2018 TSPC Zou

Transportation Research Part C 95 (2018) 165–184
Contents lists available at ScienceDirect
Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc
A two-stage taxi scheduling strategy at airports with multiple

T
independent runways
⁎
X. Zoua,b, P. Chenga, , W.D. Liub, N. Chenga, J.P. Zhangb
a
Institute of Navigation and Control, Department of Automation, Tsinghua University, Beijing 100084, PR China
b
The Second Research Institute of Civil Aviation Administration of China, Chengdu 610041, PR China
A R T IC LE I N F O ABS TRA CT
Keywords: Long taxiing times at large airports lead to fuel wastage and dissatisfied passengers. This paper
Airport investigates the 4D taxi scheduling problem in airports to minimize the taxiing time. We propose
Taxi scheduling an iterative two-stage scheduling strategy. In the first stage, all aircrafts in a current schedule
Integer programming period are assigned initial 4D routes. In the second stage, landing aircrafts that are unavailable to
Runway exit availability
fulfil their initially assigned routes are rescheduled using a shortest path algorithm based ap-
proach. In this paper, the simplified model used in most existing literature, that depicts a runway
as having a single entrance and a single exit or even sets only one point to represent both of them
has been discarded. Instead, we model the fact that a runway has multiple entrance and exit
points and use an emerging concept—Runway Exit Availability (REA)—to measure the prob-
ability of clearing a runway from a specific exit during a specific time interval so that the taxiing
scheduling model can be much higher approximation to the practical operation. An integer
programming (IP) model factoring REA is proposed for assigning 4D taxiing routes in the first
stage. The IP model covers most practical constraints faced in airport taxiing procedures, such as
the rear-end/head-on conflict constraint, runway-crossing constraint, take-off/landing separation
constraint, and taxi-out constraint. Besides, flight holding patterns at intersections are much
more realistically modelled. Furthermore, to accelerate the solving process of the IP model, we
have refined the formulation using several tricks. Simulation results by proposed scheduling
approach for operations at the Beijing Capital International Airport (PEK) for an entire day de-
monstrate a surprising taxiing time saving against the empirical data and simulation results based
on a strategy similar to what being used now days while showing an acceptable running time of
our approach, which supports that our approach may help in real operation in the future.
1. Introduction and literature review
A rapid development of air transportation has resulted in increased flight delays worldwide. Among the factors influencing the air
traffic system, airport capacity is considered as one of the bottlenecks (Neufville and Odoni, 2003). To deal with this issue, the
favoured approach in all countries is to expand the airport, i.e. to build more runways and taxiways. When an airport has a small
number of runways and a simple taxiway structure, runway capacity determines the airport capacity. However, with an increasing
complexity of the taxiway system, congestion on the taxiway too increases, so that the efficiency of taxi planning becomes another
restricting factor determining airport capacity (Cheng, 2004). Using regression analysis, Kisler and Gupta found that surface traffic
⁎
Corresponding author.
E-mail addresses: zouxiang@caacsri.com (X. Zou), chengp@tsinghua.edu.cn (P. Cheng), liuweidong@caacsri.com (W.D. Liu),
ncheng@tsinghua.edu.cn (N. Cheng), zhangjianping@caacsri.com (J.P. Zhang).
https://doi.org/10.1016/j.trc.2018.07.005
Received 25 December 2017; Received in revised form 20 June 2018; Accepted 5 July 2018
0968-090X/ © 2018 Elsevier Ltd. All rights reserved.
X. Zou et al. Transportation Research Part C 95 (2018) 165–184
has a great impact on an airport’s operational efficiency (Kistler and Gupta, 2009). The airport taxi scheduling (TS) problem has
become a focus by many researchers over the past decade and several methods to solve the TS problems have been proposed.
In this work, we sort previous literatures into three groups according to the solving method applied in the literatures. Works in
Burgain et al. (2012), Atkin et al. (2013), Yu and Qing (2017) are mainly concerned about deciding the pushback times, the take-off
sequence of departing flights, or the estimation of unimpeded taxiing time. They do not provide taxiing routes. Another kind of
methods (Baik et al., 2000; Lesire, 2010; Atkin et al., 2014; Weiszer et al., 2015; Zhang et al., 2016; Park et al., 2017) sequentially
calculate every flight’s shortest feasible taxiing route according to their chronological order along with taxiway occupation by
preceding flights, which represents constraints for subsequent flights. This kind of approaches are efficient in computational cost and
run very fast. In Atkin et al. (2014), Ravizza et al. report that it only took tens of seconds to schedule all flights for an entire day at the
Zürich Airport. However, difficulty in obtaining or the inability to obtain an optimal pushback sequence and pushback times is a
common problem encountered in this type of approaches. Most of them simply adopt a first come first served principle. The third kind
of methods (Marín, 2006, 2013; Balakrishnan and Jung, 2007; Marín and Codina, 2008; Rathinam et al., 2008; Roling and Visser,
2008; Lee and Balakrishnan, 2010; Clare and Richards, 2011; Anderson and Milutinović, 2013; Mori, 2013; Guepet et al., 2016;
Evertse and Visser, 2017) consider several operating flights in a single scheduling instance. They usually build a time–space network
representing the taxiway system and model the taxi scheduling problem as a Mix Integer Programming (MIP) or Integer Programming
(IP) problem. The method of Roling and Visser (2008) is an earlier version of the IP based method. Balakrishnan et al. published a
series of literatures (Balakrishnan and Jung, 2007; Rathinam et al., 2008; Lee and Balakrishnan, 2010) to optimize real operations in
different airports. This series assumed one or two alternative routes for each flight. Marín published a triplet of papers (Marín, 2006,
2013; Marín and Codina, 2008) considering several algorithms, such as “Fix and Relax” and “Lagrangian Decomposition” to speed up
the solving procedure. The constraints in Clare and Richards (2011) comprising runway crossing conflicts, rear-end conflicts, head-on
conflicts, varying taxiing speed, etc. are very comprehensive. However, the highly complex model resulted in a solving time too long
to be acceptable. In contrast to the approaches in Marín (2006), Balakrishnan and Jung (2007), Marín and Codina (2008), Rathinam
et al. (2008), Roling and Visser (2008), Lee and Balakrishnan (2010), Clare and Richards (2011), Marín (2013) that set discrete time
variables, the approaches in Anderson and Milutinović (2013), Mori (2013) used continuous variables to represent taxiing procedures
and proposed two MIP models. Authors in Guepet et al. (2016), Evertse and Visser (2017) set the objective of taxi scheduling to be
minimizing the emission by aircrafts’ surface movement which shows the concerning about the environmental protection. However,
there is no essential difference between the objective of minimizing taxi time and the one of minimizing the emission especially in the
view of the methods applied to solve the TS problem.
In fact, Air Traffic Flow Management (ATFM) (Bertsimas and Patterson, 1998; Bertsimas et al., 2011), Runway Configuration (RC)
(Kim et al., 2014), Departure & Arrival Scheduling (DAS) (Eun et al., 2010; Mori, 2013), Gate Assignment (GA) (Ding et al., 2005),
and Taxi Scheduling (TS) are highly dependent on each other. ATFM assigns time slots of runway operation (take-off or landing) to
flights at several airports. This time slot is macroscopic (e.g., it is set at 15 min in Bertsimas et al. (2011)), but it defines the flow
pattern of the air traffic system. With the output of ATFM, RC, DAS, GA, and TS, operations of flights at airports are made more
efficient, e.g. by deciding landing sequence and time, the runway could be operated for landing or take-off, parking position, taxiing
route, etc. However, the internal relationship among ATFM, RC, DAS, GA, and TS is beyond the scope of this paper and it was not
considered in existing literatures yet. Henceforth, in this work, we assume that the TS problem is being solved on the premise of
determined operating runways and parking positions for all flights and available earliest pushback time for departures and known
landing time of arrivals.
In this paper, we implement a discrete time IP-based method with the aim of delivering an acceptable solution.
Previous literatures have usually oversimplified runway operations, which represents a runway with only one or two nodes. In
fact, a runway usually has multiple entrances and exits. One major contribution of this paper is to use an emerging concept called
“Runway Exit Availability” (REA) along with a two-stage scheduling strategy to deal with the problem of uncertain exit and exit
times, which in fact was introduced for the first time in one of our papers in 2014 (Cheng et al., 2014). In this paper, we strengthen
the mathematical definition comprehensively and propose an operable calculation method for REA. The introduction of REA can help
the optimization model to be much closer to the real operation.
Another main contribution of this paper is to make the pattern of holding along taxiing routes more practical. Where block or
waiting along a taxiing route is necessary, previous literatures held flights on nodes representing intersections of taxiing routes and
introduced node capacity constraints, e.g. the formulation of general network flow problems. However, in reality, taxiing flights do
not stop at intersections or on nodes. Instead, they hold around the intersections just like cars waiting for green light. Hence,
theoretically, the capacity of each node must be greater than one because two or three flights can wait at the same intersection
simultaneously. However, if we set the capacity as simply greater than one, a collision caused by two or more flights on the same side
of an intersection cannot be avoided. This is because in such a scenario, the only point that would be considered is whether the gross
number of waiting flights is below the stated capacity. In this paper, we introduce new decision-making variables, called the block-
and-hold variables, to represent holding behaviours along taxiing routes. With these variables, no node capacity constraints are
needed and no extra nodes are required, so that the size of the taxiing network does not need to be expanded. We also integrate the
“taxi-out constraint” into the IP model. This kind of constraint considers the pushback and the engine starting procedures of departing
flights. Note that, the “taxi-out constraint” in this paper is different from the concept of pushback delay that is common in existing
research works. As a matter of fact, our proposal of including such a constraint has been supported by some published works (i.e.
Reference Weiszer et al. (2014) and Coupe and Milutinovi (2015)).
Finally, the heuristics for head-on conflict check and avoidance and several tricks to refine the formulation of the IP model result
in no Branch-and-Bound operations being needed in simulation instances and that a very few numbers of cuts are applied. This
166
characteristic makes the running time of our approach acceptable in practice, especially, in the light of the dense surface flight flow in
our simulation, the complexity, and large size of the IP model, which includes the majority of the constraints and numerous details
encountered during real operations.
This paper is organized as follows. In Section 2, we outline the two-stage strategy. In Section 3, the definition and the calculation
method of REA are presented. In Section 4, a time–space network of the taxiway system is defined and the IP model for the scheduling
in stage 1 is described in detail. In Section 5, we explain the tricks we implement in order to shorten the solving time. In Section 6, we
discuss the shortest path algorithm-based heuristic applied in the rescheduling in stage 2 and in the head-on conflict check and
avoidance. In Section 7, we demonstrate the experiments, which simulate the surface operations of the Beijing Capital International
Airport for an entire day, to test our strategy. We also present some comparisons between our approach and other existing approaches
in this section to highlight our improvement. Finally, a conclusion of the work and the path to real implementation of our proposal
are discussed.
2. Two-stage taxi scheduling strategy
Considering the demand in real operation, a well-designed and efficient airport scheduling strategy must assign taxiing routes to
flights (for both departures and arrivals) before they start taxiing. This means that a departing flight must receive its taxiing route
before its pushback and an arriving flight must receive the route before or soon after clearing the runway. In previous literatures, as
the exits of a runway were simplified as a single node, no specific discussion about the runway exit point or taxiing starting point of
arriving flights was needed. However, in this paper, all runways are modelled with multiple exits as they are in practice. So, we are
confronted with problems caused by the uncertainty of runway operations due to landing of flights. Intuitively, there are two kinds of
strategies to handle this issue.
(1) Plan the taxiing route for every landing flight only once when it leaves the runway.
(2) Pre-plan the taxiing route for every landing flight before it leaves the runway and, if necessary, reassign the route.
Are these two strategies equivalent to each other in terms of efficiency? In the second strategy, as we do not know the exact exit
before the landing flight leaves the runway, the question arises how the exit in the pre-planning phase can be determined, or if the 4D
taxiing route is required, how the exit time can be predetermined? If we want to pre-plan the taxiing routes, we should begin the
preplanning as close to the taxiing start time of our target flights as possible. Therefore, the running time of the approach used to pre-
plan the routes is critical. Moreover, irrespective of the planning in the first policy or the rescheduling in the second policy, the
algorithm to determine or re-determine the taxiing route for a landing fight is much more time critical because the flight has already
begun its taxiing after exiting the runway.
To deal with such problems, we propose a two-stage scheduling strategy. Similar to the optimization methods that solve real
problems, we also implement the “Rolling Horizon” architecture to limit the size of the model.
The flow diagram of the strategy is shown in Fig. 1.
Stage 1: The strategy runs periodically. In every running instance, we built an IP model to pre-plan 4D taxiing routes for landing
flights ready to touch down and departing flights that are estimated to pushback during a specific time horizon after the end of the
running instance. We call the duration of such a time horizon as a “Schedule Period”. To ensure the solvability of the model, we set
the schedule period to be much smaller than the decision time horizon (the horizon of the time variable in the IP model) of the IP
model. The relationship between the starting and ending times of one running instance, the schedule period, and the decision time
horizon is shown in Fig. 2.
After we solve the IP model and resolve the head-on conflicts, initial 4D routes of all relevant flights in the current schedule period
are generated.
Stage 2: For departures, the assigned routes remain unchanged in stage 2. However, for arrivals, as their initial routes are
assigned before they exit the runway, it is necessary to check if they can clear the runways “as the model wants”. Note that, when any
landing flight cannot leave its runway using the initially assigned exit or it cannot leave during a given time interval, this flight is
treated as “cannot fulfil the initial 4D route”. The arriving flights that leave the runway as per their initially assigned exit-time pairs
remain their initial 4D routes assigned in stage one. On the contrary, if any landing flight misses its initial exit-time pair, a re-
scheduling process will be triggered soon after the flight clears the runway. In each rescheduling call, we fix the 4D routes of other
flights in current schedule period. We therefore, only need to reassign a 4D route to a single aircraft in each rescheduling call. In every
schedule period, the arriving flights are checked sequentially in stage 2. Post that, the 4D routes of all flights to be scheduled in
current schedule period are determined and fixed. These routes will act as constraints in following periods.
Before presenting the IP model in detail, we first wish to introduce the concept and the calculation method of the REA in
Section 3.
3. Runway exit availability
As we mentioned in Section 2, if we want to pre-plan the 4D taxiing routes of landing flights before they exit the runway, the exit-
time pairs need to be pre-determined. The question that arises is whether a fixed pair can be assumed or randomly chosen? To tackle
this problem, we proposed the concept of REA in our previous work (Cheng et al., 2014). Since its proposal, we have made some
changes to its definition and calculating method. In the following section, we shall therefore detail the latest version.
167
Step into IP
next Schedule Period Initialization/Update
Solve the IP
Departures
Head-on Conflict Check Fix the 4D routes
Stage 1
Arrivals
Stage 2
Yes If all arrivals
are checked
No
Check next
landing flight
If rescheduling No
Fix 4D route
necessary
Yes
Call
rescheduling
Fig. 1. Flow diagram of the two-stage taxiing scheduling strategy.
running running running

