Professional Documents
Culture Documents
Collocation Methods To Minimum-Fuel Trajectory Problems With Required Time of Arrival in ATM
Collocation Methods To Minimum-Fuel Trajectory Problems With Required Time of Arrival in ATM
Javier García-Heras∗
Centro de Referencia de Investigación, Desarrollo e Innovación ATM — CRIDA A.I.E, 28022
Madrid, Spain
Manuel Soler†
Universidad Carlos III, 28911 Leganés, Spain
and
Francisco J. Sáez‡
Cranfield University, MK43 OAL Cranfield, United Kingdom
DOI: 10.2514/1.I010401
In the future air traffic management system, the trajectory becomes the fundamental element of a new set of operating
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
procedures collectively referred to as trajectory-based operations. Trajectory-based operations require the air traffic
management to introduce profound innovations to enable the envisioned changes. Some of these include collaborative
decision-making processes, better data and information management, and advanced decision support tools to aid
human operators. In particular, fast and accurate computation of optimal trajectories could certainly contribute to
enhance trajectory management within the future air traffic management. The trajectory optimization problem can be
solved using optimal control methods. In this paper, the existing methods for solving optimal control problems focusing
on direct collocation are discussed. In particular, pseudospectral collocation methods have shown to be numerically
more accurate and computationally much faster than other direct methods. A very relevant problem arising in air traffic
management is analyzed, that is the minimum-fuel trajectory with required time of arrival. Preliminary results illustrate
the advantages of using pseudoespectral collocation methods for trajectory optimization in air traffic management; for
the same order of numerical accuracy, they present computational times that are at least two orders of magnitude better
than other collocation methods. For similar computational times, pseudospectral collocation methods achieve a
numerical accuracy that is at least one order of magnitude better than other methods.
I. Introduction
N OWADAYS, aircraft fly following predefined routes for the horizontal profile and using flight level or isobar surfaces for the vertical profile
according to International Civil Aviation Organisation rules [1]. Moreover, operations rely on continuous tactical intervention from air traffic
control (ATC). As a result, flown trajectories are usually far from optimal, thus increasing operational cost and environmental impact (Sec. 3.5 of
[2]). In the envisioned air traffic management (ATM) system, the trajectory becomes the fundamental element of a new set of operating procedures
collectively referred to as trajectory-based operations (TBO) [3]. Improved capabilities in trajectory management (i.e., planning, sharing,
agreeing, and updating, including real-time trajectory revision and synchronization of air and ground-side systems), will result in enhanced ATM
performances in terms of capacity, efficiency, safety, and environmental impact.
The underlying idea behind TBO is the concept of business trajectory. The business trajectory is the trajectory that will meet best airline
business interests (e.g., minimize fuel burned). The TBO concept of operations and the notion of business trajectory will result in more efficient
four-dimensional (4-D) trajectories. TBO requires the ATM to introduce profound innovations to enable the envisioned changes. Some of these
include collaborative decision-making, better data and information management (sharing and usage), and decision support tools to aid human
operators.
The business trajectory can be conceived as an agreement between different stakeholders, i.e., the airline, the air navigation service providers
(ANSPs), the airports, and the network manager (Eurocontrol, in the case of Europe). The business trajectory follows a lifecycle as illustrated in
Fig. 1. In the strategic phase (long term), the airlines calculate their preferred trajectories (according to its business interests), resulting in the
business development trajectory (BDT). Months to days before the operation, all BDTs will be shared with the network manager, turning into the
so-called shared business trajectory (SBT). With these SBTs, the ANSPs can assess airspace sectorization, availability of routes, and their
allocation of resources. The network manager analyzes all SBTs (demand of flights) and ANSP resources (capacity), identifying possible capacity
and demand imbalances and consequently approving, delaying, or rerouting the different SBTs. Once all the stakeholders accept a proposed
solution, the SBT becomes the reference business trajectory (RBT), which the airline agrees to fly and the ANSPs/airports agree to provide with air
traffic services (such that safe operations are ensured). Last but not least, during flight execution, the RBT might be subject to unexpected events
(notice also that uncertainty is inherent in the system), e.g., convective weather, late/soon arrival of an aircraft that generates queuing management
near the airport, etc. Therefore, the RBT might be tactically revised, negotiated, and updated during execution.
Within this envisioned framework, and for the trajectory to follow the previously described lifecycle, a new paradigm of information sharing is
needed, involving stakeholders from across the whole European ATM network. The future ATM in Europe will rely on system-wide information
management (SWIM), which it is the most important enabler of the Single European Sky ATM Research (SESAR). It aims at assuring that the
right information will be available with the right quality to the right person at the right time. SWIM consists of standards, infrastructure, and
Received 3 June 2015; revision received 5 January 2016; accepted for publication 5 March 2016; published online 7 June 2016. Copyright © 2016 by Javier
García-Heras, Manuel Soler, and Francisco J. Saéz-Nieto. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission. Copies of this
paper may be made for personal and internal use, on condition that the copier pay the per-copy fee to the Copyright Clearance Center (CCC). All requests for copying
and permission to reprint should be submitted to CCC at www.copyright.com; employ the code 2327-3097 to initiate your request.
*ATM R&D Engineer, Avenida de Aragón, 402, Edificio Allende, 4a Planta, Madrid; jgheras@e-crida.enaire.es.
†
Assistant Professor, Bioengineering and Aerospace Engineering Department, Av. de la Universidad 30, Leganés; masolera@ing.uc3m.es.
‡
Reader, ATM/CNS, Centre for Aeronautics SATM, Bedfordshire, Cranfield; p.saeznieto@cranfield.ac.uk.
Article in Advance / 1
2 Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ
governance enabling the management of ATM information and its exchange between qualified parties via interoperable services. It covers all
ATM information, including any relevant information on the trajectory.
