Proceedings of the 6th IFAC Workshop on Lagrangian and ThBT1.

4
Hamiltonian Methods for Nonlinear Control, Universidad Técnica
Federico Santa María, Valparaíso, Chile, May 1-4, 2018

A structure-preserving Partitioned Finite

Element Method for the 2D wave equation ?
Flávio Luiz Cardoso-Ribeiro ∗ Denis Matignon ∗∗
Laurent Lefèvre ∗∗∗
∗
Instituto Tecnológico de Aeronáutica, Brazil (e-mail:
[email protected]).
∗∗
ISAE-SUPAERO, Université de Toulouse, France (e-mail:
[email protected]).
∗∗∗
Univ. Grenoble Alpes, LCIS, F-26902, Valence, France (e-mail:
[email protected])

Abstract: Discretizing open systems of conservation laws while preserving the power-balance
at the discrete level can be achieved using a new Partitioned Finite Element Method (PFEM),
where an integration by parts is performed only on a subset of the variables in the weak
formulation. Moreover, since boundary control and observation appear naturally in this
formulation, the method is suitable both for simulation and control of infinite-dimensional port-
Hamiltonian systems. The method can be applied using FEM software, and comes along with
worked-out test cases on the 2D wave equation in different geometries and coordinate systems.

Keywords: Distributed Parameter systems, Port-Hamiltonian systems, Finite Element Method,

Geometric Discretization Methods, 2D Wave equation.

1. INTRODUCTION tively the energy and co-energy variables. More recently,

general geometric pseudo-spectral methods were proposed,
The port-Hamiltonian formalism has been proven to be using high-order global polynomial (Moulla et al. (2012))
a powerful tool for the modeling and control of complex or Bessel (Vu et al. (2017)) basis functions. An exten-
multiphysics systems. In many cases, spatio-temporal dy- sion for the piezoelectric beam was proposed by Cardoso-
namics must be considered and infinite-dimensional port- Ribeiro et al. (2016). A finite volume structure-preserving
Hamiltonian models are needed. Classical academic exam- discretization method was also investigated in Kotyczka
ples such as the transmission line, the shallow water or (2016). Mixed finite element methods were also considered
the beam equations have been investigated in the port- in Farle et al. (2014a, 2013, 2014b) where one of the
Hamiltonian framework (Duindam et al., 2009). Besides, equations is kept in strong form; numerical results were
2D and 3D problems have been recently considered (Wu presented for one-dimensional systems only. A numerical
et al., 2015; Vu et al., 2016; Trenchant et al., 2017). In study of a 2D vibro-acoustic system based on the mixed
many of these examples, systems of two balance equations finite element method was performed by Wu et al. (2015).
are considered. A 2D finite difference method with staggered grids was
used to find a structure preserving scheme by Trenchant
In order to simulate and design control laws, obtaining et al. (2017).
a finite-dimensional approximation which preserves the
port-Hamiltonian structure of the original system can be In these previous work, the central idea was to define
advantageous. It may serve as a design guide such as in different discretization spaces for the energy and co-energy
Control by Interconnection (CbI) or in Interconnection variables such that the strong form of the equations was
and Damping Assignment Passivity Based Control (IDA- exactly satisfied in these finite-dimensional spaces. Defin-
PBC). Besides, preserving the underlying Dirac intercon- ing these spaces is straightforward for 1D, but seems to
nection structure results in energy conservation properties be cumbersome for higher dimensions or higher order
and associated dynamical properties (e.g. stability, pas- methods (Hiemstra et al., 2014). The kernel of the exterior
sivity, etc.). derivatives in N -D dimensional domains is not anymore
trivial and the discretization of the trace operator on
In Golo et al. (2004), the authors proposed a mixed finite boundaries with non trivial (i.e. rectangular) geometries
element structure-preserving spatial discretization for 1D often leads to dimensionality problems. As suggested in
hyperbolic systems of conservation laws, making use of the Kotyczka et al. (2018) the discretization of the weak for-
distinct low-order basis functions to approximate respec- mulation of the considered port-Hamiltonian system may
? This work is supported by the project ANR-16-CE92-0028, be a practical solution to deal with these higher dimen-
entitled Interconnected Infinite-Dimensional systems for Hetero- sional problems. We propose in this paper to follow this
geneous Media, INFIDHEM, financed by the French National approach and to perform integration by parts on one of
Research Agency (ANR). Further information is available at the two balance equations. Then, the discretization in the
https://websites.isae-supaero.fr/infidhem/the-project/.

