You are on page 1of 19

Energy 133 (2017) 424e442

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

A three-stage stochastic programming method for LNG supply system


infrastructure development and inventory routing in demanding
countries
Haoran Zhang a, Yongtu Liang a, *, Qi Liao a, Xiaohan Yan a, Yun Shen a, Yabin Zhao b
a
Beijing Key Laboratory of Urban Oil and Gas Distribution Technology, China University of Petroleum e Beijing, Beijing 102249, China
b
SINOPEC Beijing Oil Products Company, Beijing 100000, China

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

Article history: Despite of the high demand for liqueed natural gas (LNG) in demanding countries, an ideal means for
Received 13 October 2016 establishing the LNG supply system has not yet been found in many regions. In this paper, a three-stage
Received in revised form stochastic programming method has been proposed for LNG supply system infrastructure development
30 March 2017
and inventory routing in demanding countries. The minimum daily cost is set as the objective function;
Accepted 13 May 2017
Available online 17 May 2017
and the cost consists of the delivery cost, liquefaction cost, purchase cost, and construction cost. Under
the constraints of delivery mode, volume, vehicle, time, and infrastructure construction, a multi-scenario
MILP model was established and solved by a hybrid computational method (ACO-MILP), and the optimal
Keywords:
Liqueed natural gas (LNG)
infrastructure development and inventory routing were presented as the result. Finally, the method was
Supply system successfully applied to the LNG supply system along the Yangtze River in China. Furthermore, compared
Three-stage stochastic programming with the other methods, the superiority of the method was veried.
Infrastructure development 2017 Elsevier Ltd. All rights reserved.
Inventory routing

1. Introduction Difculties in optimizing LNG supply system exist in the un-


certainties of LNG price and demand [6], the multiple resources [7],
1.1. Background the variable delivery modes [8], the exible navigation routes [9],
yearly uctuating demands [10], and the simultaneous optimiza-
At present, liqueed natural gas (LNG) becomes a new major tion of infrastructure development and inventory routing. Mean-
energy source in industries, such as power generation and delivery. while, due to the complexity of LNG supply-demand structure,
According to statistics, there are 338.28 billion cubic meters natural there is certain difculty in optimizing LNG supply system, thus
gas traded as LNG around the world in 2015 [1], and the demands solving this kind of problem will be rather complicated [11].
for LNG in many countries have been increasing in recent years. To
smooth the supplyedemand gap, countries have increased their
investment in infrastructure projects of LNG [2]. However, the LNG 1.2. Related work
supply system is currently in a stage of rapid development in many
regions, but not nished yet [3]. With the difculty of optimizing It is known from the above that the optimization of LNG supply
the systems complex supplyedemand structure, the need remains system is a complicated problem on which many scholars have
to determine an optimal layout for LNG terminal storage [4]. researched. Presently, the optimization of LNG supply system is
Moreover, a minimum-cost supply plan must be drafted that sat- generally concluded into two types: the research on inventory
ises the market demand of making the LNG supply system both routing and infrastructure development of supply system.
economical and stable [5]. The LNG inventory routing problem is a special case of the
maritime inventory routing problem (MIRP) combined with several
additional conditions, namely, variable production and consump-
tion rates, contractual obligations, and berth constraints at pro-
* Corresponding author. duction terminals [12]. Andersson et al. [8] studied optimization
E-mail address: liangyt21st@163.com (Y. Liang). relating to delivery planning and inventory management in the

http://dx.doi.org/10.1016/j.energy.2017.05.090
0360-5442/ 2017 Elsevier Ltd. All rights reserved.
H. Zhang et al. / Energy 133 (2017) 424e442 425

LNG supply system, and both delivery planning and inventory are described with a pooling model and the uncertainty in the
management were formulated as mixed-integer problems. system was handled with a multi-scenario, two-stage stochastic
Grnhaug and Christiansen [13] presented arc-ow and path-ow recourse approach. Then, Li and Barton [22] proposed a two-stage
models of a real planning problem. They tested the models stochastic programming for optimal design and operation of en-
through a series of operational planning cases with a 60-day time ergy systems, and used nonconvex generalized Benders decom-
horizon. Afterwards, Nhaug, et al. [14] proposed a branch-and-price position to exploited the model for efcient global optimization.
method for the discrete-time path-ow formulation of the LNG Recently, Li, et al. [23] applied the method proposed by Li and
MIRP. Subsequently, Rakke, et al. [15] proposed a mixed integer Barton [22] to solving the infrastructure development of natural gas
programming formulation of the annual delivery program planning supply system, which could improve the solving efcient. The
problem based on a priori generation of all possible scheduled above research considered multiple uncertainties in natural gas
voyages within the planning horizon. Due to the size and supply system and put forward more efcient solving algorithms;
complexity of the problem, a rolling horizon heuristic is proposed. whereas, the construction time planning for infrastructures along
Goel et al. [16] presented an arc-ow model based on the model of with the demanding increase were not taken into account. And
Goel et al. [11] and proposed new construction and improvement these references only referred the natural gas pipeline delivery
heuristics to optimize LNG ship-schedule and inventory manage- mode.
ment at both production and terminals. Stlhane et al. [17] devel- In summary, there are few scholars optimizing the infrastruc-
oped a modied construction and improvement heuristic ture development and inventory routing of LNG supply system
algorithm to solve the problem of mass shipping and storage of simultaneously, as well as considering the diversity of LNG delivery
LNG, thereby designing an optimization plan. Jokinen et al. [9] and the uncertainties of LNG price and demand.
proposed a mixed-integer linear programming (MILP) model
involving the LNG supply system in small coastal areas. Most of 1.3. Contributions of this work
scholars focused on the LNG maritime delivery inventory routing
while few considered the land delivery (LNG tanker trailer) and the C A three-stage stochastic programming method is proposed
supplementary delivery modes such as natural gas pipeline com- for LNG supply system.
bined with delivered gas liquefaction as for the inventory routing C The multiple sources such as imported LNG, inland produc-
problem of LNG supply system; meanwhile, the uncertainties of ing LNG and pipeline gas are taken into consideration.
LNG price and demand are not discussed as well. C Variable delivery modes, the exible routes, the expanding
On the other hand, many scholars have devoted themselves into demands, and the simultaneous optimization of the infra-
the infrastructure development of LNG supply system. Lin et al. [18] structure development and inventory routing are taken into
addressed the number of LNG receiving terminals needed to be consideration.
constructed before 2020 in China. Kaplan and Yang [19] concluded C A real case study in China is given to demonstrate the
the technical problems involved in the construction of LNG methods practicality.
receiving terminals. Nowadays, the research on the infrastructure
development of LNG supply system is to macroscopically plan the
construction scheme or discuss technical constraints of the con- 1.4. Paper organization
struction based on demanding statistical analysis. There is less
work for the construction planning optimization on the infra- In section 2, the problem to be studied is elaborated, and the
structure development of LNG supply system. objective, known parameters, decision variables and model
However, many scholars investigated the infrastructure devel- assumption of the optimization problem are displayed from the
opment of natural gas supply system, which provides a signicant perspective of mathematical model. In section 3, a proposed
reference for further research. Zhang et al. [20] analyzed the in- three-stage stochastic programming method is introduced. In
uence of natural gas price and demand in Chinese market on the section 4, a LNG supply system along the Yangtze River in China
infrastructure development of natural gas supply system. Li et al. is given as case study. Section 5 analyzes the case study results in
[21] proposed a stochastic pooling problem optimization formu- detail and provides a comparison between the proposed and the
lation to address optimal design and operation of natural gas pro- previous algorithms. Finally, conclusions are provided in
duction network, where the qualities of the owed in the system section 6.

Fig. 1. LNG supply system with multiple sources in demanding countries.


