Computers & Industrial Engineering 174 (2022) 108772

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

An improved column generation algorithm for the disrupted flight recovery

problem with discrete flight duration control and aircraft
assignment constraints
Jie Li a, Kunpeng Li a, Qiannan Tian b, *, P.N. Ram Kumar c
a
School of Management, Huazhong University of Science and Technology, Wuhan 430074, China
b
Hubei Logistics Development Research Center, Hubei University of Economics, Wuhan 430205, China
c
Quantitative Methods & Operations Management Area, Indian Institute of Management Kozhikode, Kerala, India

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

Keywords: The disrupted flight recovery problem is well-studied in the literature owing to its significant impact on airlines
Flight recovery problem and passengers. In this work, we consider the disrupted flight recovery problem with two new realistic aspects, i.
Flight duration e., a new implementation for changing flight duration as a recovery option and considering the aircraft
Assignment constraints
assignment constraints. Firstly, we develop a new mixed-integer quadratic programming model encapsulating a
Column generation
functional relationship between the reduced flight duration and the changes in fuel consumption using a
piecewise function. Secondly, we propose an improved column generation approach to solve some scenarios
based on actual flight data obtained from an airline. Lastly, the experimental results show that the improved
column generation algorithm can obtain the optimal solution for all scenarios. Compared with not considering
changing the flight duration, considering changing the flight duration can save about 24% of recovery costs on
average. In addition, we analyze the effect of enforcing aircraft assignment constraints and different fuel prices
on recovery costs. Based on the experimental results, we discuss how to support practitioners in choosing the
appropriate combination of recovery options.

1. Introduction routes serviced, destinations, aircraft types, and fleet size to improve
operational efficiencies and profitability.
According to a McKinsey study, the year 2020 saw the revenues of One of the airline industry’s most challenging problems is disrup­
the airline industry totalling $328 billion, which is just about 40 % of the tions to airline schedules due to mechanical failures, bad weather, or
previous year. In nominal terms, this value is the same as in 2000. The crew absenteeism. Any of these events can have a devastating impact on
study expects the industry to be smaller in the coming years, with traffic the operations. Some direct and immediate consequences are capacity
volumes not expected to return to 2019 levels until 2024 (Jaap et al., reduction, build-up of delays, cancellations, and changes in crew ros­
2021). Airlines worldwide had to tide off the COVID-19 crisis by tering decisions. These disruption costs are roughly 5 % of airline rev­
borrowing vast sums of money to make ends meet and deal with daily enue or about $35 billion globally. And this figure is revised up to $60
cash burn. Most of them survived by taking advantage of state aid, lines billion when we factor in the cost of lost passenger productivity and lost
of credit and bond issuance. The International Air Transport Association business in support industries like hospitality, business services and
(IATA) estimates that the industry accumulated more than $180 billion tourism (Amadeus, 2020). When these unexpected events occur, airlines
in debt in 2020 (Alexandre, 2020). One of the crucial factors that drive immediately initiate recovery operations. In practice, typically, recovery
the future of the aviation sector is customer demand. A Deloitte report operations happen sequentially in stages. Airlines usually first focus on
forecasts that leisure trips primarily drive customer demand as fewer flight schedule recovery, followed by crew and passenger recovery.
corporates are likely to make business trips because of remote work and This work is motivated by a real-life problem faced by an airline
other flexible working arrangements (Anisha, 2021). Given this sce­ headquartered in China. When an unexpected event causes an inter­
nario, airlines are under intense pressure to re-evaluate the economics of ruption, the airline must recover the flight schedule within a stipulated

* Corresponding author.
E-mail addresses: [email protected] (J. Li), [email protected] (K. Li), [email protected] (Q. Tian), [email protected] (P.N. Ram Kumar).

https://doi.org/10.1016/j.cie.2022.108772
Received 8 June 2022; Received in revised form 13 September 2022; Accepted 24 October 2022
Available online 30 October 2022
0360-8352/© 2022 Elsevier Ltd. All rights reserved.
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

time. To solve the flight recovery problem, we employ a variety of re­ costs, delay costs, swap costs, cancellation costs and penalties for
covery operations such as flight delays, maintenance delays, flight violation of aircraft assignment constraints:
swaps, and flight cancellations. Some recent studies, such as Arıkan
et al., 2017, address the cruising speed’s recovery option. Most early • Extent of departure delays of different flights & maintenances
studies considering cruise speed recovery options were based on the • Optimal duration of different flights
EUROCONTROL aircraft performance model BADA (Base of Aircraft • Flights to be cancelled
Data), a fuel estimation method based on physical model. Due to the • Reassignment of the available aircraft to the flights
unavailability of aircraft engine and model information, especially new
engine and aircraft models (Yanto and Liem, 2018), and commercial Based on the above, the contributions of this paper are summarized
sensitivity (Zhang and Mahadevan, 2020), this method may not be as follows. First, we propose a new mathematical model for changing
feasible in practice. Another fuel estimation method is based on data- flight duration as a recovery option. By expressing the nonlinear func­
driven model. In recent years, many studies have used machine tional relationship between flight duration reduction and fuel con­
learning methods to establish data-driven model for fuel estimation sumption change as a piecewise function, this model can be compatible
(Khan et al., 2019; Khan et al., 2021; etc.). Our approach can take with data-driven fuel estimation models in recent studies. Furthermore,
advantage of the results based on data-driven models as an alternative this model can be used when methods by means of physics-based fuel
when physical based models such as BADA are not available. For the estimation models such as BADA are not available in existing studies. We
sake of generality, we use a piecewise function to describe the discrete propose a mixed integer quadratic program that can be solved using
functional relationship between the reduction of flight duration and the commercial solvers. In order to solve larger-scale problems, an improved
changes of fuel consumption. This step obliviates the need for physical column generation algorithm including a master problem and multiple
models. subproblems is proposed. This algorithm avoids the limitation of the
Another crucial aspect that has received little attention in the liter­ column generation technique proposed by Liang et al. (2018) that only
ature is the importance of the aircraft assignment constraints. The connections can be made when the departure time of the first flight is
aircraft assignment constraints means that the number of aircraft at each earlier than that of the second flight. In our algorithm, even if the de­
airport after the recovery period shall not be less than the required parture time of the first flight is not earlier than the second flight, it is
number. Enforcing the aircraft assignment constraints can ensure that still possible to obtain a feasible connection by delaying the departure
there are enough aircraft to carry out the flights after the recovery time of the second flight to avoid missing the optimal solution. A
period. The desired fleet size obviates the need to cancel or delay other modified label-setting algorithm is used to solve the pricing subproblem,
flights after the recovery period. Fig. 1 illustrates an example of a re­ where the amount of flight duration change is determined during label
covery plan that does not consider aircraft assignment constraints. When expansion by solving a small-scale linear programming. The branch-and-
the recovery plan ends, there are no aircraft at Airport 3, but there are bound framework is used after obtaining the optimal solution to the
flights departing from Airport 3 in the flight plan for the next day. Hence master problem to ensure that the optimal integer solution is obtained.
flight f2 and f3 may be cancelled due to lack of aircraft. In essence, the In the experimental part, the performance of the algorithm is explored
flight disruption is not solved by the recovery plan and spreads to the and analyzed, and the sensitivity analysis of some parameters involved
next day. On the other hand, if we consider the aircraft assignment in the fuel consumption function is carried out.
constraints, flight f1 should be canceled so that the flight plan for the Second, in both mixed integer quadratic programming model and
next day can operate normally without cancelling flight f2 andf3 . Note master problem model of column generation, aircraft assignment con­
that aircraft assignment constraints differ from minimizing changes in straints are considered and added to the objective function in the form of
flight plans because changes that can reduce costs are encouraged as penalty costs. When the modified label-setting algorithm obtains col­
long as aircraft assignment constraints are met. The aircraft cannot umns with negative reduced cost for each aircraft, the destination
arrive/depart when the airport is closed due to inclement weather or airport information of the corresponding aircraft after the end of the
other reasons. Therefore, a small number of severely closed airports recovery plan is passed to the master problem for constraining in the
chosen by the airline do not consider aircraft assignment constraints in master problem. In experiments, we report and analyze the effects of
this paper. aircraft assignment constraints.
Given this background, we restrict the scope to the flight recovery The remainder of the paper is as follows. Section 2 reviews the
problem in the following context: For a given flight schedule, a set of pertinent literature and highlight the research gap. In Section 3, we
disruptions, a set of maintenances, and a discrete function between the provide the problem description and mathematical formulation. Section
reduced flight duration and the changes in fuel consumption, the 4 explains the column generation approach coupled with different ac­
problem is to address the following issues while minimizing the total celeration strategies. We report the computational results in Section 5,
costs, i.e., the sum of changed fuel consumption costs, carbon emission followed by conclusions and ideas for future research in Section 6.

Fig. 1. Example for aircraft assignment constraints.