Trajectory exchange among various stakeholders prediction requires formal and structured flight descriptions to be managed by the decision
support tools without ambiguity. It is necessary to express trajectory information following fixed rules, so that it can be interchanged, negotiated,
and processed by automatic systems. This motivated the Aircraft Intent Description Language, which was introduced in [4] and further extended
in [5]. Moreover, to enhance real-time ground–air revision and update of the trajectory, fast and accurate computation of (preferably optimal)
trajectories are mandatory to enhance trajectory management within the future ATM. In this paper, we focus on the latter.
The trajectory optimization problem can be studied as an optimal control problem applied to a dynamic system in which the goal is to find the
trajectory and the corresponding control inputs that steer the state of the system between two configurations, satisfying a set of constraints on the
state and/or control variables while minimizing an objective functional.
Typically, optimal control problems are highly nonlinear, and it is very difficult to find analytical solutions even for the simplest cases. The
common practice is to use numerical methods. There are three fundamental approaches to numerically solving continuous-time optimal control
problems: dynamic programming (DP) methods, the optimality criteria of which in continuous time is based on the Hamilton–Jacobi–Bellman
partial differential equation [6]; indirect methods, which rely on the necessary conditions of optimality derived from the Pontryagin’s maximum
principle [7]; and direct methods, based on a finite-dimensional parameterization of the infinite-dimensional problem [8].
In the scope of commercial aircraft trajectory optimization using optimal control, all these methods have been explored. For instance, in [9], the
authors analyze the minimum-fuel trajectory of an aircraft at constant altitude and with fixed arrival time. Singular arc trajectories have been also
analyzed for climb performances [10] and complete flights [11]. Solutions provide smooth profiles, very suitable from an operational perspective.
However, their application is still very limited (e.g., they do not consider constrained arcs, i.e., situations in which any of the inequality constraints
is saturated, and finding singular arcs analytically to more involved problems might be intractable). Dynamic programming has been also used
very recently to solve a minimum-fuel vertical profile [12]; however, results showed to be computationally very intense, and thus the use of these
methods should be restricted to problems with a small number of variables. Also, authors in [13] use DP in a two-stage optimization algorithm to
obtain optimal aircraft horizontal paths and flight levels in the presence of winds.
Direct methods have been shown to be more suitable for solving more realistic commercial aircraft trajectory optimization problems in a highly
constrained environment as it is ATM. For instance, since the 1990s, Hermite–Legendre–Gauss–Lobatto (HLGL) direct collocation methods
[14,15] have been used to solve different commercial aircraft trajectory planning problems. To name a few, in [16,17], minimum-fuel vertical
profiles under ATM procedure constraints are analyzed. Also, minimum-fuel three-dimensional (3-D) profiles have been solved [18]. Moreover, a
recent approach has shown that binary decision variables can be combined into optimal control problems, resulting in a so-called multiphase
mixed-integer optimal control problem, which can be solved using HLGL collocation methods [19]. Therefore, these methods are very mature in
solving commercial aircraft trajectory optimization problems in an ATM environment. Nevertheless, typical computational times are relatively
high, making them more suitable for strategic flight planning rather than real-time trajectory updating.
Recently, the scientific community has shifted the attention to the so-called pseudospectral collocation methods. The basis of these methods
lays on spectral methods [20]. Some of its fundamental milestones include the formulation of the Legendre pseudospectal method [21] and the
Chebyshev method [22], later extended in [23]. Further developments include the Gauss pseudospectral method in [24] and the Radau
pseudospectral method in [25,26]. All these methods lay on different interpolation quadratures, namely, Legendre–Gauss–Lobatto and Legendre–
Gauss–Radau quadratures. Two recent, thorough publications on pseudospectral optimal control are [27,28]. Overall, their spectral convergence
makes them very attractive for real-time operations. Also, the development of two commercial off-the-shelf software packages (GPOPS [29] and
DIDO [30]) in a user-friendly and widely used interface has spread its use, particularly with applications to space mission problems and unmanned
air vehicles.
However, to the best knowledge of the authors, there are certain issues that remain unresolved. First, it is important to test whether these
pseudospectral collocation methods are also suitable for the class of trajectory planning problems arising in ATM (highly constrained). Second,
assuming they are, it would be interesting to determine their advantages and disadvantages with respect to HLGL collocation methods.
Therefore, the main contribution of this paper is to employ and compare pseudospectral collocation methods in a relevant type of problem
arising in ATM, i.e., that of finding the minimum-fuel trajectory with a required time of arrival (RTA). The aim is at illustrating how these methods
could be used to enhance accurate and fast information management within the envisioned ATM relying on the TBO concept. Note that, for the
TBO concept to deliver the envisioned benefits in terms of capacity, efficiency, safety, and environmental impact, planned trajectories will be
required to fulfill time adherence constraints (i.e., aircraft will be required to accomplish with an RTA over certain prescribed waypoints along its
trajectory). However, deviations between the actual and the planned trajectories will be unavoidable (due, for instance, to inherent uncertainty).
Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ 3
Thus, the need to recalculate (revise and update) efficient trajectories that meet time constraints in real time is paramount. A description of the
problem is given in Sec. VI.
Two examples are herein solved and discussed: first, a proof-of-concept example aimed at comparing the previously discussed collocation
methods and illustrate accuracy and computational efficiency; second, a real trajectory through flight data recorder (FDR) information of a Cairo–
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
Madrid flight, considering cruise and descent phases. BADA 4.1 [31] is used to characterize the aircraft performance model.
The paper is structured as follows. First, in Sec. II, we state the optimal control problem. In Sec. III, we present numerical methods for solving
optimal control problems. Hermite–Legendre–Gauss–Lobatto collocation methods are described in Sec. IV. Pseudospectral collocation methods
are presented in Sec. V. The two illustrative examples are discussed in Sec. VI. Finally, some conclusions are drawn in Sec. VII.