426 H. Zhang et al. / Energy 133 (2017) 424e442

2. Problem description In order to build and solve the model effectively, assumptions
are made as follows:
The main components in LNG supply system is: LNG receiving
terminal (LNG RT), LNG delivery vehicle, LNG reserve storage (LNG C The supply and marketing network from LNG RSs to terminal
RS), some of which is equipped with liquefaction equipment (LE), users is also a complicated scheduling system, but since
inland natural gas liquefaction plant (NGLP) and natural gas pipe- scheduling is generally undertook by the third-party logistics
line (NGP), as shown in Fig. 1. company, this process is not considered in the paper. In this
LNG RS directly connects with LNG terminal consumers, which paper, LNG RS is considered as the terminal of LNG supply
plays a role as short-term storage of LNG coming from LNG RT and system.
NGLP, and LNG liqueed from NGP, then directly sells to consumers C Daily demand in each region and cost items are constants
in local demanding region (DR). within a given year.
As the large LNG storage station, LNG RT is often built on port. It
mainly receives the imported LNG from large oceangoing LNG
3. Three-stage stochastic programming method
carriers, and then LNG is delivered to LNG RS in each DR by LNG
tanker trailers or small LNG carriers. There are many existing sizes
In the model, there are uncertain parameters. To deal with this
of LNG carriers, so the delivery scale of different LNG carrier sizes
problem, an effective method is proposed by stochastically gener-
should be taken into account.
ating series of uncertain parameters based on the Monte Carlo
In demanding countries, NGLP is often built inside an inland gas
theory, thereby establishing the multi-scenario MILP model. The
eld or near large network of NGP. NGLP can also be directly con-
MILP model can be recast in the following compact form:
nected to terminal users in local regions, or delivered to LNG RS in
other DR by LNG tanker trailers or small LNG carriers.   
min F E fOS dB ; oDS ; oCS ; pUS
The source of NGP is often mainland gas eld, and gas will be dB ;oDS ;oCS
sent to downstream cities quantitatively every year. If LNG RS only s:t: fD dB 0
receives LNG from LNG RT and NGLP, then the corresponding scale jD dB  0 (1)

of LNG storage tank can be built only based on handling capacity. fO dB ; oDS ; oCS ; pUS  0
If LNG RS needs to download natural gas from NGP, then the jO dB ; oDS ; oCS ; pUS  0
relevant liquefaction equipment (LE) should be allocated.
LNG RT, NGP, and site selection of NGP are mainly determined by where F represents the objective function of the model; dB repre-
supply source locations and geological conditions of construction sents the vector of the design binary decision variables (i.e. the
areas. However, besides of that, the selection of LNG RS should also decision of stations and LE construction sites, as well as the decision
take demanding variety, delivery cost at source and delivery mode of design scales in this model); oDS represents the vector of the
into account. So when building LNG supply systems, some possible operating discrete decision variables under the scenario s (i.e. the
LNG storage sites (LNG PRSSs) should be previously determined and number of delivery equipment during operation in this model); oCS
then select the optimal LNG RS site and construct plan based on represents the vector of the operating continuous decision vari-
detailed verication. ables under the scenario s (i.e. the purchase and delivery volume in
this model); pUS represents the vector of the uncertain parameters
(i.e. the purchase cost and demanding volume in this model); fOS
2.1. Model requirements represents the objective function under the scenario s; fD and jD
represent the design equality and inequality constraints; and fO
Given: and jO represent the operating equality and inequality constraints.
Certainty To solve the model effectively, many scholars put forward the
two-stage stochastic programming method. But as for the compli-
C Studying horizon. cated system where there are lots of design decision variables, the
C Locations of LNG RT, NGLP, NGP and LNG PRSS. two-stage stochastic programming method is poor in the calcula-
C Constraints of every delivery plan and vehicle data. tion and convergence. Particularly, when the construction time
C Construction cost of LNG RS; equipment cost of liquefaction planning for infrastructures is concerned, the model scale will be
system and delivery cost of each delivery plan. substantially increased since the time item is introduced into the
corresponding design decision variables. And the traditional two-
Uncertainty stage stochastic programming method may be not proper for this
kind of problems. Therefore, we propose a three-stage stochastic
C Demanding data of every DR. programming method as below.
C Purchase cost of LNG and gas at every node. The rst stage: We combine the predictive uncertain parame-
ters into the model to solve the MILP model under certain condi-
Determine: tions. Then the sensitive analysis for each uncertain parameter
should be performed, which means to generate several groups of
C Infrastructure development of each LNG RS: location, scale, MILP models under certain conditions by changing one uncertain
process and time. parameter every time, and to work out the design binary decision
C Inventory routing at each node: deliver method, deliver variables of each group. From the sensitive analysis, if the same
routing, deliver volume and vehicle numbers. design binary decision variable of each group equals with one
another, the variable will be regarded as certain design decision
Objective: variable dBO . While if the same design binary decision variable of
each group differs from one another, the variable will be regarded
Minimize the costs of delivery, handling, purchase and infra- as approximate certain design decision variable dBU .
structure development of LNG supply system during the study The next two stages: We can employ the two-stage stochastic
horizon under various operational and technical constraints. programming modelling mode to divide the model into a master
H. Zhang et al. / Energy 133 (2017) 424e442 427

Table 1
Calculation procedure of ACO-MILP.

analysis: firstly we obtain the result from the analysis and divide the

design decision variable into certain design decision variable d BO and approximate certain

design decision variable d BU .

ACO colony generate km ants ( k is used for the ant number)

and km groups of design decision variable vectors mapping to the of each

ant. The d BOk should be valued according to the analysis and the d BUk

optimal solution min fS s d BO k , d BU k , oD s , oC s , pU s of each group.


oD s , oC s

Fk : define the food density

Fk f M d BO k , d BU k E fS s d BO k , d BU k , oD s , oC s , pU s
1
f M d BO k , d BU k min fS s d BO k , d BU k , oD s , oC s , pU s
sm s
oD s , oC s

While calculation time is less than CALn

Relocate all ants to the best BPOn

For all # km

While the new location vector of the artificial ant doesnt satisfy D and D

Select several the design decision variables mapped to the ants

End while

Solve the sub-model

End for # km

End while
428 H. Zhang et al. / Energy 133 (2017) 424e442

Table 3
Predicted demanding volume of each region in each year (m3).

Demanding region 2017 2018 2019 2020 2021

R1 6585 8002 9963 10,992 11,802


R2 589 981 1067 1415 1808
R3 2026 3762 4452 5788 7021
R4 863 1437 1795 2191 2529
R5 1789 2626 3827 5114 6139

which is only related with design decision variables in the objective


function fOS ; fSS represents the function formed by the sum of all the
items which is only related with operating decision variables in the
objective function fOS .
In this paper, a hybrid computational method (ACO-MILP) which
is an MILP method coupled with an improved ant colony optimi-
zation (ACO) to handle the above problem. The coordinate vectors
of articial ants locations are used for mapping to the vectors of
design decision variables in the model (the binary variables of
station infrastructure development and of delivery mode between
nodes), and the ACO food density is used for representing the
objective function of the model. Firstly, the sensitive analysis of the
rst stage can determine the classication of the certain design
decision variable and the approximate certain design decision
variable. Set the certain design decision variable to be the same as
case study result during sensitive calculation and the approximate
certain design decision variable to be stochastic. In this way, the
initialization of ACO can be nished. Then, combine the vectors of
the obtained articial ants locations into the sub-model and nally
Fig. 2. LNG supply system along the Yangtze River in China.
work out the objective function of the model (i.e. the food density
of each location). And go to the iteration by ACO rules to work out
the optimal solution. The detailed calculation procedure of this
method is shown in Table 1.
model and a sub-model, as follows:

   4. Case study
Master min F fM dBO ; dBU E fSS dBO ; dBU ; oDS ; oCS ; pUS
dBO ;dBU
s:t: fD dBO ; dBU 0 This paper provides a LNG supply system along the Yangtze
jD dBO ; dBU  0   
River in China as an example, as shown in Fig. 2. In this case study,
Sub fSS dBO ; dBU ; oDS ; oCS ; pUS min fSS dBO ; dBU ; oDS ; oCS ; pUS LNG is for the delivery energy consumption of LNG tanker trailers
oDS ;oCS
  and carriers, as well as the production energy consumption of
s:t: fO dBO ; dBU ; oDS ; oCS ; pUS  0 nearby plants. There are LNG RTs, NGLPs and NGP in this supply
jO dBO ; dBU ; oDS ; oCS ; pUS  0 system. Among LNG RTs, six nodes exist: D1 (2018), D2 (2015), D3
(2) (2018), D4 (2018), D5 (2021), and D6 (2017). The numbers in pa-
rentheses are the respective years of commission. These LNG RTs
where fM represents the function formed by the sum of all the items use either LNG tanker trailers or LNG carriers to delivery LNG. The

Table 2
Distances between pairs of nodes (km).

