Reliability Engineering and System Safety 214 (2021) 107644

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

Dynamic network flow optimization for real-time evacuation reroute

planning under multiple road disruptions
Ayda Darvishan , Gino J. Lim *
Department of Industrial Engineering, University of Houston, Houston, TX, USA

A R T I C L E I N F O A B S T R A C T

Keywords: During the course of an evacuation, evacuees often encounter unexpected incidents interrupting their plans for
Short-notice evacuation evacuation. Roads may not be accessible due to flooding, wild-fire propagation, accidents, the collapse of
Dynamic network flow problem highway structures, and various other reasons. The evolving disturbances to the evacuation plan due to road
Network disruption
disruptions may prolong the evacuation process and lead to chaos, injuries, and loss of life unless a quick,
efficient recovery plan is implemented. In this work, we aim to provide a rerouting approach for an evacuation
network that undergoes road disruptions. Unlike previous studies, it is assumed that incidents can occur on
multiple roads and that the time of each occurrence can differ from the time of other occurrences. Flow opti­
mization techniques are used to represent evacuation traffic flow on the transportation network. A dynamic
traffic flow rate is considered in which the evacuation flow rate can change over time during the planning ho­
rizon. The variation in the flow rates enables a better projection of the traffic dynamics and consequences caused
by disturbances. Furthermore, a path-based dynamic network flow optimization formulation is proposed to make
the model scalable for large evacuation networks. Two preprocessing algorithms are introduced to calculate
specific parameters associated with road disruptions and topology of the evacuation network. The use of these
parameters enables us to transform the original optimization model into a linear model to reduce the compu­
tational burden. Numerical experiments are made to show the performance of the proposed model. Furthermore,
the effects of specific features such as disruption time, disturbance location, and the plan updating time on the
evacuation process are investigated. Results indicate that when more incidents occur later or when incident
information is received earlier, the magnitude of the rerouting completion time is lessened.

1. Introduction of 1,000 or more Americans rises about every three weeks. In another
report, FEMA [8] declared that approximately 45 to 75 major evacua­
Every year, numerous hazardous incidents have affected millions of tions occur annually in the United States. The federal government di­
people worldwide (see Fig. 1). Natural hazards are the common concern rects state officials to provide evacuation plans using guidelines such as
for many communities, including Geo-physical (Geological) incidents, the Robert T. Stafford Disaster Relief and Emergency Assistance Act [17]
such as earthquakes, volcanic disruptions, and tsunamis, or weather- and the FEMA Comprehensive Planning Guide 101 [7]. Meanwhile, as a
related events, such as storms, hurricanes, droughts, and tornados. result of variations in state populations, geographic characteristics, and
Various case-scenarios regarding technological failures and intentional transportation infrastructure, evacuation plans provided in different
malevolence, such as nuclear meltdowns, hazardous material spills, and states tend to differ from one another.
terrorist attacks, can also wreak havoc on communities. There are two phases of evacuation planning and management. The
One of the most critical elements in response to a major disaster is the first phase is to develop a route plan and schedule for the evacuation
evacuation of people in the affected area, as it is directly associated with prior to the arrival of an adversarial event. Decisions regarding route
protecting human lives. An evacuation refers to the mass physical assignments, time and schedule of evacuees’ departure, resource allo­
movement of people from endangered areas to safe shelters prior to the cation, and the operations planning horizon should be made during this
onset of, during, or after a hazardous event. According to a 2005 report phase. Such a process entails assessing the topology of the transportation
by the Nuclear Regulatory Commission (2005), the need for evacuation network, available human resources and facilities, dynamics of traffic

* Corresponding author.
E-mail address: [email protected] (G.J. Lim).

https://doi.org/10.1016/j.ress.2021.107644
Received 6 July 2020; Received in revised form 8 January 2021; Accepted 22 March 2021
Available online 15 April 2021
0951-8320/© 2021 Elsevier Ltd. All rights reserved.
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

pertaining to the minimum amount of evacuation completion time.

Investigating a transportation network performance in the face of
disastrous events has been studied by several authors. Lin et al. [15]
analyzed the performance of a stochastic-flow network under multiple
correlated failures of physical lines and routers. Chen and Miller-Hooks
[3] introduced a network resilience metric to measure recovery capa­
bility of a freight transportation network under the possibility of dis­
ruptions. Bier and Hausken [2] investigated the impact of blockage of
specific roads of transportation networks due to intentional attacks.
Faturechi and Miller-Hooks [5] provided a bi-level three-stage sto­
chastic program in which response strategies are defined in the upper
level to enhance the network resilience and the affected users choose
best possible routes for themselves in the lower level based on the
upper-level decisions.
Fig. 1. People affected by hazards each year (Source: EM-DAT). Our literature review reveals that the problem of the real-time plan
adjustment for evacuation has received significantly less scholarly
attention. Lim et al. [14] provided a preprocessing algorithm that uti­
flow, and road congestion level. The second phase, which is the focus of
lizes a path-based network flow optimization approach to reassign paths
this paper, monitors the progress of the evacuation, detects deviations
for evacuees affected by an incident, assuming the availability of
from the plan, and makes adjustments, if needed, based on real-time
real-time traffic information. However, there are two limitations on
information. Therefore, it is critical to have precise information about
their work. First, their model is limited to a single road incident. Second,
the real-time progress of evacuation and transportation infrastructure,
evacuation vehicles are assumed to join the evacuation routes at a
including the residual capacity of the roads in the evacuation network.
constant flow rate. Both assumptions are less realistic because multiple
During the course of an evacuation, the capacity of the roads can be
road disruptions can occur at any time, and the vehicle flow rate joining
blocked or severely compromised due to the occurrence of many un­
major evacuation routes can vary over the evacuation planning horizon.
foreseen events. Faturechi and Miller-Hooks [6] presented a compre­
Therefore, the purpose of this paper is to provide an approach to relax
hensive literature review related to the performance of transportation
both of the assumptions within the context of a dynamic network flow
systems in the face of hazardous events, the potential physical damages
optimization, in which a variable vehicle flow rate is assigned to each
to the infrastructure, its vulnerability, reliability, robustness, flexibility,
time interval of the planning horizon, and disruptions can occur on
survivability, and resilience. Roads may become impassable due to
multiple roads of the network. The number of evacuees leaving the in­
flooding, sinkholes, railway barriers at railroad crossings, debris fallen
termediate nodes into new pathways in the network can vary in each
on road surfaces, the collapse of highway structures caused by high
time interval, which results in a more realistic representation of evac­
winds, and subways submerged with storm water. These Real-Time
uation flow dynamics and congestion.
Events (RTE) [1] can disrupt the evacuation plan at hand and create
The paper is organized as follows. Section 2 provides a detailed
an unreasonable amount of traffic congestion. Furthermore, RTEs can
description of the problem assumptions and the developed dynamic
heavily jeopardize the safety of evacuees and yield high ground delays.
network flow optimization model formulation. Section 3 introduces
Timely and effective decisions on modifying the initial route plan to
preprocessing algorithms to calculate specific parameter values and a bi-
have a full, fast, and complete evacuation can help alleviate the risk of
section algorithm to calculate the evacuation completion time. In Sec­
chaos and prolonged delays. Therefore, the purpose of our study is to
tion 4, different disruption scenarios are discussed, and the performance
provide emergency managers with a real-time recovery scheme (i.e.,
of the proposed approach is investigated both on a sample network and
new evacuation routes and schedule) in the case of a transportation
on a real, large-scale metropolitan evacuation network. The same sec­
network road disruption. We propose to develop such plans in a way that
tion discusses computational results to examine the effects of the inci­
the total time required to clear the network is minimized.
dent time and incident location along with the plan updating time.
To date, various strategies and methods have been proposed to
Section 5 concludes the paper with a short summary and a future
develop efficient evacuation plans. Both simulation methods [4,16,19]
research direction.
and optimization techniques [10,12,18], and [13] have been widely
used to study and improve evacuation decisions along with predictions

Fig. 2. A directed graph representing an evacuation network.

