View
0
Download
0
Category
Preview:
Citation preview
FINITE ELEMENTS IN TIME and GAUSS PSEUDO-SPECTRAL METHODS
Massimiliano Vasile
Deptartment of Mechanical & Aerospace Engineering
University of Strathclyde, Glasgow, UK
JAXA
9/2/2012
Direct Transcription with GPSM
Some Examples
Finite Perturbative
Elements
Direct Transcription
with FET
Basics of Nonlinear
Programming
OPTIMAL CONTROL
Optimal Control Problem
Minimise:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
0 0( ), ( ), , ( ( ), ( ), )
ft
f f
t
J t t t t L t t t dt x x x u
And algebraic constraints
f
ii
ttt
nitttFtx
0
,,...,1 ),),(),(()( ux
0( ( ), ( ), ) 0, 1,..., 2i f ft t t i k n x x
fi tttmitttG 0 ,,...,1 ,0)),(),(( ux
Subject to a set of differential constraints
Boundary constraints
Example of Optimal Control Problem
Minimise:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
And algebraic constraints
Subject to a set of differential constraints
Boundary constraints
0
;
sp
x v
T Dv g
m
Tm
g I
1
2 max -
G T
G T T
1 0 0
2 0 0
3 0 0 0
4 0
( ( ), ( ), ) ( ) 0
( ( ), ( ), ) ( ) 0
( ( ), ( ), ) ( )
( ( ), ( ), ) ( )
f f
f f
f f
f f f f
t t t h t
t t t v t
t t t m t m
t t t h t h
x x
x x
x x
x x
0 0( ), ( ), , ( )f f fJ t t t t v t x x
Pontryagin Maximum Principle
Consider the optimal control problem:
If is optimal, with response , being and the spaces of piecewise continuous and piecewise continuous and differentiable functions respectively. Then, there exist and such that:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
0 0
0
min ( , ) ( ), ( ), , ( ( ), ( ), )
:
( ) ( , ( ), ( ))
( ( ), ( )) 0
ft
f f fU
t
f
J t t t t t L t t t dt
subject to
t t t t
t t
uu x x x u
x F x u
ψ x x
* *
0 0ˆ( , ) [ , ] [ , )q
ft C t T t T u * 1
0ˆ [ , ]nC t Tx C
1* * 1
0 0ˆ, ,
nC t T
λ
* * *
0 0
* * * * *
0 0 0
0 0
* * * *
0
*
0
( ), ( ) [0,0],
( ) 0, ( ) ( ( ), ( ), ( ), ( )), . . [ , ]
( ( ), ( ), ( ), ( )) ( , , ) ( , , )
arg min ( ( ), ( ), ( ), ( ))
( ) 0, (
f
f
T
U
t t t t t
t t H t t t t a e t t
with H t t t t L t t
H t t t t
t const H
x
u
λ
λ x u λ
x u λ x u λ F x u
u x u λ
x
* * * *
0
*
( ), ( ), ( ), ( )) 0
( )
0, 0
f
f
T
t f f f t tt t
T
f x xt t
T
t t t t
t
u λ ν ψ
λ ν ψ
ν ψ ν
* bnν
1C
First Order Necessary Optimality Conditions
Euler-Lagrange Equations:
with optimality condition
and transfersality conditions:
First order conditions only guarantee that the solution is locally stationary.
Legendre-Clebsch Condition:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
),,(),,( 00 ff
T
ff tt xxxx
),,( tFH
x i
i
i ux
; 1T
i
i i i
H Li ,...,n
x x x
F
0
0 1,...,T
j j j
f
H Lj l
u u u
t t t
F
f
f
ttt
tt
T
TT
xf
H
xxt
)(0
)(
00, uu fH t t t
Example 2: quadratic control
Minimise
Subject to:
With boundary conditions:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
21
2
ft
t
J u dt
x v
v x u
0
0
( ) 1
( ) 0
( ) 0
( ) 0
f
f
x t
v t
x t
v t
Example 2: quadratic control
Hamiltonian
with adjoint equations:
optimality condition:
And transversality conditions:
From the adjoint equations and the optimal control condition we get:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
21( )
2x vH u v x u
x v
v x
H
x
H
v
v
Hu
u
( )
( )
x f x
v f v
t
t
( ) sin cos
( ) sin cos
cos sin
x
v
t A t B t
t B t A t
u A t B t
Example 2: quadratic control
and if one replaces the control law into the dynamic equation:
which is a harmonic oscillator with a periodic forcing term:
The solution is of the form:
And the four constants can be determined so that the boundary conditions are satisfied.
9/2/2012 O
PT
IMA
L CO
NT
RO
L
cos sin
x v
v x A t B t
cos sinx x A t B t
cos sin cos sin2 2
A Bx C t D t t t
Optimal Control Problem with Mixed Path Constraints
Consider the optimal control problem:
Form the Lagrangian:
with Lagrangian multipliers m
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
0 0
0
min ( , ) ( ), ( ), , ( ( ), ( ), )
:
( ) ( , ( ), ( ))
( ( ), ( )) 0
( , ( ), ( )) 0
ft
f f fU
t
f
J t t t t t L t t t dt
subject to
t t t t
t t
t t t
uu x x x u
x F x u
ψ x x
G x u
0 0
( , )
:
( ( ), ( ), ( ), ( )) ( , , ) ( , , )
T
f
T
t H
with
H t t t t L t t
u μ G
x u λ x u λ F x u
Maximum Principle with Maxed Path Constraints
If is optimal, with response , being and the spaces of piecewise continuous and piecewise continuous and differentiable functions respectively. Then, there exist , and such that:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
*
* * * *
0 0
* * * * *
0 0
* * * * * *
( , ) 0 0
* * * *
0
* *
( ), ( ), ( ) [0,0],
( ) 0, ( ) ( ( ), ( ), ( ), ( ), ( )),
( ) ( ( ), ( ), ( ), ( ), ( )), . . [ , ]
arg min ( ( ), ( ), ( ), ( ))
( , )
f
f
U
t t t t t t
t t t t t t t
t t t t t t a e t t
H t t t t
U t
x
λ
u
λ μ
λ x u λ μ
x x u λ μ
u x u λ
x
*
* *
* * * * *
0 0
*
| ( , , ) 0
( ) 0; 0;
( ) 0, ( ( ), ( ), ( ), ( )) 0
( )
0, 0
f
f
q
T
T
t f f f t tt t
T
f x xt t
T
t
t
t const H t t t t
t
u G x u
μ G μ
x u λ ν ψ
λ ν ψ
ν ψ ν
* *
0 0ˆ( , ) [ , ] [ , )q
ft C t T t T u C1C
* bnν
1* * 1
0 0ˆ, ,
nC t T
λ
* 1
0ˆ [ , ]nC t Tx
0ˆ ,
mC t Tμ
First Order Necessary Optimality Conditions with Mixed Path Constraints
Euler-Lagrange equations:
with Optimality condition:
and second order condition:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
),,(),,( 00 ff
T
ff tt xxxx
( , , )i i
i
x F t
x u
T - ; 1T
i
i i i i
Li ,...,n
x x x x m
F G
0 1,...,T T
j j j j
Lj l
u u u u m
F G
f
f
ttt
tt
T
TT
xf
H
xxt
)(0
)(
Transversality conditions
0uu
Example of Solution
The extended Hamiltonian is
where only one of the two conditions on the thrust is true at any one time. The dynamic equations are:
and the adjoint equations are:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
;
x
v
m sp
x v
T Dv g
m
Tm
g I
1 2 max
0
( )x v m
sp
T D Tv g T T T
m g I m m
2
1 1( ) ; ( ) =- ; ( ) = x v v x v m v
D D T Dt t t
x m x v m v m m
Example of Solution
The optimal control condition writes
where again only one of the two m is true at any one time, and the transversality conditions are:
If one looks at the optimality condition can define the switching function as:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
1 2
0
0 v m
spT m g I
m m
( )
( ) 1
( ) 0
x f x x
v f
m f
tx
tv
t
2 1 max
0
1 2
0
0
0 0, 0
0 0, 0 0
0 singular arc
v m
sp
v m
sp
v m
sp
I T Tm g I
I Tm g I
Im g I
m m
m m
Treatment of Singular Arcs
Along singular arcs we have that:
Is satisfied for any admissible control. A solution is to differentiate a minimum of p times with respect to time until one obtains:
The second equation provides the required control law while p-1 is the order of the singular arc.
