International Journal of Disaster Risk Reduction 22 (2017) 431446

Contents lists available at ScienceDirect

International Journal of Disaster Risk Reduction

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

Hierarchical earthquake shelter planning in urban areas: A case for MARK

Shanghai in China

Huiyong Lia, Laijun Zhaob,c, , Rongbing Huangd, Qingmi Huc
a
School of Management, Shanghai University, Shanghai 200444, PR China
b
Sino-US Global Logistics Institute, Shanghai Jiaotong University, Shanghai 200030, PR China
c
Antai College of Economics & Management, Shanghai Jiaotong University, Shanghai 200030, PR China
d
School of Administrative Studies, York University, 4700 Keele Street, Toronto, Canada M3J 1P3

A R T I C L E I N F O A BS T RAC T

Keywords: Proper pre-disaster emergency shelter locations and the corresponding victim allocation contribute to
Emergency shelter planning mitigating disaster loss in densely populated urban areas. The number of victims and their needs in emergency
Earthquake shelters change over the duration of the post-earthquake period, as evidenced by both actual earthquake records
Refuge demand and theoretical analysis. To match the time-varying demand with the shelter planning, the hierarchical shelters
Shelter service
that provide dierent level services should be projected. We rst estimate the varying number of the victims in
Cross-entropy method
shelters during the post-earthquake period, and then model the locations for emergency shelters with a nested
hierarchy, and also model the allocation of victims among shelters. Furthermore, we employ an ecient hybrid
cross-entropy method to solve the location model and develop a better victim allocation scheme with a swap
move to overcome the drawbacks of other allocation schemes. The empirical results from application to the
Xuhui District in Shanghai of China show that emergency shelter planning based on a time-varying demand can
reduce the construction cost of shelters and the averaged evacuation distance traveled by the victims, compared
to the current policy based on the unvarying demand.

1. Introduction establishment of territorial claims, emotional security and privacy [21].

The costs caused by natural hazards are rapidly increasing in urban
areas, particularly in developing countries, where a number of mega-
cities are growing. By 2050, 66% of the world population is expected to
live in cities, where various human activities are concentrated. Thus,
cities are more and more vulnerable to disasters, particularly to
earthquakes, which can strike any city suddenly without warning.
Once an earthquake takes place in a large city, the damage can be
tremendous both in social and economic terms. Without proper
disaster operations management, even an intermediate earthquake
would become a destructive disaster to a city. Although it is not
possible to prevent or predict the next earthquake, its adverse eects
can be mitigated through various actions and strategies, e.g., operating
emergency shelters. By planning emergency shelters eectively at the
mitigation and preparedness phases, the exposure of aected popula-
tion to earthquakes can be reduced dramatically and their resiliency to
the disaster also can be improved [5].
According to the United Nations Disaster Relief Co-ordination
Oce, emergency shelter serves several vital functions: protection
against the disaster, storage of belongings and protection of property,
In the 2011 Thoku earthquake, over 2,000 emergency shelters were
put into use immediately, providing over 360,000 victims with safe
places and essential relief resources, and made better living conditions
for the victims when the earthquake destroyed their houses [19]. There
are various forms of shelters, for example, xed temporary shelters,
self-built shelters, replacing shelters, temporary outdoor shelters, and
community or collective shelters [20]. Existing infrastructure in urban
areas such as schools, gyms, arenas, green space, and fair market can
be transformed to emergency shelters through adapting emergency
facilities. These shelters can be classied into two main kinds:
temporary shelters, which are used immediately after an earthquake
(including tents accompanied by the provision of food, water and
medical treatment); and temporary houses which are constructed in a
temporary location, allowing for a return to normal daily activities (this
can take the form of prefabricated housing) and are usually intended to
be used for a longer time following an earthquake [22].
Because emergency shelters are important infrastructure in post-
earthquake period, their planning has received increasing interest from
both researchers and practitioners. Most research on emergency
shelter planning considers that the total number of victims does not

Corresponding author at: Sino-US Global Logistics Institute, Shanghai Jiaotong University, Shanghai 200030, PR China.
E-mail address: ljzhao70@sjtu.edu.cn (L. Zhao).

http://dx.doi.org/10.1016/j.ijdrr.2017.01.007
Received 14 October 2016; Received in revised form 15 January 2017; Accepted 15 January 2017
Available online 16 January 2017
2212-4209/ 2017 Elsevier Ltd. All rights reserved.
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

change throughout the post-earthquake period. For instance, HAZUS is Table 1

widely used to estimate the number of people who will seek shelters in Characteristics used for classifying the literature.
earthquake scenarios (http://www.fema.gov/hazus-software). The
Demand F Fixed
HAZUS methodology assumes that the number of people who will UF Unfixed
require short-term housing is a function of income, ethnicity, home Objective EC Evacuation Cost
ownership, and age. However, HAZUS may underestimate the shelter CC Construction Cost
needs [28], also may fail to realize that the number of people who Coverage FC Full Coverage
Non-specified
evacuate to shelters actually changes at dierent phases after an Assignment G General case
earthquake. According to the Central Disaster Prevention Council C Closest case
(CDPC)s report on the Niigata Earthquake, Japan in 2004, the
number of aected population who evacuated to shelters was peaked
at over 100,000 on the fourth day after the earthquake and decreased Table 2
to about only 10,000 people by the end of the rst month (http://www. Summary characteristics of modeling efforts.

bousai.metro.tokyo.jp/taisaku/1000902/index.html). The variation of

Arthur year Demand Objective Coverage Assignment
number of victims in shelters is aected by both physical and
behavioral factors. The physical factor focuses on the ability of [26] F EC G
buildings to withstand an earthquake and on the seismic intensity; [8] F CC FC G
[15] UF CC G
the behavioral factor focuses on peoples choices under emergency
[18] UF CC/EC G
scenario: some people would prefer to return to their homes, live with [11] F CC FC G
friends or relatives, or reside in a hotel, rather than evacuate to a public [17] UF EC C
shelter [7]. In urban areas, interrupted water supplies and non- [6] F EC FC G
functioning elevators besides the damaged buildings would aect the [12] UF CC/EC G
[27] UF CC/EC G
seeking-shelter decision of the aected people, and the system dynamic [13] F CC FC C
method is proposed to analyze comprehensively all the factors to [34] F EC FC G
estimate the varying number of victims in shelters [32]. Besides that
the number of victims in shelters varies, their needs for relief supplies
and accommodation facilities would not be the same during the post- For models without coverage constraint, no specied coverage
earthquake period. For example, victims in shelters show their low- constraint exists.
level needs (e.g., instant food, camping tents) at the initial time and Assignment: This relates to the allocation of victims, as previously
high-level needs (e.g., regular food, prefab housing) when they begin to referred, it uses either the solution generated from location proce-
recover. In this study, we consider the changing needs and varying dure (general case) or the solution with Closest Allocation con-
number of the victims in shelters simultaneously, and view them as the sideration (closest case).
time varying refuge demand (includes the needs and number).
Mathematical programming models and optimization models are Given the above modeling attributes (see Table 1), characteristics of
widely used in shelter planning. They can be applied to the shelters the literature are presented in Table 2. The presented models are sorted
location problem and the victims allocation problem. Firstly, determin- based on their year of publication in ascending order to provide better
ing where to locate emergency shelters in urban areas can be insights on the evolution of the eld.
performed using Facility Location Models (FLMs). In an actual The classication of literature presented in Table 2 presents some
evacuation period, emergency shelters perform dierent functions gaps in current literature that our work intends to t in: (1) Neither
and oer dierent levels of service to meet the varying needs of xed demand or unxed demand consider the varying refuge demand
victims. Thus, Hierarchical Facility Location Models (HFLMs) that deal of victims comprehensively. The varying refuge demand indicates the
with multi-level congurations of facility location can be adopted [24]. changing needs and varying number of victims in the shelters, and all
HFLMs have been widely applied to various elds such as educational the references consider the number of victims only. (2) The assignment
systems [29], medicine [25,3], emergency systems [33], and waste used in current references to allocate victims wouldnt achieve the
management [1,14]. To the best of our knowledge, only one study trade-o between distance and capacity consideration, because they
applies HFLMs to model the hierarchical feature of emergency shelters either use the allocation generated by location procedure or the
[6]. Secondly, allocating victims to selected shelters is essentially an solution with closest assignment consideration.
assignment problem. Assignment problems have been extensively Shanghai is a mega-city with more than 24 million people and is the
studied by researchers using FLMs [4]. When FLMs have a capacity largest economic center in China. By 2015, the GDP of Shanghai
constraint, they usually use the allocation solution generated in reached 379.2 billion dollars, with an average growth of 22.3 billion
location procedures [9]. When FLMs do not have a capacity constraint, dollars per year since 2003. Shanghai has widespread slowly accumu-
the most-used type of assignment method is the closest assignment lated soft clay layers, making it a typical soft-soil area with a fragile
constraint [10]. Moreover, the path assignment is adopted to achieve geographic environment. Many tall buildings proliferate in its central
the trade-o between the capacity and the closest distance considera- area, and the surrounding environment of the city lies on several
tion [29]. Next, we intend to provide a framework outlining the moderate-scale earthquake faults. A small area, huge population,
characteristics of the modeling eorts from literature. The character- intensive new construction and a high concentration of economic
istics using which we classify the models include the followings activities can characterize Shanghai; meanwhile it also has great
(summarized in Table 1). seismic risk. Therefore, we select a central district of Shanghai-the
Xuhui district as case study, aiming to give some insights to guide the
Demand: This includes xed and unxed patterns. Fixed demand practice of shelter planning.
The remainder of this article is organized as follows. Section 2
means the number of victims is viewed as a constant, while unxed
demand represents the number is uncertain or stochastic. introduces the methodology used in this paper. Firstly, we propose the
Objective: Two types of objective function are selected: construction estimation method of the changing needs and varying number of
victims in shelters after an earthquake. It is therefore necessary to
cost and evacuation cost (time or distance).
Coverage: For models with coverage constraint, every community determine how shelters should be structured to accommodate these
variations. Secondly, we propose a hierarchal emergency shelters
should be covered by at least one shelter, that is, the full coverage.