2
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 3. Rerouting in RPBM.

2. Problem statement rerouted to a path at another time (see Fig. 3).

For developing the mathematical formulation, the following nota­
Traffic flow and congestion on the evacuation network are repre­ tion is used:
sented through a dynamic network flow problem. A dynamic network is Sets:
composed of multiple static networks in which each static network de­
picts the status of the network at a specific time [9]. Let us consider a N Set of all nodes
directed network G = (N , A ) consisting of a set of nodes N and a set of T Set of all time slots
arcs A (see Fig. 2). The nodes are categorized as origin nodes (N o ), pSet of all paths
intermediate nodes, and destination nodes (N d ). Intermediate nodes
typically represent intersections of roads in the evacuation network. The Decision Variables:
planning horizon is divided into discrete time intervals represented by
set T = {0, 1, …, T}. A traffic assignment on this network relies upon a rpnt Disturbed flow of node n ∈ N that is rerouted and assigned to
representation of traffic as a series of vehicle flows at each time interval alternate route p at time t ∈ T
as per topology of the network. The transit time of arc a ∈ A is shown by hcnt Remaining disrupted flow on node n ∈ N at time t ∈ T
τa , and the time it takes to reach arc a ∈ A from the origin of path p ∈ p
is denoted by θpa . Parameters:
Our approach is based on a path-based approach (see Appendix A) to fpt Flow that starts on path p ∈ p at time t ∈ T (pre-disruption plan)
expedite solving the evacuation rerouting problem. In a path-based Hpnt Disturbed flow on node n ∈ N of path p ∈ p at time t ∈ T
approach, all possible paths between each origin-destination (O-D) of θpa Transit time from the origin of path p ∈ p to arc a ∈ A
the evacuation network are generated and enumerated a priori, and Ca Capacity of arc a ∈ A
Dn Demand of source node n ∈ N
candidate paths are selected and defined as an input (i.e., set p) to the
ln Capacity of destination node n ∈ N
optimization model. Consequently, it reduces the number of variables τa Transit time on arc a ∈ A
and constraints in the optimization model and reduces the computa­ Lpn Takes value 1 if node n ∈ N is the source node of path p ∈ p, and otherwise 0
tional complexity of optimizing evacuation planning models. of the Kpn Takes value 1 if node n ∈ N is the destination node of path p ∈ p, and
model. This is an advantage over other well-known methods such as a otherwise 0
Transit time from the origin of path p ∈ p to node n ∈ N
Cell Transmission Model (CTM) in simulating traffic speed and
′
θpn
congestion [13]. δpa Takes value 1 if arc a ∈ A belongs to path p ∈ p, and otherwise 0
Takes value 1 if node n ∈ N is the upstream (origin) node of arc a ∈ A , and
Parameter DTa is used to denote the disruption time on each arc γna
otherwise 0
a ∈ A . The aim of the rerouting path-based model (RPBM) is to opti­ φpmn Takes value 1 if node n ∈ N is not behind node m ∈ N on path p ∈ p, and
mally utilize the residual capacity of the road networks to reassign the otherwise 0
disturbed evacuees to new paths such that the overall evacuation reroute Wpnt Takes value 1 if the flow on path p ∈ p starting at time t ∈ T reaches the
completion time is minimized. merging arc from node n ∈ N before the disruption time of the arc
Vpnt Takes value 1 if the flow on path p ∈ p starting at time t ∈ T is disturbed and
The RPBM assigns the disturbed flow to both unaffected paths as well
stuck behind node n ∈ N , and otherwise 0
as partially damaged paths, or called affected paths. An affected path is a ηptmn Takes value 1 if the flow on path p ∈ p starting at time t ∈ T is not affected
path involving one or more disrupted arcs. Although a path can be through node m ∈ N and also is not disturbed between node m ∈ N and node
interrupted, the residual capacity of its intact arcs can still be used to n ∈ N (m is behind n), and otherwise 0
reroute the flow and push it forward through the evacuation network. ∂pn Takes value 1 if there is no disturbed arcs on path p ∈ p after node n ∈ N , and
otherwise 0
When the reassigned flow reaches a damaged arc of an affected path, it
will gather and wait behind the associated node and can be sequentially

3
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 4. Representation of rerouting variable rpnt .

Fig. 5. Case 1 and Case 2 presentation for flow f p(t− θpa ) .

Mathematical properties of the PBM model do not allow direct rep­ updating time interval (Tupdated ). The flow-route assignments before
resentation of flow departing from intermediate nodes. Variables tupdating are equal to zero as in Constraint (2).
denoting the flow are always related to the evacuees leaving the source /
node of a path rather than the intermediate node. However, a method is rpn(t− θ′ ) = 0, ∀p ∈ p, ∀a ∈ A , n ∈ N , t ∈ T Tupdated . (2)
pn
needed for rerouting the flow in order to reflect the flow departing from
an intermediate node (see Fig. 4). To resolve the issue in the optimiza­ When calculating the amount of disturbed flow on node n ∈ N at
tion model formulation, the rerouting variable rpnt is introduced to time t ∈ T, we take into account both the amount of the disturbed flow
denote the amount of flow departing from the origin of path p ∈ p at from the original plan (denoted by Hpnt ) and the amount of the rerouted
time t ∈ T. Nevertheless, in our constraints, we ignore the values of rpnt flow from nodes (m ∈ N ) that were disturbed while passing through the
∑
associated with preceding nodes (or arcs) to node n ∈ N . In this case, the alternative pathway (W pnt ηptmn rpmt ). This is stated in Constraint (3).
m∈N
preceding arcs are considered dummy arcs. This is done by introducing ∑
three sets of parameters W pnt , V pnt , and ηptmn which reflect the effect of HCpn(t+θ′ ) = Hpn(t+θ′ ) + Wpnt ηptmn rpmt , ∀p ∈ p, n ∈ N , t ∈ T. (3)
pn pn

disruptions on the evacuation flow with respect to the arc incident times,
m∈N

sequence of arcs in the set of paths as well as topology of the network. Parameter Hpnt used in Constraint (3) can be calculated as follows:
Using these parameter a mathematical model with a linear structure can
Hpn(t+θ′ ) = Vpnt fpt , ∀p ∈ p, n ∈ N , t ∈ T.
be developed. Note that a flow departing at time t ∈ T takes θpn time pn

units to reach node n ∈ N . Hence, based on the disturbed flow infor­ When flow f pt is blocked on node n ∈ N (denoted by V pnt = 1), and
mation rpn(t− θpn ) , we can address the flow reassignment from node n ∈ N
since the flow has started from the origin of the path at time t ∈ T, the
onto route p ∈ p during time interval t − θpn + θpn = t.
time at which it reaches and accumulates on node n ∈ N is t + θpn . Note
′