instance 1 instance 2 instance 3
schedule schedule schedule time

period 1 period 2 period 3
decision time horizon 1
Fig. 2. The relationship between 3 time concepts.
3.1. Definition of the runway exit availability
REA is a time-exit mixed concept and is flight dependent. It measures the probability of a landing flight clearing the runway from
a specific exit in a defined time interval.
A flight can clear the runway from a specific exit in a defined time interval when the following two events occur: Event A is that
the flight decelerates down to the safe speed before it reaches the exit; Event B is that the flight reaches the exit during the specific
time interval. What is important is that the occurrence of Event B should premise on Event A. We can then denote REA as Puti
168
Puti = P (A) ∗P (B|A) (1)

where i stands for flight, u stands for exit and t stands for time interval.
To calculate P (A) , we define the stochastic variable Sui as the speed of flight i at exit u and f S i (s ) as its probability density
u
function. We define S Thres as the safe speed threshold of flight i clearing the runway from exit u and PSui as the probability of flight i
clearing the runway from exit u, i.e. P (A) . PSui can therefore be represented as follows.
SThres
PSui = ∫0 f S i (s ) ds
u (2)
Next, we investigate the probability of a flight clearing the runway inside a specified time window on the premise of satisfying the
requirement of clear-off speed. We define a stochastic variable TDT i as the landing time of flight i and define ROTui as its runway
occupation time from touchdown to reaching exit u. The time when the flight i reaches exit u is then ETui .
ETui = TDT i + ROTui (3)
Setting the probability density function of ETui as f ET i (t ) and the possibility of flight i clearing from exit u during time interval t as
u
PTuti , we get
PTuti = ∫τ∈t f ET (τ ) dτ
i
u (4)
However, the probability calculated in Eq. (4) does not represent P (B|A) because we do not place a speed constraint on it. If,
however, during the analysis of f ET i (t ) we only consider the samples in which the flight has slowed down to the safe speed before
u
reaching the specific exit, the generated probability density function, fs ET i (t ) , will become its “speed-constrained version”. Similarly,
u
we define PTsuti as the probability that flight i clears the runway from exit u during the time interval t on the premise that its speed is
not greater than its safe speed, i.e. P (B|A) .
PTsuti = ∫τ∈t fs ET (τ ) dτi