432
H. Li et al.
location model and a victim allocation model, to minimize the total
construction cost of the emergency shelters and the total evacuation
travel distance of the victims. Thirdly, we develop a set of algorithms to
solve the models. Section 3 describes an empirical study of emergency
shelter planning for the Xuhui District of Shanghai, and provides a
comparative study between a time-varying demand and an unvarying
demand estimation. Section 4 of this article presents the conclusions,
and suggests directions for future research.
2. Methodology
The current emergency shelter planning research seldom focuses on
the varying demand of victims in shelters after an earthquake. In this
study, we rst observe that the demand varies over time in actual
earthquake evacuation situations, as evidenced by the type of relief
supplies and the housing facilities needed, in dierent post-earthquake
phases. Then, we use an estimation method to quantify the changing
trend of refuge demand, considering the building and behavioral
factors in post-earthquake period. Using the maximum number of
refuge demand at dierent phases as input, we model the hierarchical
shelters location problem and the victim allocation problem by
optimization method. Finally, a novel algorithm is conducted to form
a systematic search of the optimal solution.
2.1. Estimation method of the time-varying demand of victims
In this study, two implications are used to explain the time-varying
feature of the demand of victims in shelters after an earthquake: the
needs of victims in shelters for relief supplies and accommodation
facilities, and the number of them varies over at the dierent phases of
the post-earthquake period. In the WenChuan Earthquake of 2008,
Sichuan province, China, the primary need in rst three days during
evacuation was camping equipment, while more complex household
necessities, such as shower facilities and kitchens, came to the forefront
in the following week [6]. In the YuShu Earthquake of 2010, China,
about 82% of the aected population chose convenience food (e.g.,
instant noodles and biscuits) as the rst meal on the rst day, while less
than 6% ate regular food (e.g., vegetables and meat), and in the
subsequent two weeks, the percentage of the population who wanted
regular food went up to 33% and the demand for convenience food
decreased to 11% [31]. This variation implies a hierarchy of refuge
demand after earthquakes, showing low-level needs (e.g., convenience
food, camping tents) at the initial period and high-level needs (e.g.,
regular food, prefab housing) by the end of the period. In the Niigata
Earthquake of 2004, Japan, according to the CDPCs report, the
number of people who evacuated to shelters kept changing during
post-earthquake period, which is illustrated in Fig. 1. There are logical
reasons for the temporal variance of refuge demand, in our observa-
tions. In the early post-earthquake period, the people will evacuate to
shelters because of harsh conditions in their homes or shortages of
services (e.g., water and power supplies), which are the immediate
consequences of earthquake damage. As the critical infrastructure
repairs are completed at the recovery phase, though, people will return
to their homes and resume their pre-quake activities, except for those
whose homes are destroyed in the earthquakes; they will remain in
shelters awaiting the provision of temporary housing, where they can
survive the long post-disaster period.
There are two main factors aecting the decision made by the
people about whether choose to stay in their homes or move to
emergency shelters: the degree of damage to their houses and the loss
of access to safe water and other key services [32]. Furthermore,
according to the CDPCs report, it is estimated that about 35% of the
residential population will travel to outside the aected areas rather
than evacuate to nearby shelters, contributing to the shrinking number
of the victims in shelters at the late phase of the post-quake period.
They provide some parameters that can be used to estimate the number
433
International Journal of Disaster Risk Reduction 22 (2017) 431446
Fig. 1. Variations in the number of victims after an earthquake in Niigata prefecture,
Japan, 2004.
of victims in shelters aected by a hypothetical earthquake with an
epicenter in the Tokyo metropolitan area, which are selected for
reference in our study: the proportion of the population who will seek
refuge as a function of their homes post-disaster condition and the
proportion of the population who will be evacuated to emergency
shelters during the evacuation phase, as a function of time. In this
study, we adopt the analysis on two factors that aect the action of
victims [32] and the related parameters provided by the CDPCs report,
then propose a new method to calculate the time-varying demand
during the post-earthquake period.
After an earthquake, the damage degree of houses can be categor-
ized simply into three kinds: completely destroyed, partially damaged,
and intact. We set some values to represent the above damage degree,
1 for the ratio of completely destroyed houses, 2 for the ratio of
partially damaged houses, and 3 for the ratio of intact houses. In the
following month after an earthquake, the partially damaged houses are
not to be repaired because the limited recovery eort is concentrated
on lifeline systems (e.g., water, power, gas, etc.) and people who
evacuate from their own partially damaged houses would not go back to
their home because the houses are not suitable for living yet. In this
study, the change of refuge demand for shelters is mainly caused by
recovery of lifeline systems. Therefore, only the number of people
whose houses remain intact with the shortage of essential services is
considered as time-varying. The choice of evacuation made by the
victims also aects the number of people in the shelters. According to
the report compiled by CDPC on the Niigata Earthquake, Japan in
2004, 100% of victims whose houses are completely destroyed would
choose to evacuate, while 50.3% of victims whose houses are partially
damaged would choose to evacuate. We thus use 1 and 2 to represent
the two proportions above, respectively. Because we consider the
number of victims whose houses remain intact with the shortage of
essential services as time-varying, 3 (t ) is proposed to represent the
proportion of these victims that choose to evacuate. 3 (t ) is related to
the shortage of essential services (e.g., water, power, etc.) and the
individual degree of intolerance for this type of shortage. Based on the
practice of Wenchuan Earthquakes recovery and their simulation on
Xuhui district, we use the data of Wang et al. [32] to estimate the
shortage of essential service function (sh (t )) and the individual intol-
erance function (in (t )) by exponential simulation. Then, we give a
formula for predicting the resulting time-varying demand as follows:
di (t ) = (1 1 Popi + 2 2 Popi + 3 3 (t ) Popi ) (1)
where di (t ) represents the refuge demand associated with commu-
nity i at time t ; represents the proportion of aected people who
choose an emergency shelter as their destination instead of traveling to
H. Li et al.
the homes of friends or relatives or to hotels outside the aected areas;
Popi represents the population of community i .
According to Wang et al. [32], the proportions of the shortage of
essential services are 87.07% when the earthquake happens, 58.64%
when one day after earthquake, 57.54% when two days after earth-
quake, and 43.53% when seven days after earthquake. Based on the
practice of Wechuan Earthquake, the proportions of population
aected by the shortage of essential service are 18.8% two weeks after
earthquake and 0% one month after earthquake. We use the above
proportion data to estimate the shortage function sh(t), as shown below,
sh (t ) = 0. 94 exp (0. 15t ) (2)
and the percentage of explained variance is 98%.
Based on the data of the Niigata Earthquake, Japan, the propor-
tions of victims who choose to evacuate as a result of the shortage of
essential service is 36.2% two days after the earthquake and 91% four
days after earthquake. Because people cannot tolerate having no water
for more than seven days, we set the individual intolerant proportion
approximately to 100% ve days after the earthquake. Similarly, We
use the above proportion data to estimate the individual intolerant
function in (t ), as shown below,
in (t ) = min {2 exp (3. 5/t ), 1} (3)
and the percentage of explained variance is 99%. Then,
3 (t ) = sh (t ) in (t ).
Considering the hierarchy of needs of victims in shelters, we divide
the one-month post-earthquake period into three successive phases:
the immediate phase, which we call Phase I (1 day); the short-term
phase, which we call Phase II (2 to 10 days); and the long-term phase,
which we call Phase III (11 to 30 days). Accordingly, we categorize the
time-varying refuge demand into three levels based on this temporal
scale, namely immediate demand (Phase I), short-term demand (Phase
II) and long-term demand (Phase III). When time t belongs to a phase
k , we use the maximum number of time-varying refuge demand at site i
during phase k to represent the level k demand at site i , i.e.,
dik = max di (t ).
t phase k
After classifying the changing needs and varying number of the
victims in shelters into dierent levels demands, it is necessary to
construct dierent types of shelters to respond to these demands.
Therefore, we further divide shelters into three types, based on their
service periods and functions, namely immediate shelter (IS), short-
term shelter (SS), and long-term shelter (LS). According to the national
standard established by the China Earthquake Administration (CEA)
(http://www.csi.ac.cn/publish/main/847/1114/index.html), a higher-
level shelter must be equipped with more comprehensive facilities,
provide a larger area, a larger service radius and a longer service time
than a lower-level shelter. In our settings, a higher-level shelter can
provide all the functions as a lower-level one, but not vice versa. For
example, LS can simultaneously function as both SS and IS, but neither
the SS nor the IS can provide some of the functions of an LS. We also
emphasize that the service hierarchy of shelters must keep up with the
temporal variance in refuge demand. For example, the LS and SS can
provide the same services as the IS during the initial post-earthquake
period, and the immediate demand can be upgraded to a short-term
demand as time progresses. However, the IS cannot provide higher-
level services; the LS and SS would have to take over the short-term
demand. Likewise, when short-term demand upgrades to long-term
demand, the SS cannot supply higher-level services; only the LS will be
left to satisfy the long-term demand.
This hierarchical design of shelters is necessary to implement post-
earthquake evacuation strategies. As time progresses, when the de-
mand cannot be met in the assigned shelter type, some will be directed
to a higher-level shelter where their emerging higher-order needs can
be met, while others will choose to travel outside the aected areas or
just return to their own homes. As for specic seek-shelters move-
ments: immediately after an earthquake (Phase I), harsh living
434
International Journal of Disaster Risk Reduction 22 (2017) 431446
conditions and shortages of basic supplies will cause the aected
people to be evacuated from their places of residence to one of the
nearby shelters (IS, SS, or LS); one day after an earthquake (Phase II),
when IS can no longer accommodate their emerging needs for relief
resources, these people will be transferred to a higher-level shelter (SS
or LS). Ten days after an earthquake (Phase III), similarly, those in a SS
will be relocated to a LS. Along with this evacuation process, some
types of behavior (e.g., travel to the non-aected areas or returning
home with the recovery of basic supplies) also can result in a change in
the number of victims in shelters.
2.2. Two-stages optimization model for hierarchical shelter planning
When a disaster happens, the evacuation assignment can be divided
into two stages: the rst stage is to decide which shelters to go to and
the second stage is assigning the aected people [30]. The selection of
the shelters involves an emergency-facilitated location problem, and
the assignment of the victims involves allocating them to shelters. In
this section, we propose an emergency shelter location model and a
victim allocation model, to address the construction cost and the
evacuation distance problems in actual shelter planning. Because the
budget for building shelters is limited and the construction and
maintenance cost is very high, we rst consider the objective of
minimizing the total construction cost of shelters, given coverage,
capacity and allocation constraints. After the shelter locations are
determined, planners want to ensure prompt delivery of shelter
services, so that the aected people can resume daily activities and
begin the process of recovery as quickly as possible. We also realize that
the aected people would require dierent services at dierent phases
of the post-earthquake period. Therefore we suggested building a
hierarchical shelter system to address the time-varying refuge demand.
Fig. 2 shows the precise three-level structure of shelters and their
hierarchical coverage for serving the dierent levels of demand of
victims.
Before introducing the models for selecting sites for shelters with
this hierarchical structure and designing an allocation scheme for
victims, we make three assumptions:
1) The demand for a given community is concentrated at the physical
site of its community committee. The community committee is the
most basic administrative unit in China and can be representative
of its administrated community. Moreover, the area occupied by
one community is very small compared to the whole city; the
average area of a community in Shanghai is about 0.1 km2,
approximately 1/60,000 of the entire city area.
2) The residents would follow the planners recommended allocation
scheme when they need to be evacuated to an assigned shelter.
Once the emergency shelters are designated, the government would
organize rehearsals for the residents, to educate them on earth-
quake emergency procedures, so that the victims would follow the
planned scheme rather than haphazardly choosing where to go after
an earthquake.
3) Every resident in a given community will be assigned to only one
shelter. For ecient and orderly evacuation to shelters, it is
necessary to view the residents of a community as a whole group,
to be assigned to one shelter.
In order to provide adequate services for the victims, the solutions
obtained from any model should comply with the following two
constraints:
1) Coverage constraint: to ensure that the people are evacuated to
shelters as quickly as possible, it is necessary to evacuate them to a
nearby shelter, with reasonable distance away from their resi-
dences. Hence, the distance between each community and its
assigned shelter should be within the shelters service radius
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