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0 uH
0
0
p
up
p
up
dH
dt
dH
u dt
Optimal Control Problem with State Path Constraints
Consider the optimal control problem:
In this case the path constraint can not be directly used to form a Lagrangian function. In fact suppose that G≥0 is a path constraint on the position. This constraint does not provide any information on the velocity and acceleration at the contact point G=0. As a consequence there is no information on the correct control to be applied. This is particularly important when pure state path constraints are introduced via a direct approach.
The correct way to introduce a pure state path constraint is to differentiate p times with respect to time till the control appears explicitly:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
0 0
0
min ( , ) ( ), ( ), , ( ( ), ( ), )
:
( ) ( , ( ), ( ))
( ( ), ( )) 0
( , ( )) 0
ft
f f fU
t
f
J t t t t t L t t t dt
subject to
t t t t
t t
t t
uu x x x u
x F x u
ψ x x
G x
0p
p
d
dt
G
u
Optimal Control Problem with State Path Constraints
Once the control appears explicitly in the constraints, one can form the Lagrangian (indirect adjoining approach):
For example for p=1 and G≥0 one has:
:At a contact point t the second constraints causes a discontinuity in adjoint variables and Hamiltonian:
with q another Lagrange multiplier
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0 0
( , )
:
( ( ), ( ), ( ), ( )) ( , , ) ( , , )
pT
f p
T
dt H
dt
with
H t t t t L t t
Gu η
x u λ x u λ F x u
*
*
( , )0
( , ) 0
d t
dt
t
G x
G x
* * *
*
( ) ( ) ( ) ( , )
( ) ( ) ( ) ( , )
x
t
h
H H h
t t q t t
t t q t t
x
x
Maximum Principle with State Path Constraints If is optimal, with response , being and the spaces of piecewise continuous and piecewise continuous and differentiable functions respectively. Then, there exist , , , and such that:
9/2/2012 O
PT
IMA
L CO
NT
RO
L
0
*1
* * * *
0 0
* * * * *
0 0
* * * * * *
( , ) 0 0
* * * *
0
* *
1
( ), ( ), ( ) [0,0],
( ) 0, ( ) ( ( ), ( ), ( ), ( ), ( )),
( ) ( ( ), ( ), ( ), ( ), ( )), . . [ , ]
arg min ( ( ), ( ), ( ), ( ))
( ,
f
f
U
t t t t t t
t t t t t t t
t t t t t t a e t t
H t t t t
U
x
λ
u
λ η
λ x u λ η
x x u λ η
u x u λ
x
* *
* * *
* * * * *
0 0
*
) | ( , , ) 0, ( , , ) 0
( ) 0; 0; 0
( ) 0, ( ( ), ( ), ( ), ( )) 0
( )
0, 0
f
f
q
T
T
t f f f t tt t
T
f x xt t
T
dt t t
dt
t
t const H t t t t
t
Gu x u G x u
η G η η
x u λ ν ψ
λ ν ψ
ν ψ ν
* *
0 0ˆ( , ) [ , ] [ , )q
ft C t T t T u C1C
* bnν
1* * 1
0 0ˆ, ,
nC t T
λ
* 1
0ˆ [ , ]nC t Tx
0ˆ ,
mC t Tη ( ) cnt θ
* * *
*
* * *
( ) ( ) ( ) ( , )
( ) ( ) ( ) ( , )
( ) ( , ( )) 0; ( ) 0; ( ) ( )
T T T
x
T
tH H
t t t t
t t t t
t t t q t q t t
λ λ θ G x
θ G x
θ G x
Some References
Bryson A E, Jr. & Ho Y C. Applied optimal control: optimization, estimation, and control. Waltham, MA: Blaisdell, 1969. 481 p.
Betts J.T. Practical Methods for Optimal Control and Estimation Using Nonlinear Programming. Advances in Design and Control, SIAM 2001
9/2/2012 O
PT
IMA
L CO
NT
RO
L
FINITE ELEMENTS IN TIME
9/2/2012 ● 22
Previous Works on Finite Elements in Time for Optimal Control
Since early ’90s Finite Elements in Time used for Indirect Transcription: • Hodges, D. H. and Bless, R. R., “Weak Hamiltonian Finite Element Method for
Optimal Control Problems,” Journal of Guidance, Control, and Dynamics, Vol. 14, No. 1, 1991, pp. 148–156.
• Hodges, D. H., Bless, R. R., Calise, A. J., and Leung, M., “Finite Element Method for Optimal Guidance of an Advanced Launch Vehicle,” Journal of Guidance, Control, and Dynamics, Vol. 15, No. 3, 1992, pp. 664–671.
• Bottasso, C. and Ragazzi, A., “Finite Element and Runge-Kutta Methods for Boundary-Value and Optimal Control Problems,” Journal of Guidance, Control, and Dynamics, Vol. 23, No. 4, 2000, pp. 749–751.