Node S1 S2 S3 S4 S5 S6 S7 S8

D1 1020 (1350) 880 (1614) 1070 (1818) 1150 (1914) 1270 (2078) 1180 (2347) 1330 (2578) 1320 (2973)
D2 600 (775) 500 (1039) 680 (1243) 760 (1339) 900 (1503) 1000 (1772) 1250 (2003) 1200 (2398)
D3 420 (720) 360 (984) 550 (1188) 630 (1284) 800 (1448) 900 (1717) 1080 (1948) 1200 (2343)
D4 600 (585) 700 (849) 680 (1053) 700 (1149) 700 (1313) 950 (1582) 1050 (1813) 1300 (2208)
D5 2020 (2265) 1920 (2529) 1740 (2733) 1650 (2829) 1430 (2993) 1420 (3262) 1200 (3493) 1430 (3888)
D6 1640 (1615) 1540 (1879) 1350 (2083) 1280 (2179) 1050 (2343) 1160 (2612) 950 (2843) 1210 (3238)
F1 1830 (-) 1600 (-) 1520 (-) 1430 (-) 1300 (-) 1040 (-) 1080 (-) 730 (-)
F2 1920 (-) 1700 (-) 1660 (-) 1580 (-) 1450 (-) 1250 (-) 1200 (-) 920 (-)
F3 1660 (-) 1430 (-) 1360 (-) 1280 (-) 1150 (-) 940 (-) 880 (-) 630 (-)
F4 1550 (-) 1330 (-) 1240 (-) 1150 (-) 1000 (-) 830 (-) 760 (-) 480 (-)
S1 e 250 (264) 409 (468) 490 (564) 725 (728) 800 (997) 995 (1228) 1110 (1623)
S2 e e 222 (204) 300 (300) 500 (464) 583 (733) 782 (964) 890 (1359)
S3 e e e 96 (96) 330 (255) 436 (529) 610 (760) 790 (1150)
S4 e e e e 238 (164) 345 (433) 520 (664) 695 (1059)
S5 e e e e e 215 (269) 340 (500) 565 (895)
S6 e e e e e e 245 (231) 390 (626)
S7 e e e e e e e 320 (395)

Note: Overland distance is given rst, followed by waterway distance in parenthesis.


H. Zhang et al. / Energy 133 (2017) 424e442 429

Table 4
Maximum supply volume of each source.

Node D1-D6 F1 F2 F3 F4 S1 S2 S3 S4 S5 S6 S8

Maximum supply volume (m3/y) 2.4  106 3.3  105 3.4  105 3.4  105 3.95  105 2.5  108 0.5  108 0.6  108 5.0  108 17.8  108 4.0  108 0.7  108

Note: For LNG RTs and NGLPs, the volume shown is for LNG; for LNG PRSSs, the volume shown is for natural gas loading from NGP.

Table 5
Basic parameters for each type of delivery vehicle.

Delivery vehicle Production berth time Average velocity Average effective traveling time Loading capacity
(h) (km/h) (h) (m3)

LNG tanker trailers 3.5 70 21 50


10,000-m3 LNG carriers 7 24 21 9850
30,000-m3 LNG carriers 20 31 21 29,550

Table 6 the LNG sources to regions east of S2. The locations of all nodes in
Design scale grade of LE. the presented system are shown in Fig. 2.
Grade The proposed model was used to evaluate the infrastructure
Design scale (105m3/d) 7 10 16 21 26
development of the LNG supply system and LNG delivery from 2017
CPLS (109CNY) 3.5 5.0 8.0 10.5 13.0 to 2021. The overland distance and waterway distance between
pairs of nodes are shown in Table 2.
We assume the demands of the demanding regions conform to
Table 7 the normal distribution. The demanding volume (mean value) of
Design scale grade of LNG RS. LNG in each demanding region predicted by the model are shown
in Table 3. The standard deviation of the demanding volume is 10%
Grade
of its mean value. The maximum supply volumes for each source
Design scale (104m3) 3 4 6 8 10 12 15 18 20 and terminal are shown in Table 4. The basic parameters for each
CSL (106CNY) 8.0 10.7 16.0 21.3 26.7 32.0 40.0 48.0 53.3
vehicle type are shown in Table 5.
The model used 0.17 as the depreciation rate of vehicles and 0.14
as the depreciation rate of LE and basic equipment construction.
group of NGLPs consists of plants F1, F2, F3, and F4, all of which are The LNG wastage during delivery is assumed to be 0.3%. The
in operation. These plants can only delivery LNG by tankers. liquefaction ratio of the equipment is set at 97.9%, which is the ratio
Considering the demand in regions R1, R2, R3, R4, and R5, as well as of the LNG output to the natural gas input. The wastage during the
the construction and delivery conditions, eight locations along the input and output movement is ignored in this model. The turnover
Yangtze River are selected in this study as LNG PRSSs: S1 (R1), S2 rate of LNG RS is set as 15.
(R1), S3 (R2), S4 (R2), S5 (R3), S6 (R4), S7 (R5), and S8 (R4). The The scale of LE and LNG RS falls into 5 grades, and the design
respective demanding regions are indicated in parentheses. All scale and equipment cost corresponding to each grade are shown in
storage locations except S7 are in proximity to a natural gas pipe- Tables 6 and 7. The natural gas and LNG purchase price is uncertain.
line, which makes it possible to load the gas directly into LNG tanks According to the economic analysis and data statistics, we provide
after purication and liquefaction. the price probability of natural gas and LNG, shown as Fig. 3 where
In addition, some LNG carriers have a volume of 10,000 m3, and the maximum is the predicted price. Since the constant develop-
others have a 30,000 m3 volume. The 30,000 m3 LNG carriers are ment of LNG supply chain system leads to the decrease of LNG
limited to certain water resource facilities and thus can only deliver

(a) LNG purchase price (b) Natural gas purchase price

Fig. 3. Purchase price of natural gas and LNG.


430 H. Zhang et al. / Energy 133 (2017) 424e442

Table 8
Other costs of the model.

CPL CPS CSS CSP for LNG tanker CSP for 10,000m3 LNG CSP for 30,000m3 LNG CST for LNG tanker CST for 10,000m3 LNG CST for 30,000m3 LNG
trailer carrier tanker trailer trailer carrier tanker trailer

0.5 2.03  108 1.40  108 1.5  104 30.0  104 40.0  104 7  103 100  103 200  103

Table 9
Calculation results.