u (5)
The mathematical definition of REA is then as follows:
SThres
Puti = P (A) ∗P (B|A) = PSui·PTsuti = ∫0 f S i (s ) ds·
u
∫τ∈t fs ET (τ ) dτ
i
u (6)
∞
The larger the value of Puti , the higher the probability that flight i can clear the runway from exit u. In the scenario
∑t = 0
∞ ∞
∑t = 0 Puti = 0 , the corresponding flight can never leave the runway from exit u, while ∑t = 0 Puti = 1 means that a flight always leaves the
runway from that exit.
3.2. Calculation method of REA
From Eq. (6), we see that if we want to analytically calculate the REA, we must work out the mathematical forms of f S i (s ) and
u
fs ET i (t ). However, in this paper, instead of obtaining exact distributions, we use the frequency of occurrences of relevant events to
u
approximate the calculation of REA. Since it is difficult to collect a large number of practical samples recording the real exits and
exiting times of landing flights, we chose to make use of the simulation method.
We use the Monte Carlo approach to simulate the operations after crossing the runway threshold of landing aircrafts. In the
simulation, the threshold crossing speed, the touchdown speed, the touchdown location, and the deceleration after touchdown, are all
set as stochastic variables. We obtain their distributions by referring to literature (Kim et al., 1996) and some record data from the
North China Air Traffic Management Bureau or those from airlines. We set the distribution functions of the above four variables to the
same value for aircrafts in the same class so that the REA values of aircrafts in the same class are all the same. The count of simulation
runs for single aircraft (also represents one class of aircrafts) was set to N and the number of occasions to clear the runway from
u1, u2…uk (k is the identification number of exits on the runway) as nu1, nu2…nuk . Therefore, PSuij = nuj / N , j = 1, 2…k .
Furthermore, for an appointed exit uj and time interval t , we assume the cases that the aircraft clears from uj occur nuj times. In
u
these cases, we set the number of samples in which an aircraft exits the runway during t , as nt j . Therefore,
i uj
PTsuj t = nt / nuj , j = 1, 2…k . The approximate value is then represented as
u
Pui j t = PSuij ·PTsui j t = nt j / N , j = 1, 2…k (7)
Note that, the REA values we calculate are not used to directly predict the exit-time pairs of landing flights, but to limit the choices
of exit-time pairs in the results of the IP model in stage 1 of our scheduling strategy. We will discuss this issue in the constraints part in
the upcoming section.
4. The IP model
The initial 4D route assignment problem in stage 1 was modelled as an IP problem based on the time–space network of the airport
taxiway system. All practical constraints such as, the rear-end conflict, the head-on conflict, the minimum hold time for departures
after pushback, the runway crossing, the departure-departure separation and the departure-arrival separation have been considered
169
in the model. The solution of the model can cover the taxiing route assignment, the pushback time decision, and the take-off
sequencing.
4.1. Assumptions
We made six assumptions to simplify the model while keeping it practical. To explain the following statements, we first introduce
3 definitions on taxiing routing proposed by (Mori, 2013).
Single path: each scheduled flight only has one spatial taxiing route.
Multipath: each scheduled flight has multiple predetermined spatial paths to choose from.
Path free: no spatial routes are predetermined, with flights being free to choose any sequence of consecutive taxiway segments to
form their spatial taxiing routes.
Firstly, we assumed that an aircraft, which has landed, would not take off again during a single decision time horizon. This implies
that all flights can be strictly classified into arrivals or departures in a single running period of the scheduling strategy. Since the
decision time horizon is not too long (in this paper, decision time horizon is set to half an hour), the assumption is reasonable.
Secondly, we assumed that any flight can taxi at three different speeds only, i.e. gate speed (GS) in gate area, apron speed (AS) in
apron area, and normal speed (NS) on other taxiing routes, where GS < AS < NS. This assumption is a trade-off between constant
taxiing speed (Marín, 2006, 2013; Marín and Codina, 2008; Weiszer et al., 2015) and an entirely varying speed (Clare and Richards,
2011; Mori, 2013). The reasons for this assumption are explained below.
Thirdly, we assumed that not more than one departing flight is permitted on the same runway at the same time and that flights
may not use the runway to taxi.
Fourthly, we considered two or three neighbouring parking positions as a single node. According to surface controllers of PEK, in
real-life operations, in such a group of neighbouring parking positions, no consecutive pushbacks or arrivals are allowed to happen
during a short time period, owing to several considerations, such as limited availability of tractors and safety factors. Thus, this
simplification is considered reasonable.
Fifthly, we assumed that the touchdown times of landing flights are known, which in fact is common in relevant literatures such as
Marín (2006), Balakrishnan and Jung (2007), Marín and Codina (2008), Rathinam et al. (2008), Roling and Visser (2008), Lee and
Balakrishnan (2010), Lesire (2010), Clare and Richards (2011), Anderson and Milutinović (2013), Marín (2013), Atkin et al. (2014).
We use discrete time intervals instead of time point to mitigate the impact of uncertainty.
The last but one, we simplified the process of acceleration and deceleration during taxiing into prompt move and stop. We also
disregarded the speed change during turning. In fact, the detailed process during acceleration, deceleration, and turning is modelled
in some literatures (Mori, 2013; Weiszer et al., 2014, 2015). However, these works were based on singular handling of flights.
Lastly, we assume that proposed taxiing schedule can be achieved by the aircrafts operating on the airport surface. We claim this
assumption under the fact that surface moving speed is limited and the time stamp in our 4D taxiing routes is in the form of discrete
time intervals (with duration of 10 s)
4.2. Network definition
We built the geometrical structure of the airport taxiway system as a directed graph G = (V , E ) where V is the set of spatial nodes
and E is the set of directed edges. Nodes represent parking positions, runways, runway entrances, runway exits, runway crossing
points, air side virtual nodes, and taxiway intersections, while edges stand for geometric or logical connections between nodes.
Nodes are sorted to eight types, i.e. V = N RF ∪ N ER ∪ N AR ∪ N CR ∪ N P ∪ N O ∪ N AF ∪ N TO , where N RF is the set representing for
runways, N ER for the set of runway exits, N AR for the set of runway access points, N CR for the set of runway crossing points, N P for the
set of parking positions, N O for the ordinary intersections of taxiways, N AF for the air-side virtual nodes, with each element of N AF
corresponding to a runway, and N TO for the positions where departing flights are pushed back. We call the elements in N TO as taxi-out
nodes. Note that, N O ∩ N TO ≠ ∅. Elements in N TO may be different in different schedule periods and the size of N TO is not greater
than the size of departing flights set in the schedule.
We also categorize edges into different types. E cross represents the set of edges crossing a runway, i.e. for any edge (u, v ) ∈ E cross ,
u ∈ N CR , and v ∈ N CR . E enter represents the set of edges entering the runway, i.e. for any edge (u, v ) ∈ E enter , u ∈ N AR , and v ∈ N RF .
Similarly, E exit is the set of edges exiting a runway, i.e. for any edge (u, v ) ∈ E exit , u ∈ N RF , and v ∈ N ER . E gate, E apron , and E ord help
distinguish between taxiways in the gate area, the apron area, and the ordinary area. To match real operations, two special sets of
edges are defined. They are E withstop , for edges with block-and-hold positions, and E bidirection , for edges authorized to taxi in both
directions.
To discretize the time, we divided the whole decision time horizon into a series of discrete time intervals t (t = 1, 2, …, T ) with
equal lengths of time, where T represents the final interval of the decision horizon. For a current schedule period, we set its beginning
time interval as tstart . Integrating the time into the spatial network G , we get the time–space network, defined as N = (G, {1, 2, …, T }) .
Besides, I (u) represents the set of preceded nodes of u, i.e. I (u) = {v|(v, u) ∈ E } . O (u) represents the set of subsequent nodes of u ,
i.e. O (u) = {v|(u, v ) ∈ E } . Rw (i) represents the runway serving flight i , while To (i) represents the taxi-out node of the departing flight
i.
170
4.3. Flight set definition
We defined F as the set of flights, D as the set of departing flights and A as the set of arriving flights, where F = D ∪ A and
D ∩ A = ∅. We sort all flights planned to depart from one airport in single day into D and all flights planned to arrive in single day
into A . Note that, although several flights may correspond to one aircraft in this flight set definition, it won’t raise any problems when
solving the IP model because the situation that two or more flights corresponding to the same aircraft occurs in single scheduling
instance is impossible because the rolling time horizon scheme can obtain updated earliest pushback time for departing flights.
Additionally, Fsched was defined as the set of flights whose taxiing routes have been fixed at the beginning of the current schedule
period and Finsche as the set of flights that will be assigned 4D route in the current schedule period. Fsched is set as empty set at the
beginning. In each scheduling instance, flights whose estimated earliest moving time lying within the schedule period are set as Finsche .
After the two-stage scheduling, flights in current Finsche are inserted into Fsched and Finsche is cleared. Ftoland was defined as the set of
arriving flights whose estimated landing time intervals are within the current decision time horizon.
To mark flights in Fsched , two sets ExistRoute (u, v, t ) and ExistBlock (u, v, t ) were defined. If any flight i ∈ Fsched is scheduled to move
from node u to node v in time interval t , such a (u, v, t ) triad is added to set ExistRoute (u, v, t ) . Similarly, if any flight i ∈ Fsched is
scheduled to be blocked and held at the position on edge (u, v ) in time interval t , such a (u, v, t ) triad is added to set
ExistBlock (u, v, t ) .
4.4. Decision variables
Decision variables of the model are as the follows.
Zuti ,: This variable is not defined for any flight–node pair. For i ∈ Finsche ∩ D , u ∈ N P ∪ N RF ∪ N AF ∪ N TO or for
i ∈ Finsche ∩ A, u ∈ N P ∪ N RF ∪ N AF , if a flight i is held at node u for one time unit in time interval t , its value is 1, else it is 0.
i
Buvt , (u, v ) ∈ E withstop : 0–1 variables. If a flight i is held at the block-and-hold position on edge (u, v ) in time interval t , this has a
value of 1, else it is 0.
i
Xuvt : 0–1 variables. If a flight i moves from node u to node v in time interval t, the value is 1, else it is 0.
Sij , i, j ∈ Finsche ∩ D , and (i) = Rw (j ) : if a flight i is scheduled to take-off before another flight j , Sij = 1, else, Sij = 0 .
4.5. Auxiliary variables
TX i : the total taxiing time of a flight i , which is expressed by the number of time intervals.
4.6. Other notations of the model
Other notations and parameters of the model are as follows.
tuv : the taxiing time from node u to node v defined in number of time intervals.
usi : the source node of a flight i , usi ∈ N P for i ∈ D or usi ∈ N AF for i ∈ A .
udi : the sink node of a flight i , udi ∈ N AF for i ∈ D or udi ∈ N P for i ∈ A .
LDT i : the estimated landing time interval of an arriving flight i .
POT i : the estimated earliest pushback time interval of departing flight i .
tsi : the earliest time interval that a flight i can move in the taxiing system, where tsi = POT i , i ∈ D or tsi = LDT i , i ∈ A .
tdi : the permitted latest time interval of a flight i to arrive at its destination, tdi = tstart + T for i ∈ Finsche .
tstart : the starting time interval of a current schedule period.
Puti : Runway Exit Availability, i.e. the possibility of a flight i clearing runway from exit u during the time interval t .
Rmin i
: the lower bound of Runway Exit Availability of a flight i .
sep
Ttype (u, v ) : the safe time separation for two adjacent flights moving through the same node and measured in time intervals.. Its value
remains constant in the same taxiing area and is derived from the Minimum Taxiing Separation (MTS) and the taxiing speed in
that area. As we categorize taxiing areas into three types, the enumeration values of type (u, v ) are “gate”, “apron”, and “ord” with
sep sep sep
increasing taxiing speed in that three areas. So, we have Tgate > Tapron > Tord .
Ttype (i, j) : the time separation of two consecutive take-offs. type (i, j ) ∈ {HH , HL, HS, LH , LS, LL, SH , SL, SS } where flight i is ahead
dd
of flight j . H stands for heavy flights, L stands for large flights, while S stands for small flights.
T da : the minimum time separation of two consecutive runway operations, of which the leading one is a take-off and the followed
one is a landing.
Tto : the minimum time intervals that departing flights must be held at its taxi-out node.
4.7. Block-and-hold
To approximate entirely varying taxiing speed with multiple constant speeds and to imitate real holding manners along taxiing
routes, we set several block-and-hold positions on taxiways, as shown in Fig. 3.
On each taxiway segment with length not less than 2∗MTS, we set block-and-hold position(s) near two end points of the segment.
171
Block-and-hold
MTS position MTS
Fig. 3. Illustration of block-and-hold positions.
The distance between such a position and its nearest intersection is set as MTS.
As flights can be scheduled to wait at these positions, any low-speed taxiing can be equalized with normal speed accompanied
with a block lasting several time intervals. With this idea, there is no need to define a node’s capacity and to set capacity constraints.
4.8. Constraints
4.8.1. Flow conservation constraints

We provided the constraints of flow conservation ensuring equalization between inflow and outflow at any node in any time
interval. Different types of nodes have different expressions for flow conservation equations, as shown in Eqs. (8)–(10).
∑ Xvi, u, t − tvu + ∑ Bvi, u, t − 1 − T sep = ∑ Bvi, u, t − T sep + ∑ Xui , v, t

type (v, u) type (v, u)
v ∈ I (u) v ∈ I (u): v ∈ I (u): v ∈ O (u)
(v, u) ∈ Ewithstop (v, u) ∈ Ewithstop (8)
∀ t ∈ [tstart , tstart + T ],
∀ i ∈ Finsche ∩ A, ∀ u ∈ V /(N RF ∪ N AF ∪ N P )
∀ i ∈ Finsche ∩ D , ∀ u ∈ V /(N RF ∪ N AF ∪ N P ∪ N TO )
∑ Xvi, u, t − tvu + Ziu, t − 1 = Ziu, t + ∑ Xui , v, t

v ∈ I (u) v ∈ O (u) (9)
∀ t ∈ [tstart , tstart + T ], ∀ i ∈ Finsche , ∀ u ∈ N RF ∪ N P ∪ N AF
Xvi, u, t − tvu + Ziu, t − 1 = Ziu, t + ∑ Xui , vv, t

vv ∈ O (u) (10)
∀ t ∈ [tstart , tstart + T ], ∀ i ∈ Finsche ∩ D ,
u = To (i), v = usi
Eq. (8) corresponds to nodes with block-and-hold positions around them. Eq. (9) corresponds to nodes on which flights can hold,
e.g. parking-position nodes or runway nodes. Eq. (10) is specified for taxi-out nodes for departures.
4.8.2. Taxi-out constraints