2
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

2. Literature review to fuel cost per unit mass. The fuel costs were based on Jeppesen’s flight
planning engine. A significant drawback of this work is the consider­
We select a few representative articles published in this field in the ation of fuel consumption costs for flights whose duration is only more
last ten years as a point of reference. For a thorough review of the than 3 h.
literature, interested readers can refer to the latest review work of Most of the literature uses the fuel estimation model of physical
Hassan et al. (2021) on airline disruption management for the years model (Aktürk et al., 2014; Gürkan et al., 2016; Arıkan et al., 2016;
2009 – 2020 and Clausen et al. (2010) for earlier years. We divide the Arıkan et al., 2017; etc.). These studies base their work on the EURO­
literature review section into three sub-sections: First, we review the CONTROL aircraft performance model BADA (Base of Aircraft Data),
studies that do not consider cruise speed control as a recovery option. which may have some limitations. The speed schedules that are part of
Second, we review based on the solution methodology and last, we the BADA model may differ from those used in reality (Turgut et al.,
analyze studies that accounted for the cruise speed as a recovery option. 2014; Wasiuk et al., 2015; Wasiuk et al., 2016). In practice, parameters
When developing a recovery plan, the typical recovery options may be unavailable due to missing information (Trani and Wing-Ho,
include flight delays (Brunner, 2014, etc.), flight cancellations (Hu et al., 1997). Information on some aircraft engines and types may not be
2015, etc.), aircraft swaps (Hu et al., 2016, etc.), and flight creation available, especially new engines and types (Yanto and Liem, 2018).
(Jozefowiez et al., 2013; Sinclair et al., 2014; Zhang et al., 2016). In Unavailability of data may result in inaccurate default values for global
addition to this, some studies have considered airport closures (Eggen­ parameters rather than proper local parameters, which may cause the
berg et al., 2010; Lee et al., 2020; Vink et al., 2020) or airport flow results to be erroneous (Pagoni and Psaraki-Kalouptsidi, 2017; Khan
control (Zhang et al., 2016; Sinclair et al., 2016; Liang et al., 2018; et al., 2019). In the practice of an airline, the result of using physical
Huang et al., 2022). This work considers flight delays, flight cancella­ model to estimate fuel consumption deviates from the actual situation
tions, aircraft swaps and maintenance delays as possible recovery op­ (Khan et al., 2021). Due to commercial sensitivity, real-time aircraft
tions apart from cruise speed. We do not consider flight creation a status information cannot be obtained, which will make it difficult to
recovery option due to the required complex approval process. implement the physical model (Zhang and Mahadevan, 2020). Our
We can broadly divide the literature concerning the solution meth­ approach can use the fuel estimation results based on data-driven
odologies into heuristic and exact methods. The majority of the studies models (Yanto and Liem, 2018; Khan et al., 2019; Zhang and Mahade­
consider heuristic techniques. Eggenberg et al. (2010) proposed a van, 2020; Khan et al., 2021; etc.), which can be used as an alternative
heuristic-based column generation algorithm and discretized continuous when physical models (such as BADA) are unavailable or have large
timeline into time windows. Jozefowiez et al. (2013) introduced a deviations.
heuristic method for aircraft recovery and passenger assignment. Simi­ To summarize, this work fills a significant research gap in the liter­
larly, Sinclair et al. (2014) studied integrated aircraft and passenger ature by addressing two critical aspects. Firstly, without relying on the
recovery problems using a large neighbourhood search (LNS) algorithm. BADA model, we propose a new flexible approach for accounting for fuel
Zhang et al. (2015) proposed a two-stage heuristic for the aircraft and consumption by discretizing the flight duration in units of one minute.
crew recovery problem based on a multi-commodity network model. Hu This provision makes it convenient for the airlines, irrespective of the
et al. (2016) developed a heuristic methodology named “GRASP algo­ model, to obtain fuel consumption parameters based on historical flight
rithm” for the aircraft and passenger recovery problem. Liang et al. operation data analysis, i.e., a data-driven model. This can also be used
(2018) proposed a column generation-based heuristic algorithm to solve as a substitute when physical models for fuel estimation, such as BADA
the aircraft recovery problem. Vink et al. (2020) proposed a selection model, are unavailable on new aircraft or under other circumstances.
algorithm with MILP. Lee et al. (2020) developed an approach to opti­ Secondly, this work is the first to study the importance of aircraft
mize recovery decisions by exploiting partial and probabilistic pre­ assignment constraints. Quantifying the total recovery costs with and
dictions of future disruptions and proposed an approximation algorithm. without the aircraft assignment constraints helps practitioners arrive at
Compared with heuristics, relatively, very few articles study the prob­ the most appropriate set of recovery options.
lem using exact methodologies. Brunner (2014) presented a linear
integer programming to solve the inbound flights’ reassignment due to 3. Problem description and formulation
capacity shortage. Similarly, Hu et al. (2015) introduced an integrated
integer programming model for aircraft and passenger recovery. In this Before describing the problem, we describe some of the elements
work, we use exact methodologies by proposing a mixed-integer involved.
quadratic programming model and an improved column generation Airports: The airports are the connecting nodes for the aircraft to
algorithm. carry out a set of consecutive flights. When an airport is unavailable, no
Lastly, a few studies consider cruise speed control as a recovery aircraft can fly the original flight schedule from that airport. A closed
option. Aktürk et al. (2014) pioneered the usage of cruise speed as a airport means that the airport is closed or restricted for a period of time.
decision variable while taking environmental constraints and cost into Aircraft: Each aircraft has a start and end time regarding its avail­
the route recovery optimization model. The model allows studying the ability. After completing a flight, the plane will arrive at an airport and
trade-offs between flight delays and recovery costs. Based on conic depart from the same airport for its next flight. We do not consider the
mixed-integer programming, the authors formulate a nonlinear recovery option of ferrying the aircraft to other airports.
optimization model that considers one swap per aircraft during the re­ Flight: The aircraft travels from the departure airport to the arrival
covery period. Gürkan et al. (2016) incorporated cruise time control into airport to complete a flight and has a scheduled departure and arrival
a comprehensive model of scheduling, fleet assignment and aircraft time. Flights are of limited duration and delays are allowed.
routing problem within the scope of daily planning. Arıkan et al. (2016) Maintenance: This task requires a particular aircraft to stay at a
developed a fuel cost function for flights based on Airbus’s technical specified airport with start and end times. The maintenance of an
report. Based on cruising speed, they comprehensively consider aircraft aircraft is identical to a specific type of flight whose origin and desti­
and passenger recovery options. The downside is that the proposed nation airports are the same with zero passengers. Some maintenance
mathematical formulation is difficult to extend the model to other re­ can be performed by replacing the aircraft, while others cannot (Vink
covery actions flexibly. Arıkan et al. (2017) developed a flight network et al., 2020).
model based on aircraft, crew, and passenger flow for the integrated The development of a flight recovery plan should broadly follow the
airline recovery problem. Besides, Marla et al. (2017) used a cost index- flight operation mechanism. When two flights are flown sequentially by
based flight planning mechanism to change the flight speed on long-haul the same aircraft, the arrival airport of the previous one and the de­
flights. The authors define the cost index as the ratio of cost per unit time parture airport of the next one must be the same. In addition, after

3
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

completing the last flight, the aircraft needs to remain on the ground for Tinf Set of discrete values of allowable reduced flight duration of
⃒ ⃒
a minimum turnaround time to fly the next flight. A limited fleet of flightf, and ⃒Tinf ⃒ is the number of discrete values of allowable reduced
heterogeneous aircraft with different available time windows is located flight duration of flightf.
at different airports. Each aircraft has additional swap costs when per­ Indices.
forming flights other than the original plan. These costs may include the a Airport index;
passenger cancellation costs because of the aircraft capacity restrictions. p Aircraft index;
Different airlines may have different maintenance schedules. All the g Fleet index;
aircraft must comply with the regulations on regular maintenance and f,f1 , f2 Flight or maintenance index;
complete the maintenance within the specified time. Here, we assume i,j Sequence of flights index;
that the aircraft’s maintenance requirements are known a priori during t Discrete values index.
the recovery period. Once a disruption occurs, one or a combination of Parameters.
the following five recovery options are activated to generate a recovery clf Cancellation cost of the flight f ,∀f ∈ F;
plan. hag Number of aircraft required of type g at the airport a at the end
Changing flight duration: flight duration can be changed through of the recovery period;
aircraft speed control. Speeding up aircraft can reduce the actual ehag Number of aircraft of type g at the airport a at the beginning of
duration of the flight, thereby reducing delays. As shown in Fig. 2, the original flight plan;
maintenance conflicts with the flight departure time. By delaying the spa A boolean indicator equals 1 if airport a is the start airport of the
flight departure time and reducing the flight duration, the actual arrival
aircraft p in the original plan and 0 otherwise,∀p ∈ P, a ∈ A;
time of the flight will not be delayed. Flight duration varies with aircraft
dtf Departure time of the flightf, the meaning of dtf is the start time
speed, along with changes in fuel consumption and carbon emissions.
of the maintenance if f indicates maintenance;
Flight delays: The flight can be delayed appropriately within the
atf Arrival time of the flightf, the meaning of atf is the end time of
maximum delay time constraint to avoid more significant costs caused
the maintenance if f indicates maintenance;
by the immediate cancellation of the flight.
gtp Minimum turnaround time between two flights for aircraftp;
Maintenance delays: Maintenance can be delayed appropriately
ftp Type of aircraftp;
within the permissible time windows.
dafa A boolean indicator equals 1 if airport a is the departure
Flight swaps: As per the original flight plan, each flight is performed
by specific aircraft. Other aircraft can execute any scheduled flight with airport of the flight f and 0 otherwise,∀f ∈ F, a ∈ A;
additional swap costs in the recovery process. Since maintenance can be aafa A boolean indicator equals 1 if airport a is the arrival airport of
considered as a special flight at the same departure and arrival airports, the flight f and 0 otherwise,∀f ∈ F, a ∈ A;
the flight swaps include flexible maintenance swaps. Tfup Maximum flight arrival delay allowed;
Flight cancellations: When the cost of cancelling a flight is lower than [stp , etp ] Time window of the availability of the aircraftp;
executing it with different recovery options, the airline can cancel the Tcrf Duration of the flight or maintenance f in the original plan;
flight. However, flight passengers are entitled to adequate compensation Tcrlf Minimum allowed duration of the flight or maintenance f in
in such cases. the original plan, if f indicates maintenance, thenTcrlf = Tcrf ;
Thus, using the five recovery options, the problem is to minimize the ismf A boolean indicator equals 1 if f represents maintenance and
additional fuel consumption and carbon emission costs, costs associated 0 otherwise;
with delays, costs of swapping the aircraft, flight cancellation costs and faf Aircraft that is to perform flight f in the original plan, if f in­
penalty costs. The next section proposes a mixed-integer quadratic dicates maintenance, the meaning of faf is the aircraft due to for
programming (MIQP) model.
maintenancef;
Mathematical model (MIQP).
SCpf Swap cost of aircraft p to perform flightf , equals 0 if aircraft p
Sets.
is initially assigned flightf;
A Set of all airports;
gf Passenger delay cost of the flight f per unit time;
A Set of airports considering aircraft assignment constraints,A ⊂A;
′ ′

cfuel Fuel price;

P Set of all aircraft;
cco2 Unit carbon emission cost;
G Set of fleets;
rc Carbon dioxide emission constant;
F Set of all flights and maintenances;
fvpft Changes in fuel consumption of the aircraft p performing flight
M Set of all maintenances,M⊂F;
f with reduced flight durationt;
Index Set of the sequence of flights visited by the same
crucf Maximum flight duration allowed to be reduced for flightf;
aircraft,Index = {1, 2, ..., |F| };

Fig. 2. An example of changing flight duration.