# of cont. var. # of disc. var. # of con. CPU time (s) f1 (104 CNY) f2 (104 CNY) f3 (104 CNY) f4 (104 CNY) F (104 CNY)

22,950 10,570 30,485 442.46 104.19 190.88 3300.56 27.65 3623.28

Notes: # of cont. var. represents the number of continuous variables. # of disc. var. represents the number of discrete variables. # of con. represents the number of constraints.

production and supply costs, the predicted LNG purchase price will modes of 10,000 m3 LNG carriers, 30,000 m3 LNG carriers, LNG
drop down. The pipeline delivery distance inuences the pipeline tanker trailers, and gas pipelines. If the LNG PSS node is perceived
gas price, so the its price differs at different places. And in virtue of as the terminal, it needs to receive the LNG from other places,
the increase of energy and labor costs, natural gas pipeline delivery otherwise to construct LE for liquefaction of the delivered pipeline
cost may increase correspondingly, leading to pipeline gas price gas and thereby LNG RS is necessary at this place. If LNG RS is un-
going up. Other costs in the model are shown in Table 8. necessary, it will be denoted as grey parts. The starting point and
end point in the gure denote the delivery start and end and the
5. Results and discussion data denote delivery volume. Moreover, the gure can display the
delivery equipment number required by each delivery task of each
5.1. Calculation result in certainty condition year.
In 2017, LNG RS will be built at sites S1, S4, S5, S6, and S7. LNG RS
We set the uncertain parameters to be the expected values and at S4, S5, and S6 will include LE. Because natural gas cannot be
transformed the uncertain optimization problem into a certain delivered to S7 through pipelines, LNG carriers should be used for
problem. GUROBI is used for solution. The calculation result is delivering liqueed pipeline gas from S5 to S7. In 2020, owing to an
shown in Table 9. expansion in the market demand, LNG RS will be constructed at S8.
The resulting optimal infrastructure development and inventory
routing for the LNG supply system is shown in Fig. 4. Volumes 5.2. Sensitivity analysis for uncertain parameters
delivered by LNG tanker trailers or carriers are of LNG, and volumes
delivered through pipelines are of natural gas. In the gure, the 5.2.1. Sensitivity analysis for LNG price
yellow, blue, red and green parts respectively denote the delivery Due to the increase of demands and the development of supply

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6604.8
2017 S5 1 1794.4
IS of NGP - 373013.3 2419456.6 546537.3
D3 2 2381.2 572.8 1342.3
2018 D4 2 5645.0 1000.1
F4 7 291.5
IS of NGP - 621266.6 2382471.9 547945.2
D2 1 232.8 932.6 418.2
D3 2 6375.4 269.6
2019 D4 3 3053.7 690.4 2900.9
F4 6 249.9
IS of NGP - 675730.3 2383561.6 547945.2
D2 2 4701.2 783.6
D3 4 709.3 334.5 2124.7 3476.6
2020 D4 2 5614.6 146.2 869.3
F3 3 146.5
F4 17 1037.2
IS of NGP - 684931.5 2383562.0 547945.2
D1 1 6622.0
D2 1 1531.4 728.7 221.3
D3 3 3311.4 1851.7 1481.9
2021 D4 4 372.7 1415.4 388.2 4454.2
F3 5 244.2
F4 17 1036.9
IS of NGP - 684931.5 2383562.0 547945.2
10,000-m3 LNG carrier 30,000-m3 LNG carrier LNG tanker NGP Absebce of LNG RS

Fig. 4. Optimal infrastructure development and inventory routing for the supply system.
H. Zhang et al. / Energy 133 (2017) 424e442 431

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6604.8
2017 F4 15 624.6
S5 23 590.7 1169.7
IS of NGP - 2397965.1 546537.3
D3 2 4600.6 971.8
2018 D4 3 3425.4 984.0 573.5 1662.1
IS of NGP - 2382471.9 547945.2
D2 2 1070.2 690.4 1143.1
2019 D3 2 6375.4 269.6
D4 3 3286.5 932.6 2425.9
IS of NGP - 2383561.6 547945.2
D1 1 6622.0
D2 1 1359.3
2020 D3 3 1068.2 1419.3 2124.7 1329.9 83.3
D4 3 3334.8 3310.2
F4 9 376.6
IS of NGP - 2383561.6 547945.2
D1 1 6622.0
D2 1 6645.0
2021 D3 2 1799.1 1813.4 1236.4
D4 3 2194.0 2030.6 1669.3 751.1
S2 2 5406.4
IS of NGP - 2383561.6 547945.2

Fig. 5. LNG import price lowers by 3%.

system infrastructure construction, LNG import price may be LNG import price uctuation impacts on the infrastructure
lower, and at this point, the sensitivity analysis of the lowering development plan as below. When price lowers by 0%e3%, the
LNG price inuence on supply system construction is performed. infrastructure construction of supply system is shown as Fig. 4. In
When LNG import price lowers by 3% and the construction de- 2017, LNG RS will be built at S1, S4, S5, S6 and S7, among which S5
viates to the east accordingly. In R2 area, S3 is selected to and S6 have LE; then in 2020, the LNG RS will be built at S8. When
construct LNG RS and in 2021, S2 will be chosen to construct LNG price lowers by 3%e9%, the infrastructure construction of supply
RS. Use 3000 m3 tankers to deliver LNG. The detailed optimal system is shown as Fig. 5. In 2017, LNG RS will be built at S1, S3, S5,
infrastructure development and inventory routing for LNG supply S6 and S7, among which S5 and S6 have LE, and in 2021, LNG RS will
system is shown in Fig. 5. When LNG import price lowers to 91%, be built at S2. When price lowers by 9%e10%, the infrastructure
LNG RS at S5 will decrease gas delivery from pipeline, and begin to construction of supply system is shown as Fig. 6. In 2017, LNG RS
construct LNG RS at S2 one year ahead. The detailed optimal will be built at S1, S3, S5, S6 and S7, among which S5 and S6 have
infrastructure development and inventory routing for LNG supply LE; then in 2020, the LNG RS will be built at S2.
system is shown in Fig. 6. Shown as Fig. 7, when the LNG import price declines, this system

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6604.8
2017 F4 15 624.6
S5 23 590.7 1169.7
IS of NGP - 2397965.1 547945.2
D2 1 66.8 2633.9
2018 D3 2 6645.0
D4 3 1314.2 984.0 3773.3 573.5
IS of NGP - 547945.2
D1 1 6622.0
D2 1 1359.3
2019 D3 4 3022.1 932.6 2479.2
D4 2 3039.9 1070.2 1443.3
IS of NGP - 547945.2
D1 1 6622.0
D2 1 6645.0
2020 D3 2 2571.5 2353.9
D4 3 1576.5 1419.3 2161.1 1488.1
S2 2 1384.7 1329.9 3641.3
IS of NGP - 547945.2
D1 1 6622.0
D2 1 1857.1 4787.9
2021 D3 3 2557.4 3703.4 384.2
D4 4 800.9 549.6 5294.4
S2 1 1813.4 797.4 1669.3 478.9
IS of NGP - 1257555.4 547945.2

Fig. 6. LNG import price lowers by 9%.


432 H. Zhang et al. / Energy 133 (2017) 424e442

(a) (b)

(c) (d)

(e)
Fig. 7. Impact of declining LNG import price on total cost (a) Delivery cost changes with the declining price of imported LNG (b) Processing cost of pipeline gas changes with the
declining price of imported LNG (c) purchase cost of LNG and natural gas changes with the declining price of imported LNG (d) Construction cost of LNG RS changes with the
declining price of imported LNG (e) Total operating cost of the system changes with the declining price of imported LNG.

will prefer to purchase LNG from LNG RT rather than deliver from development is performed. When the pipeline gas price goes up by
pipelines. The total delivery cost will rise but the processing cost of 2%, in the supply system, purchasing LNG from LNG RT is preferred.
pipeline gas will decrease. When the price lowers by 3% and 9%, In R2 area, S3 should be selected to construct LNG RS and in 2021,
LNG RS constructing plan will also change. The total operating cost S2 should be chosen to construct LNG RS. Use 3000 m3 LNG carriers
of the whole system will decrease gradually. When the price lowers to deliver LNG. The detailed optimal infrastructure development
by 9%, with the increase of deployed 30,000 m3 LNG carriers, the and inventory routing for the LNG supply system is shown in Fig. 8.
delivered volume of pipeline gas will decrease signicantly. Thus, When pipeline gas price goes up by 10%, the infrastructure con-
the total operating cost of the whole system will decline rapidly. struction of supply system should be the same as when the price
increases by 2%. The detailed plan is shown in Fig. 9.
Pipeline gas price variation impacts on the infrastructure
5.2.2. Sensitivity analysis for pipeline gas price
development plan as below. When the price increases by 0%e2%,
Since the increasing of energy and labor cost, natural gas pipe-
the infrastructure construction of supply system is shown as Fig. 4.
line delivery cost may increase correspondingly, leading to pipeline
When the price lowers by 2%e10%, the system infrastructure con-
gas price going up. At this point, the sensitivity analysis of increased
struction is as Fig. 8. In 2017, LNG RS will be built at S1, S3, S5, S6 and
pipeline gas price impacting on supply system infrastructure
H. Zhang et al. / Energy 133 (2017) 424e442 433

