
centage of statewide flows that can be refueled hovers barely
above 0% because all stations happen to be located at intersections
on roads with minimal flow.
The performance of the FRLM solutions on the p-median objec-
tive is considerably better, but still far from optimal (Fig. 5). The
average demand-weighted distance for the p-median stations is
68% lower than the demand-weighted distance for the FRLM sta-
tions. The FRLM results show two dramatic drops (that is, improve-
ments): one when going from 2 to 3 stations, the other when going
from 12 to 13 stations. The first drop occurs when the first station
outside the Miami metro area is introduced in Tampa. The second,
smaller, drop is a result of the introduction of a station in Jackson-
ville, the first station north of Orlando.
Fig. 6 shows the Florida network with the FRLM and p-median
solutions for eight stations. The p-median solution disperses facil-
ities throughout the state in an attempt to cover as much of the
population as possible. In contrast, the FRLM stations are concen-
trated in two distinct clusters in the southern and central parts
of the state, building up connected networks that cover local and
some long-distance flow. This behavior is driven by two aspects
of the FRLM: multiple refueling and distance decay. Any trip with
a one-way distance longer than the vehicle range of 100 miles will
require at least two refuelings to reach its destination. This, in turn,
Fig. 4. Performance of the FRLM and p-median facility locations on the FRLM
objective for the Florida network.
 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758 755

requires pairs of alt-fuel stations less than 100 miles apart. Only
three pairs of the p-median stations are less than 100 miles apart,
giving very few opportunities for longer distance trips. The dis-
tance decay phenomena is a product of the gravity model used to
generate the underlying trip data, and is in turn an accurate reflec-
tion of the prevalence of shorter trips in real life. Nearby nodes will
generate more traffic between them than more distant cities. The
high volume of short distance trips leads the model to serve large
urban areas as comprehensively as possible. This is particularly
evident in the Miami metropolitan area. The FRLM locates four sta-
tions in the Miami area, two to the north in Delray Beach and Fort
Lauderdale and two to the south in Hialeah and Miami proper.
These four stations provide a fairly comprehensive coverage of
trips within the metropolitan area. The p-median model, on the
other hand, locates a single station in the Fort Lauderdale area. This
location is well-suited for minimizing demand-weighted distance,
but it will only capture a handful of shorter flows from the north-
ern half of the metro area to the southern half, completely missing
Fig. 5. Performance of the FRLM and p-median facility locations on the p-median
objective for the Florida network.
inter-city flows within the two halves of the region. As in the

Jacksonville

P-Median Stations
FRLM Stations
Orlando
Candidate Sites
(population) Tampa

7022 - 87366
87367 - 217955
217956 - 365196
365197 - 582653 Delray
Beach
582654 - 937276
Fort
Major Roads Hialeah Lauderdale
Miami
(flow)
0 - 6000000
6000001 - 18000000
18000001 - 50000000
50000001 - 120000000
120000001 - 220000000 0 25 50 100 Miles

Fig. 6. Map of the Florida network with p = 8 solutions.

756 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758
Orlando example, many of the p-median stations are located on
minor routes without much traffic flow. The FRLM stations, in con-
trast, are all on freeways, often at major intersections between
them to capture flow on both routes.
6. Comparison of different scales of analysis
The dramatic differences between the Orlando area and state-
wide results are primarily a function of vehicle range and distance
decay of flow volumes. At the metropolitan scale, the differences
between the FRLM and p-median models are less dramatic than
at the statewide scale, where vehicle range and distance decay
both come into play to a much greater degree. At the metropolitan
scale, the FRLM is effectively operating as a flow-interception mod-
el because almost no round trips are long enough require two refu-
eling stops. Therefore, at this scale the FRLM can concentrate on
locating facilities at heavily traveled nodes that capture unique,
non-duplicative flows. The p-median model is attempting to cen-
trally locate facilities where they are near clusters of large popula-
tion centers. Thus, while not an explicit objective or constraint of
either model, both models tend to disperse their facilities over a
wide area. The p-median model works to decrease the distance
to population nodes, while the flow-interception model only needs
to capture each flow once, making facilities at nearby nodes on the
same major roadways a low-return proposition. The fact that both
models benefit somewhat from dispersing facilities at this scale
leads to broad similarities in how spread out the stations are,
and as a result the performance of each model on the other’s objec-
tive is not as poor as at the state scale.