© 2018, IFAC (International Federation of Automatic Control) 119

chosen basis for the the energy and co-energy variables From the definition of the variational derivatives, The
(and the associated test functions) leads directly to a full time-derivative of the Hamiltonian can be computed as:
rank representation for the finite dimensional Dirac in- Z  
δH δH
terconnection structure. Besides boundary conditions are Ḣ = αq ·
α̇ + α̇p dΩ , (4)
Ω αq
δα δαp
naturally handled, even in the case of higher order finite
element basis. Finally, differently from Farle et al. the rewriting it in terms of the co-energy variables, we get:
Z
use of weak-form in both equations enables to work Ḣ = (α̇ x, t) · eq (x
αq (x x, t) + α̇p (x
x, t)ep (x
x, t)) dΩ , (5)
with ready to use finite-element software to perform the Ω
proposed discretization scheme, and consequently to con- thus, from (1):
sider more involved problems and geometries. For instance, Z
computations in the present paper have been done using Ḣ = (−∇∇ep (xx, t) · eq (x
x, t) − ep (x
x, t)div eq (x
x, t)) dΩ ,
the FreeFem++ software (Hecht, 2012). ZΩ

This paper starts with the presentation of the wave equa- = ep (x n · eq (x

x, t) (−n x, t)) ds ,
∂Ω
tions in the partitioned weak form in Section 2. Sec- (6)
ondly, we show that the discretized equations lead to a
finite-dimensional port-Hamiltonian system in Section 3. where the term n is the vector normal to the boundary
Thirdly, the numerical scheme is applied to the simulation curve ∂Ω. Let us define the boundary input as:
of the wave equation taking into account boundary exci- u∂ (s, t) := −nn · eq (x
x(s), t)) s ∈ ∂Ω . (7)
tation in Section 4. Finally in Section 5, we discuss how to Its power-conjugated boundary output is thus given by:
make use of the method in other coordinate systems, and x(s), t) s ∈ ∂Ω ,
y∂ (s, t) := ep (x (8)
how to extend it to nonlinear (polynomial) systems, then
and the power-balance reads:
some conclusions are drawn. Z
Ḣ = y∂ (s, t)u∂ (s, t) ds . (9)
2. PARTITIONED WEAK FORM FOR THE WAVE ∂Ω
EQUATION
2.2 Partitioned weak-form representation of the system of
In this section, we first recall how to write the strong conservation laws
form of the wave equation as a port-Hamiltonian system
in § 2.1. Secondly, we propose a partitioned weak-form Instead of using vector variables in (1), let us rewrite
representation for this system in § 2.2, which will be the equations using its components: αp (x x, t) = α1 (x, y, t),
T
discretized in the following section. x, t) = [α2 (x, y, t) α3 (x, y, t)] . The equations become:
α q (x
∂ ∂
2.1 Wave equation as a system of conservation laws α̇1 (x, y, t) = − e2 (x, y, t) − e3 (x, y, t)
∂x ∂y
The wave equation can be used to describe different ∂
α̇2 (x, y, t) = − e1 (x, y, t) (10)
phenomena like liquid sloshing, elasticity, propagation of ∂x
electromagnetic waves in transmission lines, etc. ∂
α̇3 (x, y, t) = − e1 (x, y, t)
∂y
In vectorial form, the wave equation can be written as:
x, t) = −div eq (x
α̇p (x x, t) , Let us take the arbitrary test functions v1 (x, y), v2 (x, y)
(1) and v3 (x, y), and rewrite the previous equation using the
αq (x
α̇ ∇ep (x
x, t) = −∇ x, t) ,
where x ∈ Ω is the position vector, eq and ep are the co- following integral form:
Z Z
energy variables, obtained from the variational derivatives v1 α̇1 dx dy = − v1 (e2,x + e3,y ) dx dy ,
of the system Hamiltonian with respect to the energy ZΩ ZΩ
variables α q and αp , respectively.
v2 α̇2 dx dy = − v2 e1,x dx dy , (11)
In the case of the linear wave equation, the Hamiltonian ZΩ ZΩ
is given by: v3 α̇3 dx dy = − v3 e1,y dx dy
Z
1 Ω Ω
αp2 + |α
αq |2 dΩ ,