RPBM (Rerouting Path-Based Model) formulation:

that θpn is the transit time from the origin of path p ∈ p to node n ∈ N .
′

We now describe the proposed dynamic network flow optimization

model formulation. When an arc disruption occurs, disturbed evacuees Accordingly, Hpn(t+θ′ ) represents the number of evacuees disturbed on
pn

are assumed to be accumulated on the tail of the affected arc (i.e., a node n ∈ N at time (t + θpn ).
′

behind the affected road in the evacuation network) for the purpose of The total number of remaining interrupted evacuees at time (t +1)
rerouting them to alternative paths. The objective function of RPBM equals its previous amount at time t ∈ T, plus the newly interrupted
aims to minimize the total number of disturbed evacuees remaining in ∑
evacuees ( HCpn(t+1) ), minus the amount of rerouted evacuees at time t,
the evacuation network by the end of the planning horizon T. p∈p
∑ as expressed in the following equation.
min hcnT ∑ ∑∑ ( )
n∈N hcn(t+1) = hcnt + HCpn(t+1) − ℑpna 1 − Wpn(t− θpa ) rpn(t− θpa ) ∀ n
Constraints are explained as follows. The planning horizon set T =
p∈p a∈A p∈p

{0, 1, …, T} covers both the pre-disruption schedule (f pt ) as well as the ∈ N , t ∈ T. (4)

post-disruption schedule (R pnt ). The time periods at which the previous Parameter ℑpna in Constraint (4) reflects the topology of the network
plan is updated are shown by set Tupdated = {tupdating , …, T}. At time t = and can be calculated as:
0, there are no disturbed evacuees in the system. Hence, the total
amount of associated flow on all nodes is set to zero as in Constraint (1). ℑpna = δpa γna ∀p ∈ p, a ∈ A , n ∈ N .

hcn(t=0) = 0, ∀n ∈ N . (1) Considering constraints (2), (3), and (4), before the updating time,
hcn(t+1) only equals the previous amount of the remaining flow plus the
The process of a plan revision can only take place during the newly interrupted flow (i.e. hcn(t+1) = hcnt +
∑
HCpn(t+1) ). However,
p∈p

4
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 6. Case 1, Case 2, and Case 3 presentation for rerouted flow rpm(t− θpa ) .

when the rerouting of the disturbed flow begins, the assigned disturbed to arc a ∈ A
flow rpn(t− θpa ) at time t ∈ T to alternate paths is no longer stalled behind When node m ∈ N comes after arc a ∈ A on path p ∈ p, parameter ∅pmn
node n ∈ N and is subtracted from the remaining disturbed flow of the equals zero and rpm(t− θpa ) is excluded in the capacity constraint.
next time interval hcn(t+1) . We use hcnt as a measure to be used to later Case 2: Rerouted flow is interrupted before reaching arc a ∈ A
study the performance of our MIP model. If node m ∈ N is directionally placed before arc a ∈ A on path p ∈ p
∑∑ ∑ [( ) ( ) (i.e. ∅pmn = 1), but the rerouted flow is disrupted before arriving to the
ηp(t− θpa )mn δpa γna 1− Vpn(t− θpa ) Lpm fp(t− θpa ) + 1− Wpn(t− θpa ) arc (ηp(t− θpa )mn = 0), then rpm(t− θpa ) is disregarded in the constraint.
p∈p n∈N m∈N
] Case 3: Rerouted flow can travel through arc a ∈ A
∅pmn rpm(t− θpa ) ≤Ca ,∀a∈A ,t∈T. (5) If node m ∈ N is directionally placed before arc a ∈ A on path p ∈ p
(i.e. ∅pmn = 1), and the reassigned flow can reach the arc without any
Constraint (5) ensures that the total flow from different paths interruptions (ηp(t− θpa )mn = 1), then R pm(t− θpa ) is included in the arc ca­
reaching a shared arc does not exceed the capacity of the arc (Ca ). This
pacity constraint.
flow includes (i) the pre-disruption flow schedule f pt , and (ii) the post-
Constraint (6) ensures that all evacuees, including those who fol­
disruption flow schedule rpnt . Let us consider flow f p(t− θpa ) departing
lowed the pre-disruption plan and those that have been rerouted, are
from the origin of the path at time t− θpa . Two cases can occur regarding restricted by the capacity of the destination node (ln ) when entering the
the share of this flow in the capacity usage of arc a∈A (see Fig. 5). shelter area.
Case 1: Flow is disturbed before reaching arc a ∈ A [ ]
Case 1 occurs when there is at least one disruption on the preceding arcs ∑∑ ∑ ∑
Kpn ηptmn Lpm fpt + ηptmn rpmt ≤ ln ∀n ∈ N d . (6)
before reaching arc a ∈ A and when the disruption time of an associated arc p∈p t∈T m∈N m∈N
is less than the time required for the flow to reach and pass through the arc.
o

Hence, if there is at least one arc a ∈ A in which the following condition holds Evacuation flow f pt is considered in Constraint (6) only if the
true, following condition holds:
( ) ( )
t − θpa + θpa + τ a > DT a ∃a ∈ A preceding to a ∈ A on p ∈ p t − θpa + θpa + τ a ≤ DT a

∀a ∈ A between the source node and destination node of path p ∈ p

then, parameter ηp(t− θpa )mn equals zero and f p(t− θpa ) is not considered in
the capacity Constraint (5). Note that m ∈ N represents the origin node
of the path if Lpm = 1. Also, n ∈ N represents the head of arc a ∈ A if γna Algorithm 1
= 1. Note that the details of calculating the amount of parameter Determination of disturbed flow.
ηp(t− θpa )mn using an algorithm is explained in Section 3. Inputs:
Case 2: Flow can pass through arc a ∈ A An evacuation network G consisting of a set of nodes N and a set of arcs A .
When the flow is not disturbed on preceding arcs to arc a ∈ A and rea­ Disruption Time of Arcs
ches and passes through arc a ∈ A , it occupies the capacity of arc a ∈ A until Calculating Vpnt :
it completely passes through the arc. This occurs when the following condition for all paths p ∈ p do
holds: for all time slots t ∈ T do
( ) for all arcs that belong to path p ∈ p do
t − θpa + θpa + τ a ≤ DT a if t + θpa + τa − DTa > 0 then
Wpnt = 1 (n is upstream node of arc a)
&V pn(t− θpa ) = 0 ∀a ∈ A preceding to a ∈ A on p ∈ p else if t + θpa + τa − DTa ≤ 0 then
W pnt = 0
When the value of (t − θpa ) + θpa + τ a is less than the disruption time
end if
(DT a ) of a preceding arc a ∈ A , the flow can reach arc a ∈ A , i.e., ηp(t− θpa )mn for all preceding nodes m ∈ N to arc a on path p do
= 1. At this point, if the flow is not disturbed on arc a ∈ A (i.e. V pn(t− θpa ) if
∑
Wpmt = 0 and Wpnt = 1 then
= 0), it can flow through the arc, and the arc capacity is reduced accordingly. m∈N
Vpnt = 1 (n is upstream node of arc a)
Note that Algorithm 2 explained in Section 3 is used to calculate the amount else then
of parameter ηp(t− θpa )mn to be used in the optimization model. V pnt = 0
Now, let us consider situations that may arise regarding post- end if
end for
disruption flow rpm(t− θpa ) (see Fig. 6). Note that this flow is placed on
end for
path p ∈ p through node m ∈ N . end for
Case 1: Reassigned flow has been added to the path from a preceding node end for

5
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 7. Assumptions used to define parameters Wpnt and vpnt .

decision variables, respectively.