Fig. 2. Visualization of the three levels of emergency shelters location and victim allocation.

(coverage). This is the most important constraint, and should be 2.2.1. Emergency shelter location model (ESLM)
given the highest priority.
2) Capacity constraint: to meet the basic needs of daily life, the area
min j ukyjk
jS kK (4)
required by the victims in shelters should not exceed the shelters
planned area. Because people may tolerate more crowding under an
s. t. xijk = 1, iI , k K
emergency scenario, some limited congestion of a shelter can be
jJ (5)
allowed; however, appropriate limits of such congestion must be
set.
yjk 1, jJ
kK (6)
The notation related to the two-stages model is given below.
k
j
xijk dik yjh,jJ , kK
Index sets: iI
k h =1 (7)
I Set of residential communities
k
J Set of candidate shelters
xijk cisk ysh, iI , jJ , kK
K Set of hierarchical levels (8)
s J h =1
Parameters:
aij Distance between site i and site j
xrjkcijk yjk ,iI , jJ , kK
D k Coverage distance of a shelter during phase k rI (9)
cijk 1 if aij D k ; 0 otherwise
j yjl = 0, jJk , l , k K (l < k , k 2) (10)
Area of candidate shelter j
k Required area per one victim who needs level k shelter service
Construction cost per unit area to build level k shelter yjk , xijk {0, 1},jJ , k K (11)
uk
dik Number of victims from residential community i during phase k Eq. (4) is the objective function of the emergency shelter location
Decisive variables: model, as it indicates that the purpose of the placement is to minimize
y k 1 if site j is located to build level k shelter; 0 otherwise
j
the total construction cost for building the hierarchical shelters. Eq. (5)
helps to ensure that all the aected population of each community is
xijk 1 if site i is allocated to shelter j during phase k ; 0 otherwise
assigned to only one candidate shelter for hierarchical refuge service at
each phase. Eq. (6) provides that there can be at most one level of
shelter constructed at each candidate site (i.e., the co-location of
The mathematical programming formulation of the model for dierent levels of shelters on the same site is prohibited). Eq. (7) puts
emergency shelter locations is dened by the objective function (4) constraints on the holding capacity of the shelters, ensuring that the
and Eqs. (5)(11). served time-varying refuge demand cannot exceed the capacity of all