H= (2) integrating the first equation by parts, we get:
2 Ω Z Z
and the co-energy variables are computed as: v1 α̇1 dx dy = (v1,x e2 + v1,y e3 ) dx dy
δH δH Ω
ep = = αp , e q = = αq . (3) ZΩ  
δαp αq
δα e2 (x, y, t)
− v1 ·
n ds ,
∂Ω
e3 (x, y, t)
The wave equation written as (1) is also known as a system Z Z (12)
of conservation laws. The spatial integral of the energy v2 α̇2 dx dy = − v2 e1,x dx dy ,
variables represent conserved quantities (e.g., total volume ZΩ ZΩ
and linear momentum, in the shallow water equations, to- v3 α̇3 dx dy = − v3 e1,y dx dy .
tal charge and flux in the transmission line). Furthermore, Ω Ω
the co-energy variables also exhibits physically relevant  
meaning (e.g., pressure and volumetric flow in the shallow e2 (x(s), y(s), t)
Note that the term −n n· is the boundary
water equations, voltage and current in the transmission e3 (x(s), y(s), t)
line equations). input, previously defined as u∂ (s, t) in (7).

© 2018, IFAC (International Federation of Automatic Control) 120

From (12), we note that the variables of index 1 (v1 and The equations can be rewritten as:
e1 ) are derived once with respect to x and y, thus, they α1 =Dxe2 + Dy e3 + Bu
M1α̇ u∂ (t) ,
must be of class H 1 (Ω). The variables of index 2 and 3 T
α2 = − Dx e 1 ,
M2α̇ (18)
should be of class L2 (Ω) and should also have boundary
values, thus belong to H 1/2 (Ω). α3 = −
M3α̇ , DyT e 1
In the following section, we discretize the partioned weak- where M1 , M2 and M3 are square matrices (of size N1 ×N1 ,
form (12), and we show that the resulting system is a finite- N2 × N2 and N3 × N3 , respectively). Dx is an N1 × N2
dimensional port-Hamiltonian system and thus preserves matrix, Dy is an N1 × N3 matrix and B is an N1 × N∂
the power-balance of the original system (9). matrix.
The time derivative of the continous Hamiltonian (5) can
3. STRUCTURE PRESERVING FINITE ELEMENT be rewritten using the coordinate variables as:
Z
DISCRETIZATION
Ḣ = (α̇1 e1 + α̇2 e2 + α̇3 e3 ) dx dy . (19)
Ω
Let us approximate the energy variables α1 (x, y, t) using
the following basis with N1 elements: Using the approximated variables:
N1
X αT1 M1e 1 + α̇
Ḣ ≈ α̇ αT3 M3e 3 ,
αT2 M2e 2 + α̇ (20)
α1 (x, y, t) ≈ α1ap (x, y, t) := φi1 (x, y)α1i (t)
i=1
(13) Let us define new energy variables as:
T
= φ 1 (x, y) α 1 (t) α1 := M1α 1 , α̃
α̃ α2 := M2α 2 , α̃ α3 := M3α 3 , (21)
The variables e1 and v1 are also approximated using such that (20) becomes:
φ 1 (x, y). α˙ T e 1 + α̃
Ḣ ≈ α̃ 1 α˙ T e 2 + α̃
α˙ T e 3 ,
2 3 (22)
Similarly, the other energy variables are approximated as: We define the discretized
h Hamiltonian as:
T
N2
X α1 , α̃
Hd (α̃ α2 , α̃
α3 ) := H α1 (xx, t) = M1−1α̃
α1 (t) φ1 (x
x) ,
α2 (x, y, t) ≈ α2ap (x, y, t) := φi2 (x, y)α2i (t)
(14) T
x, t) = M2−1α̃


i=1 α2 (x α2 (t) φ2 (x
x) ,
T
= φ 2 (x, y) α 2 (t) , −1

T i
x, t) = M3 α̃
α3 (x α3 (t) φ3 (x
x) .
N3
X (23)
α3 (x, y, t) ≈ α3ap (x, y, t) := φi3 (x, y)α3i (t)
i=1
(15)
The time-derivative of this discretized Hamiltonian is
= φ 3 (x, y)T α 3 (t) . given by:
Furthermore, e2 and v2 are approximated using φ 2 (x, y), ∂Hd ∂Hd ∂Hd
α˙ T1
Ḣd = α̃ α˙ T2
+ α̃ α˙ T3
+ α̃ . (24)
e3 and v3 are approximated using φ 3 (x, y). α1
∂α̃ α2
∂α̃ α3
∂α̃
Finally, the boundary input can be discretized using any Since we want that both power balances (20) and (24)
one-dimensional set of basis functions, say ψ = [ψ i ]: coincide, the following constitutive relations must hold:
N∂ ∂Hd ∂Hd ∂Hd
X e1 = , e2 = , e3 = . (25)
u∂ (s, t) ≈ uap
∂ (s, t) := ψ i (s)ui∂ (t) = ψ (s)T u ∂ (t) . (16) ∂α̃α1 α2
∂α̃ α3
∂α̃
i=1
Rewriting the finite-dimensional equations (18), we get the
The finite-dimensional equations become: following finite-dimensional port-Hamiltonian system:
α˙ 1
   