Algorithm 2 rpnt ∈ Z+ , hcnt ∈ Z+ , ∀p ∈ p, n ∈ N , t ∈ T (7)
Determination of unaffected flow.
Inputs: 3. Algorithms to calculate key model parameters and network
An evacuation network G consisting of a set of nodes N and a set of arcs A , and vpnt .
clearance time
Calculating ηptmn :
for all paths p ∈ p do
As explained in Appendix B, developing an optimization model for
determine set φp as sequence of nodes of path p the problem that can be solved in a timely manner was challenging.
for all time slots t ∈ T do However, upon examining the problem characteristics, we realized that
for all preceding nodes m ∈ φp to node n ∈ φp do a mixed integer linear program can be developed if some key parameter
∑ 1
if n− k=m vpkt = 0 then values are determined a priori. Hence, this section introduces two pre­
ηptmn = 1 processing algorithms to calculate the parameter values, which makes it
else then possible to solve the proposed RPBM model.
ηptmn = 0
end if
end for 3.1. Methodology to calculate parameters associated with disruptions
end for
end for
Parameter Vpnt indicates whether or not flow fpt on path p ∈ p
departing at time t ∈ T would be stalled on node n ∈ N . Algorithm 1 is
developed to calculate the value of Vpnt . First, we calculate Wpnt to define
whether fpt is stopped on node n ∈ N regardless of any possible distur­
Algorithm 3
bances that may have surfaced during the flow passage up to node
Calculating clearance time of the reroute plan.
n ∈ N . Accordingly, we initialize disruption times of preceding arcs to
Inputs: node n ∈ N to be infinity. We disregard disturbances of the preceding
RPBM and X.
arcs and only take into account the time of incident (DTa ) of the arc that
put CT of the evacuation network with no disruption as LBRCT
emerges from node n ∈ N , arc transit time (τa ), and the time to transport
put UBRCT = 2 LBRCT
from the origin of the route to arc a ∈ A (θpa ). If a flow departs on path
Calculating a tight bound for RCT:
p ∈ p at time t ∈ T, it reaches arc a ∈ A at time t + θpa . Also, it takes τa
put T = {0, 1, …, X} as the planning horizon
unit of time for the flow to completely pass through the arc. Therefore, if
solve RPBM
∑∑ t + θpa + τa is greater than the disruption time of the arc (DTa ), the flow
if Int > 0 then
n∈N t∈T is interrupted, and Wpnt takes value 1 (Fig. 7).
LBRCT = X
Next, we calculate the value of Vpnt and subsequently take into ac­
else then
UBRCT = X
count incident times of preceding arcs to arc a ∈ A . The value of Vpnt
end if equals 1 only if the flow fpt experiences no interruption while moving
∑
Calculating RCT Value: toward node n ∈ N (i.e., Wpmt = 0) and is disturbed on node n ∈ N
m∈N
put T = {0, 1, …, T} as the planning horizon
of path p ∈ p (i.e., Wpnt = 1). Else Vpnt = 0.
= 1 do
While ψ ∕
A numerical example for calculating the value of V pnt :
put T = (LBRCT + UBRCT )/2
The following example is used to illustrate the calculation of
Solve RPBM for T = {0, 1, …, T} and T = {0, 1, …, T − 1}
parameter V pnt . Consider the path, disruption times, and arc transit
if one objective value is a positive value and the other is equal to zero then
put RCT = T and ψ = 1
times shown in Fig. 8. When an evacuee departs from node i at time t =
else if both objective values are positive and T > LBRCT then
1, the evacuee can pass through arc (i, j) because the disruption on arc
put LBRCT = T
DT (i,j) = 4 occurs after the flow has reached node j (i.e., time t = 2). This
else if both objective values are equal to zero and T − 1 < UBRCT then evacuee can also pass through arc (j, k) with no interruption, as the time
put UBRCT = T − 1 it arrives at node k (i.e., time t = 2 + 5 = 7) is earlier than the time of the
end if incident on arc DT (j,k) = 8. Accordingly, this flow is not disturbed on
end while either of the arcs (i, j) or (j, k) and V pi(t=1) = V pj(t=1) = 0. Now, let us
consider a flow starting on the path at time t = 3. It will arrive at the
origin of the arc (j, k) at time t = 4 with no disturbance. The arc transit
Hence, starting from the origin and moving towards the destination,
time for arc (j, k) is 5 units of time. Thus, it cannot pass through the arc
if the flow of an evacuee is not interrupted (ηptmn = 1 where m ∈ N o and
because the incident occurred prior to the projected arrival to the
n ∈ N d ), it is counted in Constraint (6). Similarly, the rerouted flow rpmt
location, i.e., 4 + 5 > 8. Associated values of V pnt will be V pi(t=3) = 0 and
is considered to use the capacity of the destination node only if no
V pj(t=3) = 1, as the disturbed flow is only affected on node j. Similarly,
incident affects the flow from node m ∈ N (the point it had been
for the departure time t = 5, we would have V pi(t=5) = 1 because the
inserted on the path) to the destination node n ∈ N d (ηptmn = 1).
flow was not affected as the incident occurred prior to the departure
Finally, Constraints (7) reflects the non-negativity and integrality of
time.

6
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 8. Example for parameter vpnt .

Fig. 9. Evacuation test network.

Next, Algorithm 2 is developed to determine whether a flow can pass 3.2. Rerouted clearance time calculation
through a specific location in the network and can reach another loca­
tion without any interruptions. For any path p ∈ p, we first derive the This section explains the procedure to calculate a performance
sequence of nodes composing the path called φp . Then, for any combi­ measure rerouted clearance time (RCT), which is defined as the minimum
nation of node m ∈ N and n ∈ N in the set φp (when m is a precedence time to safely evacuate people from harm’s way to safe destinations
to node n), we calculate the summation of n−
∑ 1 ∑n− 1 considering the route adjustments during the evacuation. One way to
k=m V pkt . If k=m V pkt = 0,
calculate RCT is to solve the RPBM model in sequence of a planning
we can conclude that the flow on path p ∈ p at time t ∈ T is neither
horizon T = {0, 1, …, T}. The smallest value in T whose objective
interrupted on node m ∈ N nor is disturbed between node m ∈ N and
value is zero is the RCT for the problem. However, such a trial-and-error
node n ∈ N . If this condition holds, then ηptmn = 1. Otherwise, ηptmn = 0.
approach is tedious and time consuming. Hence, we introduce a two-

Table 1
An initial evacuation plan for the sample network (fpt ).
Time Slots
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Paths P9 5 2 5 5
P10 5 5 5 5 5 5 5 5 5 5 5 5 5
P11 5 5 3 5 5 5
P16 5
P17 5 5
P18 5
P19 2 5 5
P21 3
P22 5 5 5 5 5 5 5
P23 2 2 5 5 5 5 5 5
P24 5 3 5
P27 3
P28 5
P30 3
P31 3 5 5 5 5 5
P33 5 5 5
P35 5
P36 5 5 5 5 5 5 5
P37 2 5 5 5 5 5 5
P39 5
P40 5
P41 2 5 5 5 2 5 5
P42 5

7
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Table 2
Disturbed flow (Hpnt ).
Time Slots
4 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Total

Paths P17 n2 5 5 25
P18 n2 5
P19 n2 5 5
P22 n4 5 5 5 5 5 5 5 107
P23 n4 5 5 5 5 5
P31 n4 5 5 5 5 5
P41 n4 5 2 5 5
P42 n4 5

Table 3
Rerouting plan (rpnt ).
Time Slots Total
13 15 18 19 20 21 22 23 24 25 26 27 28 29 30 31