Subject to∶
xt
_ fxt; ut; p; dynamic equations;
0 gxt; ut; p; algebraic equations;
xt0 x0 ; initial boundary conditions;
ψxtf 0; terminal boundary conditions;
ϕl ≤ ϕxt; ut; p ≤ ϕu ; path constraints
The variable t ∈ t0 ; tf ⊂ R represents time, and p ∈ Rnp is a vector of parameters. Notice that the initial time t0 is fixed, and the final time tf
might be fixed or left undetermined. xt: t0 ; tf ↦ Rnx represents the state variables. ut: t0 ; tf ↦ Rnu represents the control functions, also
referred to as control inputs, assumed to be measurable. The objective function J: t0 ; tf × Rnx × Rnu × Rnp → R is given in Bolza form. It is
expressed as the sum of the Mayer term Etf ; xtf and the Lagrange term
Zt
f
Lxt; ut; p dt
t0
Functions E: t0 ; tf × Rnx → R and L: Rnx × Rnu × Rnp → R are assumed to be twice differentiable. The system is a Differential Algebraic
Equation system in which the right-hand-side function of the differential set of equations f: Rnx × Rnu × Rnp → Rnx is assumed to be piecewise
Lipschitz continuous, and the derivative of the algebraic right-hand-side function g: Rnx × Rnu × Rnp → Rnz with respect to z is assumed to be
regular. x0 ∈ Rnx represents the vector of initial conditions given at the initial time t0 , and the function ψ: Rnx → Rnq provides the terminal
conditions at the final time, and it is assumed to be twice differentiable. The system must satisfy algebraic path constraints given by the function
ϕ: Rnx × Rnu × Rnp → Rnϕ with lower bound ϕl ∈ Rnϕ and upper bound ϕu ∈ Rnϕ . The function ϕ is assumed to be twice differentiable.
3) Direct methods: They can be applied without deriving the necessary condition of optimality. Direct methods are based on a finite-
dimensional parameterization of the infinite-dimensional problem. The finite-dimensional problem is typically solved using an optimization
method, such as nonlinear programming (NLP) techniques. NLP problems can be solved to local optimality relying on the so-called Karush–
Kuhn–Tucker (KKT) conditions, which give first-order conditions of optimality. These conditions were first derived by Karush in 1939 [33], and
some years later, in 1951, independently by Kuhn and Tucker [34].
B. Indirect Methods
Indirect methods rely on Pontryagin’s maximum principle [7]. Typically, the optimal control problem is turned into a two-point boundary-value
problem containing the same mathematical information as the original one by means of necessary conditions of optimality. Then, the boundary-
value problem is discretized by some numerical technique to get a solution. Thus, indirect methods follow a “first optimize, then discretize”
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
scheme. Numerical techniques for solving this two-point boundary-value problem can be classified as gradient methods [36], indirect shooting
and indirect multiple shooting [37], and indirect collocation [38].
The practical drawbacks of indirect methods are as follows [8].
1) Proper formulations of the necessary conditions of optimality in a numerically suitable way must be derived. Because this formulation is
rather complicated, significant knowledge and experience in optimal control is required by the user of an indirect method.
2) To handle active constraints properly, their switching structure must be guessed.
3) Suitable initial guesses of the state variables and, with special relevance, to the adjoint variables must be provided to start the iterative method.
State variables have physical meaning, but adjoint variables do not, so that giving a proper initial guess might be hard, and a nonproper one usually
leads to nonoptimal solutions. Even with a reasonable guess for the adjoint variables, the numerical solution of the adjoint equations can be ill-
conditioned.
4) Changes in the problem formulation (e.g., by a modification of the model equations) imply formulating again the optimality conditions of the
problem.
5) Finally, model functions with low differentiability properties are difficult to tackle with indirect approaches.
Because of these practical difficulties, indirect methods are not suitable to solve highly constrained trajectory planning problems. In fact, rather
than indirect approaches, direct methods have been extensively used for solving aerospace trajectory optimization problems in spite of the fact that
they present less accuracy than indirect methods [39]. Two comprehensive surveys analyzing direct and indirect methods for trajectory
optimization are [40,41].
C. Direct Methods
The so-called direct methods do not use the first-order necessary conditions of the continuous optimal control problem. They convert
the infinite-dimensional problem into a problem with a finite set of variables and then solve the finite-dimensional problem using optimiza-
tion methods. Direct methods thus follow a “first discretize, then optimize” approach. A typical strategy is to convert the infinite problem into
an NLP problem, which is solved using mathematical programming techniques [42]. The most important direct numerical methods are
direct shooting [43], direct multiple shooting [44], and direct collocation [45]. A good reference on the practical importance of direct methods
is [8].
The direct single shooting method has been broadly used because it allows optimal control problems to be easily converted into an NLP problem
with a small number of variables even for very large problems. In single shooting, only initial guesses for the control NLP variables are required. In
contrast, it is very sensitive to small perturbations on the initial condition. The direct multiple shooting method reduces some of the problems that
single shooting has. However, the multiple shooting approach increases the size of the problem because additional variables and constraints have
to be included. When the problem includes inequality constraints, there is the additional disadvantage that the sequence of unconstrained and
constrained arcs has to be specified in advance. The direct collocation methods do not suffer from most of the drawbacks mentioned previously,
and therefore they are the most suitable for aerospace trajectory optimization problems [8,40,41].
A taxonomy of optimal control methods for trajectory optimization is given in Fig. 3. Notice that this taxonomy is not necessarily exhaustive.
Direct collocation methods: Collocation methods enforce the dynamic equations through quadrature rules or interpolation [14,45]. A suitable
interpolating function, or interpolant, is chosen such that it passes through the state values and maintains the state derivatives at the nodes spanning
one interval, or subinterval, of time. The interpolant is then evaluated at points between nodes, called collocation points. At each collocation point,
a constraint equating the interpolant derivative to the state derivative function is introduced to ensure that the equations of motion are
approximately satisfied across the entire interval of time [15].