α̃ 0 Dx Dy "e 1 # "B #
Z Z
T
α1 = φ 1,xφ T2 dx dy e 2
φ 1φ 1 dx dy α̇ α˙ 2  =−DxT 0 0  e 2 + 0 u ∂ ,
α̃
Ω
| {z } |Ω {z } α˙ 3
α̃ −DyT 0 0 e3 0 (26)
M1 Dx
Z y ∂ =B T e 1 ,
+ φ 1,y φ T3 dx dy e 3
Ω
where y ∂ is the conjugated output of the discretized
| {z } system.
Dy
Z From (24) and (26), the time derivative of the discretized
− φ 1 (x(s), y(s))ΨT (s) ds u ∂ (t) , Hamiltonian is given by:
∂Ω
| {z } α˙ T1 e 1 + α̃
Ḣd = α̃ α˙ T2 e 2 + α̃
α˙ T3 e 3 ,
B
T T T T
= e 2 Dx + e 3 Dy + u T∂ B e 1 − e T1 Dx e 2 − e T2 Dy e 3 ,
Z Z

φ 2φ T2 dx dy α̇
α2 = − φ 2φ T1,x dx dy e 1 ,
Ω Ω = y T∂ u ∂ .
(27)
| {z } | {z }
M2 T
Dx
Z Z Note that u T∂ y ∂
is the discrete analog of the continu-
φ 3φ T3 dx dy α̇
α3 = − φ 2φ T1,x dx dy e 1 . ous power-balance equation (9). Furthermore, this power-
|Ω {z } |Ω {z } balance is exactly preserved in the finite dimensional ap-
M3 DyT proximation spaces. From the definition of the B ma-
(17) trix (17), the definition of the approximated boundary

© 2018, IFAC (International Federation of Automatic Control) 121

input uap T
∂ (s, t) := ψ (s) u ∂ (t) and approximated co-energy
eigenvalues obtained from the numerical model were com-
ap
variable e1 (x(s), y(s), t) := φ 1 (x(s), y(s))T e 1 (t), we get: pared to the exact eigenfrequencies of the linear wave equa-
Z tion with constant coefficients. The inputs of (26) (given
T
Ḣd = e 1 φ 1 (x(s), y(s))ΨT (s) ds u ∂ , by the influx through the boundaries) were considered to
Z ∂Ω
(28) be zero, i.e.:
= eap
1 (x(s), y(s), t)uap
∂ (s, t) ds . u∂ (x, y, t) = 0 , [x, y] ∈ ∂Ω , (31)
∂Ω or, in the finite-dimensional case, simply u ∂ = 0.
Remark 1. Note that using classical finite-elements 1D
discretization basis for the boundary input, u ∂ (s, t) pro- Recall that the variables with index 1 (α1 , e1 , v1 ) must be
vides the values of the influx (−n n · eq ) at the boundary discretized with polynomials of order at least one (since
nodes. For instance, in the case of shallow water equations, they are derived once in (12)). Fig. 1 shows the relative
these are the values of volumetric influx into the system. error of the first modal frequency for three different choices
The conjugated output y ∂ is related with the curve integral of polynomial approximations. P1 P0 P0 stands for first
of e1 (x(s), y(s), t) along the elements. The co-energy vari- order polynomial for the variable 1, and order zero for the
able e1 (x(s), y(s), t) is the pressure, thus the discretized other two variables. P1 P1 P1 uses first-order polynomial
outputs y ∂ are related to the forces per unit length applied for the three variables. Finally, P1 P2 P2 uses first-order
along the external boundary. polynomial for the first variable, and order two for the
Remark 2. With this choice of partition of variables, other other two variables.
boundary inputs and outputs can easily be dealt with: it
10 2
is enough to define ũ u∂ := Σ u ∂ , and ỹy ∂ := Σ y ∂ for some P1P0P0
orthogonal matrix Σ of dimension N∂ × N∂ . Indeed, the P1P1P1
P1P2P2
power balance (27) is preserved thanks to orthogonality, 10 1

