i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

Available online at www.sciencedirect.com

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

Design and operation of a stochastic hydrogen supply chain

network under demand uncertainty

A. Almansoori a,*, N. Shah b

a
Department of Chemical Engineering, The Petroleum Institute, P.O. Box 2533, Abu Dhabi, United Arab Emirates
b
Centre for Process Systems Engineering, Department of Chemical Engineering, Imperial College London, South Kensington Campus,
London SW7 2AZ, UK

article info abstract

Article history: The design of a future hydrogen supply chain (HSC) network is challenging due to the: (1)
Received 27 September 2011 involvement of many echelons in the supply chain network, (2) high level of interactions
Received in revised form between the supply chain components and sub-systems, and (3) uncertainty in hydrogen
13 November 2011 demand. Most of the early attempts to design the future HSC failed to incorporate all these
Accepted 15 November 2011 challenges in a single generic optimization framework using mathematical modeling
Available online 12 December 2011 approach. Building on our previous multiperiod MILP model, the model presented in this
paper is expanded to take into account uncertainty arising from long-term variation in
Keywords: hydrogen demand using a scenario-based approach. The model also adds another echelon:
Hydrogen supply chain fueling stations and local distribution of hydrogen. Our results show that the future HSC
MILP stochastic model network is somewhat similar to the existing petroleum infrastructure in terms of
Demand uncertainty production, distribution, and storage. In both situations, the most feasible solution is
Risk analysis centralized production plants with truck and rail delivery and small-to-large storage
Scenarios-based approach facilities. The main difference is that the future hydrogen supply has the benefits of using
Great Britain distributed forecourt production of hydrogen at local fueling stations via several produc-
tion technologies. Finally, the performance of the studied models was evaluated using
sensitivity and risk analyses.
Copyright ª 2011, Hydrogen Energy Publications, LLC. Published by Elsevier Ltd. All rights
reserved.

1. Introduction Even though hydrogen is a more diverse, sustainable, and

environmentally friendly fuel, its widespread as a transport
Hydrogen fuel is envisioned to become a key element of the fuel necessitates the development of a sustainable supply
future energy mix, especially in the transportation sector chain network. As with the existing petroleum supply system,
[1e4]. Unlike gasoline, hydrogen can be harnessed from a wide the future hydrogen supply chain (HSC) ought to include
range of energy sources including non-renewable sources production sites, storage facilities, transportation options,
such as hydrocarbons and renewable sources such as wind, dispensing stations and end-use applications. The HSC should
biomass, or solar energy. Such advantage enables countries to be as ubiquitous as the one present now with refueling points
diversify their energy portfolios and to secure fuel supplies. supporting daily and seasonal demand fluctuations. The
Moreover, since hydrogen is a good energy carrier, it can be infrastructure should also be safe, cost-effective, and
stored in large amounts and converted readily to electricity eco-friendly. In addition, the future supply chain network
when needed with almost zero emissions. should exploit some components of the current petroleum

* Corresponding author. Tel.: þ971 2 607 5583; fax: þ971 2 607 5200.
E-mail address: [email protected] (A. Almansoori).
0360-3199/$ e see front matter Copyright ª 2011, Hydrogen Energy Publications, LLC. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijhydene.2011.11.091
3966 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

infrastructure such as steam methane reforming plants and include the logistics network of energy sources, the concept of
depleted gas fields or aquifers. importing these resources from abroad if there is a shortfall,
Most of the literature reviews in the area of HSC design and the effect of demand uncertainty on the hydrogen
focus on evaluating the performance of a pre-determined infrastructure.
hydrogen supply pathway using a steady-state simulation Li et al. [19] extended the work of Hugo [17,18] and included
approach. The work of Simbeck and Chang [5], the US National numerous combinations of potential technologies within the
Research Council [6], Thomas et al. [7], Ogden [8,9], Amos [10], HSC, which are necessary in the strategic decision-making
and Guy [11] are extensively based on adopting the simulation process for building up the future hydrogen infrastructure.
approach. Others focus on studying individual components of Their work also takes into consideration all possible hydrogen
the supply chain, such as production or storage technologies. alternatives and the interactions and trade-offs between the
Very little effort has been invested in realizing the capability various supply chain components. Another noticeable study is
of mathematical programming techniques to assist in identi- the work of Kim et al. [20] who developed a steady-state,
fying the optimum configuration of the future HSC. The aim of stochastic MILP model to take into account the effect of
this work is to develop a comprehensive optimization demand uncertainty in hydrogen activities. Their model
framework that can support strategic decisions in HSC design examined the total cost of the hydrogen network for various
and operation for vehicular use. The proposed optimization configurations using a two-stage stochastic programming
framework will make use of a scenario-based approach to approach. Even though this work can be seen as the first
capture uncertainty in demand through a mixed integer linear stochastic approach to optimize an HSC, the work neglects the
programming (MILP) model. The developed model is robust evolution of the hydrogen network over a long-term future
and capable of mapping out all possible configurations of the planning horizon leading to phased infrastructure develop-
future HSC under demand uncertainty, and applicable to any ment. In addition, the multiple and diverse primary energy
geographical setting. The model is also capable of determining feedstocks as well as the determination of logistics of fueling
the number of fueling stations and the layout of the distri- stations were not incorporated.
bution network. Shortly after their first work, Kim and Moon [21] extended
The key factor in differentiating abovementioned studies is their previous mathematical formulation and introduced
the performance criteria such as cost, environmental impact, a multi-objective optimization approach for the strategic
energy efficiency or safety. Most of these studies analyze each design of a hydrogen infrastructure taking into consideration
component of the supply chain individually and then collec- cost and safety. Guillén-Gosálbez et al. [22] also presented
tively form a specific hydrogen pathway. Each of these pre- a bi-criterion MILP optimization approach for planning and
defined pathways are then simulated and compared to designing an HSC network considering both economics and
select the “best” possible configuration based on a key environmental concerns. The work focuses on a detailed
performance indicator, where cost being the dominant life-cycle environmental analysis and a novel algorithm to
measure. Also, most of these analyses are limited in their reduce the computational cost associated with solving the
general applicability since they are based on a number of large-scale problem. More recently, Sabio et al. [23] have
assumptions concerning the level of demand, distribution developed a decision-support tool to address the strategic
distances, and geographic location. In addition, the studies do planning of hydrogen networks with risk control under
not consider the local availability or logistics network of uncertainty in the operating costs. The main purpose of their
energy sources, demand uncertainty and different sizes of work is to determine the optimal design of the HSC for
production plants and storage facilities. None of these studies a pre-defined hydrogen demand profile for vehicle use in
look at integrating the components of the HSC within a single Spain. Their optimization problem is formulated as a multi-
optimization framework to determine the optimal hydrogen objective, multi-scenario, stochastic MILP model accounting
supply configurations. for the minimization of the expected total discounted cost and
To date, significant attention has been given to the role of the worst case value. Thus, their model incorporates the
optimization techniques in designing and operating a future trade-off between risk and cost at the decision-making level.
HSC network. A number of mathematical models for planning Our work makes use of a scenario planning approach to
and designing the future HSC have appeared in the literature. capture uncertainty in hydrogen demand over a long-term
The work of Agnolucci [12], Karlsson and Meibom [13], Tzimas planning horizon. The variation of demand will be repre-
et al. [14], Ingason et al. [15], and Lin et al. [16] are some sented by a moderate number of scenarios for which a prob-
examples of these attempts. The main objective of these ability of occurrence will be specified. A multi-stage stochastic
optimization frameworks is to compare and evaluate different MILP model is proposed to outline plausible configurations of
alternatives of the hydrogen pathways and then integrate the future HSC network. The goal of our formulation is to
them within the existing energy supply chains. In an early represent the possible future realization of unknown problem
attempt to design a hydrogen infrastructure, Hugo [17,18] parameters through a set of scenarios, and to consider the
developed an optimization-based formulation that investi- range of scenarios in determining a robust solution. The
gates different hydrogen pathways in Germany. The model decisions incorporated in the presence of the stochastic
identifies the optimal infrastructure in terms of both invest- behavior are “here-and-now” decisions and “wait-and-see”
ment and environmental criteria for many alternatives of decisions [24]. “Here-and-now” decisions have to be imple-
hydrogen configurations. The model did not consider the mented prior to the realization of the uncertainty and they are
issue of stationary storage of hydrogen to accommodate daily used to describe the first-stage variables. These decisions are
and seasonal variations of demand. The model also does not associated with predicting the structure of the network, for
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7 3967