However, once the distances grow large enough that vehicle
range becomes a factor, the behavior of the FRLM changes consid-
erably. Rather than dispersing facilities to various high-traffic
nodes, the FRLM is now working to keep facilities relatively close
together, less than the vehicle range apart, so that vehicles can
use a chain of facilities to refuel longer trips. Traffic flows are also
much heavier between nearby nodes because of distance decay, so
stations may be placed close together so as not to miss any of the
really large flows in the network between nearby large cities. The
network only slowly grows into the outlying areas, instead adding
additional stations in areas already covered to handle as many trip
combinations as possible. The p-median model, on the other hand,
covers the entire state even with very few stations, yet these sta-
tions are so far apart that refueling long-distance trips is largely
impossible. Rather than coming to similarly dispersed solutions
from different directions, once the network becomes larger than
the vehicle range, the two functions are now working towards en-
tirely different goals.
7. Conclusions
The most significant finding of this paper is that, at both scales
of analysis, the FRLM facility locations perform better on the p-
median objective than the p-median facilities do on the FRLM
objective. This indicates that the FRLM does a better job at fulfilling
both theories of fueling station location (stations near people’s
homes and stations convenient to people’s trips) than the p-med-
ian model does. This statement comes with some caveats, how-
ever. The two objective functions are quite different: one is a
maximization problem trying to capture as much of a finite
amount of flow as possible, while the other is a minimization prob-
lem that works to reduce the demand-weighted distance from
facilities to population. Direct comparison of performance on such
different objectives is difficult. The flow-refueling objective is also
considerably less forgiving than the p-median objective. A facility
that is located in a slightly suboptimal position for the p-median
objective will result in a somewhat higher demand-weighted dis-
tance as some people have to travel further to reach the facility.
For the FRLM objective, a facility located just one node away from
the optimal position may be just off a major route or slightly be-
yond a vehicle’s maximum range, leading to a large drop in the
objective. Close does not count for the FRLM, but it does for the
p-median. This conclusion is so strong that it suggests that
researchers using the p-median model may want to exclude candi-
date sites that are not located at high-flow nodes.
A second major conclusion is the critical importance of the scale
of the network in relation to the vehicle’s range and the amount of
distance decay of flows. There is a stark difference in the perfor-
mance and behavior of the two models as the scale grows from
the Orlando urban area to the statewide level. At the metropolitan
scale, the two algorithms produce somewhat similar results in
terms of the dispersion of facility locations. Neither performs opti-
mally on the other model’s objective function, but they each turn
in a respectable performance. At the statewide scale, however,
the models produce very different patterns of station location. Per-
formance on the other model’s objective is extremely poor. Also,
the gap between each model’s performance on its own objective
compared with the other’s objective is much larger at the state
scale than at the metropolitan scale (compare the gaps between
curves in Figs. 1 and 4 and between Figs. 2 and 5). At the statewide
scale, the p-median model in particular does a very poor job of
intercepting flows. Because the p-median model is a location-allo-
cation model, each facility is being sited to be central to the subset
of nodes that are allocated to it, which can often place that facility
in the middle of nowhere (flow-wise, that is).
Finally, the third important conclusion is that the optimal loca-
tions chosen by the FRLM are far more stable than those selected
by the p-median model as the infrastructure is built out and the
number of stations increases. Especially at the metropolitan scale,
the stations that are optimal for small numbers of stations tend to
remain optimal when more stations are added. Such is not the case
for the p-median solutions, in which stations come in and out of
the optimal solutions as p increases. Because the p-median model
is less influenced by the demand at the exact node where the sta-
tion is located, the p-median solutions are less stable. This should
be extremely important for decision-makers who want to use an
optimization model to choose the first set of stations and want
these locations to remain good when more stations are added later.