435
H. Li et al.
the selected shelters at each phase. Eq. (8) indicates that the time-
varying refuge demand of each community must only be served by the
same or higher level shelter within the shelters hierarchical service
radius. Eq. (9) indicates that each selected shelter must serve at least
one communitys time-varying refuge demand within its hierarchical
coverage radius. Eq. (10) suggests that the site of a low-level candidate
shelter set cannot be chosen to build a higher-level shelter. Eq. (11)
gives the binary denitions of these variables.
The solutions for the above model can be divided into two types:
location decisive variable (yjk ) and allocation decisive variable (xijk ).
After yjk is chosen, xijk can be obtained. However, because the ESLM
does not take the optimization of evacuation distance into account,
victims would travel longer evacuation distances to shelters if the
allocation decisive variables of this model were put into eect. In order
to ensure that victims can be evacuated promptly to selected shelters
during dierent post-earthquake phases, using the ESLM, we should
further consider the total evacuation distance traveled by all these
people as an optimization objective, while also complying with the
coverage and capacity constraints. When the solutions are obtained at
the rst stage, we use j* to represent the set of selected shelters, which
the victims of every community should be allocated to.
2.2.2. Victim allocation model (VAM)
min aij dik xijk
iI jJ* kK (12)
s. t. xijk = 1, iI , k K
jJ* (13)
j Its foundations are importance sampling and the Kullback-Leibler
cijk x ijk dik ,jJ *, k K
iI
k (14)
xijk {0, 1},iI , jJ *, k K (15)
Eq. (12) is the objective function of the allocation model, which
minimizes the total evacuation travel distance in the dierent phases.
Eqs. (13) and (14) are similar to those for the ESLM, and represent
coverage and capacity constraints; the major dierence is that the
candidate shelters are already selected in the VAM (i.e. each j belongs
to a j*). Eq. (13) expresses that each community must be assigned to
only one selected shelter, and Eq. (14) ensures the selected shelters
capacity should exceed the overall time-varying refuge demand that is
within its service radius allocated to it. Eq. (15) yields the binary
constraints.
Neither of these two models feasibility could be guaranteed if we
stuck strictly to all the constraints. If the numbers and sizes of the
candidate shelters were small, even the combined candidate shelters
capacities would not satisfy the total demand. And even if the
combined capacities of all the candidate shelters were greater than
the total time varying demand over all phases of the post-earthquake
period, the models, which are subject to the coverage and capacity
constraints, would also generate infeasible solutions. In fact, though, in
a real emergency scenario, a limited amount of congestion is allowed in
the emergency shelters, in order to hold excessive refuge demand. But
since the demand is sensitive to evacuation distance, the coverage
constraint should be given high priority. To accommodate all these
possible scenarios, we need to relax the capacity constraints of the
ESLM and VAMmaking them soft, meaning that they could be
violated, while the coverage constraints of both models must remain
hard, meaning that they should be strictly met. However, we also
want to minimize such violations, and this problem is addressed in the
following solution procedure.
436
International Journal of Disaster Risk Reduction 22 (2017) 431446
2.3. Solution procedure
Researchers have implemented various approaches to solve the
hierarchical location problem, and these can generally be divided into
two types. The rst uses a commercial solver (e.g., CPLEX, XPESS) to
tackle a small-scale problem [29], but solving a large-scale problem
takes much longer time. The second approach uses a non-exact
randomized optimization method like a genetic algorithm and Tabu
search [16,3], which can attain near-optimal solutions in large-scale
scenarios. In our study, we nd that these randomized algorithms
would result in a high ratio of infeasible solutions when they are
applied to the proposed models, costing extra computational eort.
Therefore, we decide to develop a novel randomized algorithm based
on the cross-entropy (CE) method, which would reduce the infeasibility
through its internal mechanisms. Next, we would explain how the CE
method was applied to ESLM, and we also develop an allocation
heuristic to optimize the total evacuation distance of the VAM.
2.3.1. The CE method for emergency shelter location model
The CE method is a relatively novel heuristic algorithm in the realm
of optimization methodologies. It is rstly developed as a method for
estimating the probabilities of rare event in complex stochastic
networks, then applied to tackle combinatorial optimization, triggering
a number of other operation research applications [23]. In its most
basic form, the CE method consists of the repeated execution of the
following two steps:
1) Generate a random sample from a pre-specied probability dis-
tribution function.
2) Use this sample to modify the parameters of the probability
distribution so as to produce a better sample in the next iteration.
distance (or CE), which will be detailed in the following execution of
the CE method. We employ a CE method hybridized with a bin-packing
method to solve the ESLM. Steps include solution representation,
feasibility modication, performance calculation and CE method
execution.
2.3.1.1. Solution representation. ESLMs variables can be divided into
location variables yjk , j J , k K and allocation variables xijk ,
i I , j J ,k K . Among these, we choose yjk as the main solution
because it can decide the allocation variable by using the solution
procedure. Real-number encoding is an easy way to represent the
solution, which generally uses the hierarchical level number k to
represent that the candidate site is selected for a level-k facility [3].
This encoding implies that all the candidate sites are regarded as the
same with the respect to their level choices. In ESLM, however, this
encoding introduces many infeasible scenarios, such as low-level
candidate sites being used for higher-level shelters; this is not
acceptable, based on the coverage constraint (10). To avoid this
impediment to our hierarchical approach, we used hierarchical
binary number encoding instead of real-number encoding; the
dierence between two methods are shown in Fig. 3.
In this gure, the long vertical dashed lines represent the bound-
Fig. 3. Real-number encoding and hierarchical binary number encoding.
H. Li et al.
aries of the dierent levels of shelters and the short vertical dashed
lines separate the dierent solutions in both encoding methods. By
using hierarchical binary encoding, a Level-3 candidate shelter selec-
tion status can be represented by a binary triplet, where the rst entry
represents whether it is selected as a Level-1shelter (1 represents
selected, 0 not selected) and the second and third entries represent the
selections of Level 2 and Level 3 shelters, respectively. This encoding
suggests that a lower-level (e.g., Level 1) choice is always in front of a
higher-level (e.g., Level 3) choice in a given selection status (e.g., a
binary triplet), and the number of level-k choices is related to the
number of entries in the triplet. In the same way, we use pairs of entries
to represent the selection status of a Level 2 candidate shelter and a
single entry to represent whether a Level 1 candidate shelter is to be
built. Although the hierarchical binary number encoding has more
dimensions than real-number encoding, the CE method is insensitive
to the cardinality of the solution, so that a large-scale multi-dimen-
sional solution can be easily tackled with the CE method.
2.3.1.2. Feasibility modication. The solutions generated by the
solution procedure must be checked for their feasibility. If a solution
co-located a shelter (i.e., k K yjk >1), we choose to implement the cut
move, cutting the lower-level choice and retaining the higher-level
choice of shelter to ensure that k K yjk =1. This result can be achieved
by discarding the lower-level choice at every level of examination.
Furthermore, we propose two other requirements to ensure the
feasibility of solutions, while retaining the coverage and capacity
constraints. We adopt a top-down siting scheme to evaluate every
solution, beginning at the highest level (Level 3) and moving down to
the lowest level (Level 1), checking the requirements at every level.
We let Jk J be the set of selected shelters at level k . At every level
of evaluation, we rst make sure that the overall area of the selected
shelter exceeds the overall required area by the victims at this level,