4
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

boundpft Dividing points of piecewise function of changes in fuel

(5)
′

consumption for aircraft p performing flightf, andboundpf0 = 0; QAag ⩾0, ∀a ∈ A , g ∈ G

W1 Penalties for violation of aircraft assignment constraints. When
zpf1 i + zpf2 (i+1) ⩽1, ∀p ∈ P, f1 ∈ F, f2 ∈ F, f1 ∕
= f2 , i ∈ Index, i < |F|, aaf1 a
its value is large, it means that the constraints are enforced, and when it
is0, it means that the constraints are ignored; = daf2 a
∕ (6)
W A large number. ( )
Decision variables. df − Tcrf − tcrf ⩽Tfup , ∀f ∈ F (7)
zpfi A boolean indicator equals 1 if f is the i th flight or maintenance ∑
executed by aircraftp, and 0 otherwise; stp zpfi ⩽dtf + df , ∀p ∈ P, f ∈ F (8)
df Delay time of departure of flightf; i∈Index

tcrf Actual duration of flightf; ( )

∑ ∑
wfuelpf Actual changes in fuel consumption of aircraft p executing atf + df − Tcrf + tcrf ⩽etp zpfi + W 1 − zpfi , ∀p ∈ P, f ∈ F (9)
∑
flightf , with the condition i∈Index zpfi = 1 holding true; i∈Index i∈Index

QAag Shortage in the number of aircraft of type g ∈ G at the airport ∑∑

a that does not meet the aircraft assignment constraints; Tcrlf zpfi ⩽tcrf ⩽Tcrf , ∀f ∈ F (10)
fuelpf Changes in fuel consumption of aircraft p executing flightf;
p∈P i∈Index

fuelypft A boolean indicator equals 1 if aircraft p performs flight f

df ⩾0, ∀f ∈ F (11)
with reduced flight duration t and 0 otherwise.
Objective function and constraints.

( ) ( )
∑ ∑∑ ∑ ( ( ) )∑ ∑ ∑ ∑
Min clf 1 − zpfi + gf df − Tcrf − tcrf zpfi + (cfuel + cco2 rc)wfuelpf
p∈P i∈Index p∈P i∈Index p∈P f ∈F,ismf ⩽0
(1)
f ∈F f ∈F
( )
∑∑ ∑ ∑∑
+ SCpf zpfi + W1⋅ QAag
p∈P f ∈F ′
i∈Index a∈A g∈G

∑
zpfi = 1, ∀p ∈ P, f ∈ F, ismf = 1, p = faf (12)
i∈Index
∑ ∑
dafa zpfi = spa zpfi , ∀p ∈ P, a ∈ A, i = 1 (2) ( )
(
∑∑
)
f ∈F f ∈F df − Tcrf − tcrf ⩾W⋅ zpfi − 1 , ∀f ∈ F (13)
p∈P i∈Index

∑
zpfi ⩽1, ∀p ∈ P, i ∈ Index (14)
( ( ) ) ( ) ( ) f ∈F
atf1 + df1 − Tcrf1 − tcrf1 + gtp − dtf2 + df2 ⩽W 2 − zpf1 i − zpf2 (i+1) , ∀p
∑ ∑
∈ P, f1 ∈ F, f2 ∈ F, f1 ∕
= f2 , i ∈ Index, i < |F| zpfj ⩽ zpfi , ∀p ∈ P, i ∈ Index, j ∈ Index, i < j (15)
(3) f ∈F f ∈F

( ) ( ) ∑∑
∑ ∑ ∑ ∑ ∑ ∑ tcrf ⩽W⋅ zpfi , ∀f ∈ F, ismf = 0 (16)
ehag + aafa zpfi − dafa zpfi + QAag ⩾hag , ∀a p∈P i∈Index
p∈P,ftp =g f ∈F i∈Index p∈P,ftp =g f ∈F i∈Index
∑∑
′
∈ A ,g ∈ G df ⩽W⋅ zpfi , ∀f ∈ F, ismf = 0 (17)
p∈P i∈Index
(4)

∑ ∑∑
⎧ f vpf 0 zpfi Tcrf zpfi − tcrf = boundpf 0
⎪
⎪ i∈Index p∈P i∈Index
⎪
⎪ ∑ ∑∑
⎪
⎪ f vpf 1 zpfi boundpf 0 < Tcrf zpfi − tcrf ⩽boundpf 1
⎪
⎪
⎪
⎨ i∈Index p∈P i∈Index
... ∑...∑
wfuelpf = ∑ , ∀p ∈ P, f ∈ F, t ∈ Tinf , t⩽crucf (18)
⎪
⎪ f vpft zpfi boundpf (t− < Tcrf zpfi − tcrf ⩽boundpft
⎪
⎪
1)
⎪
⎪ i∈Index p∈P i∈Index
⎪
⎪ ... ∑ ...
⎪
⎩ ∑∑
f vpf |Tinf |− 1 zpfi boundpf (|Tinf |− 2 ) < Tcrf zpfi − tcrf ⩽boundpf (|Tinf |− 1 )
i∈Index p∈P i∈Index

5
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

and the change of flight duration is in the interval corresponding

tofuelypft = 1. Constraints (22) and (23) enforce the interval constraints
of the change of flight duration. Constraint (24) and Constraint (25)
zpfi ∈ {0, 1}, ∀p ∈ P, f ∈ F, i ∈ Index (19)
represent the actual changes in fuel consumption for the flight f actually
The Objective function (1) minimizes the sum of the costs associated performed by the aircraftp. Constraint (26) indicates the binary nature
with flight cancellations, flight delays, changes in fuel consumption & of the variables. Constraint (27) states that the actual fuel consumption
carbon emission costs, the swap costs of executing unplanned flights, change is non-negative.
and the penalty costs for violation of aircraft assignment constraints.
Constraint (2) stipulates that every aircraft departs for its first flight 4. Improved column generation
from its start airport. Constraint (3) forbids conflicts in flight schedules.
When an aircraft executes a series of flights, these flights will not conflict After preliminary experiments, the proposed model could only solve
with each other in time, which ensures the feasibility of the aircraft small problem instances to optimality in real-time. We develop a column
executing the flight plan. Equalities (4) and (5) report the shortage be­ generation-based solution methodology to overcome this drawback for
tween the number of aircraft at each airport after the recovery plan ends quickly solving even large problem instances. We model the master
and the actual number required. Constraint (6) ensures flow conserva­ problem as a set-partitioning problem, with the subproblem being a
tion. When the aircraft takes two flights in sequence, the arrival airport variant of the shortest path problem. The detailed flow chart of the
of the former flight should be the same as the departure airport of the improved column generation algorithm is illustrated in Fig. 3.
latter flight. Constraint (7) implies that the delay associated with any
flight is within its permissible limits. Constraints (8) and (9) account for
4.1. Master problem formulation
the availability time windows of an aircraft. The upper and lower limits
of any flight duration (actual) are accounted for by constraint (10).
Sets.
Constraint (11) ensures that no aircraft departs earlier than its stipulated
R Set of routes, where a route refers to a series of flights and
time. Constraint (12) forbids maintenance tasks from cancellation.
maintenance executed by the same aircraft.
Constraint (13) prohibits the early arrival of any flightf. Constraints (14)
Indices.
and (15) enforce time–space constraints by assigning only one flight f to
r Route index.
an aircraft p at any point in time. Constraints (16) and (17) imply that a
Parameters.
cancelled flight’s actual flight duration and departure delay are zero.
αpfr A boolean indicator equals 1 if flight f belongs to the route r of
Constraint (18) represents the functional relationship between the
the aircraft p and 0 otherwise;
reduced flight duration and the changes in fuel consumption in terms of
βpar A boolean indicator equals 1 if the end airport of the aircraft p
a piecewise function. The piecewise function is fitted according to the
estimated fuel consumption data of the aircraft to determine the split along the route r is airport a and 0 otherwise;
point and function value. The reduced flight duration is the difference cpr Cost of the route r of the aircraftp.
between the initially planned duration and the exact duration of the Decision variables.
flight. Constraint (19) indicates the binary nature of the variables. xpr A boolean indicator equals 1 if the aircraft p performs route r
Since Constraint (18) is nonlinear, we linearize it to obtain the and 0 otherwise;
following constraints so that the model can be solved by commercial yf A boolean indicator equals 1 if the flight f is cancelled and
solver. 0 otherwise.
∑ qag Shortage of aircraft that the recovery plan does not meet the
fuelypft ⩽1, ∀p ∈ P, f ∈ F, ismf ⩽0 (20) aircraft assignment constraints.
Model.
t∈Tinf ,t⩽crucf

∑ ∑ ∑ ∑∑
fuelpf = fvpft fuelypft , ∀p ∈ P, f ∈ F, ismf ⩽0 (21) Min cpr xpr + clf yf + W1⋅ qag (28)
t∈Tinf ,t⩽crucf r∈R ′
f ∈F a∈A g∈G