P21 n2 5 5 5 5 5 25
P2 n4 2 132
P8 5
P9 5
P10 5 5 5 5 5 5 5 5
P15 5 5 5
P16 5
P25 5 5 5 5
P26 5 5 5 5 5 5 5
P33 5

step procedure to expedite the process of finding the RCT value. The first ∑∑
step is to quickly approximate the RCT for RPBM. Using this value as a Kpn fpt ≤ ln , ∀n ∈ N d , (12)
starting solution, Algorithm 3 finds the exact value of RCT as the final
p∈p t∈T
∑ ∑
step. Let Dnt = Lpn Hpmt denote the demand (evacuees) on source
p∈pm∈N M upt ≥ fpt ∀p ∈ p, ∀t ∈ T, (13)
node n ∈ N o of path p ∈ p at time t ∈ T. We aim to reroute these
evacuees through the marginal (residual) capacity of the roads and find fpt ∈ Z+ , upt ∈ {0, 1}∀p ∈ p, ∀t ∈ T, (14)
the shortest amount of time to reach safe shelters. The residual capacity Variable fpt is the amount of flow from the source node of path p ∈ p
of the arcs (\capticat ) and the destination nodes (ln ) are calculated as
at time t ∈ T. The objective function aims to minimize the maximum
follows:
time in which a flow reaches its destination. Constraints (10)-(14) are
∑∑ ∑
\capticat = Ca −
( )
ηp(t− θpa )mn ℑpna 1 − Vpn(t− θpa ) Lpm fp(t− θpa ) ∀a used to reroute Dnt from source nodes N o through the residual capacity
p∈p n∈N m∈N of the network. According to Constraint (13), binary variable upt re­
∈ A , t ∈ T&t < DTa (8) ceives value 1 if the flow is reassigned to path p ∈ p at time t ∈ T. For
upt = 1, the associated flow reaches the destination dp unit of time later
( ) during time interval (t + dp ), where dp is the duration time of the path.
t − θpa + θpa + τ a > DT a ∃a ∈ A preceding to a ∈ A on p ∈ p (9)
Let X represent the solution of the above optimization model. Time
The solution of the following optimization model gives an approxi­ value of X is used in Algorithm 3 to expedite the RCT calculation pro­
mated RCT. cess. The algorithm has two main steps: RCT bound generation and
∑∑
( ( )) ∑∑ ∑ value calculation. In bound generation, if Int > 0, it means there
Min Maxp∈p,t∈T upt t + dp S.t. Lpn fpt ≥ Dnt , ∀n ∈ N o , (10) n∈N t∈T
p∈p t∈T t∈T
still remains some disturbed evacuees in the network, and the system is
∑∑
∑ not cleared; hence, it provides a lower bound to RCT. If Int ≤ 0, it
∂pn Lpn δpa fp(t− θpa ) ≤ \capticat , ∀a ∈ A , ∀t ∈ T, n ∈ N (11) n∈N t∈T
p∈p means the system has been cleared. However, we are aiming to find the
minimum amount of time required to clear the system; hence, we

Table 4
Total disturbed flow from the initial and the rerouted plan (HCpnt ).
Time Slots
4 6 7 8 9 10 11 12 13 14 15 16 17 18 19 24 25 27 28 31 Total

Paths P17 n2 5 5 25
P18 n2 5
P19 n2 5 5
P21 n4 5 5 5 5 5 132
P22 n4 5 5 5 5 5 5 5
P23 n4 5 5 5 5 5
P31 n4 5 5 5 5 5
P41 n4 5 2 5 5
P42 n4 5

8
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Table 5 bound of RCT to LBRCT = T if T > LBRCT . If both objective values

Test problems. equal zero, it means that the system has been cleared at T − 1. This
C1 C2 C3 C4 triggers updating the upper bound to UBRCT = T − 1 if T − 1 < UBRCT .
The algorithm continues until RCT value is found.
Disrupted (2,5, 4,5), (5,4, 5,7), (2,5, 4,5), (2,5, 4,5, 4,6),
arcs (4,7, 6,7) (6,8, 7,10) (5,7, 8,7) (5,6, 8,7, 7,10)
4. Computational results
Disruption {9,16,14,8} {17,8,10,19} {12,15,14,7} {15,6,8,10,9,12}
time The computational experiments presented in this study focus on the
Updating 16 19 15 15
impact of network arc disruptions on evacuation plans and the perfor­
time mance of the proposed recovery strategy. First, the performance of
RPBM is investigated on a small sample network in generating alter­
native routes. Then, the same experiment is conducted on a real evac­
consider X as a lower bound to RCT. In the second part of the algorithm uation network involving a large metropolitan area. The optimization
to calculate RCT, the RPBM is solved considering T and T − 1 as the models are solved using CPLEX 12.5.1, while experiments are performed
planning horizon, where T = (LBRCT + UBRCT )/2. If the objective value on a PC with a 3.07 GHz Intel Core i7 processor having 24GB RAM and
at T − 1 is positive (i.e., the system is not cleared) and becomes zero at running Ubuntu 10.04.3.
time T (i.e., the system is cleared), the minimum time to clear the
network is achieved at T; RCT = T. If both objective values are positive,
the system cannot be cleared at time T. Hence, we update the lower

Fig. 10. Accumulated remaining flow (hcpnt ) under different sample problems.

Fig. 11. Effect of arc disruption on evacuation process.