This finding is consistent with previous research on FILMs and
p-median models. For instance, Zeng et al. (2010b) found that
FILMs for real-world transport systems tend to produce fairly sta-
ble solutions because major flows tend to be concentrated on a
small number of paths, and a small number of nodes are able to
intercept most flows. For p-median models, Current et al. (1997)
recognized that the number of facilities to build may not be known
for certain, and if a different number than p end up being built, the
solution may be substantially suboptimal. They developed two
methods for locating p facilities to minimize the expected opportu-
nity loss or the maximum regret if a different number of facilities
are eventually built.
In summary, we refer back to the oft-cited empirical survey of
drivers by Kitamura and Sperling (1987). They interviewed drivers
while they refueled, and found that 74.8% were on trips to or from
home, including 7% on single-purpose trips from home. The same
study, however, found 28.6% refueled on trips to or from work
and reached the following conclusion (p. 243, with our italics):
In particular the following is important for development of an
initial distribution network for a new fuel: drivers tend to refuel
in areas that are detailed in their mental maps, e.g. in the vicin-
ity of home and workplace; drivers tend to interrupt their jour-
ney near the beginning or end, rather than in the middle, in
 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758
order to refuel; refueling is frequently performed in commuting
trips, while it is linked with shopping with less-than-expected
frequency; and single- purpose refueling is not uncommon.
When combined with existing databases that offer the distribu-
tion of trips in a metropolitan area, these findings will comprise
an essential set of relationships for establishing the spatial dis-
tribution of refueling locations in the area that will be most
acceptable to drivers of nonpetroleum vehicles. As strategic
locations for initial outlets, the analysis at this stage points to
locations along high-volume commuting routes.
This quote, based on an empirical survey of driver behavior
rather than performance of prescriptive models, presages develop-
ment of a location model similar to the FRLM based on paths and
trip tables.
8. Future work
This study clearly shows the effects of vehicle range relative to
the size of the network using the Orlando and Florida networks.
The precise point at which this shift occurs remains unknown. It
would be possible to examine this by looking at the performance
of the FRLM and p-median station locations on the FRLM objective
for an assortment of different vehicle ranges. The relationship be-
tween vehicle range, average trip length, and the proportion of
flows refueled by the p-median facilities would help characterize
the scale at which the behavior changes and the p-median model
no longer works well for flow refueling.
In this study, we optimized each objective and then simply cal-
culated how well the solution would perform on the other objec-
tive. A logical next step would be to build a true multiobjective
model and generate Pareto-optimal tradeoff curves. It could well
be that by putting a very small weight on the flow-refueling objec-
tive and adding it to the p-median objective, the multiobjective
model would slightly shift the p-median locations onto higher flow
routes so that the FRLM objective improves dramatically and the p-
median objective degrades only slightly. In addition, using a mul-
tiobjective model would allow decision-makers to place more or
less emphasis on these objectives. For instance, governmental deci-
sion-makers might put more weight on a more geographically
equitable pattern that serves most major regions, whereas the pri-
vate sector will be driven more by profit maximization.
As mentioned in the previous section, the FRLM objective func-
tion is quite unforgiving when it comes to station locations. Many
drivers, particularly early adopters of alternative fuels in urban
areas, may be willing to detour some distance off their intended
route in order to refuel their vehicles. The basic FRLM does not al-
low for any detour from the shortest path, regardless of how short
a detour it might be to access a fueling station. We are working on
a FRLM extension that allows for alternate routes that trade
slightly longer travel distances for access to a fueling station. This
proposed FRLM-detour model would more realistically model dri-
ver’s behaviors. While this change would probably greatly improve
the estimated performance of the p-median model’s station loca-
tions on the FRLM objective, it might also substantially improve
the estimated performance of the FRLM’s stations, which would
be located to take advantage of possible detours. Therefore, we
cannot say at this time whether the large gaps between the FRLM
stations and the p-median stations on the flow-refueling objective
would shrink if detours were considered.