∑∑ ∑ ∑∑
Tcrf zpfi − tcrf ⩽ boundpft fuelypft , ∀p ∈ P, f ∈ F, ismf ⩽0 αpfr xpr + yf = 1, ∀f ∈ F (29)
p∈P i∈Index t∈Tinf ,t⩽crucf p∈P r∈R

(22) ∑ ∑
(30)
′
βpar xpr + qag ⩾hag , ∀a ∈ A , g ∈ G
∑∑ ∑
p∈P,ftp =g r∈R
Tcrf zpfi − tcrf > boundpf (t− 1) fuelypft , ∀p ∈ P, f
p∈P i∈Index t∈Tinf ,1⩽t⩽crucf ∑
∈ F, ismf ⩽0 (23) xpr ⩽1, ∀p ∈ P (31)
r∈R

wfuelpf ⩽fuelpf , ∀p ∈ P, f ∈ F, ismf ⩽0 (24) xpr ∈ {0, 1}, ∀p ∈ P, ∀r ∈ R (32)

( )
∑ yf ∈ {0, 1}, ∀f ∈ F (33)
wfuelpf ⩾fuelpf − W⋅ 1 − zpfi , ∀p ∈ P, f ∈ F, ismf ⩽0 (25)
i∈Index
(34)
′
qag ⩾0, ∀a ∈ A , g ∈ G
fuelypft ∈ {0, 1}, ∀p ∈ P, f ∈ F, t ∈ Tinf , t⩽crucf , ismf ⩽0 (26)
The objective function (28) minimizes the route costs, flight
cancellation costs and the penalty costs for violation of aircraft assign­
wfuelpf ⩾0, ∀p ∈ P, f ∈ F, ismf ⩽0 (27)
ment constraints. The route costs account for flight delay costs, swap
Constraint (20) means that the value of the change of flight duration costs, changes in fuel consumption and carbon emissions costs. Equality
can only be in one interval. Constraint (21) represents the changes in (29) stipulates that each flight must be included in just one route when
fuel consumption that will result when the aircraft p performs flight f not cancelled. Constraint (30) enforces aircraft assignment constraints.

6
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Fig. 3. The flow chart of the improved column generation algorithm.

7
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Constraint (31) forbids any aircraft from flying more than one route. Def Departure delay times of the flightf;
Constraints (32), (33) and (34) indicate the variables value constraints. Crf Actual duration of flightf;
fupf Changes in fuel consumption of aircraft p executing the flightf.
4.2. Pricing Condition (1). Consider the first condition rt⩽dtj with no delays on
the flightj. The aircraft does not need to accelerate to reduce the flight
Relaxing the integer constraints leads to a linear master problem time, and hence there is no changes in fuel consumption. In this case, the
(LMP). Let π1f be the dual variable of the Constraints (29) for flightf, π 2ag relationships between the state S = (i, rt, e, Nv, De, Cr, fu) and the new
be the dual variable of the Constraints (30) for the airporta, π3p be the state S = (j, rt , e , Nv , De , Cr , fu ) are as follows:
′ ′ ′ ′ ′ ′ ′

dual variable of the Constraints (31) for aircraftp. For the route r of
aircraft p of typeg, the reduced cost c̄pr is. 1. rt = atj + gtp ;
′

2. e = a, ∀a ∈ A, aafa = 1;
′
∑ ∑
c̄pr = cpr − αpfr π1f − βpar π 2ag − π3p (35) 3. Nv = Nv ∪ {j};
′

f ∈F a∈A
′ { ′}
4. Dej = 0, De = De ∪ Dej ;
′ ′

∑( )
{ ′}
cpr = cswap + cdelay + cfuel αpfr (36)
5. Crj = Tcrj , Cr = Cr ∪ Crj ;
′ ′
pf pf pf
f ∈F

6. fu = fu;
′

In Equation (36), the cost cpr is the sum of the swap costs, delay costs,
7. c = c + SCpj ;
′

changes in fuel consumption and carbon emission costs. For a given

8. c̄ = c̄ + SCpj − π 1j .
′

aircraftp, cswap
pf is the swap costs, cdelay
pf is the delay costs and cfuel
pf is the
additional fuel consumption and carbon emission costs of the flightf.
Condition (2). We consider the second condition, i.e.,rt > dtj . Let pre
Inserting Equation (36) into Equation (35), we have.
be the visited flight node before nodei. If i = 0 andati + gtp ⩽dtj +
∑ ( swap ) ∑
c̄pr = αpfr cpf + cdelay
pf + cfuel 1
pf − π f − βpar π2ag − π3p (37) crucj + Tjup , or i ∕
= 0 andatpre + gtp + Tcrli ⩽dtj + crucj + Tjup , the
following linear program for aircraft p and route r needs to be solved to
f ∈F ′
a∈A

Therefore, the objective function of the subproblem is. determine the variables of the new state.
∑ ∑ ( ( ) ) ∑∑ LP (PR).
Min αpfr SCpf + αpfr gf df − Tcrf − tcrf + αpfr (cfuel ∑ ∑
f ∈F f ∈F p∈P f ∈F Min gk (Dek − Tcrk + Crk ) + (cfuel + cco2 rc)fupk (39)
∑ ∑
(38)
k∈Nv∪{j} k∈Nv∪{j}
+ cco2 rc)wfuelpf − αpfr π1f − βpar π2ag − π 3p
f ∈F ′
(40)
a∈A
s.t. atk + Dek − Tcrk + Crk + gtp − dtl − Del ⩽0, ∀k, l ∈ Nv ∪ {j}, k < l

4.3. Solving pricing Dek − Tcrk + Crk ⩽Tkup , ∀k ∈ Nv ∪ {j} (41)

The column generation process is to iteratively solve the pricing dtk + Dek ⩾stp , ∀k ∈ Nv ∪ {j} (42)
problem, finding routes with negative reduced cost to add to the
restricted linear master problem (RLMP). Since there are |P| aircraft, we atk + Dek − Tcrk + Crk ⩽etp , ∀k ∈ Nv ∪ {j} (43)
solve an equal number of subproblems at each iteration. We continue the
iterations until all the found routes have a positive reduced cost. At the Crk ⩾Tcrlk , ∀k ∈ Nv ∪ {j} (44)
same time, we solve the RLMP optimally. The subproblems are formu­
lated as a particular instance of shortest path problems on a connected Crk ⩽Tcrk , ∀k ∈ Nv ∪ {j} (45)
network that a modified label-setting algorithm can solve.
Quite a few studies use labelling algorithms for flight recovery Dek − Tcrk + Crk ⩾0, ∀k ∈ Nv ∪ {j} (46)
problems (Eggenberg et al. 2010, Sinclair et al. 2016, Liang et al. 2018). ∑
Drawing inspiration from them, we developed a modified label-setting fuelypkt = 1, ∀k ∈ Nv ∪ {j}, ismk ⩽0 (47)
algorithm based on dynamic programming to solve multiple sub­ t∈Tink ,t⩽cruck

problems. In our proposed modified label-setting algorithm, we first ∑

consider a directed networkG(V, E). For any aircraftp, nodes i and j can fupk = fvpkt fuelypkt , ∀k ∈ Nv ∪ {j}, ismk ⩽0 (48)
be directly connected if the arrival airport of flight i and the departure
t∈Tink ,t⩽cruck

airport of flight j are identical. Similarly, nodes 0 and i can be directly ∑

Tcrk − Crk ⩽ boundpkt fuelypkt , ∀k ∈ Nv ∪ {j}, ismk ⩽0 (49)
connected if the start airport of the aircraft p is the same as that of the t∈Tink ,t⩽cruck
departure airport of the flight i.
∑
The state S = (i, rt, e, Nv, De, Cr, fu) represents a partial path that ar­ Tcrk − Crk > boundpk(t− 1) fuelypkt , ∀k ∈ Nv ∪ {j}, ismk ⩽0 (50)
rives at the node i with the current landing airport e and is available to t∈Tink ,1⩽t⩽cruck
the next node at a timert. Nv is the set of flights that are covered. De is the
set of departure delay times of the covered flights and Cr is a set of the fuelypkt ∈ {0, 1}, ∀k ∈ Nv ∪ {j}, ∀t ∈ Tink , t⩽cruck (51)
duration of the flights represented by the visited nodes. fu is the changes
The objective function (39) minimizes the costs associated with flight
in fuel consumption. Furthermore, c denotes the cost of the stateS =
delays and additional fuel consumption and carbon emission costs.
(i, rt, e, Nv, De, Cr, fu), and c̄ indicates the reduced cost. When an aircraft
Constraint (40) stipulates the time required between two adjacent flight
visits the next flight j via arc (i, j) from the current stateS =
nodes. Constraint (41) ensures that the arrival delay of a flight does not
(i, rt, e, Nv, De, Cr, fu), we deduce the new stateS =
′

exceed the maximum delay. Constraints (42) and (43) ensure that all the
(j, rt , e , Nv , De , Cr , fu ). Moreover, c is the cost of the new stateS , and
′ ′ ′ ′ ′ ′ ′ ′
flights are completed within the available time of the aircraft. Con­
c̄ is the reduced cost. We propose the extending algorithm with the straints (44) - (46) enforce the lower and upper limits of any flight
′

following two conditions to solve the subproblem. Some additional de­ duration. Constraints (47) - (50) linearize the discrete function between
cision variables are introduced as follows: the reduced flight duration and the changes in fuel consumption.
Decision variables.

8
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Fig. 4. Example for LP (PR).

Fig. 5. Four feasible solutions for the example.