example the size, location, and capacity of production plants, the HSC network for each of these scenarios will be deter-
before the actual demand to some future event is identified. mined by the optimization procedure. Having multiple
On the other hand, “wait-and-see” decisions correspond to scenarios gives a better representation of what the future
those decisions made after the uncertainty is unveiled and hydrogen network may look like, as each scenario contributes
they relate to second-stage variables of the model. Examples towards the final network cost.
of these decisions are production and transportation rates as The stochastic behavior of the hydrogen demand is rep-
well as network retrofits and capacity expansion, which can resented by a three-stage stochastic optimization problem.
often be revised when most up-to-date information is The first-stage assumes demand to be known (deterministic)
revealed. while the second-stage takes into account demand variations
(uncertainty) over the planning horizon. These two stages are
formulated as an MILP model. The model of interest is con-
2. Problem description structed in a way that it can handle numerous scenarios. This
formulation will allow the model to behave well under all
In our previous work [25,26] the aim was to plan and design an demand fluctuations happening at some point during the life
HSC network, consisting of hydrogen production plants, of the HSC network. In this paper, nine different demand
storage facilities and transportation modes. The earlier model scenarios are considered during the three time periods of
was designed to satisfy the demand of hydrogen for a deter- interest. Each time period represents a 6-year interval starting
ministic profile over a long-term planning horizon. The from 2005 and ending in 2022. These demand scenarios were
previous attempt also neglects to include hydrogen amalgamated to form one and three unique scenarios during
dispensing technologies (fueling stations) and the logistics the first and second time periods, respectively. The formula-
associated with distributing hydrogen. Deterministic demand tion of the tree structure was accomplished through using
is seldom the case in real life applications. Demand usually a condition known as non-anticipativity. This principle will be
undergoes numerous variations relative to manufacturers’ addressed later in the paper.
and customers’ needs. These variations, i.e. stochastic The network design decisions concerning each scenario
behavior, could largely affect the design and operation aspects within the model framework include: (1) Allocation of primary
of the future HSC. Therefore, this paper considers the case energy sources, (2) Assignment of location, number, type, size
where demand is not known exactly but subject to some of different production and storage facilities, (3) Establish-
uncertainty. The paper also explores the concept of incorpo- ment of different types of transportation links (modes), and (4)
rating fueling stations into previous work, as well as the Determination of production rate, average stored inventory,
distribution of hydrogen from the production plants or storage and flow rate of hydrogen and primary energy sources. The
facilities to these fueling stations. design decisions associated with the first-stage is made prior
This work makes use of a scenario-based approach to to the realization of uncertainly and is usually referred to as
capture uncertainly in the hydrogen demand. This uncer- “here-and-now” decisions. The future decisions correspond to
tainty is characterized by a set of distinct realizations gener- those decisions made after the uncertainty is unveiled and are
ating what is known as a scenario tree. A scenario tree typically known as “wait-and-see” decisions. The model also
comprises a number of nodes at which branching occurs. Each explores the addition of a new echelon, i.e. fueling station, and
node determines the start and the end points of a particular reveals a number of new decisions. These decisions include
time period. Paths from the root to the leave nodes (i.e. from the determination of the total number of fueling stations
first to last time period) represent a scenario. Each scenario required to be installed into a particular grid, as well as the
encompasses a number of decisions that needs to be resolved. cost associated with operating these stations. Another deci-
Using an S-shape demand trajectory [6,27e29], a scenario tree sion is estimating the investment needed for the distribution
of the form shown in Fig. 1 was obtained. Within the figure, of hydrogen within a grid. The objective of the multi-stage
only three time periods (stages) and nine distinctive scenarios modeling under uncertainty is to choose the optimal deci-
were considered. Each scenario has a definite demand value sion variables that will minimize the cost of the first-stage and
and probability of occurrence. Based on these two measures, the expected cost of the subsequent stages.

10
Low 3. Model structure
Hydrogen demand (million kg/d)

8 Medium
Stage 3
High
The superstructure of the proposed model consists of 7 main
6 components: grid squares, production plants, storage facili-
ties, transportation modes, product physical forms, primary
4 energy sources, and fueling stations. A brief description of
Stage 2 each of these components is summarized below.
2
Stage 1 3.1. Grid squares
0
2005–2010 2011–2016 2017–2022
The grid squares represent the geographical location required
Time period (yr)
to map out all the possible configurations of the future HSC.
Fig. 1 e Demand scenario tree of interest. Each grid square will cover an equal size of land area and may
3968 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