Late ’90s Finite Elements in Time on Spectral Basis for Direct Transcription:
• Vasile M., Bottasso C.L., Finzi A.E., Lunar Orbital Dynamics by Finite Element in Time Method, Aerotecnica Missili e Spazio Vol. 75. Numero 3/4, Luglio-Dicembre 1996.
• Finzi A., Vasile M. Numerical Solutions for Lunar Orbits. IAF-97-A.5.08, 48th International Astronautical Congress, October 6-10, 1997/Turin, Italy
• Vasile, M. and Finzi, A., “Direct Lunar Descent Optimisation by Finite Elements in Time Approach,” Journal of Mechanics and Control, Vol. 1, No. 1, 2000.
• Vasile, M. and Bernelli-Zazzera, F., “Optimizing Low-Thrust and Gravity Assist Maneuvres to Design Interplanetary Trajectories,” The Journal of the Astronautical Sciences, Vol. 51, No. 1, 2003, January-March 2003.
His
toric
al B
ackgro
und
Direct Finite Element Transcription (DFET)
Strong solution of the differential equations:
Weak solution of the differential equations:
The solution x and control u satisfy the differential equations in a weak sense, i.e. with respect to the test functions w.
Boundary conditions are not exactly satisfied but are satisfied with respect to the test functions w.
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
0
0
( , )
f
f
tt
b
tt
dt x F x u x
0
0
( , ) ( )
f
f
tt
T T b
tt
dt w x F x u w x x
Direct Finite Element Transcription (DFET)
Strong vs. Weak solution of the differential equations:
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
0
0
( , )
f
f
tt
b
tt
dt x F x u x 0
0
( , ) ( )
f
f
tt
T w w w T w b
tt
dt w x F x u w x x
t0 tf
b
fx
( )w
fx t
0( )wx t
0
bxx
t
Direct Finite Element Transcription (DFET)
Partial integration of the differential terms:
The integral can be solved by polynomial representation of the states x, control u and test functions w and Gauss quadrature.
The choice of the test function w can be such that the projection is orthogonal.
9/2/2012
0
0
0 0
0
0
0
0
0
( , )
( , )
( , )
( , )
f
f
f
f f
f
f
f
f
tt
T T T T b
tt
tt t
T T T T b
t tt
tt
T T T T b
tt
tt
T T T b
tt
dt
dt
ddt
dt
dt
w x w F x u w x w x
w x w F x u w x w x
w x w x w F x u w x
w x w F x u w x
FIN
ITE E
LE
ME
NT
S IN T
IME
Direct Finite Element Transcription (DFET)
Polynomial representation of x, u and w:
Where xs, us and ws are discrete nodes and fs and gs are polynomials of order p-1 and p respectively.
Note that in principle fs and gs can be any arbitrary function generated on any arbitrary base.
9/2/2012
0
0
( , )
f
f
tt
T T T b
tt
dt w x w F x u w x
FIN
ITE E
LE
ME
NT
S IN T
IME
)(1
s
s
s
p
s
tfu
x
u
x )(
1
1
ss
p
s
tg ww
0
0
1 1
1 1 1 1 0
1 1 1
( , , ) ( ) ( )
f
f
T Tt p p p
s s s s s s s s p p f t t
s s st
g f g t dt w g t w g t
w x w F x u x x
Direct Finite Element Transcription (DFET)
The equation must hold true for any arbitrary ws , hence:
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
0
0
0
0
1 1
1
1
1 1
1
( ) ( ) ( ) ( , , )
...
( ) ( ) ( , , ) 0
...
( ) ( ) ( ) ( , , )
f
f
f
f
t p
s s s s t
st
t p
j s s j s s
st
t p
p s s p s s t
st
g t f t g t t dt
g t f t g t dt
g t f t g t t dt
x F x u x
x F x u
x F x u x
0
0
1 1
1 1 1 1 0
1 1 1
( , , ) ( ) ( )
f
f
T Tt p p p
s s s s s s s s p p f t t
s s st
g f g t dt w g t w g t
w x w F x u x x
Direct Finite Element Transcription (DFET)
Now assume we wanted to use Gauss quadrature formulas to solve the integrals, then we would have:
Where tk are Gauss points and tk are Gauss weights.
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
01 1
1 1
1 1
1 1
1 1
( ) ( ) ( ) ( , , )
...
( ) ( ) ( ) ( , , ) 0
...
( ) ( ) ( ) ( , , )f
p p
k k s s k k s s k t
k s
p p
k j k s s k j k s s k
k s
p p
k p k s s k p k s s k t
k s
g f g
g f g
g f g
t t t t
t t t t
t t t t
x F x u x
x F x u
x F x u x
Direct Finite Element Transcription (DFET) Discretised Weak solution of the differential equations:
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
t0 tf
b
fx
( )w
fx t
0( )wx t
0
bxx
t
0
0
( , )
f
f
tt
T T T b
tt
dt w x w F x u w x
Spectral basis: Gauss points used for integration and discretisation
Example of DFET Integration Let us consider the simple linear system with a constant u:
In weak form we can write:
And assume we use polynomials of order 0 for states and controls and of order 1 for the test functions:
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E 00
ff
tbt
T T T
tt
x v xdt
v x u v
w w w
x v
v x u
01 2 1 2
0 0 0 0
( ) ( ) 1 11; ;
( ) ( ) ( ) ( )
f
f f f f
f w w w w w wt t t t
t t t t t t t t
Example of DFET Integration The weak solution becomes:
Because they need to be true for any arbitrary weight w we can derive the two equations:
We can now take Gauss integration formulas to solve the integral:
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
0
01 2 1 2 2 1
0 0 0 0 0
( ) ( )1 1
( ) ( ) ( ) ( )
f b bt
s sf
s s sf f f ft f
x v x xw w w w dt w w
v x u v v
t t t t
t t t t t t t t
0
0
0 0 0
0
0 0
( )1
( ) ( )
( )1
( ) ( )
f
f
bt
s sf
s s sf ft
bt
s s
s s sf ft f
x v xdt
v x u v
x v xdt
v x u v
t t
t t t t
t t
t t t t
1
0
01
1
2
1
2
b
s sf
s s s
b
s s
s s s f
x v xt
v x u vt t
x v xt
v x u vt t
t t
t t
Example of DFET Integration The Gauss weight for a single point is equal to 2, therefore we have:
And if we sum the second to the first we get:
Note that for nonlinear function this is an implicit scheme.