Constraint (51) indicates the binary nature of the variables. Let L = (i, rt, e, Nv, De, Cr, fu, c, c̄) be a label to store the state. The
Next, an example is used to explain why LP (PR) needs to be solved, extending algorithm first generates labels from a dummy source node.
as shown in Fig. 4. The minimum ground time between flightsf1 , f2 and Each time, unexpanded labels with minimal reduced costs are selected to
f3 is 30 min. The departure times of flight 1 and flight 2 have been visit a new flight node to generate a new label, until all the labels are
partially delayed. After the label visit flights 1 and 2 in turn, the time expanded. The relationships between the new label L = (j, rt , e , Nv ,
′ ′ ′ ′

between flights 2 and 3 is less than the minimum ground time. De , Cr , fu , c , c̄ ) and the original label L = (i, rt, e, Nv, De, Cr, fu, c, c̄)
′ ′ ′ ′ ′

Let us suppose the flight duration cannot be reduced. In such a sce­ refer to the states S and S as mentioned before.
′

nario, the only solution is to delay the departure time of flight 3 by 10 After deriving the new labels, the dominance rules are used to
min, thereby delaying its arrival by 10 min, as shown in Solution 1 in discard the labels that cannot bear better solutions. The dominance rules
Fig. 5. This is similar to the traditional shortest path problem. However, are as follows:
if the duration of flights 1, 2, and 3 can be reduced, there may be mul­
Label L dominates L if.
′

tiple feasible solutions, of which only four are listed in Fig. 5. Delay
flight 3 departure time and reduce flight 3 duration, as shown in Solu­
1. j = i;
tion 2. Reduce the flight duration by 10 min so that flight 3 will not be
2. c̄ < c̄;
′

delayed, as shown in Solution 3, and reduce the duration of flight 1 and

flight 2, as shown in Solution 4. This involves how to assign those 10 min 3. rt ⩽rt.
′

to the delay of flight 3 or the reduced flight duration of flights 1, 2, and 3

to minimize the total cost. Therefore, we obtain the optimal path by Rule 1 guarantees that two labels both rooted at the same node. Rule
solving LP (PR). 2 confirms that the negative reduced cost of L is smaller thanL. Rule 3
′

By solving the LP(PR), we can obtain the values ofDek , Crk and fupk means that the available time to the next node in the label L is not
′

for the flight node k ∈ Nv ∪ {j} and the objective value objpr of LP (PR). greater than that in the labelL.
The relationships between the state S = (i, rt, e, Nv, De, Cr, fu) and the After a label has been expanded, the reduced cost will be updated
new state S = (j, rt , e , Nv , De , Cr , fu ) are as follows:
′ ′ ′ ′ ′ ′ ′
with π2ag andπ3p . Then if the reduced cost is negative, the label will be
stored and waiting to be selected into the column pool.
1. rt = atj + gtp + Dej − Tcrj + Crj ; Algorithm 1 presents the modified label-setting algorithm. The
′

2. e = a, ∀a ∈ A, aafa = 1; modified label-setting algorithm extends labels from the dummy source
′

3. Nv = Nv ∪ {j};
′
node and generates all the non-dominated labels with time complex­
( )
4. De = {Dek |k ∈ Nv ∪ {j} };
′
ityO n2 , where n is the number of nodes.
5. Cr = {Crk |k ∈ Nv ∪ {j} }; Algorithm 1 The modified label-setting algorithm
′

∑
6. fu = k∈Nv∪{j} fupk ;
′
Input:p, π1f , π2ag , π3p , i, j ∈ {0, 1, ..., |F| + 1},etc.
∑
7. c = objpr + k∈Nv′ SCpk ; Output:℧l ;
′

∑ 1 Initialization:
8. c̄ = c − k∈Nv′ π k .
′ ′

(continued on next page)

9
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

(continued ) time the subproblems are solved, the modified label-setting algorithm
Algorithm 1 The modified label-setting algorithm reserves three negative reduced-cost columns and adds them to the
℧←∅, ℧ is the set of labels
RLMP. Only negative reduced-cost columns are reserved, and the three
℧l ←∅ most negative are selected to be added to the RLMP.
for i = 1 to |F| do
Initialize the value of attributes in L using the(0, i). 4.5.3. Algorithm parallelization
The subproblems can be solved in parallel as they are independent of
℧←℧ ∪ {L}
end for
each other (Liang et al., 2018). Each time the subproblems are solved,
for (L in ℧) the modified label-setting algorithm will run |P| times, where |P| rep­
Select a label with the minimum negative reduced cost c̄ that has not been visited. resents the number of aircraft. This process is independent of each other
The label belongs to nodei. Let pre be the visited flight node before nodei. and can be run in parallel. Therefore, we parallelize the subproblem
for (j = 1 to |F|) do
if(e = daj , rt⩽dtj )
solving process to accelerate the convergence.
Create a new label L1 using L and(i, j).
Initialize the value of attributes inL1 .
4.6. Branch-and-price algorithm
if (e = daj , rt > dtj )
if (i = 0, ati + gtp ⩽dtj + crucj + Tjup , or i ∕
= 0, atpre + gtp + Tcrli ⩽dtj + crucj + Tup
j )
Create a new label L2 using L and(i, j). The branch-and-price algorithm is a combination of column gener­
Solve LP (PR) and initializeL2 . ation and branch-and-bound algorithm. The column generation algo­
end if rithm obtains the optimal solution of the LMP through constant
end if iterations between the RLMP and the subproblem. If the solution of the
Check if L1 or L2 is dominated. The dominated labels are discarded.
if L1 or L2 is not dominated
LMP is fractional, use the branching strategy to obtain an integer solu­
℧←℧ ∪ { L1 } or℧←℧ ∪ {L2 }. tion. If the solution is an integer, the obtained solution is optimal.
end if If the solution is fractional, we use a best-first search strategy to
∑
if Ls is dominated by L1 orL2 , where Ls ∈ ℧ explore the branch-and-bound tree and branch on r∈R αpfr xpr closest to
∑
0.5. Two child nodes are created by fixing r∈R αpfr xpr = 1 that means
℧←℧\{Ls }
end if ∑
end for the flight f must be performed by aircraftp, and r∈R αpfr xpr = 0 that
update c̄ withπ2ag , π3p . means the flight f is prohibited from being performed by aircraftp. If the
if (c̄ < 0) flight f must be performed by aircraftp, any edge (p , f) where p ∕
′ ′
= p will
be deleted, and columns related to those edges the will be removed. If
℧l ←℧l ∪ {Ll }.
end if
end for the flight f is prohibited from being performed by aircraftp, edge (p, f)
return ℧l will be deleted, and columns related to the edge (p, f) will be removed.

5. Computational experiments
4.4. Initializing columns
5.1. Experimental setup
The approach employed to initialize columns significantly influences
the computational time of the column generation algorithm. We propose
The sponsor of this consultancy project provided us with scenarios/
three ways to generate the initial columns of the RLMP - the heuristic
instances based on the actual operational data of an airline. While we
insertion algorithm, the basic route generation algorithm, and the
used the CPLEX 12.10 version to solve the RLMP, we coded the proposed
shortest path algorithm. The heuristic insertion algorithm is a greedy
column generation approach in C++. We conducted all the experiments
procedure. In this, we sort different flights chronologically based on
on a laptop with 16 GB RAM, a 2.4 GHz Intel i5-9300H CPU and a 64-bit
their departure times and sequentially assign them to aircraft without
Windows 10 operating system. Based on the computer configuration, 8
violating feasibility. During the assignment process, maintenance tasks
threads were used for parallelization. We set a computational time limit
carry higher priority. The basic route generation algorithm assigns
of 1 h (3600 s) for both approaches. Table 1 summarizes the charac­
feasible single flight routes for different aircraft and eventually recon­
teristics of the nine scenarios. The continuous closure time of the airport
ciles the flights to obtain viable solutions. Lastly, while generating col­
is 200 to 600 min. From a practitioner’s point of view, scenarios 1, 2 and
umns using the shortest path algorithm, we use the modified label-
3 are small-sized, scenarios 4, 5 and 6 are medium-sized, and scenarios
setting algorithm to solve the subproblem and set the corresponding
7, 8 and 9 are enormous.
dual variables to 0. The disadvantage of this method is that it may take a
The recovery horizon is within one day. The delay cost is related to
relatively long time.
passengers. The swap cost is a matrix that includes the penalty cost when
the number of seats in the aircraft cannot cover all passengers when the
4.5. Accelerating strategies
swap occurs. Airlines usually provide the cost parameters such as delay

4.5.1. Limiting the number of flights performed by aircraft

Table 1
Let nfp be the number of flights performed by aircraft p in the original
{ } Characteristics of the scenarios.
plan andNF = max nfp , p ∈ P . When we solve the subproblem, we set
Scenario No. of No. No. of No. of No. of No. of
an upper limit on the number of flights executed by the aircraft toNF. flights of maintenances aircrafts airports closed
The number of flights in the column obtained by this method cannot fleets airports

exceedNF. Columns with more than NF flights and negative reduced cost 1 11 1 1 2 5 0
are ignored. If the above method does not result in a column with 2 20 1 1 5 11 1
3 42 2 2 10 21 1
negative reduced cost, we remove the upper limit on the number of
4 63 3 3 14 27 2
flights and solve it again to obtain the exact solution to the subproblem. 5 72 3 3 16 28 2
6 94 3 4 22 31 2
4.5.2. Adding multiple negative reduced-cost columns 7 121 4 6 27 38 2
Adding multiple negative reduced-cost columns is a commonly used 8 147 5 6 33 41 3
9 172 5 7 38 45 3
strategy to improve column generation (Barnhart and Cohn, 2004). Each

10
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Table 2
Computational results.
Scenario CPLEX CG GAP

CP obj. CP time (s) LP obj. IP obj. CG iterations CG time (s)