While the p-median model is relatively realistic at urban scales,
it becomes much less so at the statewide scales, particularly for
small numbers of stations. For example, with five facilities, a fuel-
ing station near Palm Beach is the only one located anywhere in
the Miami metro area, despite it being more than 70 miles from
downtown Miami and 100 from the southern end of the metropol-
757
itan area. The analysis here strongly suggests that the FRLM is
more suitable for locating stations at the statewide level. If a sim-
pler, less data-intensive method is desired, a maximum covering
model with a fairly short critical distance based on some assump-
tions about how far people are willing to travel for fuel might do a
better job than the p-median model of locating stations in popula-
tion centers at the statewide level. Another alternative mentioned
earlier would be to restrict the p-median candidate sites to nodes
with heavy traffic flows, or to use some kind of multiobjective ap-
proach combining nodal traffic flows with a second objective.
References
Bapna, R., Thakur, L.S., Nair, S.K., 2002. Infrastructure development for conversion to
environmentally friendly fuel. European Journal of Operational Research 142
(3), 480–496.
Berman, O., 1997. Deterministic flow-demand location problems. Journal of the
Operational Research Society 48 (1), 75–81.
Berman, O., Krass, D., 1998. Flow intercepting spatial interaction model: a new
approach to optimal location of competitive facilities. Location Science 6 (1–4),
41–65.
Berman, O., Larson, R.C., Fouska, N., 1992. Optimal location of discretionary service
facilities. Transportation Science 26, 201–211.
Berman, O., Hodgson, M.J., Krass, D., 1995a. Flow intercepting models. In: Drezner,
Z. (Ed.), Facility Location: A Survey of Applications and Methods. Springer, New
York, pp. 389–426.
Berman, O., Krass, D., Xu, Chen.Wei., 1995b. Locating flow-intercepting facilities:
new approaches and results. Annals of Operations Research 60, 121–143.
California Environmental Protection Agency, 2005. California Hydrogen Blueprint
Plan, vols. 1 and 2.
Current, J., Ratick, S., ReVelle, C., 1997. Dynamic facility location when the total
number of facilities is uncertain: a decision analysis approach. European Journal
of Operational Research 110, 597–609.
Daskin, M.S., 1995. Network and Discrete Location Analysis. John Wiley and Sons,
New York.
Densham, P.J., Rushton, G., 1992. Strategies for solving large location-allocation
problems by heuristic methods. Environment and Planning A 24, 289–304.
Goodchild, Michael F., Noronha, Valerian T., 1987. Location-allocation and
impulsive shopping: the case of gasoline retailing. In: Ghosh, A., Rushton, G.
(Eds.), Spatial Analysis and Location-Allocation Models. van Nostrand Reinhold,
New York, pp. 121–136.
Greene, D.L., Leiby, P.N., James, B., Perez, J., Melendez, M., Milbrandt, A., Unnasch, S.,
Rutherford, D., Hooks, M., 2008. Analysis of the Transition to Hydrogen Fuel Cell
Vehicles and the Potential Hydrogen Energy Infrastructure Requirements,
ORNL/TM-2008/30.
Hakimi, S.L., 1964. Optimum locations of switching centres and the absolute centres
and medians of a grap. Operations Research 12, 450–459.
Hodgson, M.J., 1990. A flow capturing location allocation model. Geographical
Analysis 22, 270–279.
Hodgson, M.J., Rosing, K.E., 1992. A network location-allocation model trading off
flow capturing and p-median objectives. Annals of Operations Research 40 (1–
4), 247–260.
Huleatt-James, N., 2008. Hydrogen infrastructure survey 2008. <http://
www.fuelcelltoday.com/media/pdf/surveys/2008-Infrastructure-Free.pdf>
(accessed 24.07.08).
Kitamura, R., Sperling, D., 1987. Refueling behavior of automobile drivers.
Transportation Research Part A 21A (3), 235–245.
KJS Associates, 1996. Statewide Travel Demand Model: Update and Calibration,
Phase II.
Kuby, M.J., Lim, S., 2005. The flow-refueling location problem for alternative-fuel
vehicles. Socio-Economic Planning Sciences 39, 125–145.