enclose different types and sizes of hydrogen production capacity of each transportation option, the delivery distance,
plants and/or storage facilities. Also, a grid square may and the minimum and maximum flow rates of products
contain several types of primary energy sources which can be between grids. The model also determines the establishment
utilized to produce different physical forms of hydrogen of transportation links between various grids which is justi-
depending on consumers’ needs. The demand of hydrogen for fied based on the cost of the transportation mode versus the
each of these enclosed regions (i.e. grids) will vary depending cost of establishing a new production facility. Moreover, the
on the grid’s population density and fuel consumption rate. transportation decisions include whether to establish a link
Since the hydrogen demand of each grid is estimated based on between grids and what the flow rate of hydrogen and primary
a saturated hydrogen market, a penetration percentage of fuel energy sources should be.
cell vehicles is assumed for the different time periods of the
planning horizons. This penetration percentage is assumed to
follow an S-shape trajectory of the demand profile (see Fig. 1). 3.5. Product physical forms
In addition, it is unlikely that this demand profile will follow
the same pathway. Instead, the demand profile will fluctuate The production plants can produce different physical forms of
deviating from the forecasted path. These fluctuations would hydrogen depending on customers’ specifications. The two
obviously give rise to a large number of demand scenarios. common and practical forms of hydrogen are compressed gas
and cryogenic liquid. These different forms are a key factor in
3.2. Productions plants determining the transportation mode and the storage facility
necessary, as well as the total cost of the future HSC network.
Like other fuels, hydrogen can be produced from various
feedstocks utilizing different production technologies, which
can be tailored to serve a wide range of production scales. The 3.6. Primary energy sources
studied model incorporates both of these features as well as
introduces a lower and upper bound on the production Since hydrogen can be harnessed from a wide range of
capacity. The model also accounts for the existing mercantile primary energy sources, including natural gas, oil, coal,
hydrogen production facilities such as any excess reforming biomass, and renewable electricity, it is essential to identify
capacity at refineries. The establishment of a production the type and location of these energy sources when designing
facility will be determined mainly by the demand of the grid, a hydrogen infrastructure. Therefore, our model incorporates
the failure of the grid to fulfill its hydrogen needs from the geographical location, availability, quantity, and types of
neighboring grids, the trade-offs between establishing plants the primary energy sources. This consequently determines
or transportation links, and whether to establish centralized the type and size of production technologies that are selected.
or decentralized plants. The production decisions that would The main decision associated with the primary energy sour-
be determined by the model include: the number, location, ces is whether to import resources from neighboring grids or
and capacity of plant types, as well as the total production from external sources, such as another country. Another
rates of hydrogen in each grid square. decision determined by the model is how much of energy
resources will be used to produce hydrogen and how much
3.3. Storage facilities will remain unexploited for other purposes. Moreover, the
model outlines the logistics associated with distributing
Once hydrogen is produced, it must be stored in storage sites energy feedstocks over a vast geographic region.
for a certain number of days in order to serve demand and
supply fluctuations, and plant interruptions. The storage
facilities should be designed to handle different physical 3.7. Fueling stations
forms of hydrogen (i.e. liquid or gas) and a wide range of
storage capacity. The model is able to accommodate the state The key factors in the installation of fueling stations are
of the existing storage infrastructure especially the depleted characterizing the required demand and the form of product
oil or gas fields. The establishment of storage facility in each to dispense. These two factors will determine the size and
grid is essential to satisfy the grid’s local demand and may be type of fueling stations. Since we assume that all grids
independent of the production plant’s location, thus the encounter a deterministic demand represented by a distinct
storage facility serves as a distribution terminal. The storage scenario, the establishment of fueling stations is imperative to
decisions captured by the model include determining the replenish consumers with their needs. Each type of fueling
number, location and capacity of storage types, and the total station will have a fixed capacity as well as capital and oper-
average amount of hydrogen stored in each grid. ating costs. The fueling decision includes the determination of
the number of fueling stations required to be installed in each
3.4. Transportation modes grid. Due to the introduction of fueling station, the establish-
ment of transportation modes will take place between
A number of different transportation modes can be used to production plants or storage facilities and fueling stations.
deliver hydrogen from production facilities to storage sites This establishment will happen in all grids due to the presence
and finally to the fueling stations. The transportation means of production plants and/or storage facilities. The trans-
include a wide range of options such as pipeline, trucks, rail, portation modes will have a specific holding capacity, and
etc. The model of interest takes into consideration the maximum and minimum allowable flow rate.
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7 3969

X 
Aegtk ¼Aegkðt1Þ þ Iegtk þ QEeg0 gtk  QEegg0 tk
4. Model formulation g0
X (6)
 gepj Ppjigtk ce;g;t;k : tst1
A detailed description of the model notation is outlined in p;j;i
Table A.1 in Appendix A. The following section discusses the
model constraints and objective function. The first part of the where Aegkðt1Þ is the average availability of primary energy
section summarizes the constraints that are used in the sources at the end of the previous time period.
model, while the second part presents the objective function A material balance on a grid g is written for each primary
equations. energy source e during time period t and scenario k:
!
X X X
Iegtk þ QEeg0 gtk SSF ¼ QEegg0 tk þ gepj Ppjigtk ce; g;t;k (7)
4.1. Demand constraints g0 g0 p;j;i

The total demand of hydrogen will start at a low level and then The first and second terms of the left-hand-side of Eq. (7)
sharply increase until it reaches the saturation point. This represent the import of primary energy sources from over-
pattern can be described by the following equation: seas and neighboring grids, respectively. These terms are
multiplied by a safety stock factor (SSF) to store a small
DTigtk ¼ PFtk DTig (1) inventory of primary energy sources as a buffer against
unpredicted contingencies in energy sources. The first term of
where PFtk is the penetration factor of hydrogen fuel cell
the right-hand-side of the same equation indicates the export
vehicles during a particular time period t and demand
of primary energy sources to nearby grids, while the second
scenario k. This factor is multiplied by the total equivalent
term indicates the amount of primary energy sources
demand in specific region in order to estimate the demand at
consumed by a production plant. The factor gepj represents the
the corresponding scenario and time period.
utilization rate of energy resources by a production plant.
Each grid within the supply chain network will fulfill its
needs by local production facilities or by importing products
from other neighboring grids. Therefore, the local demand, i.e.
satisfied by local production, of product i in a grid g during
time period t and scenario k is expressed by the following 4.3. Production facilities constraints
constraint:
Assuming a steady-state condition during each time period t
DLigtk  PTigtk ci; g; t; k (2) and scenario k, the total production rate of product in
a particular grid is equal to:
The imported demand of product form i to a particular grid g
during time period t and scenario k is equal to the sum of all X 
PTigtk ¼ Qilgg0 tk  Qilg0 gtk þ DTigtk ci; g; t; k (8)
flow rates imported to the grid of interest by all transportation l;g0
modes l and from all neighboring grids g0 :
The total production rate is also equal to the production rate
XX
DIigtk ¼ Q ilg0 gtk ci; g; t; k (3) of all types and scales of production plants established in
l g0 a particular region:
The sum of Eqs. (2) and (3) gives the total demand of X
PTigtk ¼ Ppjigtk ci; g; t; k (9)
hydrogen: p;j

DTigtk ¼ DLigtk þ DIigtk ci; g; t; k (4) The production rate of plant type p and size j producing
product i established in grid g during time period t and
scenario k is constrained by the number of production facili-
4.2. Primary energy sources constraints
ties and the minimum and maximum production limits:

The average availability of primary energy sources e in a grid g pji NPpjigtk  Ppjigtk  PCappji NPpjigtk
PCapmin max
cp; j; i; g; t; k (10)
and scenario k at the end of the first time period t1 is given as
a sum of four terms. These are the initial average availability Distributed plants, i.e. small-scale plants, established in
of primary energy sources, the import of primary energy a region can only fulfill the demand of hydrogen in that
sources from overseas, the transportation of primary energy particular grid. Therefore, products cannot be exported to
sources between grids, and the rate of consumption of these neighboring grids but only satisfy local demand via the
sources. The terms are expressed respectively by the following constraint:
following constraint: X
Ppjigtk  DLigtk ci; g; t; k : j ¼ small plants (11)
X 
Aegt1 k ¼ A0eg þ Iegt1 k þ QEeg0 gt1 k  QEegg0 t1 k p

g0
X The following constraints make sure that plants with
 gepj Ppjigt1 k ce; g; k (5) certain production technologies cannot be built in a particular
p;j;i
size due to technical and economical limitations of the tech-
The average availability of primary energy sources at the nology [6]. The following constraints demonstrate such
next time periods is equal to: restrictions:
3970 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

where b represents the average number of days’ worth of

storage in order to account for demand fluctuations.
NPpjigtk ¼ 0 ci;g;t; k : p ¼ coal gasification & j ¼ small plants
The average inventory of product form i kept in grid g by
(12)
storage type s and size j during the time period t and scenario k
NPpjigtk ¼ 0 ci; g; t; k : p ¼ biomass gasification & j
cannot exceed certain limits:
¼ small plants (13)
sji NSsjigtk  Ssjigtk  SCapsji NSsjigtk
SCapmin cs; j; i; g; t; k
max
(25)
NPpjigtk ¼ 0 ci; g; t; k : p ¼ electrolysis & j ¼ large plants (14)
In a similar manner to the above equation, the total average
stored inventory of a product is:
4.4. Transportation constraints
X X
sji NSsjigtk  Sigtk 
SCapmin ci; g; t; k
T
SCapmax
sji NSsjigtk (26)
Local transportation of a product will only take place if the grid s;j s;j

contains at least one production plant. The distribution of the

product depends significantly on the average delivery distance 4.6. Time evolution constraints
within a grid and the local demand satisfied by the total number
of plants established. Therefore, the average local delivery During the early stages of the transition to a hydrogen
distance within a grid during time period t and scenario k is: economy, hydrogen is expected to be supplied from existing
petroleum refineries, chemical complexes or chlor-alkali
X 2LLg plants. Hydrogen could also be obtained from an existing
gtk ¼
Lave Umgtk cg; t; k (15)
m
m1 storage facility. Taking these options into account, the
number of production plants and storage facilities in
The total number of plants established in each grid g during
a particular grid and during the first time period are given by
time period t and scenario k is given as follows:
the following constraints, respectively.
X X
NPpjigtk ¼ ðm  1ÞUmgtk cg; t; k (16)
p;j;i m NPpjigt1 k ¼ NP0pjig þ IPpjigt1 k cp; j; i; g; k (27)