Error of the first natural frequency (%)

and in (26), B is replaced by B ΣT for the control, and B T
is replaced by Σ B T for the collocated observation.
10 0
Remark 3. Different choices of variables are also possible
for the integration by parts. Instead of integrating the
first equation by parts in (12), we could integrate the two 10 -1

other equations, which would also lead to another skew-

symmetric structure; the boundary inputs and outputs 10 -2
would not be the same either.
Remark 4. In the sequel, in our finite element method, we
10 -3
conveniently chose ψ (s) as φ 1 (x(s), y(s)) evaluated at the 10 1 10 2 10 3
Number of DOF
boundary. Other choices could be investigated.

4. NUMERICAL EXPERIMENTS Fig. 1. Convergence of the first natural frequency of the

2D wave equation.
In this section, several numerical experiments are pre- Remark 5. A more logical test case would be to use
sented with the goal of testing the proposed discretiza- P2 P1 P1 polynomials (instead of P1 P2 P2 , since the dif-
tion scheme. Firstly, a numerical convergence analysis is ferentiability requirement is higher on the first variable).
presented in § 4.1. Different polynomial degrees for the However, such a choice surprisingly did not lead to a
discretized variables and number of elements were con- convergent numerical scheme. It is also somewhat surpris-
sidered. Secondly, in § 4.2 the scheme is proven useful ing that using P1 P1 P1 leads to the best convergence rate
for simulating the system subjected to external boundary among the three test cases (instead of the higher-order
excitation. Finally, in § 4.3 we discuss one of the main choice P1 P2 P2 ). The convergence of the proposed scheme
advantages of this method (in comparison to other port- should be better investigated in further work.
Hamiltonian discretization methods): it can be directly
applied using available open source finite element software. 4.2 Time domain simulations
4.1 Convergence analysis
Numerical time-domain simulations were performed using
the discretized system under boundary-port excitation.
The power-preserving discretization method presented in
The following boundary conditions were used:
the previous section was implemented using quadrilateral 
elements with polynomial basis functions for a square  sin(πt) ,[x, y] ∈ ∂Ωup ,
domain. The following Hamiltonian was considered: u∂ (x, y, t) = − sin(πt) ,[x, y] ∈ ∂Ωlef t , t ≤ 1s,
0 ,[x, y] ∈ ∂Ωdown ∪ ∂Ωright .
Z
1

α2 + α22 + α32 dΩ ,


H= (29)
2 Ω 1 (32)
where Ω = [0, 1] × [0, 1]. After discretization, using (23), and u∂ (x, y, t) = 0, [x, y] ∈ ∂Ω, t > 1s. The boundary is
the Hamiltonian reads: split in four sides: ∂Ω = ∂Ωup ∪ ∂Ωlef t ∪ ∂Ωdown ∪ ∂Ωright .
1  T −1  These conditions impose a harmonic influx on one side of
Hd = α1 M1 α̃
α̃ αT2 M2−1α̃
α1 + α̃ αT3 M3−1α̃
α2 + α̃ α3 , (30) the boundary and the opposite condition at another side.
2
Snapshots of this simulation are presented in Fig. 2. The
A convergence analysis of the numerical method was done, simulations were performed using lsim, the MATLAB tool
focusing on the spectrum of the discretized system. The for linear systems simulation.

© 2018, IFAC (International Federation of Automatic Control) 122

 For instance, the 2D mesh presented in Fig. 4 is quickly
0.39 s 0.79 s done using FreeFem++ (Hecht, 2012), and obtaining the
matrices of (26) only requires a few lines of code.

1 1

0.8 0.8

1.2 0.6 1.5 0.6

1 1
0.8 0.5
0.4 0.4
1 1
0.2 0.2
0.5 0.5

0 0
0 0

1.19 s 1.59 s

1 1

0.8 0.8

1.5 0.6 1.5 0.6

1 1
0.5 0.5
0.4 0.4
1
0.2
1
0.2
Fig. 4. Square mesh with a hole in FreeFem++.
0.5 0.5