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
0
1 / 22
2
1 / 22
2
b
s s
s s s
b
s s
s s s f
x v xt t
v x u vt t
x v xt t
v x u vt t
02
2
b
s s
s s s
b
s s
s s s f
x v xt
v x u v
x v xt
v x u v
0
b b
s
s sf
vx xt
v v x u
Direct Finite Element Transcription (DFET)
• The Time domain is decomposed in finite elements leading to a polynomial development of the solution on spectral basis (Gauss Points)
• Differential constraints are expressed in weak form leading to discontinuities at boundaries
• Gauss formulas for the solution of the integral: high Integration order 2n = 2k+2
9/2/2012
t
x
t
xs
xbf
xbi
ts
Gauss
Point
Boundary
Nodes
Boundary gap db
),(11 iii
N
ittDD
0
0
( , ) ( )
f
f
tt
T T b
tt
dt w x F x u w x x
FIN
ITE E
LE
ME
NT
S IN T
IME
9/2/2012
Direct Finite Element Transcription (DFET)
• The controls are discretised using the Guass points used for integration.
• The controls can be discontinuous at the boundaries.
• The integration error within an element is absorbed into the discontinuity at the boundaries.
t
x
u
t
xs
xbf
xbi
ts
u
FIN
ITE E
LE
ME
NT
S IN T
IME
9/2/2012
DFET in Summary
Representation of controls, states and weights on spectral basis.
Same bases for the discretisation of objective function and algebraic constraints:
Differential equations are transformed into algebraic equations:
2],),(),([),,(
11
0i
kkkk
q
k
N
i
f
b
f
b tLtJ
ttt puxxx
)(1
s
s
s
p
s
tfu
x
u
x
)(1
1
ss
p
s
tg ww
)( ; )(1i
p
i DPs
gDPs
f p
02
)()()()( 111
1
b
i
Tb
i
T
pi
k
T
kk
T
kk
q
k
txwxwFwxw tttt
0)),(),(( kksks uxG
FIN
ITE E
LE
ME
NT
S IN T
IME
9/2/2012
DFET Assembling Process
:
FIN
ITE E
LE
ME
NT
S IN T
IME
1
1
1
1 1
1
1
1 1 1
1
( ) ( ) ( ) ( , , )
...
( ) ( ) ( , , ) 0
...
( ) ( ) ( ) ( , , )
i
i
i
i
i
i
t p
s s s s i
st
t p
j s s j s s
st
t p
p s s p s s i
st
g t f t g t t dt
g t f t g t dt
g t f t g t t dt
x F x u x
x F x u
x F x u x
2
1
2
1
2
1
1 1 1
1
1
1 1 2
1
( ) ( ) ( ) ( , , )
...
( ) ( ) ( , , ) 0
...
( ) ( ) ( ) ( , , )
i
i
i
i
i
i
t p
s s s s i
st
t p
j s s j s s
st
t p
p s s p s s i
st
g t f t g t t dt
g t f t g t dt
g t f t g t t dt
x F x u x
x F x u
x F x u x
The end boundary node of one element
must be equal to the beginning boundary
node of the following element
The last equation of one element
is matched to the first equation of
the following element.
Boundary nodes disappear except
for the two extremal nodes
Example of DFET Assembly Assume we have two elements:
The end of the first element must be equal to the beginning of element 2:
9/2/2012 F
INIT
E EL
EM
EN
TS IN
TIM
E
1 1
1 1 1
1 1
1 1 1
0
1
2
2
b
s s
s s s
b
s s
s s s
x v xt
v x u v
x v xt
v x u v
2 2
2 2 2
2 2
2 2 2
22
2
b
s s
s s s
b
s s
s s s f
x v xt
v x u v
x v xt
v x u v
1 1
1 1 1
1 1 2 2
1 1 1 2 2 2
2 2
2 2 2
02
2 2
2
b
s s
s s s
s s s s
s s s s s s
b
s s
s s s f
x v xt
v x u v
x v x vt t
v x u v x u
x v xt
v x u v
9/2/2012
Direct Finite Element Transcription (DFET)
The optimal control problem is transformed into a nonlinear programming problem
f
i
t
t
f
b
f
b dtLtJ ),,(),,,( 0 puxpxx
0),,,( tpuxFx
0),,,( tpuxG
0),,,(0
0 ft
t
b
f
b tpxx
)J( min y
ul byb
yc
0)(
y=[xs,us,xb
0,xb
f,t0,tf,p]
FIN
ITE E
LE
ME
NT
S IN T
IME
9/2/2012
Direct Finite Element Transcription (DFET)
Optimal Control
Problem DFET
NLP Problem
min J(y)
where y=[x,u,ti ,tf]
T
SSQP
Sparse Sequential
Quadratic
Programming
ul byb
0yc
)(
l
iisiss
b
j
Tb
j
T
p
j
is
T
ikis
T
iki
q
i
l
t
ttt
tttt
0)),(),((
02
)()()()(
)(
111
1
uxG
xwxwFwxw
yc
FIN
ITE E
LE
ME
NT
S IN T
IME
9/2/2012
Multi-Phase DFET
Discontinuous states can be inserted at the boundary of two phases.
The end states of one phase are matched to the beginning states of another phase through some interphase constraints or...
…phases can be in parallel.
),,( tJ lux
Interphase-link Constraints
),,( tluxFx
0),,( tluxG
0))(()(
)()(
1
1
1
l
f
bll
i
b
l
f
l
i
l
f
bl
i
b
tt
tt
tt
xx
xx
),,,( l
f
l
i
l
f
l
i
l
link ttxx
Phase l
Assembly
NLP Solver
0),,,(0
ft
t
b
f
b
i
l tpxx
FIN
ITE E
LE
ME
NT
S IN T
IME
FIR
ST O
RD
ER
NE
CE
SS
AR
Y CO
ND
ITIO
NS F
OR
OP
TIM
AL
ITY
The KKT Optimality Conditions
Lagrangian function of the NLP problem:
According to Karush-Kuhn-Tucker first order optimality conditions:
02/09/2012 41
FIR
ST O
RD
ER
NE
CE
SS
AR
Y CO
ND
ITIO
NS F
OR
OP
TIM
AL
ITY
With transversality conditions:
Which are equivalent to the maximum principle transversality conditions.
KKT Optimality conditions on controls and costates:
If one assumes that:
And
Then one gets the integral forms of the optimality conditions:
02/09/2012 42
FIR
ST O
RD
ER
NE
CE
SS
AR
Y CO
ND
ITIO
NS F
OR
OP
TIM
AL
ITY
Example: Minimum quadratic control with path constraints
Simple linear system with minimization of the integral of the square of the control action:
Linear differential constraints
Path constraint:
Boundary conditions:
02/09/2012 43
FIR
ST O
RD
ER
NE
CE
SS
AR
Y CO
ND
ITIO
NS F
OR
OP
TIM
AL
ITY
Example: Minimum quadratic control with path constraints
02/09/2012 44
FIR
ST O
RD
ER
NE
CE
SS
AR
Y CO
ND
ITIO
NS F
OR
OP
TIM
AL
ITY
Example: Minimum quadratic control with path constraints
02/09/2012 45
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
Discontinuity at the Boundaries
From the theory of Delfour et al. 1981 on Galerkin methods for ODEs:
And a maximum convergence rate of:
Delfour also demonstrated the accuracy of FET developed on Gauss Legendre, Radau and Lobatto points.