Departing flights must hold at the taxi-out nodes after the pushback and start the engine on these nodes. The pushback and engine
starting process always take a few minutes so that this time cannot be neglected. We factor in these constraints in Eq. (11).
∑ Zui , t ⩾ Tto, ∀ i∈ Finsche ∩ D

t ∈ [tstart , tstart + T ] (11)
where u = To (i) .
4.8.3. Taxiing safety constraints

Next, we discuss the constraints of taxiing safety. Taxiing conflicts can be divided into two kinds: rear-end conflicts and head-on
conflicts. Inequality (12) is the constraint of rear-end conflict ensuring MTS at all nodes. Considering the assumption of multiple
constant speeds, as long as MTS is kept at nodes, it can be guaranteed on all taxiways.
t
∑ Zui , t + ∑ ∑ Xvi, u, τ ⩽ 1
i : to (i) = u i : to (i) = u τ = t − t i − Tto (12)
us, u
∀ t ∈ [tstart , tstart + T ]
∀ u ∈ N O ∪ N ER ∪ N AR ∪ N CR
On the other hand, in the ‘path free’ mode where no predefined path or path set is available, if the head-on conflict constraints are
172
modelled explicitly, a large number of 0–1 variables showing the permitted direction of taxiing on bidirectional taxiways in a specific
time interval are needed. This may lead to a large gap between the solution of LP relaxation problem and the original IP and may
generate a large number of fractional variables in the LP solution. This would be very time-consuming. We do not incorporate the
constraints for head-on conflict avoidance in the model directly. Instead, we propose a heuristic algorithm to deal with head-on
conflicts. This will be presented in Section 5.
4.8.4. Runway occupancy constraints

There are only three types of situations in which a runway is occupied: a flight taking off, a flight taxiing across the runway, and a
flight landing on the runway.
Eqs. (13) and (14) are runway crossing constraints that ensure no runway crossing when the runway is occupied by take-off or
landing flights. Considering an edge (ucr , vcr ) as a cross-way across the runway urw , Eq. (13) then represents the constraints for
crossing flights and landing flights.
tsi + Tup
arr
∑ ∑ Xujcr , vcr , t = 0
t = tsi − Tdown
arr j ∈ Finsche (13)
∀ i ∈ Ftoland and Rw (i) = urw

arr
Tdown represents the airborne flying time of a landing flight for a distance Larr , which is the minimum safety distance between a
landing flight and the runway threshold. This means that no runway crossing is permitted once there is a landing flight approaching
arr
the runway and its distance to the runway threshold is already less than or equal to Larr . Tup represents the operating time between
touchdown and reaching the cross-way of a landing flight. This means runway crossing is only permitted after one landing flight
passes the intersection of the cross-way and the runway.
Inequality (14) represents the constraints for crossing flights and take-off flights.
t + T dep
Ziurw, t + ∑ Xui rw, v, t + ∑ ∑ Xujcr , vcr , τ ⩽ 1
v ∈ O (urw ) ∩ N AF τ=t j ∈ Finsche ∩ D (14)
∀ i ∈ Fsched ∪ Finsche ∩ D and Rw (i) = urw
T deprepresents the forbidden time for a runway crossing after a take-off.

Formula (15)–(17) are the constraints for time separation between two consecutive take-offs. For the reason of wake-vortex
avoidance, there is always a minimum time separation arranged between two successive take-offs, which varies depending on the
classes of the leading and the following flights. Considering two departing flights i, j that take off from same runway.
∑ t·Xui , v, t − ∑ t·Xuj, v, t ⩾ Ttype

dd
(ij ) −K · Sij
t ∈ [tstart , tstart + T ] t ∈ [tstart , tstart + T ] (15)
∑ t·Xuj, v, t − ∑ t·Xui , v, t ⩾ Ttype

dd
(ji) −K · Sji
t ∈ [tstart , tstart + T ] t ∈ [tstart , tstart + T ] (16)
Sij + Sji = 1 (17)
where u = Rw (i) = Rw (j ) , v ∈ O (u) ∩ N AF

. With formula (15)–(17), we integrate the take-off sequencing in the IP model.Eq. (18) is
the constraint of time separation between consecutive take-off and landing. For a runway node urw , we have
tsi + T ad
∑ ∑ Xujrw, v, t = 0
t = tsi − T da j (18)
∀ i ∈ Ftoland, and Rw (i) = urw
where j ∈ Finsche ∩ D and Rw (j ) = urw , v ∈ O (urw ) ∩ N AF .
4.8.5. Taxiway occupancy constraints

Besides the constraints expressed in “safety constraints”, we also need another kind of constraint to guarantee safety on a taxiway
segment when it is already occupied by a flight, which is still blocked on this segment.
∑ Bui , v, t + ∑ Xui , v, t − t sep ⩽1

uv + 2 ∗ Ttype (u, v )
i ∈ Finsche i ∈ Finsche (19)
∀ (u, v ) ∈ E withstop
where type (u, v ) ∈ {apron, ord} .
173
4.8.6. Re-taxiing avoidance constraints

Re-taxiing constraints prevent flights from re-taxiing on the same taxiway. This constraint is as shown in inequality (20).
∑ ∑ Xvi, u, t ⩽ 1
t ∈ [tstart , tstart + T ] v ∈ I (u) (20)
∀ i ∈ Finsche , ∀ u ∈ N O ∪ N P ∪ N CR
We only apply the constraints on nodes belonging to set N O ∪ N P ∪ N CR because only these nodes are connected to the bidir-
ectional edges in our taxiway network. Other types of nodes are only connected to one-way edges so that for such instances, such
constraints are not necessary.
4.8.7. Parking position and runway taxiing avoidance constraints

As stated at the beginning of this section, taxiing on runways is prohibited in this model. As such, any flight must be prohibited
from entering a runway on which they do not land or take off from. Besides, flights are prohibited to cross any parking position except
their own.
∑ ∑ Xvi, u, t = 0
t ∈ [tstart , tstart + T ] v ∈ I (u) (21)
∀ i ∈ Finsche , u ∈ N P ∪ N RF and u ≠ udi
4.8.8. Route uniqueness constraints

In flow conservation constraints, for any node u , we combined all the block-and-hold positions on the edges approaching node u in
a single equation. Therefore, an additional constraint must be proposed to avoid a situation that flight i “jumps” from one taxiway to
another, i.e. Bvi1, u, t = Bvi 2, u, t + 1 = 1. These constraints are represented in inequality (22).
tdi tdi
∑ Bvi, u, t ⩽ K ∗ ∑ Xvi, u, t
t = tsi t = tsi (22)
∀ i ∈ Finsche , ∀ (v, u) ∈ E withstop

where K is a big enough positive number.
4.8.9. Boundary constraints

Eqs. (23)–(25) are decision variable unique constraints. Eq. (23) ensures that the flow at source node is 1 at the beginning of the
decision time horizon, while Eqs. (24) and (25) ensure that the flow at all other nodes is zero at the same time.
Zui i,t
s start
+ ∑ Xui i, v,t
s start
= 1, ∀ i ∈ Finsche
v ∈ O (u) (23)
⎛ ⎞
⎜ ⎟
∑ ∑ Bui , v, tstart + ∑ Xui , v, tstart + ∑ (Zui , tstart ) = 1, ∀ i ∈ Finsche ∩ A
⎜ v ∈ O (u) ⎟
u v ∈ O (u) u ∈ N RF ∪ N P ∪ N AF
⎜ withstop
⎟
⎝ (u, v) ∈ E ⎠ (24)
⎛ ⎞
⎜ ⎟
∑ ∑ Bui , v,tstart + ∑ Xui , v,tstart + ∑ (Zui,tstart ) = 1, ∀ i ∈ Finsche ∩ D
⎜ v ∈ O (u) ⎟
u v ∈ O (u) u ∈ N RF ∪ N P ∪ N AF ∪ TO (i)
⎜ withstop
⎟
⎝ (u, v) ∈ E ⎠ (25)
Eq. (26) is the start-point constraint ensuring that any flight waits at its source point before its earliest operating time.
∑ ∑ Xui i vt = 0,
s
∀ i ∈ Finsche
t < tsi v ∈ O (usi ) (26)
Eq. (27) contains the destination constraints ensuring that a flight will no longer move once it reaches its destination. For
departures, reaching a destination means starting take-off, while for arrivals, it means arriving at parking positions.
∑ ∑ Xui i , v, t = 0,
d
∀ i ∈ Finsche
t ∈ [tstart , tstart + T ] v ∈ O (udi ) (27)
Eq. (28) is another type of a destination constraint that ensures that all flights arrive at their destinations before the required latest
time. These constraints are used to guarantee the feasibility of the model.
Zui i , t i = 1, ∀ i ∈ Finsche
d d (28)
174
4.8.10. REA constraints

This group of constraints in our model are the runway exit availability constraints. Inequality (29) ensures that the REA of the
initially assigned exit-time pair is not less than a given lower bound.
∑ ∑ ∑ Pvti Xui , v, t ⩾ Rmin

i
, ∀ i ∈ Finsche ∩ A
t u ∈ N RF v ∈ O (u) (29)
4.8.11. Existing plan constraints