9
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Table 6 = {(2, 5), (4, 5), (4, 7)}. Incidents are assumed to occur at different
Analysis of less vulnerable arcs. times: DT a(2,5) = 4, DT a(4,5) = 6, and DT a(4,7) = 8. While monitoring the
Arc (5,4) (8,7) (7,8) (7,10) (8,9) progress of the disaster, emergency management agencies received data
on the network condition and attempted to develop revised plans
Total disrupted 35 10 30 0 0
Flow disruption time [9,17] [11,14] [9,18] 0 0 accordingly. As the process of collecting information and rerouting plan
Arc capacity 5 15 15 15 15 generation takes time, it is assumed that agencies can implement the
Rerouting node n5 n8 n7 - - revised schedule immediately after the plan updating time tupdating = 10.
Rerouting time [11,17] [11,17] [11,20] 0 0
Hence, before t = 10, no rerouting itinerary is planned and the corre­
Note that the amount of disrupted flow is equal to the amount of rerouted flow. sponding variable (R pnt ) remains at zero.
The rerouting node represents the upstream node of the disrupted arc from Disruptions on arcs (2, 5), (4, 5), and (4, 7) partially affect several
which the disturbed flow is rerouted. When arc (5, 4) experiences a disturbance paths {P17 , P18 , P19 , P21 , P22 , P23 , P27 , P31 , P41 , P42 }. For example, flow
at time 9, it causes the disruption of 35 evacuees during the time intervals be­ f p23,t2 = 2 was scheduled to arrive at arc (4, 7) at time t + θp23,a(4,7) = 2 +
tween 9 and 17. This amount of flow is gathered behind Node 5 and, conse­
1 = 3. It takes τ a(4,7) = 2 time units for the flow to pass through this arc.
quently, is rerouted from the same node between time 11 and 17. Hence, the
rerouting process ends before time 22, which also happens to be the CT of the Since the arc is supposed to fail at time DT a(4,7) = 8, the flow f p23,t2 = 2
pre-disruption plan. Cases for arcs (8, 7) and (7, 8) are similar. However, dis­ is not affected. Now, flow f p23,t9 = 5 is scheduled to reach arc (4, 7) at
ruptions of arc (7, 10) and (8, 9) have no effect on the pre-disruption plan and no time t + θp23,a(4,7) = 9 + 1 = 10, but the arc is already blocked at time t =
evacuees are disturbed. This is because the affected arcs are not associated with 8. So, there will be a flow accumulation on node 2 at time t = 10, and it
the paths used in the pre-disruption plan.
is denoted by Hp23,n4,t10 = 5. The magnitude, location, and interruption
time of the disturbed flow Hpnt are demonstrated in Table 2.
4.1. Numerical experiments to illustrate the proposed approach According to Table 2, 132 evacuees out of 392 are constrained at
different locations of the transportation network between time periods
Experimental studies are conducted on the sample network shown in t = 4 and t = 19. Hence, the proposed RPBM model is used to generate
Fig. 9 [11]. The test network includes three source nodes (N 1 , N 2 , and new paths to accommodate the disturbed flow.
N 3 ), five intermediate nodes (N 4 ,…, N 8 ) and two destination nodes Table 3 shows the rerouting schedule for the disturbed flow rpnt
(N 9 and N 10 ). These nodes are connected through 22 arcs. Arc transit provided by RPBM. The first row of the table shows the flow that should
times (τ a ) as well as arc capacities (Ca ) are shown above each arc of the be reassigned to P21 from node 2. The remaining rows show route as­
network. signments from node 4 onto alternative paths {P2 , P8 ,P9 ,P10 ,P15 ,P16 ,P25 ,
Demand on the source nodes (number of evacuees that are present at P26 ,P33 }. This plan was able to reroute all interrupted flow and redirect
the source nodes) are assumed to be D1 = 110, D2 = 120, and D3 = them to safe shelters within 35 time intervals; hence, the corresponding
167, and the capacity of each of the destination nodes is assumed to be RCT is 35 time units. During this time period, a total of 25 evacuees are
750. First, all possible paths between all origin and destination (O-D) evacuated from node 2, and 132 evacuees are rerouted from node 4.
pairs are enumerated using the solution pool feature of CPLEX for the Earlier, Table 2 showed that only 107 evacuees experience conges­
shortest path problem. Next, a total of 42 shortest paths are selected as tion on node 4 as a result of the incidents. However, the plan provided by
the candidate paths with 13 paths originating from source node N 1 , RPBM assigns 132 evacuees from this node onto new pathways. This is
while 14 paths (P14 -P27 ) originate from the second node, and 15 paths due to the fact that, in RPBM, disrupted flow could still reach disrupted
(P28 -P42 ) originate from the third node as listed in Table 7 in arcs after its reroute assignment and thus add to the accumulated flow
Appendix B. For instance, P12 follows the sequence of behind the failed arcs. Consequently, the overall amount of disturbances
N 1 →N 5 →N 6 →N 8 →N 10 . These candidate paths are used as input data calculated in RPBM shown in Table 4 could be higher than that of the
to the path-based evacuation model (see Appendix A) for the purpose of associated amount shown in Table 2.
generating an initial evacuation plan. According to Table 4, 132 disturbed evacuees are associated with the
The path-based evacuation optimization model provides both the pre-disruption flow (Hpn(t+θ′ ) ), and 25 disturbed evacuees highlighted in
optimal pre-disruption route assignment and flow schedule f pt for the pn

the table (from time t = 24 to t = 31) belong to the disrupted rerouted

test case as in Table 1. For example, as shown in the first row of the ∑
flow (W pnt ηptmn rpmt ). The disruptions on the rerouted flow happened
Table, five evacuees (f P9,t2 = 5) commence the evacuation by leaving m∈N
node 1 following path P9 at time t = 2. The evacuation rate varies during on Path 21 and created traffic congestion behind Node 4 during time
time t = 3, t = 4, and t = 8: f P9,t3 = 2, f P9,t4 = 5, and f P9,t8 = 5. The slots 24, 25, 27, 28, and 31. As shown in Table 3, these disturbed flows
corresponding clearance time is 22 time units. have been previously rerouted onto Path 21 (through node 2) at time
Suppose that three incidents occurred on different roads during the slots 23, 24, 26, 27, 30. Note that the node sequence of Path 21 is 2-4-5-
evacuation. The disturbed roads (i.e., arcs) are labeled as Adisturbed 6-8-10. Hence, the starting flow from Node 2 reached Node 4 within one

Fig. 12. Effect of disruption time and updating time on evacuation process.

10
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 13. City of Houston transportation network.

unit of time. Since it could not pass through arc (4,5), the flow is stalled adjacent roads, and the proximity of the incident location to either the
behind Node 4 one time unit later at times 24, 25, 27, 28, and 31. source nodes or the destination nodes. Hence, this section analyzes the
The performance of the proposed model is further investigated using effect of disruption on each arc to determine which arc makes the
four different test instances. Table 5 shows the input data for these in­ network evacuation plan the most vulnerable by fixing the values of the
stances and includes information regarding the set of disrupted arcs, other two factors. The clearance times of the rerouting plans, accounting
corresponding disruption times, and updating times for the rerouting for the disruption of each of the 22 arcs, are shown in Fig. 11.
strategy. As shown, the most vulnerable arcs affecting the pre-disruption plan
Fig. 10 highlights the amount of remaining disturbed flow (hcnt ) in are (6, 7) and (6,8). This is due to the fact that for these two arcs, it takes
the system as the evacuation progresses. The remaining disturbed flow the longest time (RCT =34 time units) to reroute the disturbed flow and
gradually increased at early stages of the planning horizon. When the clear the network following their disruptions. The second most vulner­
rerouting process began, hcnt gradually decreased until there remains no able arcs are (3, 4), (3, 5), (7, 9) and (8, 10) with a corresponding RCT
disturbed flow to be rerouted. The system was cleared when all evacuees ranging from 30 to 32. Finally, the least vulnerable arcs are (5,4), (8,7),
reached the destination nodes. In test problem C1 , the remaining (7, 8), (7, 10), and (8, 9). Disruption on these arcs disturbs part of the
disturbed flow was zero at time t = 32. But, it took additional 2 units of flow; However, these disruptions do not change the RCT of the plan. The
time for the last rerouted flow to reach the destination via the assigned rerouting plan was able to use the residual capacity of the evacuation
alternative path. Hence, the system was cleared at time 34, i.e., RCT=34. network to accommodate the disturbed flow onto alternate paths and
Similarly, the RCT for test problems C2 , C3 , and C4 are 32, 25, and 29, redirect them to safe shelters within the same clearance time of the pre-
respectively. disruption plan (CT=22 units of time). The amount of disturbed flow,
The rerouted clearance times are different for each problem instance the time range of flow disturbance as well as the rerouting time in­
because the input data for the four cases are different in terms of the tervals, are shown in Table 6.
disrupted arcs, the arc disruption times, and the updating times. All of
these factors have influence on the amount of clearance time of the 4.1.2. Effect of disruption times
network. In the following, we provide a sensitivity analysis on the The occurrence time of an incident can be an important factor
evacuation plans under different problem settings for these factors. influencing the rerouted clearance time. For instance, if an incident
occurs on an arc, it will create a disruption only to evacuees who have
4.1.1. Effect of the network topology (disruption on arcs) not yet passed the incident location in the route. Hence, this section
When an incident happens in the evacuation network, the time of studies the effect of arc disruption times on the evacuation and rerouting
completion of the rerouting process can be impacted depending on process. For this purpose, a set of disturbed arcs are chosen as
different factor such as the location of the road in the network, its level Adisturbed = {(2, 4), (4, 8)}. The disruption times of the arcs are changed
of connectivity to other roads, the residual (remaining) capacity of the in the interval [1, 21], and an update will be triggered one unit after arc

11
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

Fig. 14. Rerouting Evacuation plan for Houston transportation network.