To obtain a meaningful result, the binary variable Umgtk is

NSsjigt1 k ¼ NS0sjig þ ISsjigt1 k cs; j; i; g; k (28)
enforced by:
X The parameters NP0pjig and NS0sjig are the number of existing
Umgtk ¼ 1 cg; t; k (17)
production plants and storage facilities, respectively. The
m
variables IPpjigt1 k and ISjigt1 k are the number of new production
The local demand satisfied by the total number of plants plants and storage facilities that need to be built early during
established for each scenario k and during time period t is the first time period at any given scenario.
expressed by the following constraints: As the network evolves over time, new production plants
  and storage facilities will be invested in to meet the increased
DMLmgtk  DLigtk  p 1  Umgtk ci; m; g; t; k (18)
demand. Therefore, the number of production plants and
The above equation is written in this form to insure line- storage facilities in each grid and during the subsequent time
arity. The subscript m is a counter that is equal to the number periods is equal to the number of the previously established
of plants established in a particular grid plus one (m ¼ 1, ., facilities plus the number of new invested plants and storage
NPmax þ 1) and Umgtk is a binary variable that determines if facilities. This can be captured by the following constraints:
m  1 plants are established in grid g. The term m  1 is
introduced to allow some grids to have no production plants. NPpjigtk ¼ NPpjigkðt1Þ þ IPpjigtk cp; j; i; g; t; k : tst1 (29)
The construction of a regional distribution network is
stipulated by the following constraints: NSsjigtk ¼ NSsjigkðt1Þ þ ISsjigtk cs; j; i; g; t; k : tst1 (30)

Qilmin Xilgg0 tk  Qilgg0 tk  Qilmax Xilgg0 tk ci; l; g; g0 ; t; k : gsg0 (19) 4.7. Non-anticipativity constraints

Xilgg0 tk þ Xilg0 gtk  1 ci; l; g; g0 ; t; k : gsg0 (20) The tree structure exhibited in Fig. 1 is represented in that
form through a condition known as non-anticipativity. The
Yigtk  Xilgg0 tk ci; l; g; g0 ; t; k : gsg0 (21) non-anticipativity principle, which was introduced initially by
Wets [30] states that if a set of scenarios have the same
available information up to time period t then the values of the
Zigtk  Xilg0 gtk ci; l; g; g0 ; t; k : gsg0 (22)
variables corresponding to these scenarios are identical up to
time period t. Since the demand trajectory for the first time
Yigtk þ Zigtk  1 ci; g; t; k (23)
period is assumed to be deterministic, such non-anticipativity
4.5. Storage facilities constraints principle will lead to one discrete scenario. The decision
variables associated with this discrete scenario will be similar
The total average inventory of product i in grid g at time period up to the first time period. The following constraints guar-
t and scenario k is: antee this condition:

STigtk ¼ bDTigtk ci; g; t; k (24) Vot1 k ¼ Vot1 ðkþ1Þ co; k : k < 9 (31)
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7 3971

where V is any decision variable presented in the model. The obtained by multiplying the storage capital costs by the total
index o denotes other indices incorporated in a particular number of new storage facilities. These costs are represented
variable such as e, g, g0 , i, p, j, s, and m. Obviously, only some of by the following objective function terms.
these indices will appear depending on the variable. !!
The demand uncertainty encountered in the second time X PRk X X
FCC ¼ PCCpji IPpjig þ SCCsji ISsjig (37)
period is presented by nine scenarios which are ‘lumped’ j;i;g;t;k
LR p s

together forming three different sets of scenarios. The deci-

where the first part of the right-hand-side of Eq. (37) denotes the
sion variables involved in each of the three ‘lumped’ scenarios
capital cost of production plants while the second part denotes
are identical up to that time period. Such similarity in these
the capital cost of storage facilities. The learning rate (LR) is
scenarios is ensured via the following non-anticipative
introduced to take into account the reduction in the cost of
constraints:
production and storage exogenous technologies as experience
Vot2 k ¼ Vot2 ðkþ1Þ co; k : k < 3 (32) accumulates with time. This coefficient is defined as follows:

LR ¼ 1 þ ½kðt  1Þ (38)

Vot2 k ¼ Vot2 ðkþ1Þ co; k : 3 < k < 6 (33)
where the term t  1 represents the cost reduction with time
Vot2 k ¼ Vot2 ðkþ1Þ co; k : 6 < k < 9 (34) and k is the percentage of the corresponding reduction per
year. PRk is the probability of occurrence of the demand
In the last time period, there will be a unique set of variables scenarios in reality.
for each of the nine scenarios. These sets of variables will yield
into nine different hydrogen network configurations.
4.10. Transportation capital cost

4.8. Fueling constraints

The total transportation capital cost for all scenarios and time
periods is equal to:
The number of fueling stations within a grid g dispensing
!!
a product form i during time period t and scenario k depends P DMLmgt TMCil 2LLg
on the total equivalent demand and the installed capacity of TCC ¼ PRk þ LUTl
i;l;m;g;t;k ðm  1ÞTMALl TCapil SPLl
the fueling stations, as follows: !!
X Qilgg0 t 2LRgg0
þ PRk þ LUTl
DTigtk X i;l;g;g0 ;t;k
TMARl TCapil SPRl
NFSigtk ¼  NPpjigtk ci; g; t; k : j ¼ small plants (35) !!
FCap X
p DTigtk Wngtk 2LLg
þ PRk þ LUTl (39)
The second term of the right-hand-side of Eq. (35) is brought i;l;n;g;t;k
ðn  1ÞTMALl TCapil SPLl
in to account for the establishment of distributed plants with
The first and second terms in Eq. (39) represent the capital
on-site production and dispensing facility; otherwise the term
transportation cost associated with the distribution of
will be taken as zero.
hydrogen within a grid (local distribution) and between grids
The second part of the Section 4 discusses the cost
(regional distribution), respectively. For more detail on the
components of the objective function. As mentioned earlier,
derivation of these two cost terms, refer to our previous work
the objective of our optimization problem is to minimize the
[25,26]. The last term in Eq. (39) represents the transportation
expected value of the network cost (both capital and oper-
capital cost associated with the secondary distribution of
ating) taken over all the scenarios. This minimization is ach-
hydrogen within a grid. The term (n e 1) is introduced to
ieved by assuming that each scenario k has a known
determine the average delivery distance within a grid and is
probability of occurrence in practice denoted by PRk. The
expressed as follows:
probabilities of the different demand scenarios are expressed
X X X
as follows: NPpjigk þ NSsjigk ¼ ðn  1ÞWngtk cg; t; k (40)
p;j;i s;j;i n
X
NK
PRk ¼ 1 (36) where the first and second term of the left-hand-side of Eq.
k¼1
(40) denote the total number of production and storage facil-
ities, respectively. Wngtk is a binary variable to determine if
4.9. Facility capital cost
there are n  1 plants and storage facilities established and to
exclude all other combinations (n ¼ 1, ., NPmax þ NSmax þ 1).
The facility capital cost over the entire planning horizon is
To assure that only one combination of Wngtk is active, the
related to the establishment of new production plants and
following constraint should be written:
storage facilities at each time period t and scenario k. It was
X
assumed that no capital cost is associated with the estab- Wngtk ¼ 1 cg; t; k (41)
lishment of fueling stations as the current gasoline stations n