In each schedule period, scheduled taxiing plans need to be considered to prevent conflicts between already scheduled flights and
to-be-scheduled flights in this period. Eqs. (30) and (31) ensure no rear-end conflict with scheduled flights while Eqs. (32) and (33)
prevent head-on conflict. As stated before, we need two groups of constraints to guarantee rear-end separation. For a better un-
derstanding, Eq. (30) corresponds to inequality (12), while Eq. (31) corresponds to inequality (19). Eqs. (32) and (33) show the logic
in head-on conflict check.
sep
t + Ttype (u, v )
∑ ∑ ∑ Xui , vv, τ = 0
sep
i ∈ Finsche vv ∈ O (u) τ = t − Ttype (u, vv ) (30)
∀ (u, v, t ) ∈ ExistRoute (u, v, t )

where type (u, v ), type (u , vv ) ∈ {ord , apron , gate} .
sep
t − tu, v + 2Ttype (u, v )
∑ ∑ Xui , v, τ = 0
sep
i ∈ Finsche τ = t − tu, v + Ttype (u, v ) (31)
∀ (u, v, t ) ∈ ExistBlock (u, v, t )

where type (u , v ) ∈ {ord , apron , gate} .
sep
t − tu, v + 2Ttype (u, v )
∑ ∑ Xui , v, τ = 0
sep
i ∈ Finsche τ = t − tu, v + Ttype (u, v ) (32)
∀ (u, v, t ) ∈ ExistRoute (u, v, t )
∑ Xvi, u, t = 0
i ∈ Finsche (33)
∀ (u, v, t ) ∈ ExistBlock (u, v, t )
4.9. Objective functions
In our model, we chose to minimize the total taxiing time of the flights considered in the IP model in each schedule period. The
runway occupancy time of arriving flights and the take-off time of departing flights are not taken into consideration in the objective
function.
Eq. (34) represents the total taxiing time of an arriving flight, while Eq. (35) represents the total taxiing time of a departing flight.
⎛ ⎞
TX i = ∑ ∑ tuv Xui , v, t + ∑ Bui , v, t ∀ i ∈ A
⎜ ER ∪ N O ∪ N CR,(u, v ) ∈ E ER ∪ N O ∪ N CR,(u, v ) ∈ Ewithstop
⎟
t ⎝ u ∈ N u ∈ N ⎠ (34)
or
⎛ ⎞
TX i = ∑ ∑ tuv Xui , v, t + ∑ Bui , v, t + ∑ Zui , t ∀i∈D
⎜ ⎟
t ⎝ u ∈ NP ∪ N O ∪ N CR,(u, v) ∈ E u ∈ N P ∪ N O ∪ N CR,(u, v ) ∈ Ewithstop u ∈ rw (i) ∪ to (i) ⎠ (35)
The objective function of the model is shown in (36).
minα ∑ TX i + β ∑ ∑ t∗Xui i, v, t
s
i ∈ Finsche t ∈ [tstart , tstart + T ] i ∈ Finsche ∩ D (36)
v ∈ O (usi ), (usi , v ) ∈ E gate
where α and β are two coefficients, and α ≫ β . The first term of the objective function (36) is the total taxiing time of all flights in the
current schedule period, while the second term is the sum of pushback times of departing flights in the current schedule period. In the
case of an arriving flight that has the same parking position as a departing flight, the pushback time of the departing one cannot be
175
delayed too much because the solution of the IP model does not accept a long time hold (also considered in the total taxiing time with
the second term in (36)) outside the parking position for the arriving flight. This ensures that we do not need special constraints to
limit the maximum pushback delay (such as those in Mori (2013)) to prevent possible long take-off delays for departing flights.
5. Refinement of the IP model
As the preliminary IP model is complex and large in size, it has to be refined so as to speed up the solving process. For this
purpose, three improvements were implemented.
5.1. Heuristic for head-on conflict
As stated in “Taxiing safety constraints”, we do not explicitly incorporate the head-on conflict constraints in the IP model. Instead,
after the IP is solved, we check assigned routes of all flights in Finsche and record events that suggest that two or more flights could
potentially be involved in head-on conflicts. We put all these flights into a set Fheadon , and sort it in descending order in terms of times
involved in potential head-on conflicts. Beginning with the most frequently involved flight, we sequentially recalculate the 4D path of
flights in Fheadon . When the recalculation of one flight is executed, the routes of all other flights are fixed and are set as constraints.
After the recalculation, we update the 4D routes of the flight. We recheck potential head-on conflicts and reset Fheadon . This iterative
process continues until no potential head-on conflicts exist. The IP model with Eqs. (8)–(36) is termed IP-(1).
5.2. Refinement of taxi-out constraint
We found that the inequality representing the taxi-out constraints is one of the major sources leading to a large number of
fractional variables in the solution of the LP relaxation problem. We merged Eq. (11) into the rear-end constraint and rewrote Eq. (12)
as follows:
t t
⎛ ⎞
∑ Zui , t + ∑ ∑ Xvi, u, τ + ∑ ∑ Xui , v, τ ⩽ 1
sep
⎜ ∑ ⎟
i : to (i) = u i : to (i) = u τ = t − t i − Tto τ = t − Ttype (u, v ) i ⎝ v ∈ O (u) ⎠ (37)
us, u
The third term in Eq. (37) is responsible for the taxi-out constraint. The IP model with Eqs. (8)–(10), (12)–(37) is termed IP-(2).
5.3. Refinement of take-off/take-off separation constraint
Similar to the consideration of the head-on constraint, we abandoned the constraints in Eqs. (15)–(17). The equivalent trans-
formation form equation is as shown in Eqs. (38)–(40).
Inequality (38) ensures the time separation between flights in current schedule period within the same aircraft class.
sep
t + Ttype
∑ ∑ Xui , v, τ ⩽ 1
i with same class (i) τ=t
Rw (i) = urw, i ∈ Finsche ∩ D (38)
sep sep
where u = urw , v ∈ O (u) ∩ N AF , class (i) is the class of flight i , and class (i) ∈ {H , L, S } , if class (i) = H , Ttype = THH , if class (i) = L ,
sep sep sep sep
Ttype = TLL , if class (i) = S , Ttype = Tss .
Inequality (39) ensures the time separation between flights in the current schedule period with different classes.
sep
t + Ttype (i , j )
Xui , v, t + ∑ ∑ Xuj, v, τ ⩽ 1
j : class (j ) ≠ class (i) sep
τ = t − Ttype (j , i )
j ∈ Finsche ∩ D, Rw (j ) = urw (39)
∀ i ∈ Finsche ∩ D , and Rw (i) = urw

where u = urw , v ∈ O (u) ∩ N AF .
Eq. (40) ensures the time separation between scheduled departing flights.
sep
Se (i) + Ttype (i , j )
∑ ∑ Xuj, v, t = 0
sep
j ∈ Finsche ∩ DRw (j ) = urw t = Se (i) − Ttype (j , i ) (40)
∀ i ∈ Fsched ∩ D , and Rw (i) = urw
176
where u = urw , v ∈ O (u) ∩ N AF , Se (i) represents the scheduled take-off time interval of a scheduled flight i .
The IP model (8)–(10), (12)–(14), (18)–(40) is termed IP-(3).
6. Heuristics when updating the 4D taxiing route
In the proposed two-stage scheduling strategy, we update the 4D taxiing routes of some flights in two scenarios – one is in the
situation that probable head-on conflicts are detected, and the other is in rescheduling when any landing flight misses its initially
assigned exit-time pair. Considering on-job requirements, the updates need to be executed fast. We propose a shortest path algorithm
(SPA) based heuristic method to handle this.
Before carrying out the update, a minor modification was made to the time–space network. In Section 4.2, the block-and-hold
i
point on an edge, (u, v ) , is not modelled as a spatial node of the network, but just represented using decision variables Buvt . However,
to cover possible block operations when running SPA, we must directly model such a block-and-hold position as a spatial node and
connect it with node u and v .
As stated before, each calling of the update recalculates the 4D taxiing route for a single flight. The routes of other flights are then
temporarily fixed, so that their routes act as constraints in the update. The time–space network thus needs to be simplified by
removing edges, which would possibly cause conflicts with flights whose 4D routes are temporarily fixed. This simplified network is
denoted as N ′.The mission of the heuristic method is to find the shortest path in N ′ from node S (ncur , tcur ) to any node in
{T (ndest , tdest )|tdest ∈ [tcur + tshortest , tterm]}. For departing flights, ncur ∈ N P stands for its source node, ndest ∈ N AF for the airside virtual
node corresponding to its take-off runway, and tcur for its required pushback time. For arriving flights, ncur ∈ N ER stands for its runway
exit node, ndest ∈ N P for its sink node, and tcur for its actual runway exit time. tshortest represents the shortest time from ncur to ndest when
safety constraints are not considered, and tterm is the final time interval in the current decision horizon.
Proposition. The time complexity of the heuristic method is O (|T|log|T|(m + n)) where |T| is the number of time intervals in a single decision
horizon, m is the number of taxiway segments in the taxiway system, and n is the number of spatial nodes.
Proof. Firstly, we look for the shortest path from (ncur , tcur ) to (ndest , tcur + tshortest ) . If no path exists, a flight cannot reach its
destination in tcur + tshortest . In such a case, the binary search in {T (ndest , tdest )|tdest ∈ [tcur + tshortest , tterm]} is used to find the first feasible
path from (ncur , tcur ) to (ndest , tcur + tshortest + k ), with no path existing between (ncur , tcur ) and (ndest , tcur + tshortest + l), l < k . The
maximum number of this iteration is log (tterm−tcur −tshortest ) , which is bounded by log|T|. Besides, as N ′ is acyclic (the nodes in N ′ are
defined using spatial positions and time intervals, with time always stepping ahead so that no directed cycle exits), we can use the
reaching algorithm whose complexity is O (m′) (Ahuja et al., 1993) as the SPA where m′ stands for the number of edges in N ′.
Apparently, m′ is in O (|T|(m + n)) . Hence, the total complexity of the heuristic algorithm for a single flight is O (|T|log|T|(m + n)) .□
7. Experiment results
7.1. Experiment setting
We tested our strategy in a simulated environment representing the taxiing system of the Beijing Capital International Airport
(PEK). PEK has three terminals and three parallel runways, which operate independently. It is the second busiest airport in the world
in terms of passenger throughput. The normal flight flow density is around 95–110 take-offs/landings per hour.
As per the definition in Section 4, the model of the taxiway network has three airside virtual nodes, three runway nodes, eight
runway crossing nodes, six runway access nodes, sixteen runway exit nodes, ninety-six parking-position nodes, and two hundred and
sixty-five ordinary-intersection nodes (ninety-six of these are potential taxi-out nodes). We set the discrete time interval to be 10 s.
The schedule period was set at 5 min, while the decision horizon was set at 30 min, i.e. 180 time intervals, or T = 180 .
According to the taxiing speeds suggested in Eurocontrol Experimental Center (2012) and the analysis of actual taxiing data
sep sep sep
record from the Tower Controlling Center of PEK, we set Tgate = 3, Tapron = 2, and Tord = 1. For the time separation between con-
secutive runway operations, we utilized the regulations in Ahuja et al. (1993) wherein the time separation between a landing and a
take-off has no relationship with flight class, but depends only on the order of these two operations. Therefore, only two relevant
parameters, T da and T ad needed to be defined. For the runway crossing, we set the safe distance between a landing aircraft and the
runway threshold as 1.5 nautical miles. From this safe distance and the geometrical structure of the runways, we could set relevant
time parameters for runway crossing constraints. Besides, we set the minimum hold time at taxi-out node for the departures at 3 min,
i
i.e. Tto = 18. Default value of Rmin was set to 0.1.
In the experiment, we simulated the operations on July 21st, 2015 from 7 a.m. to midnight, except for the time between 4p.m. to
5p.m., which was affected by severe weather. There are 192 schedule periods in total. Three runways operated independently, i.e. all
of them could serve both take-off and landing operations. PEK operates in northbound pattern, that means flights land or take-off
northbound. All flight information, including the operating runways, the parking positions, the estimated landing time, and the
required pushback time, were offered by the Tower Control Center of PEK.
All simulations were run on three different PC platforms. They are equipped with Core i3-2120 CPU/4 GB RAM, Core i5-4460
CPU/4 GB RAM, and Core i7-3520M CPU/8 GB RAM. The IP models in the experiment were all solved by CPLEX 11.0.0.
177
Table 1
Results of period 12:20–12:25.
CS CT AST APT AET SET ATT STT SAV
CXA8118 D 12:21:00 12:25:20 12:55:25 12:35:50 34:25 10:30 69.5%