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6604.8
2017 F4 15 624.6
S5 23 590.7 1169.7
IS of NGP - 2397993.8 546537.3
D3 2 2990.1 573.5 1092.7
2018 D4 2 5036.0 984.0 625.1
F4 22 916.1
IS of NGP - 2382471.9 547945.2
D2 1 286.7 1070.2 630.6
D3 3 3787.2 690.4 2167.5
2019 D4 2 5588.1 302.0 754.9
F4 22 916.1
IS of NGP - 2383561.6 547945.2
D2 2 5087.4 454.0 312.1
D3 2 5937.7 707.3
2020 D4 4 257.9 2124.7 4262.4
F3 20 513.2 293.0
F4 18 41.6 1036.9
IS of NGP - 2383561.6 547945.2
D1 1 6621.4
D2 1 2580.9 297.0
D3 3 2635.2 1813.4 755.9 1440.4
2021 D4 4 2511.2 4121.8
F3 20 256.6 632.4
F4 18 41.6 1036.9
IS of NGP - 2383561.6 547945.2

Fig. 8. Pipeline gas price increases by 2%.

S7, among which S5 and S6 have LE; then in 2020, LNG RS will be 5.2.3. Sensitivity analysis for LNG demand
built at S8. When LNG demand in every place uctuates, it will affect the
Shown as Fig. 10, when the price of pipeline gas increases the whole infrastructure development of supply system. At this point,
same as the decline of LNG import price, this system will prefer to the sensitivity analysis of increasing LNG demand inuence on
buy LNG from LNG RT rather than deliver it from pipeline. The total infrastructure development of supply system is performed. The
LNG delivery cost will rise while the handling cost of pipeline gas whole infrastructure development of supply system will not be
will lower down. Even though the construction plan of LNG RS will inuenced when LNG demand increases. When demand increases
change, the total infrastructure construction cost remains the same. by 10%, the whole infrastructure development of supply system and
So the total operating cost of the system will increase gradually. inventory routing is shown as Fig. 11. When demand decreases by

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6604.8
2017 F4 26 1081.0
S5 17 590.7 713.4
IS of NGP - 2108956.4 546537.3
D2 2 5671.6 548.0
2018 D3 3 2354.5 984.0 647.2 573.5 2085.9
IS of NGP - 1973808.0 547945.2
D2 2 1308.8 2179.9
2019 D3 2 6375.4 269.6
D4 3 1977.8 1070.2 1275.3 932.6 1389.1
IS of NGP - 2014197.7 547945.2
D2 2 5376.8 505.5
D3 3 4065.3 2579.7
2020 D4 3 1582.9 1419.3 2138.6 1489.4
F3 20 513.2 293.0
F4 18 41.6 1036.9
IS of NGP - 2374773.8 547945.2
D1 1 6622.0
D2 1 2580.4 297.2
D3 3 2635.1 1813.4 750.7 1433.7
2021 D4 4 2516.7 4128.3
F3 20 256.6 632.4
F4 18 41.6 1036.9
IS of NGP - 2383340.9 547945.2

Fig. 9. Pipeline gas price increases by 10%.


434 H. Zhang et al. / Energy 133 (2017) 424e442

(a) (b)

(c) (d)

(e)
Fig. 10. Impact of the increasing pipeline gas price on operating cost (a) LNG delivery cost changes with the increasing pipeline gas price (b) Processing cost of pipeline gas changes
with the increasing price of pipeline gas (c) purchase cost of LNG and natural gas changes with the increasing price of pipeline gas (d) Construction cost of LNG RS changes with the
increasing price of pipeline gas (e) Total operating cost of the system changes with the increasing price of pipeline gas.

4%e10%, in the supply system, LNG RS at S8 will be applied next the approximate certain design decision variable. In the sensitive
year. When demand decreases by 10%, the whole infrastructure analysis, at the nodes such as S2, S3, S4, and S8, some conditions
development and inventory routing is shown as Fig. 12. decide to construct the LNG RS while some do not; therefore, the
Shown as Fig. 13, with the increase of demand, all costs are rising station construction binary variables belong to dBU . As for the
up accordingly. Even though operating cost of each part is non- sources D1, D2, D3, D4, F3 and F4, the conditions need to purchase
linearly increasing, the total operating cost as a whole is LNG at these places but there is difference in inventory routes;
increasing linearly. therefore, these delivery binary variables St;j;i;i0 belong to dBU . The
design decision variable related to the nodes represented by solid
5.2.4. Result of Sensitivity analysis spots is the certain design decision variable dBO . In the sensitive
From the sensitive analysis, we can know the classication of analysis, at the nodes such as S7, S6, S5, and S1, each condition
the certain design decision variable dBO and the approximate needs to construct the LNG RS, and S6 and S5 need to construct LE
certain design decision variable dBU , shown as Fig. 14. The design while others do not; therefore, the station construction binary
decision variable related to the nodes represented by mesh spots is variables of these nodes and of LE belong to dBO . In Fig. 14, at the
H. Zhang et al. / Energy 133 (2017) 424e442 435

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6594.1
D2 1 50.9
2017 F4 15 624.6
S4 9 533.6
S5 21 86.7 101.4 84.3 1349.2
IS of NGP - 684192.8 2438356.0 547945.2
D3 2 5781.8 392.4
2018 D4 3 3046.9 375.6 717.6 2504.9
IS of NGP - 683393.3 2383561.6 547945.2
D2 2 1103.1 92.4 682.2 1112.6 693.3
D3 2 6190.3 454.7
2019 D4 3 3334.8 3310.2
F4 5 218.9
IS of NGP - 684931.5 2383561.6 547945.2
D2 2 2718.5
D3 4 696.7 1396.8 1549.5 2882.1
2020 D4 2 4850.7 476.4 1317.9
F4 3 146.5
IS of NGP - 684931.5 2383561.6 547945.2
D1 1 6622.0
D2 1 6399.2 728.7 221.3
2021 D3 4 910.0 1409.2 1939.0 2190.6
D4 4 2562.1 4082.9
F4 12 499.7
IS of NGP - 684931.5 2383561.6 547945.2

Fig. 11. LNG demand increases by 10%.

sources that are framed, namely D5, D6, F1, and F2, each condition The scenario number determines whether the stochastic samples
doesnt need to purchase LNG; therefore, these delivery binary can accurately represent the model uncertainties. Smaller scenario
variables St;j;i;i0 belong to dBO . number will lead to uncertainties in the model solving result, while
large number will lead to the model scale redundancy which pro-
5.3. Calculation result in uncertainty condition longs calculation time. From this point of view, further analysis is
necessary for the scenario inuence on the solving result. We
Combine the sensitive analysis result into the model of equation valued the scenario number from 1 to 45 and solved the model
(2) to calculate the infrastructure development and inventory respectively. The nal result is shown in Fig. 15. It can be seen the
routing of LNG supply system. According to 2.3 section, the method calculation is stable when the scenario number is 45, which is able
is essentially based on a multi-scenario MILP model. First of all, to represent the model uncertainty. And the objective function
generate several series of uncertain parameters stochastically by nally converges to 3895.026  104 CNY.
the Monte Carlo method, thereby constituting multiple scenarios. During the ACO calculation, we pay attention to the ACO

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 5944.3
2017 S5 1 398.7 1614.9
S6 2 132.9
IS of NGP - 2438356.2 547945.2
D3 2 1578.5 885.6 211.6 1370.4
2018 D4 2 5644.9 1000.1
IS of NGP - 2282143.2 547945.2
D3 3 3109.9 752.5 2290.7
D4 2 5585.9 145.0 914.1
2019 F4 5 249.9
S4 1 98.8
IS of NGP - 670716.5 2383561.6 547945.2
D2 2 1152.5 2238.7
D3 3 3440.5 1534.7 1669.8
2020 D4 2 5329.6 192.5 1122.9
F4 17 707.9
IS of NGP - 684931.5 2383561.6 547945.2
D2 2 4593.0 828.1
D3 4 711.7 2562.9 3370.4
2021 D4 2 5349.0 547.3 748.7
F3 23 549.9 390.7
F4 18 44.1 1036.9
IS of NGP - 684931.5 2383561.6 547945.2

Fig. 12. LNG demand decreases by 10%.