1 1387.26 2.15 1387.26 1387.26 10 0.89 0.00 %

2 19703.40 4.53 19703.40 19703.40 8 1.28 0.00 %
3 20360.12 74.49 20360.12 20360.12 13 4.15 0.00 %
4 80745.42 694.82 80745.42 80745.42 12 7.45 0.00 %
5 52446.68 1653.66 52446.68 52446.68 13 9.70 0.00 %
6 – – 81251.98 81251.98 28 125.08 –
7 – – 144601.54 144601.54 23 227.06 –
8 – – 143318.22 143318.22 48 1005.53 –
9 – – 150319.99 150319.99 63 2375.70 –
Average Gap – – – – – – 0.00 %

cost per minute per passenger, swap cost for different flights, cancella­
Table 4
tion cost per flight, changes in fuel consumption for different flight
Run time of different accelerating strategies.
durations, etc. These values vary based on the airline and the geogra­
phies they operate. Owing to the non-disclosure agreement (NDA) Scenario Basic Limiting Multiple Parallelization All
flight columns
signed with the airline, we desist from presenting the actual values of
number
these parameters here. However, in line with the inputs received from
1 1.75 1.71 1.25 1.96 0.89
practitioners, we have chosen to assign the following approximate
2 2.45 2.25 1.34 1.56 1.28
values to different parameters: 3 15.03 10.54 4.52 5.29 4.15
4 26.65 23.98 14.16 11.39 7.45
• gf : Passenger delay cost per unit time – $ 5–15 per minute 5 51.48 29.66 19.78 21.79 9.70
• clf : Flight cancellation cost – $ 8,000–10,000 6 577.53 516.92 324.28 237.06 125.08
7 1302.67 1240.10 744.42 796.31 227.06
• SCpf : Swap cost – $ 8–200 for swappable aircraft and ∞ for non-
8 – – 2479.98 2397.41 1005.53
swappable aircraft 9 – – – – 2375.70
• W1: Penalties for violation of aircraft assignment constraints – $
10,000,000
• cfuel: Fuel price – $ 0.6 per Kilogram
5.2. Computational results
• rc: Carbon dioxide emission constant (amount of CO2 produced for
every Kilogram of fuel burn) – 3.15
Table 2 lists the computational results obtained. The column “CP
• cco2 : Unit carbon emission cost – $ 0.02 per Kilogram
obj.” is the objective function value of the MIQP model as reported by
• gtp : Minimum turnaround time between two consecutive flights –
CPLEX, with “CP time (s)” indicating the run time. Columns “LP obj.”
30–40 min
and “IP obj.” represent the solution of RLMP and integer solution of the
• Tfup : Maximum acceptable arrive delay time for any flight and
proposed column generation algorithm, respectively. Columns “CG it­
maintenance – 180 min by default erations” and “CG time (s)” report the number of iterations of the sub­
• Percentage of the Maximum flight duration allowed to be reduced for problem and the run time of the column generation algorithm,
any flight – 10 % − 15 % respectively. Lastly, the column “GAP” reports the optimality gap, where
GAP can be calculated with(IP obj. − CP obj.)/CP obj. × 100\% .
Airport closures directly affect flight availability. Here, the data is The results show that CPLEX converges to optimality in the stipu­
preprocessed to reflect airport closures directly on flights to reduce lated time for only 5 out of 9 scenarios. Scenarios 6–9 cannot be solved
model complexity. For example, an airport is closed from 8:00 am to with MIQP by CPLEX since the model definition phase exceeds the
12:00 am. If the original scheduled departure or arrival time of a flight is computer’s memory. On the contrary, the proposed column generation
within the airport closing time window, the departure time of the flight approach takes a tiny fraction of the CPLEX’s computational time and
will be postponed. If the resulting delay exceeds the default maximum converges for all nine scenarios in less than 1 h. We can observe that the
allowable arrival delay time, then the flight will be cancelled directly. “LP obj” and “IP obj” values are identical for all nine scenarios indicating
Otherwise, the flight will be held and scheduled, and the maximum zero optimality gap, which means that there is no branch-and-bound
allowable arrival delay time Tfup is modified accordingly. process. These results establish the strength and validity of the pro­
posed column generation approach and reflect its suitability for
handling real-life operational scenarios. Table 3 provides flight cancel­
lations costs, flight delays costs, changes in fuel consumption & carbon
emission costs and swap costs of each scenario. Evidently, in most sce­
Table 3
narios, the cancellation costs dominate the rest.
Share of each cost component in the total costs.
Scenario Cancellations Delays Fuel & carbon Swap
costs costs emission costs costs 5.3. Sensitivity analysis
1 0.00 1034.90 297.36 55.00
2 18105.00 1561.80 36.60 0.00 5.3.1. Different accelerating strategies
3 9160.00 8686.80 744.22 1769.10
In Section 4.5, we describe three accelerating strategies to expedite
4 74040.00 5765.40 587.82 352.20
5 45895.00 3878.40 353.78 2319.50 the convergence of the column generation approach. We test them
6 65060.00 11675.40 1190.88 3325.70 thoroughly to assess their relevance with a computational time limit of
7 131125.00 8869.40 1475.24 3131.90 3600 s. In Table 4, the column “Basic” implies the absence of any ac­
8 131125.00 7453.80 1220.12 3519.30 celeration strategy. Similarly, “Limiting flight number” means deploying
9 131315.00 10887.90 2021.49 6095.60
the acceleration strategy described in Section 4.5.1. “Multiple columns”

11
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Fig. 6. Run time of different accelerating strategies.

Table 5
Results of different numbers of dividing points.
Scenario 5 21 Time gap Obj. gap

Obj. Time (s) Obj. Time (s)

1 1822.61 0.67 1387.26 0.89 32.84 % –23.89 %

2 19706.99 0.81 19703.40 1.28 58.02 % − 0.02 %
3 20423.19 3.75 20360.12 4.15 10.67 % − 0.31 %
4 80756.33 4.89 80745.42 7.45 52.35 % − 0.01 %
5 52471.38 12.34 52446.68 9.70 − 21.39 % − 0.05 %
6 81315.42 88.78 81251.98 125.08 40.89 % − 0.08 %
7 144637.66 698.26 144601.54 227.06 − 67.48 % − 0.02 %
8 143346.22 2144.20 143318.22 1005.53 − 53.10 % − 0.02 %
9 150380.28 2208.81 150319.99 2375.70 7.56 % − 0.04 %
Average Gap – – – – 6.71 % − 2.72 %

Table 6
Results without and with aircraft assignment constraints.
Scenario without aircraft assignment with aircraft assignment with aircraft assignment (80 %)

Obj. Time (s) Obj. Time (s) Obj. gap Obj. Time (s) Obj. gap

1 1387.26 0.89 1387.26 0.89 0.00 % 1387.26 1.47 0.00 %

2 13208.14 1.02 19703.40 1.28 49.18 % 19703.40 1.11 49.18 %
3 20360.12 4.43 20360.12 4.15 0.00 % 20360.12 3.23 0.00 %
4 53387.08 6.27 80745.42 7.45 51.25 % 53387.08 7.05 0.00 %
5 34085.34 6.36 52446.68 9.70 53.87 % 52446.68 9.98 53.87 %
6 66536.00 137.54 81251.98 125.08 22.12 % 72077.68 125.16 8.33 %
7 91204.99 356.88 144601.54 227.06 58.55 % 117827.61 322.08 29.19 %
8 88054.39 819.12 143318.22 1005.53 62.76 % 107539.56 997.72 22.13 %
9 106022.15 2651.81 150319.99 2375.70 41.78 % 107704.39 3075.68 1.59 %
Average Gap – – – – 37.72 % – – 18.25 %

refer to deploying the acceleration strategy as described in Section 4.5.2, consumption function, we solve all the nine scenarios with 5 and 21
and “Parallelization” indicates the acceleration strategy as explained in dividing points separately. A different number of dividing points implies
Section 4.5.3. Lastly, as the name suggests, “All” indicates the use of all varying temporal discretization. We intend to study if this parameter
three acceleration strategies. The run time of the different accelerating affects the convergence time and solution quality. We tabulate the re­
strategies is shown in Fig. 6. sults of improved column generation in Table 5.
The results show that all the acceleration strategies contribute to the The results show that higher granularity results in less costly solu­
quicker convergence of the proposed approach. Among the three, the tions. In all scenarios, when there are more dividing points, the cost of
effect of multiple columns strategy is similar to that of parallelization the solution decreases to varying degrees. In terms of solution time, more
subproblems, while the strategy of limiting the flight number has the dividing points are more suitable for solving large-scale scenario.
least impact. Therefore, using fuel consumption functions with higher granularity
helps to obtain a better solution.
5.3.2. Effect of the discretization of the fuel consumption function
To study the effect of the number of discrete points in the fuel

12
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Fig. 7. Obj. gap and time gap with and without aircraft assignment constraints.

enforcing aircraft assignment constraints, the average total costs in­

Table 7
crease by almost 37.72 %, compared to 18.25 % for an 80 % satisfaction
Results of Scenario 6 in various types of costs.
rate. This situation presents a trade-off for practitioners. The obj. gap
Count index without aircraft with aircraft with aircraft and time gap with and without aircraft assignment constraints are
assignment assignment assignment (80 %)
shown in Fig. 7.
No. of 5 7 6 When we do not consider the aircraft assignment constraints, we
cancellations
carry the risk of having a lesser number of aircraft required for future
Cancellation costs 45955.00 65060.00 55765.00
No. of swaps 22 19 20
flights. For instance, at the end of the recovery plan, if there are three
Swap costs 3846.80 3325.70 3497.50 aircraft of the same type at Airport A but the next day’s flight plan re­
No. of delays 19 16 16 quires five aircraft of this type, then there will be no aircraft to perform
Delay costs 15101.30 11675.40 11624.30 some flights. Airlines usually respond to this situation by modifying the
Fuel & carbon 1632.90 1190.88 1190.88
schedule, cancelling some flights, or sometimes ferrying idle aircraft
emission costs