Kuby, M., Lim, S., 2007. Location of alternative-fuel stations using the flow-refueling
location model and dispersion of candidate sites on arcs. Networks and Spatial
Economics 7 (2), 129–152.
Kuby, M.J., Lim, S., Wang, K., 2004. A model for optimal location of hydrogen
refueling stations: an Arizona case study. In: Proceedings of the National
Hydrogen Association, Los Angeles; April 26–30, 2004.
Kuby, M., Lines, L., Schultz, R., Xie, Z., Kim, J., Lim, S., 2009. Optimal location strategy
for hydrogen refueling stations in Florida. International Journal of Hydrogen
Energy 34, 6045–6064.
Lim, S., Kuby, M., 2010. Heuristic algorithms for siting alternative-fuel stations
using the flow-refueling location model. European Journal of Operational
Research 204 (1), 51–61.
Lin, Z., Ogden, J., Fan, Y., Chen, C., 2008. The fuel-travel-back approach to hydrogen
station siting. International Journal of Hydrogen Energy 33 (12), 3096–3101.
Melaina, M.W., 2003. Initiating hydrogen infrastructures: preliminary analysis of a
sufficient number of initial hydrogen stations in the US. International Journal of
Hydrogen Energy 28 (7), 743–755.
Melaina, M., Bremson, J., 2008. Refueling availability for alternative fuel vehicle
markets: sufficient urban station coverage. Energy Policy 36 (8), 3223–3231.
Melendez, M., 2006. Transitioning to a Hydrogen Future: Learning from the
Alternative Fuels Experience. Technical Report No. NREL/TP-540-39423.
758 C. Upchurch, M. Kuby / Journal of Transport Geography 18 (2010) 750–758
Melendez, M., Milbrandt, A., 2005. Analysis of the Hydrogen Infrastructure Needed
to Enable Commercial Introduction of Hydrogen-Fueled Vehicles, CP-540-
37903.
National Research Council, 2004. The Hydrogen Economy: Opportunities, Costs,
Barriers, and R&D Needs. The National Academies Press, Washington, DC.
Nicholas, M.A., 2010. Driving demand: what can gasoline refueling patterns tell us
about planning an alternative fuel network?.
Nicholas, M.A., Ogden, J., 2006. Detailed analysis of urban station siting for
California hydrogen highway network. Transportation Research Record 1983,
121–128.
Nicholas, M.A., Handy, S.L., Sperling, D., 2004. Using geographic information
systems to evaluate siting and networks of hydrogen stations. Transportation
Research Record 1880, 126–134.
Ogden, J.M., 1999. Developing an infrastructure for hydrogen vehicles: a
Southern California case study. International Journal of Hydrogen Energy 24
(8), 709–730.
Revelle, C.S., Swain, R., 1970. Central facilities location. Geographical Analysis 2, 30–
42.
Sperling, D., Kitamura, R., 1986. Refueling and new fuels: an exploratory analysis.
Transportation Research Part A 20A (1), 15–23.
Tietz, M.B., Bart, P., 1968. Heuristic methods for estimating the generalized vertex
median of a weighted graph. Operations Research 16, 955–965.
Upchurch, C., Kuby, M.J., Lim, S., 2009. A capacitated model for location of
alternative-fuel stations. Geographical Analysis 41, 85–106.
US Census Bureau, 2010. TIGERÒ, TIGER/LineÒ and TIGERÒ-related products. <http://
www.census.gov/geo/www/tiger/>.
US Department of Energy, 2002. National Hydrogen Energy Roadmap. US
Department of Energy, Washington, DC.
Zeng, W., Hodgson, M.J., Castillo, I., 2008. The pickup problem: consumers’
locational preferences in flow interception. Geographical Analysis 41, 3–22.
Zeng, W., Hodgson, M.J., Castillo, I., 2010a. A generalized model for locating facilities
on a network with flow-based demand. Networks and Spatial Economics.
doi:10.1007/s11067-008-9073-8.
Zeng, W., Castillo, I., Hodgson, M.J., 2010b. Aggregating data for the flow-
intercepting location model: a GIS, optimization, and heuristic framework.
Geographical Analysis 42, 301–322.