Collocation methods are characterized by the interpolating function and by the nodes and collocation points they use. One of the simplest
methods of collocation is the Hermite–Simpson collocation method [14,46]. In this method, a third-order Hermite interpolating polynomial is
used locally within the entire sequence of time subintervals, each solved at the end points of a subinterval and collocated at the midpoint. When
arranged appropriately, the expression for the collocation constraint corresponds to the Simpson integration rule. A generalization of the method is
obtained using the nth-order Hermite interpolating polynomial and choosing the nodes and collocation points from a set of Legendre–Gauss–
Lobatto points defined within the time subintervals. These choices give rise to the Hermite–Legendre–Gauss–Lobatto (HLGL) collocation
method [15].
Another relevant family of collocation methods is the so-called pseudospectral, in which state and control equations are parameterized using
global polynomials, and the differential-algebraic equations are evaluated in the nodes or collocation points derived from a Gaussian quadrature
[27]. The global polynomials are the Legendre and Chebyschev polynomials. In the Legendre pseudospectral methods, the quadrature nodes have
to be computed numerically as zeros of the Lagrange polynomials. Otherwise, in the Chebyschev pseudospectral methods, the quadrature nodes
come from explicit formulas. These methods can be further classified attending at the class of collocation points, typically Gauss, Gauss–Radau,
or Gauss–Lobatto collocation points [20].
Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ 5
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
The reader is referred to [23,27,28] and references therein for recent and comprehensive reviews of pseudospectal methods for optimal control.
In each time subinterval, ti ; ti1 , xi xti , and ui uti are the values of the state and controls variables at the grid point ti , respectively. We
can also define hi tf − t0 ∕N, which is called the integration step size for step i, i 0; : : : ; N − 1. Thus, the variables of the corresponding
nonlinear programming (NLP) problem will be
fx0 ; u0 ; x1 ; u1 ; : : : ; xN ; uN g
together with other independent variables according to the integration rule and the control scheme employed. They include, in general, the values
of both state and control at the collocation points (e.g., the center point of the subinterval xi;C , ui;C and the controls at the points ui;a , ui;b ).
Then, the ordinary differential equation is replaced by a finite number of equality constraints called defect equations, which can be written in the
most general form as
Each integration scheme leads to a different formulation of this set of transcribed constraints, as explained in the following sections.
− ;
5 5
1 h
xi;C xi xi1 i fxi − fxi1 (4)
2 8
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
where xi;C is a discrete approximation for xt at ti;C ti ti1 ∕2 and i 0; : : : ; N − 1. Simpson’s system constraint is then formulated using
xi;C to evaluate the system equation, resulting in a discrete value at center point of the subinterval fi;C fxi;C . Then, by enforcing fi;C to be
equal to the first time derivative of the Hermite cubic interpolant polynomial at the center point of the subinterval (i.e., x_ i;C fi;C ), one defect
equation per subinterval is generated:
hi
cSi xi ; xi1 xi − xi1 fxi 4fxi;C fxi1 0 (5)
6
and
r
3 hi
ti;b ti;C
72
yielding
with i 0; : : : ; N − 1.
The fifth-degree Gauss–Lobatto system of constraints is formulated to approximate the differential equation at the two internal collocation
points ti;a and ti;b , resulting in the discrete values fi;a fxi;a and fi;b fxi;b , respectively. Then, fi;a and fi;b are equated with the first
derivative in time of the fifth-degree interpolant polynomial evaluated at the same collocation points, x_ i;a and x_ i;b , respectively.
8 Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ
with i 0; : : : ; N − 1.
the center ui;C ; and free control interpolation scheme, by simply adding ui;C as a new independent variable in each subinterval. In the fourth and
fifth-degree, the typical control interpolation scheme is the free one.
tf t0 tf − t0
t τ (10)
2 2
In such a form, the optimal control problem (OCP) can now be rewritten taking the following into consideration.
1) The initial time and the final time are τ −1 and τ 1, respectively.
2) The dynamic equations in OCP are
tf − t0
xτ
_ fxτ; uτ; p
2
3) The path constraint and algebraic equations in OCP are ϕxτ; uτ; p ≤ 0 and 0 gxτ; uτ; p, respectively.
XN
Wτ
xτ xk ϕk τ (11)
k0
Wτ k
where k is an index denoting the node for the global Lagrange interpolating polynomial of order N, W denotes a positive weight function, xk
denotes the state value for node k, and ϕk τ denotes the general expression for a Lagrange interpolation polynomial of degree N that satisfies
ϕk τi 0, i ≠ k and ϕk τk 1, i.e.,
Y
N
τ − τi
ϕk τ (12)
τ
i0;i≠k k
− τi
Notice that i is another index denoting the node for the global Lagrange interpolating polynomial of order N.
Collocation points in pseudospectral methods differ depending on the polynomial approximation used. In the Legendre pseudospectral
methods, they correspond to the zeros of a specific Lagrange polynomials expression; otherwise, in the Chebyschev pseudospectral methods, the
collocation points have an explicit formula.
1. Legendre Polynomials
The Legendre polynomials Lk x, k 0; 1; : : : ; N are the eigenfunctions of the singular Sturm–Liouville problem denoted by
2k 1 k
Lk1 τ τL τ − L τ (14)
k1 k k 1 k−1
2. Chebyshev Polynomials
The Chebyshev polynomials T k τ, k 0; 1; : : : ; N are the eigenfunction of the singular Sturm–Liouville problem denoted by
p 0 k2
1 − τ2 T k0 τ p T k τ 0 (15)
1 − τ2
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
Applying the trigonometric relation cosk 1θ cosk − 1θ 2 cos θ coskθ, the solution to the problem results in the following
recursive relation:
with T 0 1 and T 1 τ.