will be converted to hydrogen stations. The conversion cost of

these stations is assumed to be negligible. Nevertheless, the 4.11. Facility operating cost
capital cost of the production plants is obtained by multi-
plying the plant capital costs by the total number of new The facility operating cost for all scenarios during the entire
plants. Likewise, the capital cost of storage facilities is planning horizon is given as follows:
3972 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

!
X X X 4.14. Expected total network cost
FOC ¼ PRk UPCpji Ppjigt þ USCsji Ssjigt þUFCi NFSigtk FCapi
j;i;g;t;k p s
The expected total daily cost of hydrogen network is mini-
(42)
mized as follows:
It must be noted that the last cost term of the above equa-   
1 FCC þ TCC
tion denotes the operating cost of the fueling stations. minTDCexp ¼ min þ FOC þ TOC þ ESC (48)
NTP aCCF

4.12. Transportation operating cost

5. Case study: Great Britain
The transportation operating cost is categorized into: fuel,
labor maintenance, and general costs. The daily fuel cost for
The case study presented in our previous work [25,26] is
all scenarios and time periods is equal to:
revised here to capture uncertainty in hydrogen demand. The
! stochastic case study is also expanded to incorporate fueling
P 2LLg DMLmgt stations. In this paper, three different hydrogen supply
FC ¼ PRk FPl
i;l;m;g;t;k ðm  1ÞFELl TCapil configurations are examined, which are summarized as
!
P 2LRgg0 Qilgg0 t follows:
þ PRk FPl
i;l;g;g0 ;t;k FERl TCapil
!  Configuration 1: Production of liquid hydrogen in small,
X 2LLg DTigtk Wngtk
þ PRk FPl (43) medium, and large plants via steam methane reforming,
ðn  1ÞFELl TCapil
i;l;n;g;t;k coal and biomass gasification, and water electrolysis tech-
nologies. Tanker trucks are used to deliver hydrogen to
The labor cost for all scenarios and time periods is equal to:
three different sizes of stationary storage facilities referred
!!
P DMLmgt 2LLg to as small, medium, and large liquid hydrogen tanks.
LC ¼ PRk DWl þ LUTl  Configuration 2: Similar to the first scenario, however, both
i;l;m;g;t;k ðm  1ÞTCapil SPLl
!! tanker trucks and railway tank cars will be used for delivery.
P Qilgg0 t 2LRgg0
þ PRk DWl þ LUTl (44)  Configuration 3: Similar to the first scenario in terms of
i;l;g;g0 ;t;k TCapil SPRl
!! production, storage, and transportation technologies but
P DTigtk Wngtk 2LLg
þ PRk DWl þ LUTl fueling stations and secondary transportation are included.
i;l;n;g;t;k ðn  1ÞTCapil SPLl

The maintenance cost for all scenarios and time periods is A key outcome of this exercise is the understanding of the
equal to: data requirements for the application of hydrogen network
! design. In building the proposed model, a number of
P 2LLg DMLmgt engineering-oriented methods were applied to validate the
MC ¼ PRk MEl
i;l;m;g;t;k ðm  1ÞTCapil data as much as possible. These include comparison of similar
!
P 2LRgg0 Qilgg0 t data from alternative sources, comparison with petroleum
þ PRk MEl (45)
i;l;g;g0 ;t;k TCapil supply chain data, first principles modeling, costing conver-
!
P 2LLg DTigtk Wngtk sions and analysis, and thermodynamics and dimensional
þ PRk MEl analysis. These attempts were helpful in generating a defini-
i;l;n;g;t;k ðn  1ÞTCapil
tive data set that was used in our model. Due to the large
The general cost for all scenarios and time periods is equal amount of data, some of the input data required for the
to: abovementioned configurations are described below. Addi-
!! tional cost data of the different hydrogen technologies can be
P DMLmgt 2LLg
found in the work of Almansoori and Shah [31]. It must be
GC ¼ PRk GEl þ LUTl
i;l;m;g;t;k ðm  1ÞTMALl TCapil SPLl noted that some of the assumed data might be optimistic and
!!
P Qilgg0 t 2LRgg0 does not seem closer to the evolution of the current hydrogen
þ PRk GEl þ LUTl (46)
i;l;g;g0 ;t;k TMARl TCapil SPRl market. This is because the initial enthusiasm of introducing
!!
P DTigtk Wngtk 2LLg hydrogen vehicles has cooled down due to technical difficul-
þ PRk GEl þ LUTl
i;l;n;g;t;k ðn  1ÞTMALl TCapil SPLl ties associated with the on-board hydrogen storage, the high
cost of hydrogen delivery infrastructure, and the recent
Finally, the total transportation operating cost is equal to
decline in the price of crude oil. Also the introduction of
the sum of all cost functions in Eqs. (43e46).
alternative fuels such as methanol, ethanol, and biodiesel has
slowed down the progress of the hydrogen economy.
4.13. Primary energy sources cost
5.1. Hydrogen demand
The cost of primary energy sources for all scenarios during the
entire planning horizon is equal to: The penetration of hydrogen fuel cell vehicles for the three
! time periods was assumed to be 5%, 20%, and 50%, respec-
X X X
ESC ¼ PRk UICe Iegt þ UDCe LRgg0 QEeg0 gt (47) tively. Due to the existence of uncertainty, the last two
k e;g;t e;g;g0 ;t
percentages were assumed possibly to fluctuate, deviating
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7 3973