436 H. Zhang et al. / Energy 133 (2017) 424e442

(a) (b)

(c) (d)

(e)
Fig. 13. Impact of varied demand on operating cost (a) LNG delivery cost changes with variation of predicted LNG demand (b) Processing cost of pipeline gas changes with variation
of predicted LNG demand (c) Purchase cost of LNG and natural gas changes with variation of predicted LNG demand (d) Construction cost of LNG RS changes with variation of
predicted LNG demand (e) Total operating cost of the system changes with variation of predicted LNG demand.

convergence process and calculation stability as well. There are two The result shows that when BPOn becomes smaller, the conver-
critical parameters in ACO algorithm: the ant number km of each gence degree of each time is worse than BPOn 15, since the
colony and the distributed optimal location number BPOn every ants distributed to each location increase with BPOn decreases,
time articial ants relocate. Set the scenario number as 45, BPOn as making the ant colony search bias depth as the lead, which
15, and km as 60, 80, and 100. To verify the algorithm stability, we contributes to each calculation greatly inuenced by self-
calculated for ve times and output the convergence process, probability of ACO algorithm. For example, the forth calculation
shown as Fig. 16. It is obvious that the three-stage stochastic pro- has a fast convergence speed while some are not, which
gramming method has a stable ACO convergence and works out the strengthens the calculation difference. When BPOn is large, the
same solution for each calculation. km pays less inuence on the ant colony search biases breath as the lead, which decreases
model convergence. In this paper, km is recommended to be 60 the convergence speed but strengthens the convergence stability
since larger km results in longer ACO calculation. in the last iteration.
Set the scenario number as 45, km as 50, and BPOn as 5 and To more comprehensively discuss the proposed method, two
20. Similarly, to verify the algorithm stability, we calculated for two-stage stochastic programming methods are selected for com-
ve times and output the convergence process, shown as Fig. 17. parison. The rst one is solved by ACO coupled with MILP model
H. Zhang et al. / Energy 133 (2017) 424e442 437

superior to those without sensitivity analysis. The two-stage CM-


MILP is slightly better, but the convergence, stability and solving
efciency are worse than the three-stage method.
Finally, we decided the scenario number to be 45, ACO
parameter BPOn to be 15, and km to be 60. Considering the un-
certainties in natural gas, LNG purchase price and regional de-
mand, the proposed three-stage stochastic programming method
was used for model solution. And the optimal infrastructure
development and inventory routing of the LNG supply system
along the Yangtze River in China was carried out, shown as Fig. 19.
The blue solid arrow indicates the delivery direction of LNG car-
riers, and the red dotted arrow indicates the delivery direction of
LNG tanker trailers. If there is a year next to the node number, it
indicates the start time of station construction or delivery plan at
this node. If not, it indicates construction or delivery should start
from 2017.
The nal proposed infrastructure development plan is to
construct LNG RS of grade , , ,, and respectively at S1, S3, S5,
S6, and S7 in 2017, and LE of grade and will be needed at S5 and
S6 to liquefy the delivered pipeline gas. In 2020, the LNG demand
increase will result in constructing the LNG RS of grade at S8. The
detailed inventory routing plan of LNG supply chain ranges from
2017 to 2021. The LNG RT at D2, D3 and D4 will employ LNG carriers
for LNG delivery from the Yangtze River to S1, S3, S5, S6, and S7.
Meanwhile, in order to satisfy the demand of S7, LNG should be
delivered from the NGLP at F4 and the LNG RS equipped with
liquefying function at S6 to S7 by LNG tanker trailers. And after
2020, it should be delivered from the NGLPs at F3 and F4 to S8 by
LNG tanker trailers since there is the LNG RS constructed at S8. In
Fig. 14. Result of Sensitivity analysis. 2021, the 30,000 m3 LNG carriers will be needed for LNG delivery
from D1 to the S1 due to the demand increase.
To verify the calculation result, we set the uncertain parame-
ters to be the expected values. In the light of the calculated
infrastructure development and inventory routing, the detailed
delivery plan was obtained, shown as Fig. 20. Compared with
Fig. 4, the model solving result which takes uncertainties into
account is quite different from that which does not although the
model parameters are the same, thereby demonstrating the ne-
cessity of uncertainties.

6. Conclusions

In this paper, a three-stage stochastic programming method has


been proposed for LNG supply system infrastructure development
and inventory routing in demanding countries. The minimum daily
cost is set as the objective function; and the cost consists of the
Fig. 15. Relationship between objective function and scenario number. delivery cost, liquefaction cost, purchase cost, and construction
cost. Under the constraints of delivery mode, volume, vehicle, time,
and infrastructure construction, a multi-scenario MILP model was
(ACO-MILP) and without sensitivity analysis. The second is an two- established and solved by a hybrid computational method (ACO-
stage method (CM-MILP) based on an intelligent group algorithm MILP), and the optimal infrastructure development and inventory
with crossover and mutation, which proposed by Zhou et al. [25]. routing were presented as the result.
Similarly, in order to verify the stability of the algorithm, we repeat The method was applied to the LNG supply system along the
the calculation ve times and output the convergence process, Yangtze River in China. Aiming at the predicted uctuation of
shown as Fig. 18. ACO algorithm can get a better solving effect when LNG import price and pipeline gas price, and the variation of
some good initial solutions are already known before calculation LNG demand, this paper addressed the sensitivity of these fac-
[24]. However, if the sensitive analysis is ignored, it will be difcult tors through quantitative analysis on each factors inuence on
to ensure the stochastic initial solutions to be satisfying on account the inventory routing, infrastructure development and total
of lots of binary variables and the complex model structure. Under operation cost of the whole system. According to the sensitive
this circumstance, ACO algorithm will be of slower convergence analysis, the model design variables were classied, thereby
and worse stability. Compared with Figs. 16 and 18, after the one- guiding initialization of the proposed ACO-MILP algorithm and
stage sensitivity analysis, the initial solutions are apparently nally obtaining the solution to this system optimization
(a) km=50 (b) km=80

(c) km=100
Fig. 16. km inuence on solving result.

(a) BPOn=5 (b) BPOn=20


Fig. 17. BPOn inuence on solving result.

(a) Two-stage ACO-MILP (b) Two-stage CM-MILP


Fig. 18. Comparison between two algorithms.
H. Zhang et al. / Energy 133 (2017) 424e442 439

There is still the detailed delivery scheduling plan when LNG RSs
supply to downstream markets, which plays a role in the infra-
structure development and inventory routing of LNG supply system
to some extent. However, in this paper, we perceive the LNG RS as
the terminal and this process as the model assumption, which is
not considered. In the future, the combined optimization of LNG
supply system and downstream market sales plan will be our
further research object.