As in the previous case, Chebyshev pseudospectral methods use three different types of collocation points among those in the family of the
Chebyshev–Gauss points [20], namely:
1) Chebyshev–Gauss (CG): neither the initial point (τ −1) nor the end point (τ 1) is part of the collocation points:
2k 1π
τk cos ; k 0; : : : ; N (18)
2N 2
2πk
τk cos ; k 0; : : : ; N (19)
2N 1
πk
τk cos ; k 0; : : : ; N (20)
N
Figure 7 illustrates the nonuniform grids of collocation points with eight samples for the aforementioned pseudospectral collocation methods.
The success or failure of a pseudospectral method depends on the type of optimal control problem (e.g., type of boundary conditions, the
appropriate selection of time affine transformation and weight function, and the selection of the collocation points). This fact has been deeply
studied by Fahroo and Ross (e.g., in [48]). Table 1 summarizes the suitable choices within the previously listed items. According to the type of
problems arising in the scope of commercial aircraft trajectory optimization, i.e., optimal control problems with any kind of initial and final
boundary conditions (fixed or free) and finite time horizon, the sequel will focus on both Legendre and Chebyschev pseudospectral methods based
on Gauss–Lobatto collocation points.
dt XN XN
xτ
_ j xt
_ j ≈ xk ϕ_ k τj Djk xk (21)
dτ k0 k0
where Djk denotes the first derivative matrix of a Lagrange polynomial interpolation of order N (ϕ_ k τt).
In the case of LGL points, the Lagrange polynomial interpolation of order N is given by
LG
LGR
LGL
CG
CGR
CGL
−1 −0.5 0 0.5 1
8
> LN τj 1
>
> ; j ≠ k;
>
> L N τk τj − τk
>
>
< N 1N
Djk − ; j k 0; (23)
>
> 4
> N 1N
>
>
> ; j k N;
>
: 4
0; otherwise
In the case of CGL points, the Lagrange polynomial interpolation of order N is given by
where
2; k 0; N;
ck (25)
1; k 1; : : : ; N − 1
As mentioned before, the CGL collocation nodes are defined from 1 to −1. Therefore, a symmetric transformation is needed to make the initial
point −1 and the final point 1. As a result, both collocation nodes and the first derivative matrix have opposite sign after such transformation (i.e.,
τ^ −τ and D^ jk −Djk ).
and the planned trajectories will be unavoidable (due, for instance, inherent uncertainty). Thus, the need to recalculate (revise and update) efficient
trajectories that meet time constraints in real time is paramount.
1. Equations of Motion
For trajectory planning purposes, it is common to consider a three-degree-of-freedom (3-DOF) dynamic model that describes the point variable-
mass motion of the aircraft over a spherical flat-Earth model. We consider a symmetric flight, that is, we assume that there is no sideslip and that all
forces lie in the plane of symmetry of the aircraft. The resulting sets of differential equations that govern the motion of the aircraft are as follows:
The states are the longitudinal position x, the lateral position y, the altitude h, the true airspeed V, the heading angle χ, the flight-path angle γ, and
the mass of the aircraft (m). A parabolic drag polar and an International Standard Atmosphere (ISA) model are assumed. The coefficient of lift (CL ) is
a known function of the angle of attack and the Mach number. The engine thrust T, the coefficient of lift, and the bank angle μ are the control inputs of
the 3-DOF aircraft model. wx and wy correspond to the eastbound and northbound wind components, respectively. We use the Eurocontrol’s BADA 4
family as aircraft performance model [31].
BADA 4 family models can provide accurate modeling of aircraft performances over the complete flight envelope, including transonic regimes
(in other words, they take compressibility effects into consideration). It is also aimed at supporting complex aircraft operations in different flight
phases such as those arising from ATM operational constraints. A good analysis of BADA 4 family capabilities can be checked in [50].
Note that the maximum dynamic altitude increases as fuel is burned and that several flight envelop constraints are nonconvex.
2 3 2 Tt−DVt;C t 3
V mt
L
d4 5 4 5
xe Vt (28)
dt
m ηVt; Tt
2. Numerical Results
It has been proven that the KKT NLP necessary conditions approach the optimal control necessary conditions of optimality as the number of
variables grows. Indeed, at the solution of the NLP problem, the Lagrange multipliers can be interpreted as discrete approximations to the optimal
control adjoint variables [8]. Therefore, we first solve the trajectory optimization problem using the aforementioned direct collocation methods
for an increasing number of sample points. Results for the Hermite–Simpson, the fifth-degree, the LGL pseudospectral, and the CGL
pseudospectral methods are presented in Tables 3–6, respectively. Because of nonconvexities, the solution obtained can only be claimed as locally
optimal. IPOPT solver [49] has been used with its default settings. The computational times correspond to a Mac OS X 2.56 GHz laptop with
4 GB RAM.
Figures A1 and A2 illustrate the solutions for the true airspeed and the thrust of the aircraft. These figures appear in Appendix A for the sake of
clarity in the exposition.
For the HLGL collocation methods (i.e., Hermite–Simpson and fifth-degree), it can be observed that the objective function is very sensitive to
the number of samples. This is especially noticeable for a low number of samples. Notice that, by increasing the number of samples, the solution
converges to a value of minimum-fuel consumption (objective function), and so do velocity and thrust. Moreover, the instabilities due to the
boundary conditions are softened, as illustrated in Fig. A1. These methods need a considerable amount of sample points to achieve the optimal
solution, and thus the employed computational time is higher than in pseudospectral methods. Note that the grid of samples is uniformly
distributed, and thus, near the bounds of the interval, the function approximation error is remarkably greater than in the center points of the interval.
This is called the Ruge phenomenon.
On the contrary, the pseudospectral collocation methods (i.e., LGL and CGL) do not show much sensitivity to the number of samples nor to
instabilities near the boundaries. Note that pseudospectral methods employ a nonhomogeneous grid of sample points, locating more sample points
near the boundaries (where typically high-frequency dynamics are found); see Fig. A2.