disruption times. Fig. 12 (a) shows the RCTs when the disruption time of less than a second for running each preprocessing algorithm as well as the
the arcs changes between time 1 and time 21. As can be seen, the MIP model on the small network of Fig. 8.
rerouting CT is higher when the disruption time occurs earlier in the
planning horizon. This is not surprising because as a disruption happens
earlier on the road, more evacuees are affected, and more time is 4.2. Numerical experiments on a large-scale network
required to reroute them and clear the system.
We continue the experiments on an evacuation network of the
4.1.3. Effect of information (updating time) Greater Houston area [11]. Houston, Texas, the fourth largest city in the
Shortly after an incident occurs within the network, it starts to disrupt U.S., is known to be one of the most vulnerable metropolitan cities sit­
the evacuation flow. It takes time for the planners to (1) understand and uated on the Gulf Coast, and has been severely affected by hurricanes
analyze the situation, (2) make an appropriate decision on the rerouting and floods for the several decades. The Houston network (Fig. 13)
strategy, and (3) execute the reroute plan accordingly. To account for this comprises of a total of 42 nodes and 107 arcs. The first thirteen nodes
delay in reroute planning, we introduce an updating time tupdating to cap­ represent source nodes (N 1 -N 13 ), and the last four nodes (N 39 -N 42 )
ture the total time it took from the incidence analysis until the new plan is represent safe destination nodes.
ordered to be executed. The effect of tupdating on the evacuation process is For the purpose of demonstrating our proposed evacuation rerouting
studied and is shown in Fig. 12 (b). The more delay there is in receiving approach, we used the same input data for this network as it was re­
update information on the network situation, the longer the clearance ported in [11]. Hence, the total number of evacuees (i.e., evacuation
time of the network after disturbed flow reassignments. Note that it took vehicles) on the source nodes are assumed to be 56,600, in which each of
source nodes 1-6 has 100 evacuees, 3,500 for each of nodes 7-10, and 14,

12
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

000 evacuees each for nodes 11-13. The transit times are defined to be required to mobilize disturbed evacuees to the safe shelters. Numerical
multiplies of τ = 30 minute intervals. Using the PBM model in experiments were thoroughly conducted to study the performance of the
Appendix A, we first generate the pre-disruption evacuation plan using proposed RPBM under different problem configurations. Computational
140 candidate paths to be selected in the optimization model. The experiments were made to test computational efficient in solving the
resulting pre-disruption plan distributed the evacuation flow over 52 proposed approach. The impact of three incident-related factors have
selected paths, and it took 129τ to clear the network. Disruptions are been investigated to better understand their effects on the rerouting
triggered on arcs (22,42), (20,32), (11,27), and (35, 34) at times 56, 69, process, including the location of the disruption, the time of disruption
67 and 48, respectively. These road disruptions affect flows on 11 paths occurrence, and the plan updating time. The results showed that more
and result in 10,966 evacuees being stranded behind nodes 11, 20, 22, flow was disturbed if an incident occurs earlier during the evacuation,
and 35. Among these evacuees, 600 are on node 11, 1,675 on node 20, 7, which leads to a greater amount of time to reroute the affected flow and
331 on node 22, and 1,360 on node 35. The RPBM is used to provide a clear the system. When the time for plan updates was delayed (i.e., the
reroute plan for the disturbed flow. The total combined computation rerouting process takes place later), the clearance time of the network
time of running both Algorithm 1 and Algorithm 2 was 67.91 seconds, increased accordingly. This emphasizes the importance of making
while it took 203.60 seconds to run Algorithm 3 to calculate RCT. It took timely decisions for fast response to incidents as it is crucial in an effi­
52.41 seconds to solve the RPBM based on the RCT value. The corre­ cient evacuation rerouting plan. The proposed approach has also been
sponding reroute plan is shown in Fig. 14. In the figure, the amount of tested on a large-scale evacuation network, and the results support that
rerouted flow from each node during different time intervals are illus­ the proposed approach can handle in evacuating large metropolitan
trated. The total time taken to move all disturbed flow to safe shelters areas in a timely manner.
was 162τ. This means that our model required approximately 16 hours The proposed approach is limited to deterministic problem settings.
to adjust the plan and rearrange 251,850 disturbed evacuees to safe As a future work, one can extend this paper by considering various
shelters through the residual capacity of the Houston network after uncertainties such as variations in the number of would-be evacuees,
disruption. alternative road capacities, or evacuees’ behavior. Another venue for an
extension is to extend the proposed approach to make it as a basis for
5. Conclusion vulnerabilities analysis through developing a probabilistic mechanism
that accounts for factors including simultaneous multiple occurrences.
This paper introduced a real-time rerouting evacuation strategy that
can be applied to post disruption networks to minimize evacuation CRediT authorship contribution statement
clearance time in response to the occurrence of real-time incidents. For
this purpose, a dynamic network flow optimization model formulation Ayda Darvishan: Investigation, Methodology, Data curtion, Vali­
(RPBM) was introduced, in which variable evacuation flow rates are dation, Writing – review & editing. Gino J. Lim: Conceptualization,
considered to develop alternative paths to achieve more practical and Supervision, Writing – review & editing, Writing – original draft.
effective evacuation plans. Due to very high-level complexity for
developing a practically useful optimization model, computational al­
gorithms have been developed to calculate a few key values for specific Declaration of Competing Interest
parameters related to road disruption. As a result, it enabled us to
develop a simple and computational efficient MIP model for the evac­ The authors declare that they have no known competing financial
uation problem. A rerouting clearance time calculation algorithm is interests or personal relationships that could have appeared to influence
introduced to efficiently calculate the minimum amount of time the work reported in this paper.

Appendix A

Path-Based Model [11]: The path-based model (PBM) for the evacuation planning problem is described below (18)-(22). The objective function of
the program minimizes the sum of unmet demand βn of all source nodes n ∈ N o . Constraint (19) guarantees that the total flow that exits a source node
n ∈ N o plus the remaining flow on the source node (unmet demand) is greater than the demand on the source node.
Constraint (20) places a restriction on the total amount of flow that reaches arc a ∈ A at time t ∈ T and ensures that it is less than the capacity of the
road Ca . Parameter δpa has binary values and takes value 1 only if arc a ∈ A belongs to path p ∈ p. Variable f p(t− θpa ) showcases the flow that exits the
source node of path p ∈ p at time t − θpa . Since θpa represents the transit time of the flow from the origin of the path to arc a ∈ A , the flow reaches arc
a ∈ A at time t ∈ T. The total flow that reaches a destination node is restricted by the capacity of the node. Non-negativity and integrality conditions of
the variables are shown in Constraints (22) and (23).
Minimize
∑
βn (PBM) (18)
n∈N o

Subject to:
∑∑
fpt + βn ≥ Dn , ∀n ∈ N o , (19)
p∈p+
n
t∈T

∑
δpa fp(t− θpa ) ≤ Ca , ∀a ∈ A , ∀t ∈ T, (20)
p∈p

∑∑
Kpn fpt ≤ ln , ∀n ∈ N d , (21)
p∈p t∈T

fpt ∈ Z+ ∀p ∈ p, ∀t ∈ T, (22)

13
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

βn ∈ Z+ ∀n ∈ N o (23)