Appendix A. Mathematical formulation

A.1. Objective function

Using equally divided payments as the dynamic depreciation


method of all infrastructure construction and equipment in-
vestments, the model sets the minimum daily cost as the objective
function with consideration of the cost of construction and all
means of delivery (minF f1 f2 f3 f4 ). The cost includes the
cost of delivery using LNG tanker trailers and carriers, f1 ; the cost of
LNG delivered from the pipeline, f2 ; the cost of purchase from
sources, f3 ; and the depreciation cost of the LNG tanks, f4 .
The operating cost and depreciation during delivery are
included in the delivery cost using LNG tanker trailers and carriers.
This cost is expressed as:
 
PPPP ES
nSs;t;j;i CSTj 365j nSs;t;j;i CSPj
s t j i
f1 s2S; t2T; j2J; i2I
Fig. 19. LNG supply system infrastructure development and inventory routing. tm$sm
(A.1)
problem. Furthermore, the method proposed in this paper was The cost of handling delivered gas from pipelines includes the,
compared with the other two methods to verify the formers LNG processing and deprecation cost of liquefaction and store
superiority. equipment, and infrastructures. This cost is expressed as:

Number of End point


Starting
Year transport
point S1 S2 S3 S4 S5 S6 S7 S8
vehicles
D2 2 6604.8
2017 F4 15 624.6
S5 23 590.7 1169.7
IS of NGP - 2397993.8 546537.3
D3 2 2381.2 573.5 1092.7
2018 D4 2 5645.0 984.0 625.1
F4 22 916.1
IS of NGP - 2382471.9 547945.2
D2 1 226.7 1070.2 237.9 454.9 371.8
D3 2 6190.3 454.7
2019 D4 3 3245.0 477.7 2922.4
F4 22 916.1
IS of NGP - 2382136.3 547945.2
D2 2 5087.4 454.0 312.1
D3 2 5937.7 707.3
2020 D4 4 257.9 2124.7 4262.4
F3 20 513.2 293.0
F4 18 41.6 1036.9
IS of NGP - 2383561.6 547945.2
D1 1 6621.4
D2 1 2580.9 297.0
D3 3 2635.2 1813.4 755.9 1440.4
2021 D4 4 2511.2 4121.8
F3 20 256.6 632.4
F4 18 41.6 1036.9
IS of NGP - 2383561.6 547945.2

Fig. 20. LNG supply system infrastructure development and inventory routing.
440 H. Zhang et al. / Energy 133 (2017) 424e442

PPP
VPs;t;i CPL VPs;t;i  VPmaxt;i SPt;i s2S; t2T; i2IG (A.10)
s t i
f2
tm$sm The volume of LNG purchased at a given node on a given day
!
PP P cannot exceeds its maximum purchase volume. This is expressed
EL CPLScl BPMA ELS SPt;i CPS as:
t i cl
s2S; s2S; s2S; cl2CL
365$tm VPULs;t;i  VPULmaxt;i s2S; t2T; i2IL (A.11)
(A.2)
According to the conservation of capacity, the sum of the input
The cost of purchasing LNG from LNG RT or NGLP and the cost of volume of LNG, the volume of imported and produced LNG, and the
purchasing natural gas are expressed as: volume of natural gas after liquefaction at a given node should
! equal the output volume of LNG plus the demanding volume at the
PP P P given node. This is expressed as:
VPULs;t;i CPULs;i VPs;t;i0 CPGs;i0
f3
s t i i0
s2S; t2T; i2IL; i0 2IG X Xh i
tm$sm 1  lP VLs;t;j;i0 ;i VPULs;t;i 1  lL aVPs;t;i
(A.3) j i0
XX
The deprecation cost of building LNG PRSS includes depre- VLs;t;j;i0 ;i VDs;t;i s2S; t2T; j2J; i; i0 2I (A.12)
j i0
cation cost of LNG store equipment and infrastructure. LNG
store equipment deprecation cost is related to LNG PRSS The total consuming LNG volume in all LNG RSs nodes in one DR
throughout capacity. LNG infrastructure deprecation cost is should be equal to the total demanding LNG volume in this DR.
expressed as: X
  FDRi;k VDs;t;i VTDs;i;k s2S; t2T; i2IPS; k2K (A.13)
PP P i
ES BLPMAt;i;cr CSLcr At;i CSS
t i cr
f4 (A.4)
365$tm
t2T; j2J; i2IPS; i0 2I; cr2CR
A.4. Vehicle constraints

The number of vehicles should be an integer. This number is


A.2. Delivery mode constraints inuenced by several factors, such as the anchoring period, route,
speed, effective moving time, capacity, and daily volume of LNG
The delivery is unilateral in order to save cost. This constraint is delivered. This is expressed as:
expressed as:   
X VLs;t;j;i;i0 Taj 2Li;i0 ;j Rj
0 0  nSs;t;j;i
St;j;i;i0 St;j0 ;i0 ;i  1 t2T; j; j 2J; i; i 2I (A.5)
i0
Vtj Ntj
A maximum value, M, is introduced. According to the denition   
X VLs;t;j;i;i0 Taj 2Li;i0 ;j Rj
of the binary variable for delivery, VLt;j;i;i0 should meet the constraint 
Vtj Ntj
expressed as: i0
1 s2S; t2T; j2J; i; i0 2I (A.14)
MSt;j;i;i0  VLs;t;j;i;i0 s2S; t2T; j2J; i; i0 2I (A.6)
The total number of vehicles at every node should be less than
If the binary variable for delivery is 0, then the number of ve- the limit.
hicles under this delivery plan must be 0, expressed as:
X
X ns;t;j;i  nSSOmaxt;j FOW M s2S; t2T; j2J; i2I (A.15)
M St;j;i;i0  nSs;t;j;i s2S; t2T; j2J; i2I (A.7) i
i0

Inuenced by terrain, vehicles and delivery conditions, only


specic delivery mode can be chosen at some nodes. It is expressed
as: A.5. Transit cycle constraint

St;j;i;i0  FISj;i;i0 t2T; j2J; i; i0 2I (A.8) From other nodes to destination node, transit cycle of every
delivery plan should be no more than the turnover period of LNG
If at the source node, facilities have not been constructed and RS. The constraint is expressed as:
operated, then LNG cannot be exported. The constraint is expressed
  
as: St;j;i0 ;i Taj 2Lj;i0 ;i Rj
b t2T; j2J; i; i0 2I (A.16)
X Ntj
St;j;i;i0  FMCt;i M t2T; i2IL (A.9)
i0

A.6. LNG RS construction constraints


A.3. Volume constraints
If one node needs to receive LNG from other nodes or deliver
The delivered gas from pipelines in every LNG RS should meet natural gas from gas pipeline, then there is a must to construct LNG
the volume limit constraint. It is expressed as: RS at this node. The constraints are expressed as:
H. Zhang et al. / Energy 133 (2017) 424e442 441