We turn now the discussion to compare the performances in terms of accuracy and computational time of the different direct collocation
methods. The solutions to the different methods seem to converge toward a fuel consumption value of 2263 kg. Let us assume this value as the
baseline for comparison. Notice that the accuracy value has been calculated through the following expression:
Table 7 presents the numerical performances of the different collocation methods for number of samples of 40–50. We observe a much higher
accuracy for the pseudospectral methods at a similar computational time (10 times better when compared to the fifth-degree collocation method).
If one requires to achieve a particular accuracy, for instance of order 10−4, the amount of fuel burned has to be lower or equal than 2265.3 kg.
Table 8 presents these results. Notice that the Hermite–Simpson method would need 400 samples and 2.33 s of computational time, whereas the
fifth-degree would need 400 samples and 33.69 s of computational time.
440
35000
420
30000
400
25000
380
20000 360
340
15000
320
10000
300
Fig. 9 State variable Vt and control input Tt for Hermite–Simpson, fifth-degree, CGL, and LGL simulations. N is the number of sample points.
All methods’ solutions are depicted in Fig. 9 for the sake of comparison. The first thing that can be observed is that all four direct collocation
methods follow very similar patterns in terms of speed and throttle level position profile. In terms of aircraft performances, aircraft weight
decreases with time because fuel is burned. Therefore, according to aircraft dynamics in Eq. (28), velocity also decreases, and so does the throttle
50
45
40
Latitude (°)
35
30
25
20
−10 −5 0 5 10 15 20 25 30
Longitude (°)
Fig. 11 Cairo–Madrid horizontal profile. Triangles correspond to waypoints.
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
level position (the control input to speed up/slow down). It is important to observe that, near the final fix, the throttle level goes to its idle value, and
the velocity hits its lower bound near the stall.
C. Realistic Example
The focus herein is on the minimum-fuel trajectory planning problem for a realistic flight. We analyze a 4-D trajectory with RTA at a given final
fix based on FDR data of an Airbus 319 real flight flying from Cairo to Madrid. The RTA is adjusted based on the flight time (base on FDR data)
over the Initial Approach Fix (IAF) called ASBIN. Please refer to Fig. 10 for further information. Real wind has been approximated fitting the
obtained from FDR by using the following n-order polynomial:
X
n
w^ ai · si
i0
where s is the distance covered by the aircraft, w^ represents the along-track wind, and ai correspond to the coefficients of the polynomial that
approximate real values (we employ least squares).
Dynamic equations are based on those in Eq. (27), however simplified. The lateral path is enforced to be tracked, and thus the aircraft is not
allow to turn (i.e., back angle is set to zero and the heading established to follow the preestablished sequence of waypoints). The route is presented
in Fig. 11.
the final condition is constituted based on the pair (latitude, longitude) of absin, together with its overflight altitude and time (i.e., RTA) based on
FDR data. The objective functional is to minimize the total amount of fuel consumption.
2. Numerical Results
As in the previous example, we solve the optimal control problem (OCP) using the four collocation methods (i.e., Hermite–Simpson, fifth-
degree, LGL, and CGL). Results are presented in Tables 10–13. According to results, one can readily see that the solution converges to a value of
around 9650.70 kg, which we consider as theoretically optimal for the numerical analysis. This means that, if one would require to achieve a
numerical accuracy of order 10e − 4, the objective function value (amount of fuel burned in this problem) would have to be lower than or equal to
9660.3 kg. Numerical accuracies and computational times for the different methods are presented in Tables 14 and 15.
Horizontal route, altitude profile, velocity profile, mass profile, thrust profile, and flight path angle profile are presented in Fig. 12, where: FDR-
ATS corresponds to the FDR data simulated using the calibrated aircraft trajectory simulator, and CGL-ATS corresponds to the solution the OCP
Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ 17
using a CGL collocation method and simulated using the calibrated aircraft trajectory simulator. The solution to the problem results in a
continuous-climb cruise followed by continuous-descent approach. For the sake of comparison, Fig. 12 also includes the values extracted from
FDR. To fairly compare both trajectories, an aircraft trajectory simulator (TS) has been built using Matlab and Simulink. In this manner, both
trajectories can be simulated with the same models and parameters (i.e., aircraft model based on BADA, ISA atmosphere, wind conditions, etc.).
For more information on aircraft TS, the reader is referred to Appendix B.
Analyzing FDR data in Fig. 12, it can be observed that the aircraft cruises at constant altitude and nearly at constant true airspeed. Note,
however, that true airspeed presents a step down, probably due to an ATC advisory to avoid a potential hazard (e.g., a conflict with other aircraft or
a convective region).
240
12000
10000 220
8000 200
6000 180
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
4000 160
2000 140
FDR−ATS FDR−ATS
CGL−ATS N=160 CGL−ATS N=160
0 120
0 2000 4000 6000 8000 10000 12000 14000 16000 0 2000 4000 6000 8000 10000 12000 14000 16000
6.4
40000
6.2
30000
6
20000
5.8 10000
5.6 0
FDR−ATS FDR−ATS
CGL−ATS N=160 CGL−ATS N=160
5.4 −10000
0 2000 4000 6000 8000 10000 12000 14000 16000 0 2000 4000 6000 8000 10000 12000 14000 16000
15
1
10
0
5
−1 0
−5
−2
−10
−3
−15
FDR−ATS FDR−ATS
CGL−ATS N=160 CGL−ATS N=160
−4 −20
0 2000 4000 6000 8000 10000 12000 14000 16000 0 2000 4000 6000 8000 10000 12000 14000 16000
The optimal trajectory shows that aircraft’s best interest for saving fuel is to fly at the maximum operational altitude hu t, which increases as
fuel is burned. This is indeed a continuous-climb cruise. Moreover, the top of descent (TOD) is located 300 s before the one for the FDR trajectory,
and therefore the optimal strategy is to start descent earlier at a reduced speed and typically with idle thrust. The continuous descent from TOD
toward runway head (or any intermediate waypoint before final approach) is referred to as continuous-descent approach.