5.3.3. Effect of aircraft assignment constraints Table 9

We solve two versions of the master problem – one with aircraft Recovery costs with and without flight duration change.
assignment constraints and another without them. We intend to inves­ Scenario without flight duration with flight duration GAP
tigate the effect on the computational time and the total costs using the change change

improved column generation. We also conducted experiments where the Obj. Time (s) Obj. Time (s)
aircraft assignment constraints were not fully satisfied, requiring about 1 2186.20 0.61 1387.26 0.89 − 36.54 %
80 % of the number of aircraft of each type required at each airport. The 2 19758.00 0.39 19703.40 1.28 − 0.28 %
results are as shown in Table 6. 3 21501.10 1.87 20360.12 4.15 − 5.31 %
For all the scenarios, by enforcing aircraft assignment constraints, 4 195461.70 3.27 80745.42 7.45 − 58.69 %
5 53031.50 5.25 52446.68 9.70 − 1.10 %
the results show will lead to increased costs. The reason why the cost of 6 116201.00 52.03 81251.98 125.08 − 30.08 %
Scenario 1 and 3 did not increase, from the experimental results, is that 7 225372.70 90.29 144601.54 227.06 − 35.84 %
there are no flight cancellations in the recovery plan or all cancellations 8 223731.10 241.45 143318.22 1005.53 − 35.94 %
are round-trip flights, so it does not affect the number of aircraft at each 9 175719.60 882.62 150319.99 2375.70 − 14.45 %
Average Gap − 24.25 %
airport after the recovery period. The test results in Table 6 show that by – – – –

Table 8
Experimental results for smaller cancellation cost.
Scenario without aircraft assignment with aircraft assignment with aircraft assignment (80 %)

Obj. Time (s) Obj. Time (s) Obj. gap Obj. Time (s) Obj. gap

1 1387.26 1.42 1387.26 1.20 0.00 % 1387.26 1.17 0.00 %

2 6845.14 1.10 7029.90 1.06 2.70 % 7029.90 1.08 2.70 %
3 12723.63 2.53 12723.63 2.75 0.00 % 12723.63 3.02 0.00 %
4 21033.08 6.53 28917.42 8.23 37.49 % 21033.08 7.81 0.00 %
5 14653.34 5.61 20320.18 8.87 38.67 % 20320.18 9.03 38.67 %
6 31959.58 103.45 35709.98 87.11 11.73 % 33042.18 104.86 3.39 %
7 39140.91 304.40 52621.44 226.72 34.44 % 45682.11 230.22 16.71 %
8 36002.39 780.53 51324.62 790.17 42.56 % 41900.56 854.90 16.38 %
9 46974.39 1742.14 58165.59 1986.81 23.82 % 48710.40 1898.15 3.70 %
Average Gap – – – – 21.27 % – – 9.06 %

13
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Fig. 8. Gap in recovery costs with and without flight duration change.

from other airports to run original follow-on flights. In essence, the price to obtain the normal operation of the flight after the recovery
airlines’ response defers part of the recovery costs to the future and period.
affects the normal operation of future flights. It makes sense to pay more
for the current recovery costs when airlines believe that the normal 5.3.4. Effect of varying the flight duration
operation of flights after the recovery period is very important. A To quantify the savings in the recovery costs by incorporating the
compromise might be to compare expected future losses with current recovery option of varying the flight duration, we conduct experiments
additional recovery costs and make logical decisions on a case-by-case with and without it using the improved column generation. The results
basis, such as setting partially satisfied aircraft assignment constraints. are tabulated in Table 9, where GAP is the change in recovery costs.
As the experimental results show, the current recovery costs can be As expected, incorporating the recovery option of varying the flight
reduced as much as possible on the basis of guaranteeing the operation duration helps in saving a significant amount of recovery costs, about 24
of some important flights after the recovery period by setting partially % on average. The gap in recovery costs with and without flight duration
satisfied aircraft assignment constraints. change is shown in Fig. 8. Changing the flight duration helps execute
For a more detailed presentation, taking Scenario 6 as an example, some flights that would have had to be cancelled otherwise. The only
the changes in various types of costs are given in Table 7. It can be seen drawback with considering this recovery option is the additional
from Table 7 that the cost increase with considering aircraft assignment computational time required to solve the problem.
is mainly caused by the flight cancellation costs, but considering aircraft
assignment guarantees the normal execution of subsequent flights after 5.3.5. Analysis of different fuel prices
the end of the recovery period. Aircraft fuel prices are volatile, so the fuel prices were set to different
Next, we set a smaller flight cancellation cost of 0.3 times the original values for experiments. The fuel price is set to 0.4 and 1.0, respectively,
value. Experiments were carried out using improved column generation compared to the solution with a fuel price of 0.6. The results of the
without considering the aircraft assignment, considering the aircraft experiments using the improved column generation are shown in
assignment, and considering the incompletely satisfied aircraft assign­ Table 10.
ment (about 80 %). The experimental results are shown in Table 8. The Fig. 9 shows the difference between the solutions with fuel prices of
recovery costs considering the aircraft assignment constraints increased 0.4 and 1.0 and the solution with fuel price of 0.6, which is used as a
by about 21.27 % on average, while the recovery costs considering the benchmark. Combining Table 10 and Fig. 9, it can be seen that when the
incompletely satisfied aircraft assignment increased by about 9.06 % on fuel price increases, the total recovery costs will increase, and vice versa.
average. Therefore, when the flight cancellation cost is relatively small, Therefore, the recovery option of changing the flight duration can save
considering the aircraft assignment constraints has a smaller impact on more costs when the fuel price is low.
the total recovery costs, which means that the airlines can pay a small
5.3.6. Analysis of different aircraft assignment requirements
In this subsection, we report another aircraft assignment constraints,
Table 10 i.e., for designated airports, the number of aircraft after the recovery
Results of the experiments with different fuel prices. plan should be the same as that after the end of the original plan, rep­
Scenario fuel price = 0.4 fuel price = 1.0 resented by Constraint 2. In contrast, the aircraft assignment constraints
Obj. Time (s) Obj. Time (s) considered above are represented by Constraint 1. The modification of
the mathematical model for Constraint 2 is shown in Appendix A. We
1 1291.99 1.07 1544.41 1.38
2 19687.99 0.87 19720.85 0.89 conducted experiments and showed the results of improved column
3 20077.36 3.56 20677.46 3.70 generation in Table 11. Using Constraint 2 implies more stringent
4 80554.11 6.67 81071.52 6.98 aircraft assignment, and Scenario 5 cannot obtain a solution that sat­
5 52317.71 10.92 52599.34 8.92 isfies Constraint 2 and therefore suffers a penalty cost. Overall, the
6 80784.58 115.93 81844.99 134.49
7 144118.09 287.76 145201.29 327.84
implementation of more demanding aircraft assignment may lead to
8 142929.31 1299.09 143806.05 977.50 higher costs, as some flights may have to be cancelled to meet the
9 149644.38 2306.09 151412.77 1572.34 demanding assignment requirements. This requires practitioners to

14
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Fig. 9. Cost difference between different fuel prices.

aircraft assignment constraints ensure the normal operation of sub­

Table 11
sequent flights and prevents spillover of the recovery costs/actions
Results of different aircraft assignment requirements.
into the future.
Scenario Constraint 1 Constraint 2 GAP

Obj. Time (s) Obj. Time (s) We can extend this work in multiple ways. For example, the early
1 1387.26 0.89 1387.26 1.01 0.00 % departure of flights when all the passengers have boarded is one of the
2 19703.40 1.28 19703.40 1.28 0.00 % possible recovery options. Developing heuristics/meta-heuristics for
3 20360.12 4.15 29477.92 4.65 44.78 % large-scale scenarios that can generate good quality solutions quickly
4 80745.42 7.45 80745.42 7.71 0.00 % can be taken up further. On the basis of the research in this paper, the
5 52446.68 9.70 1061408.31 8.59 1923.79 %
6 81251.98 125.08 107180.14 140.62 31.91 %
simultaneous aircraft and crew recovery problem can be further studied.
7 144601.54 227.06 180642.28 313.19 24.92 % Otherwise, it would be interesting to study the problem in a bi-objective
8 143318.22 1005.53 179446.66 846.67 25.21 % context with total recovery costs as the first objective and penalty cost
9 150319.99 2375.70 204347.92 1875.14 35.94 % for violating aircraft assignment as the second objective.
Average Gap – – – – 231.84 %

CRediT authorship contribution statement

weigh the current cost, the expected future cost, and other factors such
as policies in actual operation. Jie Li: Methodology, Investigation, Data curation, Writing – original
draft. Kunpeng Li: Software, Data curation, Writing – original draft.
6. Conclusions and scope for future work Qiannan Tian: Methodology, Data curation, Writing – review & editing.
P.N. Ram Kumar: Validation, Writing – review & editing.
This paper distinguishes itself from the rest of the studies in the
context of disrupted flight recovery problems by considering two critical Declaration of Competing Interest
real-life characteristics: a) A new implementation of the recovery option
by changing flight duration. We use a piecewise function to describe the The authors declare that they have no known competing financial
discrete functional relationship between the reduced flight duration and interests or personal relationships that could have appeared to influence
changes in fuel consumption and b) aircraft assignment constraints - the work reported in this paper.
enough aircraft should be assigned to airports at the end of the recovery
period. Based on a series of experiments with real data, we have come to Data availability
the following conclusions:
Data will be made available on request.
• The proposed MIQP model is suitable for solving small-scale sce­
narios with commercial solvers such as CPLEX but not for solving Acknowledgments
large-scale scenarios. In contrast, the proposed improved column
generation algorithm can solve large-scale scenarios in a reasonable This work is supported by the National Natural Science Foundation
time. The experimental results show that the three strategies we of China (Grant Number 72001072); Ministry of Education of the Peo­
adopted to accelerate the column generation process are very ple’s Republic of China Humanities and Social Science Youth Founda­
effective in facilitating the faster convergence of the algorithm. tion (Grant Number 20YJC630135); Hubei Provincial Department of
• The recovery option of changing flight duration can save an average Education (Grant Number 20Q120); Hubei Provincial Excellent Young
of 24 % of recovery costs. The proposed method can generate high- and Middle-aged Scientific and Technological Innovation Team of Col­
quality solutions in real time and make it suitable for real-life leges and Universities (Grant Number T2022024); National Social Sci­
deployment. ence Foundation of China (Grant Number 21CGL019).
• Our results suggest that enforcing aircraft assignment constraints
may result in higher current recovery costs. However, implementing

15
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Appendix A

Here the MIQP model and the master problem model are modified to accommodate Constraint 2 in subsection 5.3.6. The first is the modification of
MIQP for Constraint 2. Let QALL indicates the total difference in the number of aircraft that does not meet the aircraft assignment constraints. The
objective function (1) becomes.
( ) ( )
∑ ∑∑ ∑ ( ( ) )∑ ∑ ∑ ∑
Min clf 1 − zpfi + gf df − Tcrf − tcrf zpfi + (cfuel + cco2 rc)wfuelpf
(52)
f ∈F p∈P i∈Index f ∈F p∈P i∈Index p∈P f ∈F,ismf ⩽0
( )
∑∑ ∑
+ SCpf zpfi + W1⋅QALL
p∈P f ∈F i∈Index

Constraint (4) becomes.