02/09/2012 46
x
t
xh
xb
x
db
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
Discontinuity at the Boundaries
From the theory of Delfour et al. 1981 on Galerkin methods for ODEs:
And a maximum convergence rate of:
Simple double integrator:
02/09/2012 47
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
Discontinuity at the Boundaries
The expected convergence of the discontinuity at the boundary is O(hp+2):
Example
Elliptical orbit motion:
Therefore one could use the following indicator:
02/09/2012 48
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
h-p Adaptivity Strategies
p-adaptivity
h-adaptivity
Given the residual on the error at each element:
Solve the following linear quadratic optimization problem:
02/09/2012 49
If the residuals are not zero, then all
the elements with a nonzero residual
are split in two.
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
Examples: Minimum Time to Orbit
The problem is to minimize the time to reach a given altitude with a given velocity (Bryson and Ho 1979):
With the dynamic constraints:
And the boundary conditions:
From Pontryagin maximum principle one gets the control law:
02/09/2012 50
H-
P AD
AP
TIV
ITY S
TR
AT
EG
Y
Examples: Minimum Time to Orbit
02/09/2012 51
H-
P AD
AP
TIV
ITY S
TR
AT
EG
Y
Examples: Minimum Time to Orbit
02/09/2012 52
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
Examples: Step-by-Step Adaptation Process
02/09/2012 53
H-P A
DA
PT
IVIT
Y ST
RA
TE
GY
Examples: Step-by-Step Adaptation Process
02/09/2012 54
Given a solution of the NLP problem, one has C(x)=0.
By linearizing in a neighbourhood of x one can get:
Where the matrix is the Jacobian of the constraints whose components are the gradients with respect to states and controls.
If the uncertainty on the initial state is given and no variation on the final state is required, one can solve the following linear quadratic optimisation problem:
RO
BU
ST D
ES
IGN
Robust Control
02/09/2012 55
11 12 13
1min
2
. .
0
; 0
T
s i s
i f
s t
UU U
J U J X J X
X X
If the control is considered to be fixed, the Jacobian matrix can be used to derive the state transition matrix from initial states to final states:
Starting from the bottom of the matrix one can compute:
And then proceed with::
Up to:
with
RO
BU
ST D
ES
IGN
Transition Matrix
02/09/2012 56
• Vasile M. Finite elements in Time: A Direct Transcription Method for
Optimal Control Problems. AAS, Astrodynamics Specialist Conference,
Toronto 1-5 August 2010.
RO
BU
ST D
ES
IGN
Some References
02/09/2012 57
9/2/2012
GAUSS PSEUDO-SPECTRAL METHODS (GPSM)
Consider again the differential equations:
And the polynomial representation of states and controls:
GA
US
S PS
EU
DO
-SP
EC
TR
AL M
ET
HO
DS
Pseudo-spectral Transcription
( , , ) 0t x F x u
)(1
s
s
s
p
s
tfu
x
u
x
With the first derivatives:
1
( ) p
s s
s
f t
x x
Let’s assume now that the nodes xs are chosen to be the zeros ts of
orthogonal polynomial, like Legendre polynomials, in the interval [-1,1].
Then f(t) is different from zero at each t=ts .
Then we can collocate the differential equations as follows:
Now consider the stong integral form:
GA
US
S PS
EU
DO
-SP
EC
TR
AL M
ET
HO
DS
Pseudo-spectral Transcription
,
0
( , , ) 0; s 1,...,2
M
s k k s s s
k
tD Mt
x F x u
And assume the integral term is integrated numerically with a Gauss
quadrature formula:
Then one can choose a Lagrange interpolating polynomial that
interpolates x0 plus the M Gauss points xs to represent the states (the
controls repsectively).
0
0( ) ( ) ( , , )
ft
f
t
t t t dt x x F x u
0
1
( ) ( ) ( , , )2
M
f s s s s
s
tt t t
x x F x u
The complete set of equations becomes:
GA
US
S PS
EU
DO
-SP
EC
TR
AL M
ET
HO
DS
Pseudo-spectral Transcription
,
0
0
1
( , , ) 0; s 1,...,2
( ) ( ) ( , , )2
M
s k k s s s
k
M
f s s s s
s
tD M
tt t
t
t
x F x u
x x F x u
t0 tf
fx
0( )x t
0
bxx
t
Spectral basis: Gauss points used for integration and discretisation
The controls are collocated only at the Gauss points:
GA
US
S PS
EU
DO
-SP
EC
TR
AL M
ET
HO
DS
Pseudo-spectral Transcription
,
0
0
1
( , , ) 0; s 1,...,2
( ) ( ) ( , , )2
M
s k k s s s
k
M
f s s s s
s
tD M
tt t
t
t
x F x u
x x F x u
t0 tf
( )fx t
0( )x t
0
bxu
t
Spectral basis: Gauss points used for integration and discretisation
As for DFET one can partition the time domain in segments and use a PS
transcription on each segment:
GA
US
S PS
EU
DO
-SP
EC
TR
AL M
ET
HO
DS
Multiple Phases Pseudo-spectral Transcription
1 , 1 0,
, 1 0,
( ) ( ); 1,..., 1j f j j j segment
f j j
t t j n
t t
x x
t
x
t
xs
xf
xi
ts
Gauss
Point
It is now clear that in order to restore the continuity from one segment to
the next one has to add a matching condition:
The matching condition restores the continuity on the states but not on the
control that can remain discontinuous.
Benson, D. A., Huntington, G. T., Thorvaldsen, T. P., and Rao, A. V., "Direct
Trajectory Optimization and Costate Estimation via an Orthogonal
Collocation Method," Journal of Guidance, Control and Dynamics, Vol. 29,
No. 6, November–December, 2006, pp. 1435–1440
Benson, D. A., A Gauss Pseudospectral Transcription for Optimal Control,
Ph.D. Thesis, Dept. of Aeronautics and Astronautics, Massachusetts
Institute of Technology, November 2004.
http://dspace.mit.edu/handle/1721.1/28919
Huntington, G. T., Advancement and Analysis of a Gauss Pseudospectral
Transcription for Optimal Control, Ph.D. Thesis, Dept. of Aeronautics and
Astronautics, Massachusetts Institute of Technology, May 2007.