PP IL The set number of LNG RT and NGLP


SPt;i St;j;i;i0 IPS The set number of possible LNG storage sites
j i0
At;i  t2T; j2J; i; i0 2I (A.17) j2J The set of numbers representing the delivery modes,
jm$im
where jm is the maximum number
XX k2K The set of numbers representing the demanding regions
At;i  SPt;i St;j;i;i0 t2T; j2J; i; i0 2I (A.18) t2T The set of numbers representing the years, where tm is
j i0 the maximum number
If the binary variables for LE construction and the binary vari-
ables for LNG RS at last year are both 1, then those variables in the Continuous parameters
next year must be 1. The constraints are expressed as: CPGs;i The unit price of the natural gas loaded from the pipeline
at node i in scenario s (CNY/m3)
SPt1;i  SPt;i t2T; i2IG (A.19) CPL The liquefaction cost (CNY/m3)
CPLScl The LE price in grade cl (CNY)
At1;i  At;i t2T; i2IPS (A.20) CPS The basic construction cost of LE (CNY)
CPULs;i The unit price of the LNG from node i in scenario s (CNY/
m3 )
BPMAt1;i;cl  BPMAt;i;cl t2T; i2IG; cl2CL (A.21)
CSLcr The store equipment price in grade cr (CNY)
CSPj The unit price of the vehicle using delivery mode j (CNY)
BLPMAt1;i;cr  BLPMAt;i;cr t2T; i2IPS; cr2CR (A.22) CSS The cost of the LNG RSs basic construction (CNY)
CSTj The daily delivery cost using delivery mode j (CNY)
The processing scale of a liquefaction plant should be larger than ESj The depreciation rate of vehicle of delivery mode j
that in each scenario. EL The depreciation rate of the LE
ELS The depreciation rate of the basic equipment construction
BPMAt;i;cl VPMAcl  VPs;t;i s2S; t2T; i2IG; cl2CL (A.23)
ES The depreciation rate of the LNG RS0 facilities
The design scale of a LNG RS should be larger than the maximum Li;i0 ;j The driving distance between node i and node i0 using
stored LNG of the LNG RS multiplied by the turnover coefcient in mode j (km)
each scenario. Ntj The effective moving time per day using mode j (h)
8 Rj The average speed using mode j (km/h)
< X Xh i Taj The average anchoring period using mode j (h)
BLPMAt;i;cr VPMAcr b 1  lP VLs;t;j;i0 ;i VPmaxt;i The maximum volume of natural gas loaded from the
:
j i0 pipeline to node i in year t (m3)
9 VPULmaxt;i The maximum purchase volume of LNG at node i in year t
=
1  lL aVPs;t;i s2S; t2T; i2IPS; cr2CR (m3)
; Vt The capacity of each vehicle using mode j (m3)
VTDs;i;k The daily quantity of LNG demanded in region k in year t
(A.24)
in scenario s (m3 )
VPMAcr LNG RS design scale of grade cr (m3 )
VPMAcl Liquefaction plant design scale of grade cl (m3 )
Nomenclature a The volume ratio of natural gas to LNG
b The turnover period of the LNG RS
lL The liquefaction loss of natural gas
lP The delivery loss of LNG
Abbreviations
ACO Ant colony optimization
CNY Chinese Yuan Binary parameters
DR Demanding region FMCt;i The binary parameter for the construction completion. If
LE liquefaction equipment the LNG RT at node i is operational in year t, FMCt;i 1;
IS Initial station otherwise, FMCt;i 0
LNG PRSS Possible LNG reserve storage site FDRi;k The binary parameter for the demanding region. If node i
LNG RT LNG receiving terminal is within demanding region k, FDRi;k 1; otherwise,
LNG RS LNG reserve storage FDRi;k 0
MILP Mixed-integer linear programming FISj;i;i0 The binary parameter for the delivery mode. If j delivery
MIRP Maritime inventory routing problem mode is applied between node i and node i0 , FISj;i;i0 1;
NGLP Natural gas liquefaction plant otherwise, FISj;i;i0 0
NGP Natural gas pipeline
Integer parameters
Sets and indices nSSOmaxt;j The total number of available vehicles in year t
cr2CR The set of LNG RS design scale
cl2CL The set of liquefaction plant design scale Continuous variables
i; i0 2I The set of numbers representing the nodes in the supply VDs;t;i The consumption volume of LNG at node i in year t in
system, where im is the maximum number scenario s (m3)
s2S The set of numbers representing the scenarios, where sm VLs;t;j;i;i0 The average daily volume of LNG delivered by mode j
is the maximum number from node i to node i0 in year t in scenario s (m3)
IG The set number of the place where natural gas pipeline VPs;t;i The daily volume of natural gas loaded from the pipeline
can be delivered from pipelines in year t at node i in scenario s (m3)
442 H. Zhang et al. / Energy 133 (2017) 424e442

VPULs;t;i The average daily volume of LNG that all regions purchase supply chain using liquid nitrogen for natural-gas liquefaction. Appl Energy
2016;182:154e63.
from node i in year t in scenario s (m3)
[8] Andersson H, Christiansen M, Fagerholt K. Transportation planning and in-
ventory management in the LNG supply chain. Energy Nat Resour Environ
Binary variables Econ 2010:427e39.
[9] Jokinen R, Pettersson F, Saxen H. An MILP model for optimization of a small-
At;i The binary variable for the LNG RS. If there has
scale LNG supply chain along a coastline. Appl Energy 2015;138:423e31.
constructed LNG RS at node i in year t, Ai 1; otherwise, [10] Mo C, Park HC, Coimbra CFM. Estimation of the building energy loads and LNG
Ai 0 demand for a cogeneration-based community energy system: a case study in
St;j;i;i0 The binary variable for delivery. If node i is delivering LNG Korea. Energy Convers Manag 2014;87:1010e26.
[11] Goel V, Furman KC, Song JH, El-Bakry AS. Large neighborhood search for LNG
to node i0 using mode j in year t, Stj;i;i0 1; otherwise, inventory routing. J Heuristics 2012;18:821e48.
Stj;i;i0 0 [12] Christiansen M, Ronen D. Ship routing and scheduling: status and perspec-
SPt;i The binary variable for the LE0 construction. If there is any tives. Transp Sci 2004;38:1e18.
[13] Grnhaug R, Christiansen M. Supply chain optimization for the liqueed
LE at node i in year t, StPi 1; otherwise, StPi 0 natural gas business. 2009.
BLPMAt;i;cr The binary variable for LNG RS design scale, if the [14] Nhaug R, Christiansen M, Desaulniers G, Desrosiers J. A branch-and-price
construction design scale of node i is the LNG RS of cr method for a liqueed natural gas inventory routing problem. Transp Sci
2010;44:400e15.
grade, BLPMAt;i;cr 1; otherwise, BLPMAt;i;cr 0 [15] Rakke JG, Stlhane M, Moe CR, Christiansen M, Andersson H, Fagerholt K, et al.
BPMAt;i;cl The binary variable for liquefaction plant design scale, if A rolling horizon heuristic for creating a liqueed natural gas annual delivery
the construction scale of node i is the liquefaction plant of program. Transp Res Part C Emerg Technol 2011;19:896e911.
[16] Goel V, Slusky M, Hoeve WJV, Furman KC, Shao Y. Constraint programming for
cl grade, BPMAt;i;cl 1; otherwise, BPMAt;i;cl 0
LNG ship scheduling and inventory management. Eur J Oper Res 2015;241:
662e73.
Integer variables [17] Stlhane M, Rakke JG, Moe CR, Andersson H, Christiansen M, Fagerholt K.
nSs;t;j;i The number of vehicles using delivery mode j in year t A construction and improvement heuristic for a liqueed natural gas in-
ventory routing problem. Comput Ind Eng 2012;62:245e55.
using node i as the starting location to deliver the gas in [18] Lin W, Zhang N, Gu A. LNG (liqueed natural gas): a necessary part in Chinas
scenario s future energy infrastructure. Energy 2010;35:4383e91.
[19] Kaplan A, Yang C. Design considerations for an LNG receiving terminal. In: SPE
annual technical conference and exhibition; 2003.
References [20] Zhang Q, Li Z, Wang G, Li H. Study on the impacts of natural gas supply cost on
gas ow and infrastructure deployment in China. Appl Energy 2015;162.
[1] BP. BP statistical review of world energy. 2015. www.bp.com/ [21] Li X, Armagan E, Tomasgard A, Barton PI. Stochastic pooling problem for
statisticalreview. natural gas production network design and operation under uncertainty.
[2] Chen Z, An H, Gao X, Li H, Hao X. Competition pattern of the global liqueed Aiche J 2011;57:2120e35.
natural gas (LNG) trade by network analysis. J Nat Gas Sci Eng 2016;33: [22] Li X, Barton PI. Optimal design and operation of energy systems under un-
769e76. certainty. J Process Control 2015;30:1e9.
[3] Ma Y, Li Y. Analysis of the supply-demand status of Chinas natural gas to [23] Li X, Tomasgard A, Barton PI. Natural gas production network infrastructure
2020. Pet Sci 2010;7:132e5. development under uncertainty. Optim Eng 2016:1e28.
[4] Wood DA. A review and outlook for the global LNG trade. J Nat Gas Sci Eng [24] Zhou Z, Zhang JY, Liu P, Li Z, Georgiadis MC, Pistikopoulos EN. A two-stage
2012;9:16e27. stochastic programming model for the optimal design of distributed energy
[5] Lee S, Seo Y, Lee J, Chang D. Economic evaluation of pressurized LNG supply systems. Appl Energy 2013;103:135e44.
chain. J Nat Gas Sci Eng 2016;33:405e18. [25] Zhang H, Liang Y, Liao Q, Wu M, Yan X. A hybrid computational approach for
[6] Yoo BY. Economic assessment of liqueed natural gas (LNG) as a marine fuel detailed scheduling of products in a pipeline with multiple pump stations.
for CO2 carriers compared to marine gas oil (MGO). Energy 2017;121:772e80. Energy 2017;119:612e28.
[7] Kim J, Seo Y, Chang D, Yan J. Economic evaluation of a new small-scale LNG

You might also like