that is,
j k dik , k K
j Jk iI (16)
and this is a relaxation of constraint (7). The coverage constraints,
however, are hard constraints in the ESLM; they must be satised,
with no exceptions. We decide to examine whether a given coverage
matrix (a matrix with values equal to either 0 or 1, with 1 representing
a community that can be covered by a given shelter and 0 representing
a community that cannot be covered by that shelter) has an all-zero
rank. A feasible coverage matrix cannot have an all-zero rank; if such a
ranking is generated based on these two requirements, then a quick
projection procedure is used to modify the induced solution in order
to satisfy the feasibility conditions. This projection procedure simply
selects shelters randomly from the unselected candidate shelters, until
the feasibility check is met at each level.
We emphasize that some solutions satisfying Eq. (16) could yet
violate constraint (7), which could result in infeasible condition. Recall
that when designing the ESLM and VAM, we view the capacity
constraint (7) as a soft constraint, meaning that this constraint could
be violated. Yet we want to minimize the number of any such
violations. This goal can be achieved by giving the violation a
suciently large penalty in the next performance calculation, so that
it would be discarded through CE execution.
2.3.1.3. Performance calculation. After the feasibility modication of
solutions generated in the CE execution, we need to calculate their
performance value, which consists of the objective value obtained
through Eq. (4) and the penalty value of any capacity violation. This
problem can be treated as a capacitated bin-packing problem that
determines the most ecient allocation of cargo to limited-capacity
bins [9]. In our application, we must consider the coverage constraints
437
International Journal of Disaster Risk Reduction 22 (2017) 431446
that are ignored in their research. We treat a selected shelter as a
capacitated bin with an elastic capacity, composed of a base capacity
(actual capacity) and an additional capacity (virtual capacity) that could
hold excessive demand. We then propose a new allocation heuristic
which aims at reducing the number of shelter-capacity violations.
First, all communities are sorted in ascending order based on their
number of potential available shelters. If there is a tie, they are sorted
in descending order based on demand. Under this scheme, the rst
community is the hardest one to be allocated because the fewest
shelters covered it, while the last community is the easiest one. In the
next step the communities are allocated sequentially from the rst to
the last. Each community is allocated to a shelter that could cover the
community with a maximum residual capacity. After all communities
are allocated to shelters, both the penalty value and the performance
value of the solution could be computed. In the ESLM, we dene a new
penalty variable pjk as follows:
k
pjk = max k xijk dik (t )j yjh,0, jJ ,
k K .
iI h =1 (17)
These new variables have the following construction implication: if
a shelter can provide adequate area for the allocated evacuation
demandin other words, the capacity constraint is satisedthen
pjk will be zero. Otherwise, when the overall allocated demand exceeds
the shelters capacity, the virtual capacity discussed above will be
allocated for holding this extra demand. Since in fact the area of a
candidate shelter is xed, this additional capacity is termed virtual
because it is not real space; this pjk could be seen as a congestion price
that must be paid for holding the excess quantity of victims in a shelter.
To prevent unacceptable crowding in a shelter, we introduced a penalty
factor that would increase as the congestion in a shelter rises. With
these modications, the performance value PV can be expressed as:
PV = (u k j yjk +u k pjk ).
jJ kK (18)
Finally, by calculating PV , the objective value of the ESLM as well
as the penalty value of the capacity violation can be minimized together
in the next step, CE method execution.
2.3.1.4. CE method execution. After adopting the penalty approach to
introduce the capacity constraint (7) into the objective function, the
ESLM is converted to a relaxation model that would minimize PV and
be subject to all constraints except constraint (7). The constraints of
this relaxation model constitute the feasible region of variable y , which
is represented by @ , and the ESLM is transformed into a 0-1integer
minimization problem as follows:
z* = min PV ( y)
y@ (19)
where z* represents the minimum value of this optimization problem.
The main idea of the CE method is to design an eective learning
mechanism throughout the process of searching for solutions. It
employs the associated stochastic problem (ASP) to estimate a rare
event when searching for the minimum value of the original combina-
torial optimization problem (COP). Since we want to use binary strings
to identify the selection status of candidate shelters, we use a Bernoulli
density function (,u ) to generate these strings. The ASP in our study
can then be represented as:
u (PV ( y) z ) = I{PV ( y) z} ( y, u ).
y@ (20)
where I{PV ( y) z} is the indicator function, whose value is 1 if PV ( y) z
and 0 otherwise, and u is the probability measure that a random state
Y , drawn under (,u ), has performance value less than or equal to a
given threshold value z . Then, we generate a random sample Y1,,Yk ,
H. Li et al.
Fig. 4. Flowchart of the Swap Allocation.
,YN from a dierent density function ( y, v ). In this case, an estimator
of event l that is given by the likelihood ratio estimator
N For example, if community i1 is originally allocated to shelter j1 and
1 ( y, u)
l = I{PV (Yk )z} ,
N k =1
( y, v ) (21)
where ( y, v ) constitutes a change of measure and is chosen such that
the CE between ( y, u ) and ( y, v ) is minimal.
The next objective is to determine the optimal reference parameter
associated with v (PV ( y) z ).The rule is used iteratively to generate a
sequence of reference parameters {v t } and non-increasing levels {z t } for
the estimation of v (PV ( y) z t ). At each iteration, the performance
values of Y1YN are sorted in ascending order and the threshold value z t
is chosen as the percentile of these values, where 0 < < 1 is a given
parameter. The new value of z t is then used to generate a new vector
v t+1. The new vector is, in turn, used to draw a sample population under
a dierent distribution function ( y,v t+1), which will lead to better
values of z t+1.
The process terminates when the level value z t does not change over
a number of iterations or when the vector v t converges to a binary
vector. Moreover, we implemented a smoothed version of the
updating rule, in such a way that at each iteration t :
vit +1 = vlit +1+(1 ) vit , (22)
typically with 0.7 0.9 [2]. The smoothed update prevents the
method from converging too fast to a degenerate vector, thereby
leading to a more thorough exploration of the search space.
2.3.2. The allocation heuristic for victim allocation model
In the VAM, the generated allocation variable xijk can be used to
guide the victims to shelters. This ESLM-generated allocation scheme
is called the ESLM Allocation. This allocation scheme, however, may
allocate the victims to a distant shelter rather than to a nearby one.
Another allocation scheme is called the Closest Allocation, it evacu-
ates the victims to the nearest shelter, but without considering the
shelters capacity. The new allocation scheme we propose includes a
swap move, enabling a trade-o between total evacuation travel
distance and shelter congestion. The result should be a shorter total
evacuation travel distance than that generated by the ESLM allocation,
438
International Journal of Disaster Risk Reduction 22 (2017) 431446
and less shelter congestion than that generated by Closest Allocation.
The swap move is, simply, to switch two communities between shelters.
community i2 is originally allocated to shelter j2 , then after implement
the swap move, community i1 would be allocated to shelter j2 and
community i2 would be allocated to shelter j1. We call this allocation
method with a swap move the Swap Allocation, and prescribe an
optimizing requirement for when to implement this move: if inequality
(23) holds, then the shelter allocation choicesi1 and i2 would be
swapped; otherwise the original allocation would hold.
ai1 j1 di1k (t ) + ai2 j2 dik2 (t ) (ai1 j2 di1k (t ) + ai2 j1 dik2 (t )) 0, (23)
The following procedure describes the details of applying Swap
Allocation to the VAM. First, the communities are sorted in descending
order based on their refuge demand. Communities would be allocated
sequentially to shelters from the largest demand to the smallest