CHH7182 D 12:23:17 12:25:10 12:48:30 12:33:30 25:13 08:20 67.0%
CCA1712 D 12:32:33 12:33:50 13:01:58 12:40:50 27:27 07:00 74.5%
HDA901 D 12:26:54 12:37:10 12:50:31 12:47:10 24:37 10:00 59.4%
CCA1590 A 12:21:21 ⧹ 12:24:30 12:23:50 03:09 02:29 20.6%
CES5709 A 12:21:46 ⧹ 12:26:13 12:26:50 04:27 05:04 −13.9%
TUA605 A 12:23:35 ⧹ 12:35:50 12:34:20 12:15 10:50 11.6%
CCA1306 A 12:24:29 ⧹ 12:31:44 12:31:00 07:15 06:31 10.3%
CCA862 A 12:25:05 ⧹ 12:33:05 12:31:20 08:00 06:25 18.8%
Total Taxiing time for departures 111:42 35:50 67.9%
Total Taxiing time for arrivals 35:06 31:24 10.5%
Total Taxiing time 146:50 67:14 54.2%
7.2. Example of results
In this subsection, we provide the simulation results of one schedule period, from 12:20 p.m.–12:25 p.m. In this period, nine
flights were to be scheduled, four being departures and the other five being arrivals. Besides, there were twenty-three other scheduled
flights, which were still taxiing in this period.
In Table 1, we have listed the details of the scheduling results for this period. CS represents the call sign. CT categorizes the flights
into arrivals (A) and departures (D). AST (Actual Start Time) means the actual touchdown time for arrivals and the actual pushback
time for departures. In this example, CXA8118 pushed back at its earliest possible pushback time according to the practical record.
APT means the assigned pushback time. AET (Actual End Time) means the actual take-off time for departures and the actual time
reaching parking positions for arrivals. SET stands for scheduled end time. ATT stands for actual taxiing time, while STT for scheduled
taxiing time. SAV stands for the percentage of saving in terms of taxiing time.
The scheduled end time is always in tens of seconds, because we set a single time interval as 10 s. Although this mode may
introduce some errors when calculating STT, it can be neglected in view of the saving in taxiing time. The results in Table 1 lists about
60–70% saving in taxiing time for departing flights and about 10–20% saving for arrivals. Flight CES5709 is an exception whose
taxiing time is even longer after scheduling. However, the total taxiing time of five arriving flights in this schedule period decreases
by 10.5% while that of four departing flights decreases by 67.9%. The saving in aggregate taxiing time of nine flights is 54.2%.
In Fig. 4, we illustrate the taxiing routes of three flights within this period. The taxiing route of CXA8118 is drawn in blue, while
that of TUA605 is in red and that of CCA862 is in purple. The time stamps in the boxes denote the moments at which the flights leave
their respective nodes except for those pointing to their destinations. In the boxes pointing to the destination nodes, time stamps in
these boxes represent the reaching times. Note the red star along the taxiing route of CCA862. The red box to the right-bottom side of
it shows that CCA862 is scheduled to stop at that point for 20 s to avoid conflicts with other flights crossing the nearest intersection.
7.3. Advantage analysis
7.3.1. Comparison to other IP/MIP based approaches

For the IP or MIP based approaches, it may not be suitable to directly compare the scheduling results because the constraints,
assumptions, and functions of different models may be very different. In fact, to our best knowledge, such comparisons were never
carried out in existing literatures. We compare our approach to five other IP/MIP based approaches as listed in Table 2, in which
several features regarding constraints, assumptions, functions, solving speed, size of the model, etc. are considered.
In Table 2, the row “Available paths” refers to taxiing routes that are available to choose from. “A” stands for alternate paths, i.e.
two or a set of predefined paths. “F” means path free, i.e. no predefined taxiing routes. The “Taxiing speed” row shows the speed
assumption made by the respective approaches. “Fixed” refers to the assumption that a flight taxies at constant speed, whereas
“Varying” refers to the assumption that the taxiing speed of any flight can change within a defined boundary at all times during
taxiing. “Multi” stands for the assumption set in this paper that a flight taxies at different speeds in different areas of the taxiway
system. In the rows three to seven, the head-on conflict constraint, the runway crossing constraint, the take-off sequencing, the
multiple runway exits, and the engine starting constraints are considered. “A” stands for the absence of relative constraints or
considerations while “P” stands for its presence. “Holding manner” refers to how holdings during the taxiing are modelled in the
approaches. “On” stands for those flights that are modelled to hold right on the nodes representing the intersections, while “Beside”
stands for flights that are modelled to hold beside the intersections, akin to cars waiting for a green light. The last three rows show
whether the approaches run periodically, assign complete taxiing routes or part of entire taxiing routes to flights in a single period,
and whether they are tested based on practical data. The options “Y” and “N” stand for yes and no resp.
As listed in Table 2, three previous works rely on the varying speed assumption. In Guepet et al. (2016), the alternate path
strategy was applied to limit the size of the model. In Clare and Richards (2011), although there is no need to fix taxiing routes, only a
small part of the entire taxiing route can be determined in a single schedule period, while Evertse and Visser (2017) is like the case of
178
24:50
25:40
24:00
26:20
26:40
26:50
27:50
28:10
28:30
29:10
TERMINAL1
26:40
29:40 TERMINAL2 30:40

26:50
25:20
30:20 31:30 26:00
29:50
30:30 33:30 34:50 31:20 E 27:20

27:50
RW36L 31:00 32:00 32:40 34:20 31:50

28:00
T
31:40
32:40
32:10 E 30:10 29:20
31:00 R 29:10
29:40
M 30:00
34:00 DI
34:20
N
A
34:30
L
35:00
35:20
3
C
RW01
RW36R
Fig. 4. Examples of scheduled 4D taxiing routes.
Clare and Richards (2011). These two compromises caused by the varying speed assumption partly verify our statement regarding the
need for a multi-speed assumption. The information listed in Table 3, especially the comparisons between “Multiple runway exits”,
“Engine starting”, and “Holding manner”, demonstrate the improvements of our approach in terms of better practicality.
7.3.2. Comparison to sequential scheduling approaches and practical operations