from the predicted path. This will result in nine scenarios each energy sources utilized by each type and size of the produc-
with different penetration percentages, as exhibited in Table tion technologies to generate a unit of hydrogen, i.e. rate of
1. Since this study assumes deterministic behavior in the utilization (gepj). Moreover, the table lists the capital and the
first time period, the penetration factor remains the same for unit production costs (PCCpji & UPCpji) of the hydrogen
all scenarios. The nine scenarios during the second time production technologies mentioned earlier. During the early
period were partitioned into a set of three distinct scenarios. phase of the hydrogen economy, hydrogen will be supplied
Each of these clustered scenarios has the same penetration from existing sources such as petrochemical plants or petro-
factor. For the third time period, each of the nine scenarios leum refineries. In these refineries, hydrogen will be produced
was given a different penetration factor (see Table 1). Based on via steam methane reforming since it is currently the cheap-
these penetration percentages, the average daily demand of est and most common method of production.
hydrogen during each scenario was estimated. This estima-
tion was done by multiplying the corresponding percentages
5.4. Storage facilities
by the total equivalent demand, which was calculated at 100%
market level. The values obtained in tonnes per day (t/d) are
The minimum and maximum storage capacities of liquid
also listed in Table 1.
hydrogen storage facilities for three different sizes are shown
The hydrogen demand of each grid for the nine scenarios
in Table A.4. The table also shows the capital and unit storage
over the first, second, and third time periods is tabulated in
costs for the different sizes of liquid hydrogen storage facili-
A.2 in Appendix A. It can be noted from Table A.2 that that the
ties. The storage facilities were designed to hold 10 days’
demand of hydrogen in each grid during the first time period is
worth of stock of liquid hydrogen. As with the production
equal for all scenarios due to the deterministic assumption.
plants, this study assumes a number of storage facilities to be
However, for the second time period, only the same cluster of
available at the start of the planning horizon. These storage
scenarios (k1ek3, k4ek6, and k7ek9) will have the same
facilities will be located within the same grids as the produc-
demand. On the other hand, the demand value of each grid
tion plants.
and scenario during the third time period varies according to
the penetration percentage assumed.
5.5. Transportation modes
5.2. Primary energy sources
The maximum allowable flow rate of liquid hydrogen by truck
The initial average availability of primary energy sources in or rail is assumed to be equal to the maximum capacity of
each grid was assumed based on the geographical location large plants (960 t/d). The minimum flow rate, however, is
and proximity of feedstocks from the grid. The values are assumed to be equal to a fully-loaded transport unit. Addi-
summarized in Table A.3. The cost values associated with tional parameters concerning the capital and operating costs
distributing these energy sources throughout the hydrogen estimation are listed in Table A.4. The table also shows that
network are listed in Table A.4. These values are calculated by a few parameters depend mainly on whether hydrogen is
performing a detailed evaluation analysis on the availability of distributed locally or regionally. These parameters include
primary energy sources in Great Britain (see reference [31]). transport mode fuel economy, speed, and availability. The
The evaluation analysis also includes an economic study to capital cost of truck transport consists of the tank unit cost,
estimate the cost associated with handling and distributing the truck cab cost and the undercarriage cost.
the energy sources. Since this study allows for a small
inventory of primary energy sources to be preserved, the 5.6. Fueling stations
safety stock factor was assumed to be 95%. This means that at
least 5% of the feedstock must be kept in the respective grid The hydrogen fueling station considered in this study is
for future utilization. designed based on the average size of typical current petrol
stations. The typical size of today’s petrol stations ranges from
5.3. Production plants 100,000 to 250,000 gallons per month [6]. For the purpose of
this analysis, liquid hydrogen based fueling stations will have
The minimum and maximum production capacities a capacity of 5000 kg/d and a utilization rate of 90%. Since this
pji &PCappji Þ of each plant type with respect to size are
ðPCapmin max
study assumes the conversion of the current petrol fueling
given in Table A.4. Table also shows the amount of primary stations to hydrogen, no significant capital investment will be

Table 1 e Scenario penetration factor and average hydrogen demand for different time periods.
Time period, t(yr) Scenario, k (%) & Demand, D (t/d)

k1;D k2;D k3;D k4;D k5;D k6;D k7;D k8;D k9;D

t1 (2005e2010) 05; 670 05; 670 05; 670 05; 670 05; 670 05; 670 05; 670 05; 670 05; 670
t2 (2011e2016) 15; 2009 15; 2009 15; 2009 20; 2679 20; 2679 20; 2679 25; 3349 25; 3349 25; 3349
t3 (2017e2022) 30; 4018 40; 5358 50; 6697 40; 5358 50; 6697 60; 8037 50; 6697 60; 8037 70; 9376
3974 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

required. However, there will be an operating cost which was The network structures in Fig. B.3 demonstrate the
estimated to be 0.39 $/kg [31]. demand level at 30, 50, and 70 percent, respectively. Out of the
nine demand scenarios studied, these network structures
were chosen because they represent the worst, middle and
6. Results and discussion best paths during the third time period. When the demand
grows to 30% (see Fig B.3 (a)), a new large steam methane
For each of the examined hydrogen supply configurations, the reforming plant is installed in grid 16. This plant covers the
network structure for only three demand scenarios over the demand of Wales and the South West of England instead of
three time periods of planning horizon will be presented. utilizing the plants in grid 17. Additionally, grid 14 has another
These demand scenarios are referred to as the “worst”, large plant in order to meet the required demand. Moving to
“middle”, and “best” paths. The corresponding paths are the middle network structure, it is interesting to note that the
labeled as scenario number 1, 5, and 9, respectively. For each production sources were increased from five production
of the examined configuration, there will be 7 different nodes in Fig. B.3 (a) to eight in Fig B.3 (b). Hence, the hydrogen
hydrogen network structures over the first, second and third network started to become more decentralized as demand
time periods. Despite the fact that the third time period level increased. Also at this stage, building a coal gasification
consists of 9 different scenarios, only scenarios 1, 5, and 9 will plant becomes attractive with respect to producing hydrogen
be considered. Starting with the first hydrogen supply via steam methane reforming due to the availability of large
configuration, the network structures resulting from solving quantities of coal. This example is shown in grids 5 and 25
the stochastic MILP model are shown in Figs B.1eB.3 in where each grid has a large-scale coal gasification plant. The
Appendix B. feedstock of these plants was imported from nearby coal
As stated earlier, the demand of hydrogen was assumed to mines located at grids 7, 23 and 24. The last network structure
be deterministic during the first time period. This demand describes the demand of hydrogen when it reaches 70%. As
level is the consequence of 5% penetration of the transport expected, satisfying this demand will lead to a solution with
market by hydrogen fuel cell vehicles. The network structure high a degree of decentralization (see Fig. B.3 (c)).
resulting from this penetration is shown in Fig. B.1. It is The second hydrogen supply configuration studied in this
evident from the figure that the demand in most grids is met study is related to transporting liquid hydrogen via tanker trucks
by medium-to-large steam methane reforming plants, which and railway tank cars. As with the first supply chain configura-
are assumed to exist prior to the hydrogen economy. Besides tion, only the network structures for demand scenarios 1, 5, and
these large plants, grids with low demand, such as 2, 8, 12, 20, 9 over the three time periods will be outlined. These network
and 26, satisfy their needs by establishing small-scale steam structures are shown in Figs B.4eB.6, respectively. To avoid
reforming plants. The feedstock of these plants is obtained repetition, the network structures for the second configuration
from the grid itself if this specific grid contains a natural gas will not be analyzed in depth as was done with the first network
terminal, or from nearby grids abundant with natural gas. The configurations. However, the discussion will focus on the main
figure also shows that the over all demand of hydrogen is factors that led to the differences between the single- and multi-
mainly fulfilled by three production sources located in grids 7, mode network structures. The first factor behind this dissimi-
17, and 29. This finding suggests that there is a high degree of larity is the high storage capacity of railcars in comparison to
centralization during the initial phase of the hydrogen tanker trucks. This implies that railcars would have fewer trips
market. between the production sites and storage facilities. As a result,
During the second time period, the network structures for railcars could be dispatched to customers even at low-to-
only three demand scenarios are examined. The first network moderate utilization tank capacity, whereas trucks will be sent
structure, which is depicted in Fig. B.2 (a), assumes 15% only at full tank capacity. The second factor is the low operating
penetration by hydrogen fuel cell vehicles. In this figure, the transportation cost of railcars. Thus, this would make long
network structure remained similar to the one given in Fig. B.1 delivery distances more favored. In a multi-mode network,
due to the minor increase in the demand level. To manage this truck will be utilized only for short delivery distances.
increment, the production rate of the existing plants was Examining the network structures of the single- and multi-
raised slightly without installing new production facilities. mode cases shows that the type of transportation modes has
However, when the demand reached 20%, only one small a great affect on the arrangement of production plants. In the
steam reforming plant was enough to cope with this demand single-mode network, production plants were usually located
addition (see Fig. B.2 (b)). The third network structure in a way that minimized the delivery distance between
describes the demand at 25% of hydrogen fuel cell vehicles production sources and customers. Also plant production will
entering the current transport market. At this stage, more be increased until demand cannot be met by existing capacity.
production plants were required to meet the corresponding On the other hand, these two issues are more flexible in the
demand. This can be demonstrated in Fig. B.2 (c) where five multi-mode network since delivery distance is not a critical
extra steam reforming plants were established. Of these factor due to the cheap transport cost. The previous discus-
plants, one small and one large plant were built in grid 14 to sion can be illustrated by looking at Fig. B.2 and B.5 (c). Fig. B.2
cover the demand of South Scotland and North England. (c) shows that there is no production plant in grid 5 because
Another large plant was built in grid 20 which satisfies the grid 7 together with grid 14 cover most parts of Scotland with
demand of neighbouring grids instead of importing hydrogen the minimum possible delivery distances. However, in Fig. B.5
from grid 29. The last two small plants were built in grid 29 (c), grids 5 and 14 provide neighboring grids with hydrogen
since this grid has the highest demand. even though this results in longer delivery distances.
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7 3975