demand. Under this scheme, a community with a larger demand would
have allocation priority over one with a smaller demand, and would be
allocated to the nearest shelter, in order to maximize constraint (9):
each selected shelter must serve at least one communitys demand
within its coverage radius. We let the nearest shelter JiN for each
community i constitute the nearest shelter set J N . Next, to minimize the
evacuation distance for a given unallocated community i1, while
violating the shelter capacity constraint as little as possible,
Community i1 should be preferentially allocated to the nearest shelter
with positive residual capacity. The two possible situations, with their
corresponding solutions for i1s allocation, are:
1) There is at least one shelter with positive residual capacity that can
provide service for i1 within its coverage radius. We want to nd the
closest (e.g., j1) of all the shelters that has sucient positive residual
capacity for i1. If j1 is the shelter closest to i1 (i.e., j1 =ji1N , ji1N J N ),
then i1 would be allocated to j1. Otherwise, if j1 ji1N , we would try to
determine whether the previous allocated community, i2 , and its
allocated shelter, j2 , could satisfy Eq. (23) with i1 and j1 together,
and if so, then a swap would be implemented (allocate i1 to j2 and i2
to j1); otherwise i1 would remained allocated to j1.
2) None of the shelters that can cover i1 has any positive residual
H. Li et al.
capacity, i.e., there is already congestion in all the shelters to which
i1 can be allocated. First, the shelter with the least degree of capacity
violation (i.e., the least congestion), denoted as j1C , will be selected
as the rst allocated shelter to receive victims from Community i1.
Then, to optimize the evacuation travel distance, we will try to nd a
previous allocated community, i2 , and its allocated shelter, j2 , that
can satisfy Eq. (23) with i1 and j1C together, and if there is such a
shelter, a swap will be implemented (allocate i1 to j2 and i2 to j1C );
otherwise i1 will remain allocated to j1C .
Finally, all communities will be allocated to their selected shelters.
The overall allocation procedure is depicted in Fig. 4. The detailed
performances of these three allocation schemes will be evaluated and
compared in Section 3.
3. Case study
Shanghai, which sits on the southern tip of the Yangtze Delta Plain,
is the largest city in China, with an area of 6340 km2. It has widespread
slowly accumulated soft clay layers, making it a typical soft-soil area
with a fragile geographic environment. Shanghai has great vulnerability
because of its proliferation of more than 1200 tall buildings (height
100 m) and more than 24.8 million residential population in 2013,
and the areas surrounding the city lie on several moderate-scale
earthquake faults. Although no serious earthquakes been recorded in
Shanghai, it is on the list of seismic-risk city specially supervised by the
CEA. Furthermore, the construction and planning of emergency
shelters in Shanghai have lagged behind those of other major cities
(e.g., Beijing, Wuhan city) in China. The Xuhui District lies in the
central area of Shanghai, and has a land area of 54.93 km2 (composed
of 12 sub-districts and one township) and had a residential population
of 1,085,130 in 2010. Its population density is 19,755 persons/km2
and it has more than 160 tall buildings. But it has as yet no emergency
shelter, of any level. It is critical, therefore, to plan emergency shelters
for this district. We therefore choose the Xuhui District as our study
area.
3.1. Data processing
The open spaces in the study area, such as public parks, school
playgrounds, and green spaces, can be the candidate sites for shelters.
However, to avoid secondary damage to the aected people, high-risk
open spacesthose located less than 500 meters away from disaster-
prone areas such as gas stations and chemical factoriesshould be
excluded from the list of shelter candidates. Based on the 2011 relief
resources provided by the Shanghai Civil Defense Oce (SCDO), we
compile a list of 125 candidate sites, after excluding the high-risk ones.
SCDO published the guideline for emergency shelters planning in
2008, and it constructs shelters that be divided into three levels by a
hierarchical way, which is elaborated in Table 3. In this paper, we apply
the guideline to the study area, considering the models of our study and
the fact that the Xuhui district has great population and relatively less
candidate sites for shelters. The candidate shelter set is further divided
into 17 LSs (providing level-1 to -3 services), 56 SSs (providing level-1
to -2 services) and 52 ISs (providing level-1 services) based on the site
size. In this study, the service radii for Level 1, Level 2 and Level 3
Table 3
The classification of emergency shelters in Shanghai.
Shelter level Required area Service radius Time Service facility
(m2) (m) (day)
I 15000 5000 > 30 Comprehensive
II 4000 1000 10 to 30 General
III 2000 500 < 10 Basic
439
International Journal of Disaster Risk Reduction 22 (2017) 431446
Table 4
Hierarchical shelter relevant parameters in this study.
Shelter type Required Service Unit Time Service level
area (m2) radius construction (day)
(m) cost (RMB/m2)
LS 15,000 5000 40 1 Level-1,2,3
SS 2,000 2000 15 2 to 10 Level-1, 2
IS 300 1000 5 11 to Level-1
30
service are set at 1000 m, 2000 m and 5000 m respectively.
Construction costs for IS, SS and LS are set at 5 RMB/m2, 15 RMB/
m2 and 40 RMB/m2, respectively, provided by the SCDO. We list the
relevant parameters in Table 4.
The population sizes of the residential communities are extracted
from the 2010 Economic and Social Development Statistics Yearbook
of Shanghai (http://www.stats-sh.gov.cn/data/toTjnj.xhtml?y=2010e),
and the longitude and latitude of each residential community are
obtained using Google Earth. We identify a total of 303 residential
communities. We take the maximum refuge demand for each phase (I
to III) for each residential community as its time-varying refuge
demand dik under a Richter scale magnitude 7 earthquake scenario,
as shown in Fig. 5.
In Table 5, we set the refuge ratio of the residential population
according to Fig. 5 and refer to Chen et al. [6] and CEAs standard GB
21734-2008 to determine the required area per capita during the
evacuation phase.
We digitize a map of the Xuhui District, in which we can add many
layers to present the data for communities and shelters. In version 10.1
of ArcGIS, we use the Euclidean distance to calculate the distances
between shelters and communities using the OD Cost Matrix function
of the Network Analysis module. Fig. 6a shows the resulting geographic
distribution of residential communities. The dierent-sized circles
represent the relative sizes of the residential populations. The geo-
graphic distribution of dierent levels of emergency shelter candidates
and their corresponding services are presented in Fig. 6b. A circle
represents a Level 1 service, a triangle a Level 2 service, and a square a
Level 3 service. With this arrangement, a candidate LS that can provide
all three levels of refuge service is represented by three shapes
simultaneously, and similarly, an SS candidate is represented by two
shapes, and an IS candidate by one shape.
3.2. Results and analysis
The results are acquired via Matlab 2013b and then output to ESRI
ArcGIS 10.1 for mapping purposes. By using the CE method to solve
the ESLM, among the 125 emergency shelter candidate sites in the
study area, including 17 LS, 56 SS and 52 IS candidates, 58 shelters are
nally selected to provide refuge service, shown as Fig. 7. Among these
shelters, 22 ISs, 23 SSs and 13 LSs can provide hierarchical refuge
service in the Xuhui District during the three post-earthquake phases.
The total construction cost of these selected emergency shelters is
25,108,435 RMB, and the evacuation distance traveled by the victims over
Phase I to Phase III is 1,601,908,539 meter using the allocation result
generated by solving the ESLM. The results after the ESLM allocation are
presented in Figs. 8a, 9a and 10a. The many overlapping lines indicate
that the evacuation travel distance has not been optimized, although the
obtained allocation result satises the coverage constraints. Therefore, we
try the other two allocation schemesClosest Allocation and Swap
Allocationin order to determine which allocation is best tted for this
emergency scenario. In Figs. 810, we use blue lines to represent the
victim allocation, and circular dashed lines to represent the service radii
for the dierent service levels. It can be easily seen that there are fewer
overlapping lines in these two alternative allocation schemes, meaning
that these allocations would result in shorter evacuation travel distances
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