We make comparisons to two sequential scheduling methods. One (named with SM1) schedules both the arrivals and departures
forward (from runway to apron for arrivals while from apron to runway for departures). The other one (named with SM2) is similar to
the Quickest Path Problem with Time Windows (QPPTW) proposed in Atkin et al. (2014) which scheduled the arrivals forward (from
runway to apron) and scheduled the departures backward (from runway to apron). Because neither SM1 nor SM2 were capable of
dealing with the uncertainty about the runway exit, we just assumed all arriving fight to clear the runway from the exit nearest to the
179
Table 2
Comparison between several IP/MIP based approaches.
Roling and Visser Marín (2013) Clare and Richards (2011)) Evertse and Visser Guepet et al. (2016) Ours
(2008) (2017)
Available paths A F F F A F
Taxiing speed Fixed Fixed Varying Varying Varying Multi
Head-on conflict avoidance A A P P P P
Runway crossing A A A A A P
Take-off sequencing A P P P P P
Multiple runway exits A A A A A P
Engine starting A A A A A P
Holding manner On On On On On Beside
Periodical scheduling Y Y Y Y Y Y
Complete route Y Y N N Y Y
Test with real data N Y Y Y Y Y
Table 3
Percentage of taxiing time saving through the day from 07:00 to 24:00 except 16:00–17:00.
Time SAVS1D (%) SAVS2D (%) SAVRD (%) SAVS1A (%) SAVS2A (%) SAVRA (%) SAS1V (%) SAS2V (%) SARV (%)
7–8 31.1 65.5 65.5 14.1 10.1 10.1 29.7 61.0 61.0
8–9 26.8 70.9 70.9 7.2 2.6 7.2 26.9 60.1 67.9
9–10 46.3 57.2 57.2 11.4 5.6 12.3 35.8 41.2 43.7
10–11 48.2 64.3 62.0 13.0 3.7 11.8 35.6 40.0 44.0
11–12 40.2 67.4 61.4 16.9 −2.3 17.7 32.9 38.2 50.8
12–13 44.7 65.0 61.2 19.0 −4.4 19.1 35.0 38.3 45.1
13–14 39.7 66.1 63.7 20.4 −0.8 19.4 34.2 39.9 51.2
14–15 42.3 55.4 55.4 17.8 8.1 16.2 36.6 43.1 46.3
15–16 38.1 61.8 59.8 11.8 9.2 12.4 31.1 45.0 47.9
17–18 29.0 70.0 68.6 12.6 6.9 12.6 26.0 36.9 58.1
18–19 32.7 69.0 67.6 10.0 3.2 11.9 26.3 33.3 51.1
19–20 36.6 61.7 60.7 12.8 5.5 13.3 30.7 39.9 48.8
20–21 39.8 63.8 63.1 22.8 3.4 22.8 34.3 40.4 50.0
21–22 40.0 60.5 60.5 14.9 9.1 15.4 36.8 44.7 49.3
22–23 52.8 63.4 63.4 12.5 10.5 13.6 37.8 43.6 44.9
23–24 20.0 20.0 20.0 14.6 14.6 14.6 15.4 15.4 15.4
runway threshold. Additionally, the first come first service apply to both SM1 and SM2. In SM1, the estimated landing time and the
estimated earliest pushback time were used as the reference when deciding the scheduling sequence. In SM2, as the literature (Atkin
et al., 2014) did not offer the rule to determine exact take-off time while it must be predetermined before running the algorithm, we
just used actual take-off time for departing flights.
In Table 3, we list the statistical data of the experimental results comparing our approach, SM1, SM2, and the practical data for
every hour. SAVS1D, SAVS2D, and SAVRD represent the percentage of taxiing time savings for departures implementing SM1, SM2
and REA, respectively. Similarly, SAVS1A, SAVS2A, and SAVRA correspond to the time savings for arrivals, and SAS1V, SAS2V, and
SARV are the percentage of time savings in relation to aggregate taxiing times.
It is evident from the experimental results, that our approach is much more efficient than that implementing SM1 or SM2 and
offers a considerable reduction in taxiing times in practical operations. The percentage saving in taxiing times with REA approach in a
practical operation scenario is approximately 60% for departures, while that for arrivals is approximately 10%. The aggregate time
saving considering both departures and arrivals is approximately 50%.
For SM1, we obtain about 30% time saving against real operations from. As the time saving for arrivals obtained using our
approach and SM1 is almost the same, we can claim that the difference in efficiency is caused by the different 4D route scheduling for
departures. One major reason for long taxiing times is the inefficient pushback order that authorizes pushback too early or too
frequently (Atkin et al., 2013). Moreover, as the logic used in SM1 is most similar to the active operating rule used in practice (at least
in China), the consequence of SM1 can also be set as a benchmark when testing the superiority of our approach. Comparing our
approach to SM1 can show the advantage of our optimization more directly because the extents of approximation to the practical
operation in terms of involving the uncertainties in the model are nearly the same for both of the approach, while the main difference
between them are whether the optimization is applied. We can see that our approach still obviously outperforms SM1 especially in
some departing peak hours like the hours in the morning.
Different from the SM1, SM2 scheduled the departures backward which solves the problem about early pushback in most se-
quential scheduling methods and every departing flight can reach its runway without any blocks theoretically because all the delay
can be absorbed at the parking position. The aggregate taxiing time saving of when applying SM2 is more than that of SM1. However,
we mention two facts about the SM2. One is that the take-off time for every departing flight must be known as a prerequisite. On the
contrary, our approach and some other IP/MIP based approaches can schedule the take-off sequence by themselves (see Table 2).
180
1
0.93
0.83
161
52
A B C
Fig. 5. Comprehensive performance of different cases.
Secondly, as the delay is always absorbed at the parking position in SM2, it may lead to long time occupancy of one parking position,
which may block the arriving flights taxiing to that occupied position. This negative impact can be more obvious along with heavier
airport surface traffic flow and it can be found in the 6th and 9th column in Table 3 where the taxiing time savings for arrivals are
much less than other two algorithms or even in negative numbers comparing to actual record. However, in our approach, we avoid
long time stand waiting with the second term of the objective function.
7.4. Optimality analysis of REA
In this section, we discuss the question proposed in Section 2 that concerns about which policy is more efficient, the policy with
only once taxiing route assignment for landing flights, or the policy with pre-planning and necessary rescheduling? We also discuss
the question that if the pre-planning is preferred, can the REA take any advantages in the pre-planning phase?
We make use of the simulation results and try to answer the questions above. We set three cases here.
Case A: Each landing fight is scheduled only once when it exits the runway. The IP model in scheduling stage 1 only considers
departures in the corresponding schedule period.
Case B: Each landing flight is assigned an exit-time pair with the highest REA value in the IP model of scheduling stage 1.
i
Case C: IP-(3) is applied in scheduling stage 1, Rmin is set to 0.1 for all landing flights.
The results are shown in Fig. 5. We use histograms to visualize the number of occasions of best performance in terms of the total
taxiing time. The numbers and the expressions on the top of the histograms represent the number of occasions and the corresponding
cases. Fig. 5 also shows that the total scheduled taxiing time in case C is the shortest in the simulation, with only 83% of that in case A
and 10% lower than that in case B. This result answers the questions we recalled at the beginning of this section, that methods with
pre-planning are superior to methods with a single scheduling for landing flights. Moreover, considering several exit-time pairs bound
by a REA value is better than setting a single fixed pair. However, we cannot ensure such an optimality gap for every airport.
Intuitively, we hypothesize, that the optimality gap may increase in more complex taxiway system and heavier surface traffic.
7.5. Solving speed analysis
7.5.1. Improvement through formulation refinement

In Section 5, we proposed three improvements to refine the IP model, thereby generating three alternatives IP-(1), IP-(2), and IP-
(3). To simplify the statements below, we consider the formulation that directly involves the head-on constraint as the original IP
model, IP-(O).
Until now, we have not verified any inequalities to be facets. Instead, we set four features revealing the size and complexity of the
IP model and showed the improvement using the simulation results of problems having different features.
(a) The number of flights to be scheduled in a single schedule period is no more than 5.
In this situation, all four models can be solved in a limited time, with IP-(O) being the slowest.
(b) The number of flights to be scheduled in a single schedule period is between 6 and 8.
With an increase in the number of flights, IP-(O) is incapable of being solved within 5 min. The other three models still function
well, but branch and bound (B&B) is occasionally required by IP-(1).
181
Table 4
The results about solving speed for the schedule periods between 12:00 and 13:00.
PRD FN FNS GAP (%) CUTS CPU time (s)
i3 i5 i7
1 11 20 0 0 196 122 91
2 7 17 0 0 90 57 41
3 7 19 0 0 96 50 36
4 7 23 0.3 1 140 90 59
5 9 24 0.1 2 111 80 64
6 5 18 0 0 79 46 31
7 7 16 0 0 127 80 56
8 6 14 0 0 73 45 34
9 8 15 0 0 100 65 48
10 8 17 0 0 118 73 52
11 7 20 0.5 3 156 94 63
12 5 18 0 0 53 35 27
(c) The number of flights to be scheduled in a single schedule period exceeds 10 and the ratio of departing flights is small.
In this situation, IP-(1) cannot be solved within 5 min in most instances, while IP-(2) and IP-(3) still function well. B&B is
occasionally required by IP-(2).
(d) The number of flights to be scheduled in a single schedule period exceeds 10 and the ratio of departing flights is large.
In such a situation with heavy surface flow and a large number of possible take-off permutations, IP-(2) needs a long solving time
when several B&B operations take place. However, IP-(3) can still be solved in a very limited time with no B&B and only a few or even
without any cuts.
To summarize, the improvement of the solving speed from IP-(O) to IP-(3) proves the efficacy of our improvements. Following
results regarding the solving speed correspond to IP-(3).
7.5.2. Solving time of IP-(3)

Firstly, we show the results during 12:00–13:00. In Table 4, we list the results showing solving speed of IP-(3) for twelve schedule
periods in this hour. PRD represents the period. FN is the number of flights to be scheduled in the current period. FNS is the number of
scheduled flights that still taxi in this period. GAP, which is measured in percentage, represents the difference between the value of
the objective function in LP relaxation and that of the IP model. CUTS is the number of cuts applied by CPLEX when solving the IP.
CPU time is recorded separately for simulations on i3, i5, or i7 platforms. It is the total CPU time used by CPLEX, including the model
generating time and the solving time.
We see that the CPU time may be significantly different even if FNs are equal. This is caused by various origins and destinations of
the flights involved in the IP model and the status of scheduled flights. Average solving time of the IP models with different number of
flights involved in the model, is illustrated in Fig. 6, where the results when FN is below 5 are not present.
Throughout the simulation utilizing IP-(3), CPLEX does not call B&B in any schedule period and the maximum number of cuts
applied in a single period does not exceed three. It is to be noted that the situation with FN = 14, i.e. the flow density with 14
operations in 5 min, is a very high traffic scenario for current air traffic systems. The solving time under such dense flows is just
272
248
i3-2120, 4GB RAM
230
i5-4460, 4GB RAM
205
i7-3520M, 8GB RAM
181
Solving
161 160
time of IP
(second) 147
140
130
125
121
114
111
102
96
92
83
75 73 71
62 65 64
52
47 44
33 34
25
5 6 7 8 9 10 11 12 13 14 FN
Fig. 6. Average solving time of the IP model.
182
around 140 s (less than half of the duration of single schedule period) on a PC with Core i7-3520 M and CPLEX 11.0.0. This means
that we can start the scheduling at most 2.3 min in advance for each schedule period. This fact is very helpful for us to make full use of
the latest information in the taxiway system. We may expect an even shorter solving time if we run the model on a more powerful
computer.
7.5.3. Time cost in head-on conflict check

For one hundred ninety-two schedule periods in the simulation, we record the time cost in every call of heuristics for a potential
head-on conflict, which is in fact an instance to solve the shortest path problem for a single flight in the time–space network. The
maximum solving time of a single instance is 1385 ms and the average value is 766 ms.
7.5.4. Time cost in exit-time pair check for landing flights