( ) ( )
∑ ∑ ∑ ∑ ∑ ∑ ( )
(53)
′
aafa zpfi − dafa zpfi − hag − ehag = QAag , ∀a ∈ A , g ∈ G
p∈P,ftp =g f ∈F i∈Index p∈P,ftp =g f ∈F i∈Index

Constraint (5) becomes.

∑∑⃒ ⃒
QALL = ⃒QAag ⃒ (54)
′
a∈A g∈G

Constraint (54) can be linearized as follows by introducing the variableQAag .

′

∑∑ ′
QALL = QAag (55)
′
a∈A g∈G

(56)
′ ′
QAag ⩽QAag , ∀a ∈ A , g ∈ G

(57)
′ ′
− QAag ⩽QAag , ∀a ∈ A , g ∈ G

For the master problem formulation, let qag be the surplus that the recovery plan does not meet the aircraft assignment constraints. The objective
′

function (28) becomes.

∑ ∑ ∑∑( )
(58)
′
Min cpr xpr + clf yf + W1⋅ qag + qag
r∈R ′
f ∈F a∈A g∈G

Constraint (30) becomes.

∑ ∑
(59)
′ ′
βpar xpr + qag − qag = hag , ∀a ∈ A , g ∈ G
p∈P,ftp =g r∈R

And add the following constraint.

(60)
′ ′
qag ⩾0, ∀a ∈ A , g ∈ G

References Eggenberg, N., Salani, M., & Bierlaire, M. (2010). Constraint-specific recovery network
for solving airline recovery problems. Computers & operations research, 37(6),
1014–1026.
Aktürk, M. S., Atamtürk, A., & Gürel, S. (2014). Aircraft rescheduling with cruise speed
Gürkan, H., Gürel, S., & Aktürk, M. S. (2016). An integrated approach for airline
control. Operations Research, 62(4), 829–845.
scheduling, aircraft fleeting and routing with cruise speed control. Transportation
Alexandre, J. (2020). IATA Annual Review 2020. URL: https://www.iata.org/
Research Part C: Emerging Technologies, 68, 38–57.
contentassets/c81222d96c9a4e0bb4ff6ced0126f0bb/iata-annual-review-2020.pdf.
Hassan, L. K., Santos, B. F., & Vink, J. (2021). Airline disruption management: A
Amadeus IT Group S.A. (2020). Amadeus Global Report 2020. URL: https://corporate.
literature review and practical challenges. Computers & Operations Research, 127,
amadeus.com/documents/en/resources/corporate-information/corporate-
Article 105137.
documents/global-reports/2020/amadeus-global-report-2020.pdf.
Hu, Y., Song, Y., Zhao, K., & Xu, B. (2016). Integrated recovery of aircraft and passengers
Anisha, S. (2021). Deloitte: Business Travel Set for a Slow Takeoff. URL: https://www2.
after airline operation disruption based on a GRASP algorithm. Transportation
deloitte.com/us/en/pages/about-deloitte/articles/press-releases/deloitte-business-
research part E: logistics and transportation review, 87, 97–112.
travel-set-for-a-slow-takeoff.html.
Hu, Y., Xu, B., Bard, J. F., & Chi, H. (2015). Optimization of multi-fleet aircraft routing
Arıkan, U., Gürel, S., & Aktürk, M. S. (2016). Integrated aircraft and passenger recovery
considering passenger transiting under airline disruption. Computers & Industrial
with cruise time controllability. Annals of Operations Research, 236(2), 295–317.
Engineering, 80, 132–144.
Arıkan, U., Gürel, S., & Aktürk, M. S. (2017). Flight network-based approach for
Huang, Z., Luo, X., Jin, X., & Karichery, S. (2022). An iterative cost-driven copy
integrated airline recovery with cruise speed control. Transportation Science, 51(4),
generation approach for aircraft recovery problem. European Journal of Operational
1259–1287.
Research, 301(1), 334–348.
Barnhart, C., & Cohn, A. (2004). Airline schedule planning: Accomplishments and
Jaap, B., Steven, S., Nina, W. (2021). Back to the future? Airline sector poised for change
opportunities. Manufacturing & service operations management, 6(1), 3–22.
post-COVID-19. URL: https://www.mckinsey.com/industries/travel-logistics-and-
Brunner, J. O. (2014). Rescheduling of flights during ground delay programs with
infrastructure/our-insights/back-to-the-future-airline-sector-poised-for-change-post-
consideration of passenger and crew connections. Transportation Research Part E:
covid-19.
Logistics and Transportation Review, 72, 236–252.
Jozefowiez, N., Mancel, C., & Mora-Camino, F. (2013). A heuristic approach based on
Clausen, J., Larsen, A., Larsen, J., & Rezanova, N. J. (2010). Disruption management in
shortest path problems for integrated flight, aircraft, and passenger rescheduling
the airline industry—Concepts, models and methods. Computers & Operations
under disruptions. Journal of the Operational Research Society, 64(3), 384–395.
Research, 37(5), 809–821.

16
J. Li et al. Computers & Industrial Engineering 174 (2022) 108772

Khan, W. A., Chung, S. H., Ma, H. L., Liu, S. Q., & Chan, C. Y. (2019). A novel self- Trani, A. A., & Wing-Ho, F. (1997). Enhancements to SIMMOD: A neural network post-
organizing constructive neural network for estimating aircraft trip fuel consumption. processor to estimate aircraft fuel consumption. Phase I final report. URL: Https://isr.
Transportation Research Part E: Logistics and Transportation Review, 132, 72–96. umd.edu/NEXTOR/pubs/RR-97-8.pdf.
Khan, W. A., Ma, H. L., Ouyang, X., & Mo, D. Y. (2021). Prediction of aircraft trajectory Turgut, E. T., Cavcar, M., Usanmaz, O., Canarslanlar, A. O., Dogeroglu, T., Armutlu, K., &
and the associated fuel consumption using covariance bidirectional extreme learning Yay, O. D. (2014). Fuel flow analysis for the cruise phase of commercial aircraft on
machines. Transportation Research Part E: Logistics and Transportation Review, 145, domestic routes. Aerospace Science and Technology, 37, 1–9.
Article 102189. Vink, J., Santos, B. F., Verhagen, W. J., & Medeiros, I. (2020). Dynamic aircraft recovery
Lee, J., Marla, L., & Jacquillat, A. (2020). Dynamic disruption management in airline problem-an operational decision support framework. Computers & Operations
networks under airport operating uncertainty. Transportation Science, 54(4), Research, 117, Article 104892.
973–997. Wasiuk, D. K., Khan, M. A. H., Shallcross, D. E., & Lowenberg, M. H. (2016).
Liang, Z., Xiao, F., Qian, X., Zhou, L., Jin, X., Lu, X., & Karichery, S. (2018). A column A commercial aircraft fuel burn and emissions inventory for 2005–2011. Atmosphere,
generation-based heuristic for aircraft recovery problem with airport capacity 7(6), 78.
constraints and maintenance flexibility. Transportation Research Part B: Wasiuk, D. K., Lowenberg, M. H., & Shallcross, D. E. (2015). An aircraft performance
Methodological, 113, 70–90. model implementation for the estimation of global and regional commercial aviation
Marla, L., Vaaben, B., & Barnhart, C. (2017). Integrated disruption management and fuel burn and emissions. Transportation Research Part D: Transport and Environment,
flight planning to trade off delays and fuel burn. Transportation Science, 51(1), 35, 142–159.
88–111. Yanto, J., & Liem, R. P. (2018). Aircraft fuel burn performance study: A data-enhanced
Pagoni, I., & Psaraki-Kalouptsidi, V. (2017). Calculation of aircraft fuel consumption and modeling approach. Transportation Research Part D: Transport and Environment, 65,
CO2 emissions based on path profile estimation by clustering and registration. 574–595.
Transportation Research Part D: Transport and Environment, 54, 172–190. Zhang, D., Lau, H. H., & Yu, C. (2015). A two stage heuristic algorithm for the integrated
Sinclair, K., Cordeau, J. F., & Laporte, G. (2014). Improvements to a large neighborhood aircraft and crew schedule recovery problems. Computers & Industrial Engineering, 87,
search heuristic for an integrated aircraft and passenger recovery problem. European 436–453.
Journal of Operational Research, 233(1), 234–245. Zhang, D., Yu, C., Desai, J., & Lau, H. H. (2016). A math-heuristic algorithm for the
Sinclair, K., Cordeau, J. F., & Laporte, G. (2016). A column generation post-optimization integrated air service recovery. Transportation Research Part B: Methodological, 84,
heuristic for the integrated aircraft and passenger recovery problem. Computers & 211–236.
Operations Research, 65, 42–52. Zhang, X., & Mahadevan, S. (2020). Bayesian neural networks for flight trajectory
prediction and safety assessment. Decision Support Systems, 131, Article 113246.

17