The last network configuration addressed in this chapter is 30

the extension of the first network configuration to account for

25
fueling stations and secondary distribution of hydrogen. The

Scenario probability (%)

network structures for this case study will not be shown since 20
they resemble the ones given in Figs B.1eB.3. Only the number
of fueling stations established in each grid is given here. 15

Again, the obtained results are shown for scenarios 1, 5 and 9

10
over the examined three time periods (see Table A.5). The
tabulated figures do not include the number of distributed 5
plants. It can be seen from Table A.5 that grids with low
demand, such as 8, 12, and 16, have no fueling stations due to 0

the presence of distributed plants. These plants replace the

need for installing fueling stations since they operate as
fueling source. However, as the demand of these grids Total daily cost ($/d)

increased, new fueling stations were built due to the failure of

Fig. 2 e Cost of hydrogen network for the base case.
existing plants to cope with the increased demand. In addi-
tion, the table shows that the number of fueling stations is
dependent upon the demand of individual grids. Therefore,
grids with a high demand level will have a large number of 6.1. Risk analysis
fueling stations.
For the three examined HSC configurations, the average A risk analysis was conducted to study the effect of demand
daily cost for the nine scenarios over the entire planning uncertainty in the outcomes of the total network cost. The
horizon is summarized in Table 2. The table also shows the analysis would also describe the distribution of the cost profile
expected total daily cost of the network under all demand for all demand scenarios. The shape of this profile will
scenarios. It can be seen from the table that the middle determine the degree of risk since a wider distribution means
scenario (k5) has the highest cost. This is because the middle a more risky solution. To perform the analysis, the first
scenario was assumed to have the highest probability value, configuration was chosen as the base case for this study.
as indicated by the values within the pretenses in Table 2. This Then, the probability of demand scenarios occurring during
assumption was based on the fact that the demand trajectory the network’s lifetime was plotted versus the average total
is expected to follow this path during the dawn of the daily cost at each scenario, as exhibited in Fig. 2. It can be
hydrogen economy. The studied models were formulated as noticed from Fig. 2 that the cost profile has a normal distri-
MILP problems and solved using GAMS with CPLEX v9.0 code. bution. The middle histogram corresponds to scenario 5
A Pentium 4, 1.8 GHz Dell machine was used to execute the which has the highest probability of occurrence, i.e. 25%. The
run. The computational results of these configurations are average total daily cost of this scenario is about 11 million $/d.
also summarized in Table 2. Although the model has a large The worst-case scenario for this set of data has a cost of 6
dimension due to the introduction of demand uncertainty, the million $/d with a probability of 6.25% while the best-case
optimality gap and time required for solving these configura- scenario has a cost of 14 million $/d with the same proba-
tions are promising. bility of occurrence. The last two cases represent scenarios 1
and 9, respectively. It can be inferred from Fig. 2 that the width
of the cost profile is reasonable despite the wideness of the
demand range.
In order to arrive to a sensible meaning, the average total
Table 2 e Summary of total cost and computational
results for the studied network configurations. cost of the hydrogen network for each scenario was converted
to $/t of hydrogen delivered. This conversion was carried out
Scenario, k Configuration
by calculating the average daily demand of hydrogen for each
1 2 3 scenario over the planning horizon. Then, these values were
k1 (6.25%) 0.39 0.39 1.29
k2 (12.50%) 0.99 0.99 2.19
k3 (6.25%) 0.59 0.60 1.88
k4 (12.50%) 1.09 1.03 2.25 Table 3 e Average total cost of hydrogen network.
k5 (25.00%) 2.73 2.47 4.05
Scenario, k Network cost ($/t) Network cost ($/km)
k6 (12.50%) 1.46 1.44 3.03
k7 (6.25%) 0.72 0.66 2.17 k1 2824 0.180
k8 (12.50%) 1.54 1.55 3.20 k2 2964 0.189
k9 (6.25%) 0.87 0.88 2.68 k3 3002 0.191
Total cost 10.39 10.02 12.21 k4 3038 0.193
Number of constraints 379,820 563,658 430,336 k5 3072 0.195
Number of continuous variables 200,980 231,886 202,060 k6 3074 0.196
Number of integer variables 77,724 108,630 168,606 k7 3114 0.198
Optimality gap (%) 6 3 9 k8 3214 0.204
CPU time (d) 3 3 3 k9 3262 0.208
3976 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7

30 optimal network structures for three different HSC configu-

rations are outlined. In all of these network configurations,
25 steam methane reforming technology is mostly chosen as
a means of producing hydrogen due to its economical benefits.
Scenario probability (%)