The algorithm used for rescheduling landing flights, which miss their initial assigned exit-time pairs is very similar to what is used
for head-on conflict check and resolution. The running time is also very similar. The maximum solving time of a single instance is
1219 ms and the average value is 697 ms.
8. Conclusion and future work
8.1. Summary
We presented a periodical two-stage strategy to schedule 4D taxiing routes in the taxiway system at large airports. To help model
runway operations more accurately, we introduced a concept, Runway Exit Availability (REA), to measure the possibility of arriving
flights clearing the runway in specific exit-time pairs. We constructed a time–space network to represent the taxiing system and built
an IP model based on this network. The IP model, incorporating many real-life constraints and specific considerations about the stops
along taxiing routes, is solved in stage 1 periodically. In stage 2, if any arriving flight cannot exit the runway from its initial assigned
exit-time pair in stage 1, a rescheduling program based on the shortest path algorithm is executed. To accelerate the solving process,
we carried out several refinements in the IP model, such as utilizing head-on conflict check and resolution heuristic instead of directly
building the head-on conflict constraint in the IP model, transforming the taxi-out constraint, and making equivalent changes to the
take-off/take-off separation constraint
We tested our strategy at the Beijing Capital International Airport (PEK) and analysed the performances of the methods with and
without REA. The results revealed the potential of our strategy in reducing taxiing time significantly while promising an acceptable
solving time.
8.2. Future work
As we mentioned in the introduction, the Taxiing Scheduler (TS) should be related to many other schedulers or programs, such as
Gate Assignment (GA), Runway Configuration (RC), Departure/Arrival Scheduler (DAS), Airline Operation Manager (AOM), Air
Traffic Flow Management (ATFM), and also with the hardware architecture of Communication, Navigation, and Surveillance (CNS).
Only through co-operation between these systems, can a fully-functional airport surface management system be built. However, how
should the inner relations in such a system be like? Should these sub-problems be solved together or separately? There is still a long
road to its realization.
Funding
This work was supported in part by the National Natural Science Foundation of China (Grant 61179052), in part by the National
Natural Science Foundation of China (Grant U1222202), in part by the 973 National Basic Research Program of China (Grant
2010CB731805, 2010CB731806), in part by Aviation Science Fund of China (Grant 20128058005), and in part by ATM Safety Project
from CAAC (Grant TMSA 1501/1).
Acknowledgment
We acknowledge the large amount of empirical data and the valuable suggestions from North China Air Traffic Management
Bureau of Civil Aviation Administration of China, Air China, and the Tower Control Center of the Beijing Capital International
Airport. We also appreciate the careful reviews and the precious advices from the reviewers.
References
Ahuja, R.K., Magnanti, T.L., Orlin, J.B., 1993. Network Flows: Theory, Algorithms, and Applications. Prentice-Hall Inc, Upper Saddle River, NJ, USA.
Anderson, R., Milutinovic, D., 2013. An approach to optimization of airport taxiway scheduling and traversal under uncertainty. Proc. Inst. Mech. Eng., Part G: J.
Aerospace Eng. 227 (2), 273–284.
Atkin, J.A.D., Maere, G.D., Burke, E.K., Greenwood, J.S., 2013. Addressing the pushback time allocation problem at Heathrow airport. Transport. Sci. 47 (4), 584–602.
Atkin, J.A.D., Burke, E.K., Ravizza, S., 2014. A more realistic approach for airport ground movement optimisation with stand holding. J. Sched. 17 (5), 507–520.
Baik, H., Sherali, H.D., Trani, A.A., 2000. Time-dependent network assignment strategy for taxiway routing at airport. Transp. Res. Rec. 1788, 02–3660.
183
Balakrishnan, H., Jung, Y., 2007. A framework for coordinated surface operations planning at Dallas-Fort Worth International Airport. In: The AIAA Guidance,
Navigation, and Control Conference, Hilton Head, NC, USA.
Bertsimas, D., Patterson, S.S., 1998. The air traffic flow management problem with enroute capacities. Oper. Res. 46 (3), 406–422.
Bertsimas, D., Lulli, G., Odoni, A., 2011. An integer optimization approach to large-scale air traffic flow management. Oper. Res. 59 (1), 211–227.
Burgain, P., Pinon, O.J., Feron, E., Clarke, J.-P., Mavris, D.N., 2012. Optimizing pushback decisions to valuate airport surface surveillance information. IEEE Trans.
Intell. Transp. Syst. 13 (1), 180–192.
Eurocontrol Experimental Center, 2012. User Manual for the Base of Aircraft Data (BADA) Revision 3.10.
Cheng, P., Zou, X., Liu, W., 2014. Airport surface trajectory optimization considering runway exit selection. In: 2014 IEEE Intelligent Transportation System
Conference, Qingdao, China.
Cheng, V.H.L., 2004. Surface operation automation research for airport tower and flight deck automation. In: The 7th International IEEE Conference on Intelligent
Transportation Systems.
Clare, G.L., Richards, A.G., 2011. Optimization of taxiway routing and runway scheduling. IEEE Trans. Intell. Transp. Syst. 12 (4), 1000–1013.
Coupe, William J., Milutinovi, Dejan, 2015. Integration of uncertain ramp area aircraft trajectories and generation of optimal taxiway schedules at Charlotte Douglas
(CLT) Airport. In: 15th AIAA Aviation Technology, Integration, and Operations Conference, Dallas, TX, USA.
Ding, H., Limb, A., Rodrigues, B., Zhu, Y., 2005. The over-constrained airport gate assignment problem. Comput. Oper. Res. 32, 1867–1880.
Eun, Y., Hwang, I., Bang, H., 2010. Optimal arrival flight sequencing and scheduling using discrete airborne delays. IEEE Trans. Intell. Transp. Syst. 11 (2), 359–373.
Evertse, C., Visser, H.G., 2017. Real-time airport surface movement planning: minimizing aircraft emissions. Transp. Res. Part C 79, 224–241.
Guepet, J., Briant, O., Gayon, J.P., Acuna-Agost, R., 2016. The aircraft ground routing problem: analysis of industry punctuality indications in a sustainable per-
spective. Eur. J. Oper. Res. 248 (3), 827–839.
Kim, B., Li, L., Clarke, J.-P., 2014. Runway assignments that minimize terminal airspace and airport surface emissions. J. Guidance Control Dyn. 37 (3), 789–798.
Kim, B.J., Trani, A.A., Gu, X., Zhong, C., 1996. Computer simulation model for airplane landing-performance prediction. Transp. Res. Rec. 1562, 53–62.
Kistler, M.S., Gupta, G., 2009. Relationship between airport efficiency and surface traffic. In: 9th AIAA Aviation Technology, Integration, and Operations Conference
(ATIO).
Lee, H., Balakrishnan, H., 2010. Optimization of airport taxiway operations at Detroit Metropolitan Airport (DTW). In: AIAA Aviation Technology, Integration and
Operations (ATIO) Conference, Fort Worth, TX, USA.
Lesire, C., 2010. An iterative A* algorithm for planning of airport ground movements. In: 19th European Conference on Artificial Intelligence (ECAI 2010), Lisbon,
Portugal.
Marín, Á.G., 2006. Airport management: taxi planning. Ann. Oper. Res. 143, 191–202.
Marín, Á.G., 2013. Airport taxi planning: Lagrangian decomposition. J. Adv. Transport. 47, 461–474.
Marín, A., Codina, E., 2008. Network design: taxi planning. Ann. Oper. Res. 157, 135–151.
Mori, R., 2013. Aircraft ground-taxiing model for congested airport using cellular automata. IEEE Trans. Intell. Transp. Syst. 14 (1), 180–188.
Neufville, R.D., Odoni, A.R., 2003. Airport Systems, Planning, Design and Management. McGraw-Hill, New York, USA.
Park, B.-S., Lee, H., Seon, Y.K., Lee, H.-T., 2017. Airport surface movement scheduling with route assignment using first-come first-served approach. In: 17th AIAA
Aviation Technology, Integration, and Operations Conference, Denver, USA.
Rathinam, S., Montoya, J., Jung, Y., 2008. An optimization model for reducing aircraft taxi times at the Dallas Fort Worth International Airport. In: 26th International
Congress of the Aeronautical Sciences (ICAS).
Roling, P.C., Visser, H.G., 2008. Optimal airport surface traffic planning using mixed integer linear programming. Int. J. Aerospace Eng. 2008 (1).
Weiszer, M., Chen, J., Stewart, P., 2015. A real-time active routing approach via a database for airport surface movement. Transp. Res. Part C 58, 127–145.
Weiszer, M., Chen, J., Ravizza, S., Atkin, J.A.D., Stewart, P., 2014. A heuristic approach to greener airport ground movement. In: 2014 IEEE Congress on Evolutionary
Computation, Beijing, China.
Yu, Z., Qing, W., 2017. Methods for determining unimpeded aircraft taxiing time and evaluating airport taxiing performance. Chin. J. Aeronaut. 30 (2), 523–537.
Zhang, T., Ding, M., Wang, B., Chen, Q., 2016. Conflict-free time-based trajectory planning for aircraft taxi automation with refined taxiway modeling. J. Adv.
Transport. 50 (3), 326–347.
184

2018 TSPC Zou

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

2018 TSPC Zou

Загружено:

Авторское право:

Доступные форматы

Transportation Research Part C 95 (2018) 165–184

Contents lists available at ScienceDirect

Transportation Research Part C

A two-stage taxi scheduling strategy at airports with multiple

1. Introduction and literature review

2. Two-stage taxi scheduling strategy

3. Runway exit availability

Fig. 1. Flow diagram of the two-stage taxiing scheduling strategy.

running running running

schedule schedule schedule time

3.1. Deﬁnition of the runway exit availability

Puti = P (A) ∗P (B|A) (1)

PTsuti = ∫τ∈t fs ET (τ ) dτi

3.2. Calculation method of REA

4.2. Network deﬁnition

4.3. Flight set deﬁnition

4.4. Decision variables

Decision variables of the model are as the follows.

4.5. Auxiliary variables

4.6. Other notations of the model

Other notations and parameters of the model are as follows.

Fig. 3. Illustration of block-and-hold positions.

4.8.1. Flow conservation constraints

∑ Xvi, u, t − tvu + ∑ Bvi, u, t − 1 − T sep = ∑ Bvi, u, t − T sep + ∑ Xui , v, t

∑ Xvi, u, t − tvu + Ziu, t − 1 = Ziu, t + ∑ Xui , v, t

∀ t ∈ [tstart , tstart + T ], ∀ i ∈ Finsche , ∀ u ∈ N RF ∪ N P ∪ N AF

Xvi, u, t − tvu + Ziu, t − 1 = Ziu, t + ∑ Xui , vv, t

∀ t ∈ [tstart , tstart + T ], ∀ i ∈ Finsche ∩ D ,

4.8.2. Taxi-out constraints

∑ Zui , t ⩾ Tto, ∀ i∈ Finsche ∩ D

4.8.3. Taxiing safety constraints

4.8.4. Runway occupancy constraints

∀ i ∈ Ftoland and Rw (i) = urw

∀ i ∈ Fsched ∪ Finsche ∩ D and Rw (i) = urw

T deprepresents the forbidden time for a runway crossing after a take-oﬀ.

∑ t·Xui , v, t − ∑ t·Xuj, v, t ⩾ Ttype

∑ t·Xuj, v, t − ∑ t·Xui , v, t ⩾ Ttype

Sij + Sji = 1 (17)

where u = Rw (i) = Rw (j ) , v ∈ O (u) ∩ N AF

∀ i ∈ Ftoland, and Rw (i) = urw

where j ∈ Finsche ∩ D and Rw (j ) = urw , v ∈ O (urw ) ∩ N AF .

4.8.5. Taxiway occupancy constraints

∑ Bui , v, t + ∑ Xui , v, t − t sep ⩽1

where type (u, v ) ∈ {apron, ord} .

4.8.6. Re-taxiing avoidance constraints

4.8.7. Parking position and runway taxiing avoidance constraints

∀ i ∈ Finsche , u ∈ N P ∪ N RF and u ≠ udi

4.8.8. Route uniqueness constraints

∀ i ∈ Finsche , ∀ (v, u) ∈ E withstop

4.8.9. Boundary constraints

4.8.10. REA constraints

∑ ∑ ∑ Pvti Xui , v, t ⩾ Rmin

4.8.11. Existing plan constraints

∀ (u, v, t ) ∈ ExistRoute (u, v, t )

∀ (u, v, t ) ∈ ExistBlock (u, v, t )

∀ (u, v, t ) ∈ ExistRoute (u, v, t )

∀ (u, v, t ) ∈ ExistBlock (u, v, t )

4.9. Objective functions

v ∈ O (usi ), (usi , v ) ∈ E gate

5. Reﬁnement of the IP model

5.1. Heuristic for head-on conﬂict

5.2. Reﬁnement of taxi-out constraint

5.3. Refinement of take-off/take-off separation constraint