FDR trajectory’s total fuel consumption (including cruise, descent, and landing portions of the flight) amounts 10,130.00 kg. On the contrary,
the solution to (OCP) provides 9893.33 kg of fuel consumption, resulting in 236.87 kg less of fuel burned than FDR’s trajectory. Recall that the
consumptions have been both computed with the trajectory simulator in Appendix B. Please refer to Fig. 12.
VII. Conclusions
The SESAR Master Plan proposes roadmaps of essential operational changes shifting from the current operational concept (based on a
structured airspace) to the so-called time-based operations (short term), trajectory-based operational concept (medium term), and further to the so-
called performance-based operational concept (long term). The capabilities to revise and update trajectories during execution (real time) is
paramount to the operational concept proposed by SESAR. Accurate, yet expedited calculation of the trajectory would enhance trajectory
synchronization through integrated automation support, while simultaneously meeting the goals of SESAR. Human operators will remain at the
core of the system (overall system managers and supervisors) using automated systems with the required degree of integrity and redundancy.
Based on the results and discussions presented in this manuscript, pseudospectral collocation methods have shown to behave substantially better
than Hermite–Legendre–Gauss–Lobatto collocation methods in terms of accuracy and computational time in a particular problem arising in ATM
under the aforementioned SESAR’s operational concept, that of meeting a prescribed fix at a prescribed time. For the same order of numerical
accuracy, pseudospectral collocation methods present computational times that are at least two orders of magnitude better; for similar
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
computational times, pseudospectral collocation methods achieve a numerical accuracy that is at least one order of magnitude better. The bottom
line is that, by using pseudoespectral methods, we can obtain fast and accurate optimal trajectory computation, which in turn could enhance
efficient tactical modifications on the trajectory, facilitating synchronization between aircraft and ground systems.
440
35000
420
30000
400
25000
380
20000 360
340
15000
320
10000
N=800
300 N=800
N=400 N=400
5000 N=200 N=200
N=100 280 N=100
N=50 N=50
N=25 N=25
0 260
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
440
35000
420
30000
400
25000
380
20000 360
340
15000
320
10000
300
N=1600 N=1600
5000 N=800 N=800
N=400 280 N=400
N=200 N=200
N=100 N=100
0 260
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
c) Hermite-simpson d) Hermite-simpson
Fig. A1 State variable Vt and control input Tt for different number of sample points.
Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ 19
440
45000
420
40000
400
35000
30000 380
25000 360
20000 340
15000 320
N=140 N=140
N=120 N=120
N=100 N=100
10000 300 N=80
N=80
N=60 N=60
N=50 N=50
5000 N=40 280 N=40
N=30 N=30
N=15 N=15
0 260
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
a) Chebychev-gauss-lobatto. b) Chebychev-gauss-lobatto.
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
440
40000
420
35000
400
30000
380
25000
360
20000
340
15000
320
N=140 N=140
N=120 N=120
10000 N=100 N=100
300 N=80
N=80
N=60 N=60
N=50 N=50
5000 280 N=40
N=40
N=30 N=30
N=15 N=15
0 260
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
c) Legendre-gauss-lobatto d) Legendre-gauss-lobatto
Fig. A2 State variable Vt and control input Tt for different number of sample points.
model). In the 3-DOF PMM aircraft kinematic and dynamic equation, the block system of Eq. (27) is implemented without considering winds.
This block requires aircraft performances, power plant information, and aircraft initial conditions. Finally, in the geographic reference block,
aircraft variables are evaluated referred to a geographic reference; wind values are included, and x, y are converted into longitude and latitude
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
12000
60 FDR
FDR−ATS−WC
FDR−ATS 10000
55 CGL−ATS
50
8000
45
°
40 6000
35
4000
30
25 2000 FDR
FDR−ATS−WC
FDR−ATS
20 CGL N=160
CGL−ATS N=160
−15 −10 −5 0 5 10 15 20 25 30 35 0
° 0 2000 4000 6000 8000 10000 12000 14000 16000
220 6.4
200 6.2
180 6
160 5.8
2
50000
1
40000
0
30000
−1
20000
10000 −2
0 −3
FDR−ATS−WC FDR−ATS−WC
FDR−ATS FDR−ATS
CGL N=160 CGL N=160
CGL−ATS N=160 CGL−ATS N=160
−10000 −4
0 2000 4000 6000 8000 10000 12000 14000 16000 0 2000 4000 6000 8000 10000 12000 14000 16000
20
15 −5
FDR
−10 FDR−ATS−WC
10 FDR−ATS
−15 CGL N=160
CGL−ATS N=160
5 −20
−25
0
−30
−5
−35
−10 −40
−45
FDR
−15
FDR−ATS−WC
−50
FDR−ATS
CGL−ATS N=160
−20 −55
0 2000 4000 6000 8000 10000 12000 14000 16000 0 0.5 1 1.5 2 2.5 3 3.5
6
x 10
models, ISA, and wind as a function of the distance covered by the aircraft. It is straightforward to see that the optimal control problem modeling
does not capture the FDR real behavior. Thus, for the aircraft TS to fairly reproduce the FDR trajectory, the former needs to be calibrated. In
particular, specific fuel consumption has been corrected due to discrepancies between real consumption recorded in FDR and BADA models. Real
wind has been approximated fitting the obtained from FDR by using the following polynomial:
w^ 6.798 · 10−56 · s9 4.394 · 10−48 · s8 − 5.164 · 10−41 · s7 2.194 · 10−34 · s6 − 4.435 · 10−28 · s5
4.343 · 10−22 · s4 − 1.653 · 10−16 · s3 − 1.691 · 10−11 · s2 1.547 · 10−5 · s − 35.28 (B1)
where s is the distance covered by the aircraft in meters.