http://dspace.mit.edu/handle/1721.1/42180
GA
US
S PS
EU
DO
-SP
EC
TR
AL M
ET
HO
DS
Some References
9/2/2012
BASICS OF NONLINEAR PROGRAMMING
•Find x* solution of the nonlinear problem c(x*)=0
•Given a first guess x and a linear expansion of c in x
)()(
)()( xxdx
xdcxcxc
pxxcdx
xdcxx
)()(
1
•Quadratic convergence if cC1and cxx0
•Problem: computing the derivative of c
The value of x is updated with the Newton’s step p given by
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
Simple 1D Newton Method
•Find x that minimises F(x)
•Given a value x, the function F is expanded up to the second order:
))(('')(2
1))((')()( xxxFxxxxxFxFxF
))(('')('0)(')(
xxxFxFxFdx
xdF
Sufficient condition for optimality
First order necessary condition is to have:
0)('' xF
Simple 1D Minimisation
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
•Find the vector x=[x1,…,xn]T such that:
0
x
x
xc
)(
)(
)(
1
nc
c
pxx
cGp
where G is the Jacobian matrix of nonlinear functions c
Newton’s method in N dimensions B
AS
ICS
OF N
ON
LIN
EA
R P
RO
GR
AM
MIN
G
•Find x=[x1,…,xn]T that minimises F(x):
0
x
x
x
n
ndx
dFg
dx
dFg
g
)(
)(
)(1
1
pxx
gHp
Where H is the Hessiana matrix of F
First order necessary conditions for optimality
Newton’s method for minimisation in N dimensions B
AS
ICS
OF N
ON
LIN
EA
R P
RO
GR
AM
MIN
G
•Build an estimate B* of matrix B:
),(* xch BB
xx
xxch
T
T
)( B
xx
xxh
T
T
B
BB
In this case the convergence is no more quadratic but superlinear
Update the Hessian matrix with the BFGS (Broyden-Fletcher-
Goldfarb-Shanno) formula:
Approximation of the Hessian and Jacobian matrix B
AS
ICS
OF N
ON
LIN
EA
R P
RO
GR
AM
MIN
G
•find x that minimises F(x) subject to:
0xc )(
0)(
01
xc
Gg
L
cFL i
m
i
i
T
x
Suffcicient condition for optimality is that the Hessian of the L is:
First order necessari conditions for optimality (Kuhn-Tuker):
A scalar function called Lagrangian is defined as:
m
i
ii
T cFFL1
)()()(),( xxcxx
0vHv L
T
Constrained optimisation: Equality Constraints B
AS
ICS
OF N
ON
LIN
EA
R P
RO
GR
AM
MIN
G
*c
*g
gZ T
x1
x2
2
2
2
1)( xxxF
02)( 21 xxxc
)1,1(* x
0Gg
2
1
1
2
2T
0gZ T
ZHZ L
T
Minimise:
Subject to
And solution
At the solution point
Equivalent necessary condition is that the projection of the gradient of the
function F on the constraint is zero:
Equivalent sufficient condition is that the projection of the Hessiana is
positive defined:
Constrained optimisation: Equality Constraints B
AS
ICS
OF N
ON
LIN
EA
R P
RO
GR
AM
MIN
G
•Find x that minimises F(x) subject to:
0)( xc
For x=x*:
•Some constraints are equal to zero
•Others are strictly satisfied
set active the tobelongs i.e. for 0)( Aici x
sconstraint inactive ofset the tobelong i.e. 'for 0)( Aici x
Necessari condition for optimality is that lagrangian multipliers at the solution
are:
0i
Constrained optimisation: Inequality Constraints B
AS
ICS
OF N
ON
LIN
EA
R P
RO
GR
AM
MIN
G
•Two different approaches:
•active set method
•Interior point method
Active Set
This methods looks for the active set at every optimisation step
Interior Point
Inequality constraints are penalised in the objective function
through a barrier function dependent on a penality parameter m:
)(ln)( xx cF m
The solution is forced to stay within the feasible region by the
barrier function lnc(x)
Constrained optimisation: Inequality Constraints
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
The solution of the constrained optimisation problem satisfies the following
first order conditions:
0
0)(
0)(
0
m
m
m
xc
xc
GGg
i
e
T
i
T
ex
L
L
L
)()()(),,( xcxcxxiTeTFL mm
Slack variables are introduced to insert inequality
constraints as equality constraints:
0s
sxcxsx
])([)(),,( TFL
0
0)(
0
lu
T
y
L
L
bs
xyb
yc
Gg
Constrained optimisation: The K-T system
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
•The direction of steepest descent is computed solving the K-T system:
c
Ggp
0G
GH
TT
L
)( scxGs dd T Tdd sxp ,
•The solution of the K-T system gives the direction of steepest
descent and the step. The step length is corrected by a factor a:
a
pyy
•The value a is such that the bounds on slacks s and
unknowns x are not violated
Constrained optimisation: The K-T system
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
HxxxgTT
2
1
Given the quadratic objective function
Subject to linear constraints
bAx
Solution method: starting from a first guess x0 and an estimate of the
active set A0
1)at step k solve the K-T system with the constraints in Ak as equalities
2)Maximum step length along the direction p which does not violate active
constraints
3) if the step crosses on inactive constraints the step is reduced and the
constraint is inserted among active constraints, then go to step 1 otherwise
check the sign of the associated lagrangian multipliers
4) If all the multiplier are positive then stop, otherwise if erase from the active
set the constraint associated to the most negative multipliers.
Sequential Quadratic Programming
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
x2
x1
C1<0
C2>0 Feasible region
x(1)
x*
x(0)
x(2)
2
2
2
1)( xxxF
02)( 211 xxxc
03
24)( 212 xxxc
Step x1 x2 Active set
0 4 0 A0=c2
1 2.77 1.85 2=-5.53 erase c2
2 1.2 0.8 a=0.5 add c1
3 1 1 A*=c1
Minimise
Subject to
Example
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
NOTICE!
Speed of convergence and robustness improve if Jacobian ed Hessin
are analytical
Robustness of convergence: local convergence must be
acchieved from every starting point in the solution space:
)()(2
1)(),,( scscscsx TTFM
Unidimensional minimisation in the direction of steepest descent
using a quadratic or cubic model modello of the merit function:
)(),,( min a MM sx
a q
pxx
ss
Betts 2000
where is a weight matrix
The merit function
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
• Betts J.T. Practical Methods for Optimal Control and Estimation
Using Nonlinear Programming. Advances in Design and Control,
SIAM 2001.
• Fletcher, R. Practical Methods of Optimization, Wiley, 1987.
• Gill, P.E, W. Murray and M. Wright, Practical Optimization,
Academic Press, 1981.