Appendix B

This section is to explain our motivation for the proposed evacuation reroute planning model formulation. The pre-processing algorithms help us to
develop a less complex optimization model which can be solved in a timely manner for the application discussed in this paper. Let us explain what
happens if we were going to solve the optimization without the pre-processing algorithms. The main challenge lies in calculating the amount of
disturbed flow (Hpnt ). If the proposed pre-processing algorithms are not used, the following constraint should be included in the optimization
framework to calculate the amount of disturbed flow (Hpnt ) on node n ∈ N of path p ∈ p at time t ∈ T.
⃒ ⃒ ( ⃒ ⃒ )
⃒t + θpa + τa − DTa ⃒ ∑ ⃒t + θ ′ + τ ′ − DT ′ ⃒ ⃒ ⃒
pa a a
Hnp(t+θ′ pn ) ≥ fpt γna δpa − γna′ δpa′ + ⃒Npn ⃒ M∀p ∈ p, n ∈ N , t ∈ T,
t + θpa + τa − DTa ′ t + θpa′ + τa′ − DTa′
n ∈Npn

where M is an arbitrarily large number and Npn is the set of all preceding nodes to node n ∈ N of path p ∈ p. Parameters γna and δpa are used to
represent the topology of the network (see the notation in p. 5). Flow f pt departs from the origin of path p ∈ p at time t ∈ T and reaches the end of arc
a ∈ A at time t + θpa + τ a . The term |t +θpa +τ a − DTa |/t +θpa + τ a − DTa is used to monitor arc disruption time DTa on the flow, f pt . If a disruption on
the arc occurs after the flow has passed the incident arc, then t + θpa + τ a < DTa ; hence, |t + θpa + τ a − DTa |/t + θpa + τ a − DTa = − 1. If the
disruption happens before the flow arrives at the end of the arc, then we will have |t + θpa + τ a − DTa |/t + θpa + τ a − DTa = 1. To indicate whether
flow f pt is disturbed by arc a ∈ A , we also need to take into account the effect of disruption times of the preceding arcs to arc a ∈ A . If the flow is not
∑ ⃒ ⃒
interrupted by any preceding arcs to the incident arc, we will have γna′ δpa′ |t + θpa′ + τa′ − DTa′ |/t + θpa′ + τa′ − DTa′ = − ⃒Npn ⃒. Hence, if flow f pt
n′ ∈Npn
is stopped due to the disruption on arc a ∈ A (which emerges from node n ∈ N ) and not any other preceding arcs, the following equations hold:
⃒ ⃒/
⃒t + θpa + τa − DTa ⃒ t + θpa + τa − DTa = 1
/
∑ ⃒ ⃒ ⃒ ⃒
γna′ δpa′ ⃒t + θpa′ + τa′ − DTa′ ⃒ t + θpa′ + τa′ − DTa′ = − ⃒Npn ⃒
′
n ∈Npn

This leads to Hnp(t+θ′ = f pt if the objective function is minimized, i.e., the amount of disturbed flow is minimized. For any other scenarios, the
pn )

constraint can be relaxed with an exception when the disruption occurrence time equals the time that the flow arrives at the end of the arc (DT a = t +
θpa + τ a ). In this case, we will have the following expression 00 in the constraint, which cannot be not defined.
To overcome this modeling complexity, we proposed to calculate the values of parameter Vpnt a priori using the pre-processing algorithms. As a
result, it simplifies the equation Hpn(t+θ′ ) = Vpnt f pt in the RPBM as Vpnt is not an unknown variable, but a parameter.
pn

Appendix C

Table 7

References

[1] Beroggi GE, Wallace WA. Operational control of the transportation of hazardous
Table 7 materials: an assessment of alternative decision models. Manage. Sci. 1995;41(12):
Node sequence of candidate paths. 1962–77.
[2] Bier VM, Hausken K. Defending and attacking a network of two arcs subject to
Sequences Sequences traffic congestion. Reliab Eng Syst Saf 2013;112:214–24.
Paths P1 1-4-7-9 Paths P22 2-4-5-8-10 [3] Chen L, Miller-Hooks E. Resilience: an indicator of recovery capability in
P2 1-4-6-7-9 P23 2-4-7-8-10 intermodal freight transport. Transport. Sci. 2012;46(1):109–23.
[4] De Silva FN, Eglese RW. Integrating simulation modelling and GIS: spatial decision
P3 1-4-6-8-10 P24 2-4-6-7-8-10
support systems for evacuation planning. J Oper Res Soc 2000;51(4):423–30.
P4 1-4-5-6-7-9 P25 2-4-6-8-7-9
[5] Faturechi R, Miller-Hooks E. Travel time resilience of roadway networks under
P5 1-4-5-6-8-10 P26 2-4-8-10
disaster. Transport. Res. Part B: Methodological 2014;70:47–64.
P6 1-4-5-8-10 P27 2-5-4-7-9 [6] Faturechi R, Miller-Hooks E. Measuring the performance of transportation
P7 1-4-7-8-10 P28 3-5-6-7-9 infrastructure systems in disasters: a comprehensive review. J Infrastruct Syst
P8 1-4-6-7-8-10 P29 3-5-6-8-10 2015;21(1). 04014025.
P9 1-4-6-8-7-9 P30 3-5-8-10 [7] FEMA, D. (2010). Developing and maintaining emergency operations plans, 1-124.
P10 1-4-8-10 P31 3-4-7-9 [8] FEMA. Producing emergency plans: a guide for all-hazard emergency operations
P11 1-5-6-7-9 P32 3-4-6-7-9 planning. 2008.
P12 1-5-6-8-10 P33 3-4-6-8-10 [9] Ford Jr LR, Fulkerson DR. Flows in networks. Princeton university press; 2015.
P13 1-5-8-10 P34 3-5-4-7-9 [10] Kim S, Shekhar S, Min M. Contraflow transportation network reconfiguration for
P14 2-4-7-9 P35 3-5-4-6-7-9 evacuation route planning. IEEE Trans Knowl Data Eng 2008;20(8):1115–29.
P15 2-4-6-7-9 P36 3-5-4-6-8-10 [11] Lim GJ, Zangeneh S, Baharnemati MR, Assavapokee T. A capacitated network flow
optimization approach for short notice evacuation planning. Eur J Oper Res 2012;
P16 2-4-6-8-10 P37 3-5-7-9
223(1):234–45.
P17 2-5-6-7-9 P38 3-5-6-7-8-10
[12] Lim GJ, Baharnemati MR, Kim SJ. An optimization approach for real time
P18 2-5-6-8-10 P39 3-5-6-8-7-9
evacuation reroute planning. Ann Oper Res 2016;238(1–2):375–88.
P19 2-5-8-10 P40 3-5-8-7-9 [13] Lim GJ, Rungta M, Davishan A. A robust chance constraint programming approach
P20 2-4-5-6-7-9 P41 3-4-5-6-7-9 for evacuation planning under uncertain demand distribution. IISE Trans 2019;51
P21 2-4-5-6-8-10 P42 3-4-5-6-8-10 (6):589–604.

14
A. Darvishan and G.J. Lim Reliability Engineering and System Safety 214 (2021) 107644

[14] Lim GJ, Zangeneh S, Kim SJ. Clustering approach for defining hurricane [17] Robert T. Stafford disaster relief and emergency assistance act. Public Law 2000;
evacuation zones. J Urban Plann Dev 2016;142(4). 04016008. 10:106–390.
[15] Lin YK, Chang PC, Fiondella L. Quantifying the impact of correlated failures on [18] Rungta M, Lim GJ, Baharnemati M. Optimal egress time calculation and path
stochastic flow network reliability. IEEE Trans Reliab 2012;61(3):692–701. generation for large evacuation networks. Ann Oper Res 2012;201(1):403–21.
[16] Pidd M, De Silva FN, Eglese RW. A simulation model for emergency evacuation. [19] Sheffi Y, Mahmassani H, Powell WB. Evacuation studies for nuclear power plant
Eur J Oper Res 1996;90(3):413–9. sites: a new challenge for transportation engineers. ITE J 1981;51(6):25–8.

15