Fig. 5. Hierarchical time-varying refuge demand.

Table 5 the shelters, indicating they could hold more refuge demand than their
and refuge ratio for the three phases during evacuation period. ocial capacity. Based on the variation of shelter area per person and
the CEA standard, we categorize shelter congestion into three states: no
Phase I Phase II Phase III
congestion, light congestion and heavy congestion; the classication
Required area per capita (m2) 0.5 2 3 criteria for the dierent phases are shown in Table 7. These three
Refuge ratio 99% 37% 20% congestion degrees are color-coded as follows: green represents no
congestion, yellow, light congestion and red, heavy congestion. It is
important to note that the congestion degree of a given shelter may
than in the ESLM allocation. Table 6 shows the calculated results of the
change over time because the area occupied per person will change
evacuation distance over the three phases, using the three dierent
during the various evacuation phases.
allocation schemes.
A rough overview of Figs. 810 reveals that Closest Allocation and
Since capacity violations are allowed, congestion is introduced to
Swap Allocation would result in more congestion in shelters than the

Fig. 6. Distribution maps of the residential communities and the shelter candidates.

440
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

By analyzing the above tables and gures, we can draw the

conclusion that Closest Allocation produces the least evacuation
distance and the most shelter congestion, while the ESLM allocation
creates the largest evacuation distance and the least shelter congestion.
Swap Allocation, though, can obtain a balanced result with respect to
both evacuation distance traveled by the victims and emergency shelter
congestion. We therefore recommend Swap Allocation for guiding
victim allocation during the evacuation phase.

3.3. Comparable analysis

In this section, we compare the results under two dierent

estimations of refuge demand. In Section 3.2, we view the hierarchical
refuge demand as a varying quantity through the dierent phases.
Current policy uses approximately 30% of a residential population as
the estimation of number of the residential people to plan emergency
shelters in China [35]. Fig. 12a shows the unchanging trend of refuge
demand under this estimation. Compared to the time-varying refuge
demand estimation, whose trend is presented in Fig. 12b, the unvary-
ing estimation of shelter demand has lower immediate and short-term
demands but more long-term demand.
We make unvarying refuge demand as an input to the ESLM and
run it, producing a total construction cost for emergency shelters of
35,290,075 RMB. This result is 40.55% higher than the result under
the time-varying refuge demand, and with respect to the number of
victims, the hierarchical shelters under the time-varying refuge de-
mand estimation can service more victims than under the unvarying
estimation53% of the residential population (averaged by refuge
ratio: 99%, 37% and 20% in three phases) under the time-varying
refuge demand, but only 30% of the residential population under the
unvarying refuge demand. Therefore, using the time-varying refuge
Fig. 7. Distribution map of the result of hierarchical emergency shelter planning.
demand estimation to plan hierarchical emergency shelters can pro-
duce a better result because more refuge demand is satised with lower
ESLM allocation, as more candidate sites are red or yellow for those
shelter construction cost. The results of shelter selection (12 LSs, 12
allocations. Fig. 11, however, shows a detailed comparison for shelter
SSs and 15 ISs) under both unvarying refuge demand and time-varying
congestion status among those three allocation schemes. From Phase I
refuge demand are illustrated in Fig. 13.
through Phase III, Closest Allocation always shows the lowest value for
Fig. 13 also shows that the number of selected shelters using the
evacuation distance (from the line chart), and ESLM allocation always
unvarying refuge demand is less than for the time-varying refuge
shows the least shelter congestion (from the bar chart). However, Swap
demand. Furthermore, Table 8 shows that more LS and SS areas need
Allocation produces a lower value of evacuation distance than ESLM
to be built, and the resulting higher cost per square meter causes an
allocation, and also results in less shelter congestion than Closest
overall higher construction cost while producing fewer shelters. After
Allocation.
the locations of emergency shelters are determined, we use Swap

(a) ESLM allocation (b) Closest allocation (c) Swap allocation

Fig. 8. Comparison of the three allocation schemes during phase I.

441
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

(a) ESLM allocation (b) Closest allocation (c) Swap allocation

Fig. 9. Comparison of the three allocation schemes during phase II.

(a) ESLM allocation (b) Closest allocation (c) Swap allocation

Fig. 10. Comparison of the three allocation schemes during phase III.

Table 6 Table 7
The evacuation distance for different phases. Shelter congestion status (m2/person) classication for the three phases.

Allocation Phase I (m) Phase II (m) Phase III (m) Total evacuation Phase Heavy congestion Light congestion No congestion
scheme distance (m)
I < 0.25 [0.25, 0.5] > 0.5
ESLM 598,761,672 479,481,234 523,665,633 1,601,908,539 II <1 [1, 2] >2
Allocation III < 1.5 [1.5, 3] >3
Closest 351,767,530 187,769,925 165,510,947 705,048,402
Allocation
Swap 425,550,950 222,941,426 213,217,885 861,710,261 shown in Table 9. It can be seen that the evacuation distance traveled
Allocation
by the victims under the time-varying refuge demand estimation is
greater than under the unvarying estimation during every phase except
Allocation to allocate the victims to shelters. Figs. 1416 compare the Phase III. It is obvious that the more the refuge demand, the greater
Swap Allocation results for the two refuge demand estimations from the evacuation distance. However, if this evacuation distance is
Phase I through Phase III. averaged over the refuge demand, the time-varying estimation pro-
Regarding the congestion status in shelters, the time-varying duced less per capita evacuation distance in all three phases.
estimation performs worse than the unvarying estimation except for Finally, from the comparison of the two estimations for refuge
Phase III, in which there are two heavy congestion sites. The results demand, we should select the time-varying estimation, as it is more
of computing the evacuation distance under the two estimations are representative of an actual evacuation situation. Furthermore, the

442
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

(a) Phase I (b) Phase II (c) Phase III

Fig. 11. Statistical charts of congestion status and evacuation distance for the three allocation schemes for the three phases.

(a) Unvarying refuge demand (b) Time-varying refuge demand

Fig. 12. The comparison diagram of two estimations on refuge demand.

(a) Time-varying refugee demand (b) Unvarying refugee demand

Fig. 13. Comparison of results with emergency shelter planning under two refuge demand estimations.

443
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

Table 8 coverage and capacity constraints, while the second-stage model deals
The area and cost of hierarchical shelters under two estimations. with the victim allocation problem, to minimize the total evacuation
travel distance while also considering the congestion in shelters, over
Shelter type Time-varying refuge demand Unvarying refuge demand
the dierent post-earthquake phases. These models are then applied to
Refuge Construction cost Refuge Construction cost emergency shelter planning in the Xuhui District of Shanghai city.
area (m2) (RMB) area (m2) (RMB) Analyzing the results, we compare three dierent allocation schemes to
determine which is best tted for an emergency scenario, and compare
LS 561,484 22,459,360 841,402 33,656,080
SS 152,595 2,288,925 82,948 1,244,220 the eects of the time-varying refuge demand estimation with those of
IS 72,030 360,150 77,955 389,775 the unvarying refuge demand estimation.
Total: 25,108,435 Total: 35,290,075 Our study oers four main contributions to future emergency shelter
planning. First, we take into consideration the temporal variation in needs
and quantities of the victims, when planning the emergency shelter
system. We also design a shelter system that could provide dierent levels
time-varying refuge demand estimation is a more cost-eective shelter of service after an earthquake, creating a hierarchical structure of shelters.
planning method because it reduces shelter construction cost and These two points are critical to emergency shelter planning. Second, we
evacuation travel distance, and produces more tolerable congestion in propose a hierarchical emergency shelter location model and a victim
shelters, compared to the unvarying estimation. allocation model under the time-varying refuge demand estimation, to
minimize the total construction cost of shelters and the evacuation travel
4. Conclusion distance of victims. Third, we develop a novel cross-entropy method to
tackle the location model and a Swap Allocation scheme to solve the
This study considers hierarchical emergency shelter planning for allocation model. Finally, in the empirical case of the Xuhui District of
earthquake disasters in urban areas, taking into consideration the time- Shanghai, the Swap Allocation scheme makes a trade-o between the total
varying refuge demand estimation. Because the number of victims and evacuation travel distance and congestion in emergency shelters, as a
their needs in emergency shelters change over the post-earthquake better way to allocate the victims during an emergency period, than either
period, a hierarchical system of emergency shelters should be planned, Closest Allocation or ESLM allocation. We also nd that emergency
to provide dierent levels of refuge service over dierent phases of shelter planning based on a time-varying refuge demand estimation
evacuation. Once the shelter locations and levels are determined, the would reduce construction costs, compared to using the unvarying refuge
corresponding victim allocation must then be addressed. Therefore, we demand estimation, in an empirical case. Although the time-varying
model these problems as a two-stage mathematical programming estimation results in more immediate and short-term refuge demand than
model. The rst-stage model formulates the shelter location problem the unvarying estimation, it reduces the individual evacuation travel
to optimize the total shelter construction cost, considering their distance and produces tolerable congestion in emergency shelters.