• Nocedal, J. and S. Wright, Numerical Optimization, Springer, 1998
Some References
BA
SIC
S O
F NO
NL
INE
AR
PR
OG
RA
MM
ING
9/2/2012
SOME SPACE APPLICATIONS
9/2/2012 ● 82
Direct Finite Element Transcription (DFET)
Low
-thru
st M
ulti-G
ravity
Assis
t Tra
jecto
ries
BepiColombo •3000 variables and constraints for the NLP problem •4 to 7 swingbys •resonant orbits •more than 20 switching points
9/2/2012 ● 83
Direct Finite Element Transcription (DFET)
Low
-thru
st M
ulti-G
ravity
Assis
t Tra
jecto
ries
Earth-Europa Transfer
• 6000-7000 variables and constraints for the NLP problem
• 14 swingbys
• resonant orbits
• variable thrust
• Variable reference frames
9/2/2012 ● 84
Direct Finite Element Transcription (DFET)
Low
-thru
st M
ulti-G
ravity
Assis
t Tra
jecto
ries
WSB Transfers
• 1000-1500 variables and constraints
• Highly nonlinear and unstable dynamics
• Impulsive manoeuvres
9/2/2012 ● 85
Some References
Low
-thru
st M
ulti-G
ravity
Assis
t Tra
jecto
ries
• Vasile M., Campagnola S. Design of Low-Thrust Gravity Assist Trajectories to Europa.
Journal of the British Interplanetary Society, Vol. 62, No.1, pp. 15-31, 8 January 2009.
http://arxiv.org/ftp/arxiv/papers/1105/1105.1823.pdf
• Vasile M., Bernelli-Zazzera F., Targeting a heliocentric orbit combining low-thrust
propulsion and gravity assist manoeuvres, Operational Research in Space and Air,
Kluwer Academy Press, Ed. 1-4020-1218-7, 2003.
• Vasile M., Bernelli-Zazzera F., Optimising low-thrust and gravity assist manoeuvres to
design interplanetary trajectories, The Journal of Astronautical Sciences, vol. 51(1),
2003. http://arxiv.org/ftp/arxiv/papers/1105/1105.1829.pdf
• Vasile M. Bernelli-Zazzera F., Direct Averaging for Multiple Revolution Trajectory
Optimisation. 2nd International Symposium on Low Thrust Trajectories (LOTUS2),
Toulouse 18-20 June 2002.
• Vasile M., Bernelli-Zazzera F., Combining Low-Thrust and Gravity Assist Manoeuvres to
Reach Planet Mercury. AAS/AIAA Astrodynamics Specialist Conference,30 Jul-2 Aug
2000,Quebec City, Canada
• Vasile M., Bernelli Zazzera F., Jehn R., Janin G., Optimal Interplanetry Trajectories
Using a Combination of Low-Thrust and Gravity Assist Manoeuvres. IAF-00-A.5.07,
51st International Astronautical Congress,2-6 Oct,2000/Rio de Janeiro, Brazil.
• Vasile, M. and Finzi, A., “Direct Lunar Descent Optimisation by Finite Elements in Time
Approach,” Journal of Mechanics and Control, Vol. 1, No. 1, 2000.
9/2/2012
FINITE PERTURBATIVE ELEMENTS
9/2/2012
Equations of Motion
Non-singular Equinoctial elements:
No singularities for zero-inclination
and zero-eccentricity orbits.
1
2
1
2
sin
cos
tan sin2
tan cos2
a
P e
P e
iQ
iQ
L
X
Gauss planetary equations in Equinoctial elements, under a perturbing acceleration ε in
the r-t-h frame:
2
2 1
11 2 1 2
22 1 1
2sin cos cos cos cos sin
cos cos cos 1 sin cos sin ( cos sin ) sin
cos cos cos 1 sin cos sin (
da a pP L P L
dt h r
dP r p pL P L P Q L Q L
dt h r r
dP r p pL P L P Q
dt h r r
b a b a
b a b a b
b a b a
2
2 211 2
2 221 2
1 23
cos sin ) sin
1 sin sin2
1 cos sin2
cos sin sin
L Q L
dQ rQ Q L
dt h
dQ rQ Q L
dt h
dL rQ L Q L
dt a h
b
b
b
m b
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Low Thrust Two-Points Boundary Value Problem
Objective:
Where the control vector u is:
Satisfying Gauss’ Planetary Equations and the boundary constraints:
and with:
Assumptions:
Perturbing acceleration ε is very small compared to the local gravitational acceleration:
Constant modulus and direction in the radial-tangential reference frame.
00
f
t
t
X
X X
X f
( ), ( ), ( )T
u t t t a b
0
min
ft
ut
V t dt
2r
m
, , const a b
max( )t
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
From Time to True Longitude
A system of differential equations in time is translated into a system of differential equations in true longitude:
It is possible to rewrite Gauss Planetary Equations with respect to Longitude by applying the chain rule:
Where , with the hypothesis of small, constant, perturbing acceleration:
Thus obtaining also the time equation:
2
dL h
dt r
d d dt
dL dt dL
2 3
2
1 21 sin cos
dt r h
dL h P L P L
Finite P
ERTU
RB
ATIVE ELEM
ENTS
( , , , , )ddt
f L a bX X ( , , , , )ddL
f L a bX X
dL
dt
9/2/2012
First order expansion of Equations of Motion (1)
0 1X X X
' ' ' 0 1X X X
The aim is that of obtaining a first-order expansion of the variation of
Equinoctial elements in the form with respect to a reference orbit:
With the manipulations described above, one obtains a set of
equations in the form:
Which could be integrated between L0 and L.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Excursus on Perturbation Theory
Consider the following example:
And the first order expanded solution:
If this expanded solution is substituted in the original equation one gets:
Finite P
ERTU
RB
ATIVE ELEM
ENTS
2x x x
0 1
0 1
0 1
x x x
x x x
x x x
2
0 1 0 1 0 1
2 2 2 2
0 1 0 1 0 0 1 1
( )
2
x x x x x x
x x x x x x x x
9/2/2012
Excursus on Perturbation Theory
Now assuming we are interested only in first order terms we can
collect terms with similar order of :
The first equation is a simple harmonic oscillator and once its solution is
substituted in the second equation one has a simple harmonic oscillator
with periodic forcing term.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
0 0
2
1 1 0
x x
x x x
9/2/2012
First order expansion of Equations of Motion (2)
A first order expansion could be obtained by simply integrating the ODEs with respect
to L, which could be done in closed form.