20
In the case of natural gas shortages, other production tech-
nologies such as coal gasification and water electrolysis are
15
utilized because of their local availability. These options are
10
more profitable than importing natural gas from foreign
countries. The size of these plants varied according to the
5 local demand of each grid. For instance, during the early phase
of hydrogen introduction distributed forecourt plants are
0 established for the most part, however as demand reached its
0.180 0.189 0.191 0.193 0.195 0.196 0.198 0.204 0.208
saturation point more centralized facilities emerged. The
Unit cost ($/km)
results obtained also show that the type of transportation
Fig. 3 e Cost of hydrogen network per kilometer. mode has a great effect on the structure of HSC network. For
example, the transport of hydrogen via trucks only resulted in
short transportation links with high utilization capacities. On
divided by the daily cost to obtain the cost per tonne delivered
the other hand, allowing both rail and truck for hydrogen
for each scenario, as exhibited in Table 3. Finally, the average
delivery resulted in longer transportation links with low-to-
network cost was expressed in $/km to compare it to the cost
high flow rates.
of the current petrol supply network (see Table 3). It must be
It has been shown from the examined network configura-
noted that the assumptions used to estimate the figures in
tions that the consideration of demand uncertainty may lead
Table 3 can be found in the work of Almansoori and Shah [31].
to significant changes in the structure and cost of the optimal
To visualize the shape of the cost histogram, the proba-
supply chain network. Due to this a risk analysis is carried out
bility of each scenario was plotted versus the cost per kilo-
to evaluate the degree of risk during the existence of uncer-
meter, as shown in Fig. 3. From Fig. 3, the network cost for the
tainty. The solution found suggests a reasonable cost distri-
middle-case scenario is 0.208 $/km. This value is compared to
bution profile despite the wide range of demand.
the cost of the existing petrol network. According to our
calculation, the cost of a petrol network is 0.107 $/km [31].
Using the average annual distance travelled per vehicle, the
cost of hydrogen and petrol infrastructures were estimated to Appendix. Supplementary data
be 3447 and 1780 $/yr/vehicle, respectively. It is evident from
these values that the cost of a hydrogen infrastructure is Supplementary data related to this article can be found online
roughly twice that for untaxed petrol. The main reasons at doi:10.1016/j.ijhydene.2011.11.091.
behind this are the high costs of hydrogen production,
storage, and distribution technologies. However, it is expected
that in the near future the cost of hydrogen technologies will references
decrease enough to compete with its opponent. Despite the
soaring cost of a hydrogen infrastructure, building such an
infrastructure would improve the urban air quality, mitigate [1] Romm JJ. The hype about hydrogen: fact and fiction in the
race to save the climate. Washington: Island Press; 2004.
climate change, and reduce foreign oil imports.
[2] Hoffmann P. Tomorrow’s energy: hydrogen, fuel cells and
the prospects for a cleaner planet. Cambridge: The MIT Press;
2002.
[3] Rifkin J. The hydrogen economy: the creation of the
7. Conclusion worldwide energy web and the redistribution of power on
earth. Cambridge: Polity Press; 2002.
This paper addresses the effect of uncertainty on the design [4] Chui F, Elkamel A, Fowler M. An integrated decision support
and operation of the future HSC network. A scenario-based framework for the assessment analysis of hydrogen
approach is used to capture the uncertainty in demand over production pathways. Energ Fuel 2006;20:346e52.
a long-term planning horizon. This uncertainty is represented [5] Simbeck DR, Chang E. Hydrogen supply: cost estimate for
hydrogen pathways e scoping analysis. Golden: National
by a multi-stage stochastic optimization problem formulated
Renewable Energy Laboratory; 2002. pp. 1e69.
as an MILP model. The first-stage is assumed to have a deter- [6] NRC. The hydrogen economy: opportunities, costs, barriers,
ministic behavior while the second and third stages are and R&D needs. National Research Council and Committee
subject to demand variations. For each demand scenario, the on Alternatives and Strategies for Future Hydrogen
network structure is revealed and its average total daily cost is Production and Use and National Academy of Engineering,
determined. Also, the expected cost of the HSC network over National Academy of Sciences U.S., National Academies
Press Edition; 2004.
all demand scenarios is resolved. In addition, the model
[7] Thomas CE, Kuhn F, James BD, Lomax FD, Baum GN.
accommodates fueling stations and hydrogen distribution
Affordable hydrogen supply pathways for fuel cell vehicles.
logistics as well as a number of other strategic decisions. Int J Hydrogen Energy 1998;23:507e16.
The feasibility of the derived models is tested by using [8] Ogden JM. Prospects for building a hydrogen energy
Great Britain as the backdrop for our analysis. As a result, the infrastructure. Annu Rev Energy Environ 1999;24:227e79.
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 7 ( 2 0 1 2 ) 3 9 6 5 e3 9 7 7 3977

[9] Ogden JM. Developing an infrastructure for hydrogen [21] Kim J, Moon I. Strategic design of hydrogen infrastructure
vehicles: a Southern California case study. Int J Hydrogen considering cost and safety using multiobjective
Energy 1999;24:709e30. optimization. Int J Hydrogen Energy 2008;33:5887e96.
[10] Amos WA. Costs of storing and transporting hydrogen. [22] Guillen-Gosalbez G, Mele F, Grossmann IA. Bi-criterion
Golden: National Renewable Energy Laboratory; 1998. pp. optimization approach for the design and planning of
1e216. hydrogen supply chains for vehicle use with economic and
[11] Guy KWA. The hydrogen economy e developing the environmental concerns. AIChE J 2010;56:650e67.
infrastructure. 6th World Congress of Chemical Engineering: [23] Sabio N, Gadalla M, Gullién-Gosálbez G, Jimenez L. Strategic
Melbourne; 2001. pp. 1e10. planning with risk control of hydrogen supply chains for
[12] Agnolucci P. Hydrogen infrastructure for the transport vehicle use under uncertainty in operating costs: a case
sector. Int J Hydrogen Energy 2007;32:3526e44. study of Spain. Int J Hydrogen Energy 2010;35:6839e52.
[13] Karlsson K, Meibom P. Optimal investment paths for future [24] Guillén G, Mele FD, Bagajewicz MJ, Espuña A, Puigjaner L.
renewable based energy systems e using the optimization Multiobjective supply chain design under uncertainty. Chem
model Balmorel. Int J Hydrogen Energy 2008;33:1777e87. Eng Sci 2005;60:1535e53.
[14] Tzimas E, Castello P, Peteves S. The evolution of size and cost [25] Almansoori A, Shah N. Design and operation of a future
of a hydrogen delivery infrastructure in Europe in the hydrogen supply chain e snapshot model. Chem Eng Res Des
medium and long term. Int J Hydrogen Energy 2007;32: 2006;84:423e38.
1368e80. [26] Almansoori A, Shah N. Design and operation of a future
[15] Ingason H, Ingolfsson H, Jensso P. Optimizing site selection hydrogen supply chain: multi-period model. Int J Hydrogen
for hydrogen production in Iceland. Int J Hydrogen Energy Energy 2009;34:7883e97.
2008;33:3632e43. [27] Hart D, Bauen A, Fouquet R, Leach MA, Pearson PJ,
[16] Lin Z, Ogden J, Fan Y, Chen C. The fuel-travel-back approach Anderson D. Hydrogen supply for SPFC vehicles. London:
to hydrogen station siting. Int J Hydrogen Energy 2008;33: Imperial College London; 2000.
3096e101. [28] Ogden JM. Where will the hydrogen come from? System
[17] Hugo A. Environmentally conscious process selection, design considerations and hydrogen supply. In: Sperling D,
and optimization. Ph.D. dissertation. London, UK: Cannon JS, editors. The hydrogen energy transition: moving
Department of Chemical Engineering, Imperial College toward the post petroleum age in transportation. London:
London; 2005. Elsevier Academic Press; 2004. p. 73e92.
[18] Hugo A, Rutter P, Pistikopoulos EN, Amorelli A, Zoia G. [29] Hart D, Leach MA, Fouquet R, Pearson PJ, Bauen A. Methanol
Hydrogen infrastructure strategic planning using infrastructure e will it affect the introduction of SPFC
multiobjective optimization. Int J Hydrogen Energy 2005;30: vehicles? J Power Sources 2000;86:542e7.
1523e34. [30] Wets RJB. The aggregation principle in scenario analysis and
[19] Li Z, Gao D, Chang L, Liu P, Pistikopoulos EN. Hydrogen stochastic optimization. In: Wallace SW, editor. Algorithms
infrastructure design and optimization: a case study of and model formulations in mathematical programming.
China. Int J Hydrogen Energy 2008;33:5275e86. Berlin: Springer-Verlag; 1989. p. 92e113.
[20] Kim J, Lee Y, Moon I. Optimization of a hydrogen supply [31] Almansoori A. Design and operation of future hydrogen
chain under demand uncertainty. Int J Hydrogen Energy supply chain. Ph.D. dissertation. London, UK: Department of
2008;33:4715e29. Chemical Engineering, Imperial College London; 2006.