Then, all the observed differences between FDR trajectory and aircraft TS variables are corrected, as can be observed in Fig. B3, where FDR
denotes the real FDR data. FDR-ATS-WC corresponds to the FDR data simulated using the noncalibrated aircraft trajectory simulator, FDR-ATS
corresponds to the FDR data simulated using the calibrated aircraft trajectory simulator, and CGL-ATS corresponds to the solution the OCP using
a CGL collocation method and simulated using the calibrated aircraft trajectory simulator. The result of this calibration process gives the calibrated
aircraft TS. For the simulation step, one just needs to independently (for both FDR data and optimal solution to the OCP) introduce the reference
inputs to the calibrated aircraft TS. Finally, both trajectories can be compared in Fig. B4, where FDR denotes the real FDR data. FDR-ATS-WC
corresponds to the FDR data simulated using the non-calibrated aircraft trajectory simulator, FDR-ATS corresponds to the FDR data simulated
using the calibrated aircraft trajectory simulator, and CGL-ATS corresponds to the solution the OCP using a CGL collocation method and
simulated using the calibrated aircraft trajectory simulator, as shown in Sec. VI.C.2.
References
[1] “Rules of the Air,” 10th ed., International Civil Aviation Organization, TR 42, Canada, 2005.
[2] “Performance Review Report: An Assessment of Air Traffic Management in Europe During the Calendar Year 2011: Eurocontrol PRR 2011,” Eurocontrol,
Performance Review Commission, Belgium, May 2012, http://www.eurocontrol.int/prc [retrieved Jan. 2012].
[3] “Single European Sky ATM Research Master Plan,” 2nd ed., Single European Sky ATM Research, Belgium, 2012, http://ec.europa.eu/transport/modes/air/
sesar [retrieved May 2013].
[4] López-Leonés, J., “The Aircraft Intent Description Language,” Ph.D. Thesis, Univ. of Glasgow, Glasgow, Scotland, 2007.
[5] Besada, J., Frontera, G., Crespo, J., Casado, E., and López-Leonés, J., “Automated Aircraft Trajectory Prediction Based on a Formal Intent-Related Language
Processing,” IEEE Transactions on Intelligent Transportation Systems, Vol. 14, No. 3, 2013, pp. 1067–1082.
22 Article in Advance / GARCÍA-HERAS, SOLER, AND SÁEZ
[6] Bellman, R., Dynamic Programming, Princeton Univ. Press, 1962, pp. 1–365.
[7] Pontryagin, L. S., Boltyanskii, V., Gamkrelidze, R. V., and Mishchenko, E. F., The Mathematical Theory of Optimal Processes, Interscience Publ., Glasgow,
1962, pp. 1–360.
[8] Betts, J. T., Practical Methods for Optimal Control and Estimation Using Nonlinear Programming, Advances in Design and Control, Soc. for Industrial and
Applied Mathematics, Philadelphia, 2010, Chap. 4.3.
[9] Franco, A., Rivas, D., and Valenzuela, A., “Minimum-Fuel Cruise at Constant Altitude with Fixed Arrival Time,” Journal of Guidance, Control, and
Dynamics, Vol. 33, No. 1, 2010, pp. 280–285.
[10] Nguyen, N., “Singular Arc Time-Optimal Climb Trajectory of Aircraft in a Two-Dimensional Wind Field,” AIAA Guidance, Navigation, and Control
Conference and Exhibit, AIAA Paper 2006-6598, Aug. 2006.
[11] Franco, A., and Rivas, D., “Optimization of Multiphase Aircraft Trajectories Using Hybrid Optimal Control,” Journal of Guidance, Control, and Dynamics,
Vol. 38, No. 3, 2015, pp. 452–467.
[12] Miyazawa, Y., Wickramashinghe, N. K., Harada, A., and Miyamoto, Y., “Dynamic Programming Application to Airliner Four Dimensional Optimal Flight
Trajectory,” Guidance, Navigation, and Control (GNC) Conference, AIAA Paper 2013-4969, 2013.
[13] Hok, K., Ng, B. S., and Grabbe, S., “Optimizing Aircraft Trajectories with Multiple Cruise Altitudes in the Presence of Winds,” Journal of Aerospace
Information Systems, Vol. 11, No. 1, 2014, pp. 35–47.
[14] Hargraves, C. R., and Paris, S. W., “Direct Trajectory Optimization Using Nonlinear Programming and Collocation,” Journal of Guidance, Control, and
Dynamics, Vol. 10, No. 4, 1987, pp. 338–342.
[15] Williams, P., “Hermite–Legendre–Gauss–Lobatto Direct Transcription Methods in Trajectory Optimization,” Advances in the Astronautical Sciences,
Vol. 120, No. 1, 2005, pp. 465–484.
[16] Betts, J. T., and Cramer, E. J., “Application of Direct Transcription to Commercial Aircraft Trajectory Optimization,” Journal of Guidance, Control, and
Dynamics, Vol. 18, No. 1, 1995, pp. 151–159.
[17] Soler, M., Zapata, D., Olivares, A., and Staffetti, E., “Framework for Aircraft 4D Trajectory Planning Towards an Efficient Air Traffic Management,” Journal
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401
[48] Fahroo, F., and Ross, I., “On Discrete-Time Optimality Conditions for Pseudospectral Methods,” AIAA/AAS Astrodynamics Specialist Conference and
Exhibit, AIAA Paper 2006-6304, Aug. 2006.
[49] Wachter, A., and Biegler, L. T., “On the Implementation of an Interior-Point Filter Line-Search Algorithm for Large-Scale Nonlinear Programming,”
Mathematical programming, Vol. 106, No. 1, 2006, pp. 25–57.
[50] Poles, D., Nuic, A., and Mouillet, V., “Advanced Aircraft Performance Modeling for ATM: Analysis of BADA Model Capabilities,” Proceedings of the 2010
IEEE/AIAA 29th Digital Avionics Systems Conference (DASC), IEEE Publ., Piscataway, NJ, 2010, pp. 1.D.1-1–1.D.1-14.
H. Balakrishnan
Associate Editor
Downloaded by MONASH UNIVERSITY on June 23, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.I010401