This requires finding the primitives of the integrals in the form:
0
0
0
1 3
10 20
3
10 20
3
10 20
1
1 sin cos
cos
1 sin cos
sin
1 sin cos
F
F
F
L
FL
L
C FL
L
S FL
I L dLP L P L
LI L dL
P L P L
LI L dL
P L P L
For example:
2010 20
2 2 2 2
10 10 20 10 20 10 20 10
3/22 2
10 20
2 2 2 2
10 10 20 10 20 10 20 10
3/22 2
10 20
20 10 20
sin
1 sin cos
ln 1 cos 1 sin 1
1
ln 1 cos 1 sin 1
1
1 cos sin 1
L
S
LI L dL
P L P L
P P P P L P L P P P
P P
P P P P L P L P P P
P P
P L P L P
2 2
10 20 20 101 cos sin 1P P P L P L
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Analytical Solution of the Equations of Motion
Thus the first order approximate solution of perturbed Keplerian motion takes the
form:
2 2 2 2
0 1 0 0 0 20 2 0 10 2 0 0 0 11 0
1 10 11
2 20 21
1 10 11
2 20 21
0 1
( ) 2 cos cos ( , ) ( , ) 22 cos sin ( , )
( )
( )
( )
( )
( )
s ca L a a a h a P I L L P I L L h a I L L
P L P P
P L P P
Q L Q Q
Q L Q Q
t L t t
b a b a
A complete set of analytic equations parameterised on the Longitude is thus available
to propagate the perturbed orbital motion, in the form:
, , , ,0 0(L L) f (L ) L a b X X
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Some remarks on accuracy
The accuracy of the approximation is dependent on:
The ratio between the local gravitational acceleration and the perturbation. Thus:
Error increases according to the thrust level works better with small
thrust magnitude (<1 N).
Error increases with the distance from the central body works better very
close to the attractor.
The amplitude of the trajectory arc. Error increases superlinearly with the
amplitude of the propagation interval.
Tests have revealed that in most practical applications thrust and local gravitational
force still present a favourable ratio.
The error in amplitude is mitigated by dividing a single arc into sub-arcs of smaller
amplitude.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
FPET method
A Direct Transcription Method based on Finite Perturbed Elements in Time (FPET) has been designed using the Perturbative approach.
Each transfer trajectory is divided into n subarcs:
, , , ,2
, , , ,2
Lf
Lf
a b
a b
m
m
X X
X X
2
L
Amplitude of arc is ∆L. Perturbed motion propagated
using analytical solution. Constant thrust vector in the r-
t-h reference frame. Reference node for propagation
is the midpoint of the arc. Motion is propagated
analytically backwards and forwards by from the midpoint to obtain boundary nodes. (better accuracy compared to a single sided propagation)
Xm
X+
X-
Reference
point
Upper boundary
point
Lower boundary
point
Forward
propagation
Backward
propagation
∆L
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
FPET method
2
L
The subarcs are then matched to each other at the boundaries to obtain the complete trajectory.
Conceptually similar to Sims & Flanagan Direct Transcription Method which used Keplerian arcs with ∆V discontinuities at the boundaries.
In the FPET method, thrust is continuous, albeit constant within each element.
In both cases, orbit propagation is analytical.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Accuracy and CPU time (1)
To assess accuracy and CPU cost, orbital motion under constant thrust was propagated within a given time interval.
A range of values for the number sub-arcs was considered.
Accuracy was calculated in terms of relative error on final state with respect to numerical integration of the exact equations of motion.
CASE 1, Heliocentric orbital motion: 0.5 N continuous thrust on a 2000 kg spacecraft. ε=2.5 10-5 m/s2
Departure from Earth. 1.5 years.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
The LT Boundary problem with FPET
CASE 2, Geocentric orbital motion: 0.5 N continuous thrust on a 2000 kg spacecraft. ε=2.5 10-5 m/s2
Departure from LEO 5 days (≈50 revolutions).
FPET method is 2 orders of magnitude faster than numerical integration. Even with few elements per revolution, accuracy is high for geocentric orbits. In heliocentric orbits, accuracy is still adequate. The FPET method is particularly advantageous with a relatively low number of
sub-arcs.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
The Low Thrust Two-Points Boundary Value Problem with FPET
The FPET transcription method is used to solve the LT boundary problem:
Decision variables for each of the n FPET:
Position of the reference point (5 scalars).
Acceleration magnitude, azimuth and elevation (3 scalars).
8n decision variables and 5(n+1)+1 scalar constraints.
The problem is efficiently solve with a gradient-based local optimizer (fmincon active-set).
1
1 0
1
1
min
. . , 2, , 1 0
, 1, ,
FPET
FPET
FPET
n
i i
i
i i FPET
n f
n
i
i
i max FPET
J t
s t i
tF
n
i n
To
u
eq
X X
C X X
X X
Fixed time of flight
Arrival conditions
Departure conditions
Inter-element matching conditions
Total ∆V
Upper bounds on acceleration
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Earth-Mars Direct transfer (1)
Boundary problem:
Departure from Earth at 5600 MJD2000.
Rendezvous with Mars after a transfer time of 3 years.
2 complete revolutions.
Maximum acceleration: 2.5 10-5 m/s2.
40 FPET.
Initial guess for the local optimizer: constant thrust profile.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Earth-Mars Direct transfer (2)
Results:
Total ∆V: 5.63 km/s.
Relative error 10-3.
Solution found with DITAN: 5.71 km/s.
Hohmann Transfer: 5.49 km/s.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Earth-Mars Direct transfer: Multi-Objective Problem (1)
Bi-objective optimization problem: ∆V and Time of Flight are minimised.
Each time to objective function is called, a boundary problem needs to be solved.
Solved with EPIC, a hybrid-memethic stochastic optimizer.
8000 function evaluations.
Lower Upper
T0 [MJD2000] 5000 5779.94
ToF [days] 100 1500
nrev 1 3
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Earth-Mars Direct transfer: Multi-Objective Problem (2)
Bi-objective optimization problem: ∆V and Time of Flight are minimised.
An interesting comparison could be made with an analogous Biimpulsive Transfer to Mars.
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012
Some References
Zuiani, F., Vasile, M., Avanzini, G., Palmas A. Direct transcription of low-thrust trajectories with finite trajectory elements. (2012) Acta Astronautica, 72. pp. 108-120. ISSN 0094-5765
http://www.sciencedirect.com/science/article/pii/S0094576511002852
http://strathprints.strath.ac.uk/33788/
Zuiani F., Vasile M. Preliminary Design of Debris Removal Missions by Means of Simplified Models for Low-Thrust, Many-Revolution Transfers. International Journal of Aerospace Engineering, Volume 2012 (2012), Article ID 836250, 22 pages, doi:10.1155/2012/836250
http://www.hindawi.com/journals/ijae/2012/836250/
Finite P
ERTU
RB
ATIVE ELEM
ENTS
9/2/2012 ● 106
Massimiliano Vasile
massimiliano.vasile@strath.ac.uk
Dept. of Mechanical & Aerospace Engineering
University of Strathclyde, Glasgow, UK
http://www.strath.ac.uk/space/research/optimisation/
Recommended