Fig. 14. Comparison of the two refuge demand estimations using Swap Allocation during Phase I.

444
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

Fig. 15. Comparison of the two refuge demand estimations using Swap Allocation during Phase II.

Fig. 16. Comparison of the two refuge demand estimations using Swap Allocation during Phase III.

445
H. Li et al. International Journal of Disaster Risk Reduction 22 (2017) 431446

Table 9 Prev. Manag. 16 (2007) 235244.

Statistics on evacuation distance for the different phases under the two estimations. [9] M.S. Daskin, P.C. Jones, A new approach to solving applied location/aIIocation
problems, Comput. Aided Civ. Inf. 8 (1993) 409421.
[10] R.A. Gerrard, R.L. Church, Closest assignment constraints and location models:
Phase Time-varying refuge demand Unvarying refuge demand
Properties and structure, Locat. Sci. 4 (1996) 251270.
[11] F. Hu, W. Xu, X. Li, A modied particle swarm optimization algorithm for optimal
Evacuation Per capita Evacuation Per capita (m/ allocation of earthquake emergency shelters, Int. J. Geogr. Inf. Sci. 26 (2012)
distance (m) (m/person) distance (m) person) 16431666.
[12] Z.-H. Hu, J.-B. Sheu, L. Xiao, Post-disaster evacuation and temporary resettlement
I 425,550,950 467.41 131,503,863 471.81 considering panic and panic spread, Transp. Res. B-Methodol. 69 (2014) 112132.
II 222,941,426 648.57 216,992,221 778.53 [13] F. Klc, B.Y. Kara, B. Bozkaya, Locating temporary shelter areas after an earth-
III 213,217,885 1,147.59 367,704,140 1,319.26 quake: a case for Turkey, Eur. J. Oper. Res. 243 (2015) 323332.
[14] T. Kulcar, Optimizing solid waste collection in Brussels, Eur. J. Oper. Res. 90
(1996) 7177.
[15] A. Kulshrestha, D. Wu, Y. Lou, Y. Yin, Robust shelter locations for evacuation
We also perceive three main directions for future research. First, planning with demand uncertainty, J. Transp. Saf. Secur. 3 (2011) 272288.
because the psychological condition of victims can aect their behavior, [16] J.M. Lee, Y.H. Lee, Tabu based heuristics for the generalized hierarchical covering
we will study how their social behavior, such as rumor spreading and location problem, Comput. Ind. Eng. 58 (2010) 638645.
[17] A.C.Y. Li, L. Nozick, N. Xu, R. Davidson, Shelter location and transportation
the herding eect, aect the refuge demand; shelter location planning
planning under hurricane conditions, Transp. Res. E-Log. 48 (2012) 715729.
and victim allocation based on more accurate refuge demand estima- [18] L. Li, M. Jin, L. Zhang, Sheltering network planning and management with a case
tion can be further rened. Second, because of lack of real road data of in the Gulf Coast region, Int. J. Prod. Econ. 131 (2011) 431440.
our study area, we use the Euclidean distance as alternative. After real [19] H. Nakanishi, J. Black, Residential relocation, school relocation and childrens
transport: The eect of the Great East Japan Earthquake and Tsunami 2011, Int. J.
road distance data are collected, it will be possible to make more Disaster Risk Reduct. 18 (2016) 232243.
accurate evacuation plan in a future study. Third, emergency shelter [20] M. Nappi, J. Souza, Disaster management: hierarchical structuring criteria for
planning is considered only at the community scale in our empirical selection and location of temporary shelters, Nat. Hazards 75 (2015) 24212436.
[21] Oce of the United Nations Disaster Relief Co-ordinator, Shelter after disaster:
study. Considering that a community comprises several blocks, we will Guidelines for assistance. United Nations, New York, 1982.
further extend our study to the block scale by collecting more detailed [22] E.L. Quarantelli, Patterns of sheltering and housing in US disasters, Disaster Prev.
data and information, to achieve yet more accurate emergency shelter Manag. 4 (1995) 4353.
[23] R. Rubinstein, The cross-entropy method for combinatorial and continuous
planning. optimization, Methodol. Comput. Appl. 1 (1999) 127190.
[24] G. ahin, H. Sral, A review of hierarchical facility location models, Comput. Oper.
Acknowledgments Res. 34 (2007) 23102331.
[25] G. ahin, H. Sral, S. Meral, Locational analysis for regionalization of Turkish Red
Crescent blood services, Comput. Oper. Res. 34 (2007) 692704.
This study is supported by grants from the National Natural Science [26] H.D. Sherali, T.B. Carter, A.G. Hobeika, A location-allocation model and algorithm
Foundation of China (Project No. 71373155), and the Shanghai for evacuation planning under hurricane/ood conditions, Transp. Res. B-
Methodol. 25 (1991) 439452.
Pujiang Program (Project No. 14PJC073).
[27] J.-B. Sheu, C. Pan, A method for designing centralized emergency supply network
to respond to large-scale natural disasters, Transp. Res. B-Methodol. 67 (2014)
References 284305.
[28] U. Tamima, L. Chouinard, Development of evacuation models for moderate seismic
zones: a case study of Montreal, Int. J. Disaster Risk Reduct. 16 (2016) 167179.
[1] A.I. Barros, R. Dekker, V. Scholten, A two-level network for recycling sand: a case [29] J.C. Teixeira, A.P. Antunes, A hierarchical location model for public facility
study, Eur. J. Oper. Res. 110 (1998) 199214. planning, Eur. J. Oper. Res. 185 (2008) 92104.
[2] P.-T. Boer, D.P. de, Kroese, S. Mannor, R.Y. Rubinstein, A tutorial on the cross- [30] C.-H. Tsai, Y.-L. Yeh, The study of integrating geographic information with multi-
entropy method, Ann. Oper. Res. 134 (2005) 1967. objective decision making on allocating the appropriate refuge shelters: using
[3] B. Boey, D. Yates, R.D. Galvo, An algorithm to locate perinatal facilities in the Kengting National Park as an example, Nat. Hazards 82 (2016) 21332147.
municipality of Rio de Janeiro, J. Oper. Res. Soc. 54 (2003) 2131. [31] D. Wang, B. Huang, Y. Li, et al., Planning and construction of emergency shelter: a
[4] R. Burkard, M. DellAmico, S. Martello, Assignment Problems, Society for study based on surveying data of Yushu earthquake, J. Nat. Disasters 1 (2012)
Industrial and Applied Mathematics, Philadelphia, 2009. 6670.
[5] W. Chen, G. Zhai, C. Fan, et al., A planning framework based on system theory and [32] J. Wang, L. Zhao, C. Gu, System dynamical study on needs of emergency shelters
GIS for urban emergency shelter system: a case of Guangzhou, China, Hum. Ecol. for earthquakes, Chin. Saf. Sci. J. 23 (2013) 121128.
Risk Assess. 0 (2016) 116. [33] M.J. Widener, M.W. Horner, A hierarchical approach to modeling hurricane
[6] Z. Chen, X. Chen, Q. Li, J. Chen, The temporal hierarchy of shelters: a hierarchical disaster relief goods distribution, J. Transp. Geogr. 19 (2011) 821828.
location model for earthquake-shelter planning, Int. J. Geogr. Inf. Sci. 27 (2013) [34] J. Xu, X. Yin, D. Chen, et al., Multi-criteria location model of earthquake evacuation
16121630. shelters to aid in urban planning, Int. J. Disaster Risk Reduct. 20 (2016) 5162.
[7] S. Chien, L. Chen, S. Chang, et al., Development of an after earthquake disaster [35] Z. Yin, S. Li, Z. Zhao, S. Yang, A method for prediction and classication of
shelter evaluation model, J. Chin. Inst. Eng. 25 (2002) 591596. earthquake disasters, Earthq. Res. Chin. 7 (1991) 919.
[8] J. Dalal, P.K.J. Mohapatra, G.C. Mitra, Locating cyclone shelters: a case, Disaster

446