0
0
0
0
As we did with our own codes in § 4.1, a convergence
analysis of the first mode was also performed for the square
domain using FreeFem++. The results for the error of the
Fig. 2. Snapshots of simulation for a harmonic excitation first natural frequency are presented in Fig. 5 and follow
at two of the boundaries of the domain. The variable the same trend that we obtained with our code.
α1 is shown.
2
10
P1P0
Finally, the convergence of the time domain simulation P1P0P0
P1P0P0
P1P1
P1P1P1
P1P2P2
P1P2
with boundary conditions given by (32) was verified. Fig. 1
P1P1P1
P1P2P2
10
3 shows the L2 norm of the error, taking into account the
state of the system after one second of simulation. The
numerical result obtained using 1600 elements was taken 0
10
error (%)

as reference to compute the error. First-order convergence

is observed for a choice of P1 P0 P0 polynomials. −1
10

10 -2 −2
10

−3
10
1

0 1 2 3 4
10 10 10 10 10
Number of states
L 2 norm error of variable

10 -3
Fig. 5. Convergence of the first natural frequency of
the 2D linear wave equation, discretization using
FreeFem++.

10 -4
5. PERSPECTIVES AND CONCLUSIONS

10 0 10 1 10 2
Two extensions of the PFEM are proposed before conclud-
Number of DOF ing.

Fig. 3. Convergence of the L2 error as a function of the 5.1 Extension to other coordinate systems
number of degrees of freedom of the discretization
scheme. The error is computed after 1 second of
The method presented above was written in cartesian
simulation. Polynomials of order 1, 0 and 0 were
coordinates. For other geometries, it can be easier to deal
used as approximation functions for the variables
with other curvilinear systems, like the polar coordinates.
(P1 P0 P0 ). A first-order convergence is observed.
In this latter case, the methodology to obtain the dis-
cretized equations remain the same: for example, in (26),
4.3 Using FEM software for more complex geometries the same kind of algebraic structure is obtain for J, with
matrices Dρ and Dθ and their transpose. For the choice
One of the main advantages of the method proposed in of the basis functions, polynomials are suitable for the ρ
this paper compared to previous work on power-preserving variable, but care must be taken to use periodic functions
discretization methods is that it proves compatible with for the θ variable: trigonometric polynomials will solve this
classic FEM software. problem, as detailed in (Boyd, 2001, chap. 18).

© 2018, IFAC (International Federation of Automatic Control) 123

5.2 Extension to nonlinear equations Duindam, V., Macchelli, A., Stramigioli, S., and Bruyn-
inckx, H. (2009). Modeling and Control of Com-
In the previous section, we performed the discretization plex Physical Systems: The Port-Hamiltonian Approach.
of a linear wave equation with a quadratic Hamiltonian Springer Berlin Heidelberg, Berlin, Heidelberg.
(29). Nonlinear equations like the shallow water equations Farle, O., Klis, D., Jochum, M., Floch, O., and Dyczij-
(SWE), exhibit the same interconnection structure as (1) Edlinger, R. (2013). A port-Hamiltonian finite-element
with power balance (9). The difference lies in the fact that formulation for the Maxwell equations. Proceedings of
the Hamiltonian is non-quadratic (and, for this reason, the the 2013 International Conference on Electromagnetics
constitutive equations are nonlinear). in Advanced Applications, ICEAA 2013, (4), 324–327.
Farle, O., Baltes, R.B., and Dyczij-Edlinger, R. (2014a).
In the case of the 2D irrotational SWE, the Hamiltonian A Port-Hamiltonian Finite-Element Formulation for the
can be written as:   Transmission Line. In Proceedings of 21st International
Z
1 1 Symposium on Mathematical Theory of Networks and
α3 α12 + α22 + ρgα32 dΩ .


H= (33)
2 Ω ρ Systems, 724–728. Groningen, The Netherlands.
Farle, O., Baltes, R.B., and Dyczij-Edlinger, R.
In the case of non-quadratic Hamiltonian, the derivation (2014b). Strukturerhaltende Diskretisierung verteilt-
of the discretized Hamiltonian is not as straightforward as parametrischer Port-Hamiltonscher Systeme mittels
in the quadratic case (that is, obtaining (30) from (29)). finiter Elemente. at - Automatisierungstechnik, 62(7),
Nevertheless, it is possible to approximate the Hamiltonian 500–511.
using integration by numerical quadrature. Indeed, since Golo, G., Talasila, V., van der Schaft, A.J., and Maschke,
the variables are approximated using polynomial basis B. (2004). Hamiltonian discretization of boundary
functions, and since the integrand of the Hamiltonian (33) control systems. Automatica, 40(5), 757–771.
is also a polynomial, the discretized Hamiltonian can be Hecht, F. (2012). New development in FreeFem++. J.
exactly computed by quadrature. Numer. Math., 20(3-4), 251–265.
Hiemstra, R., Toshniwal, D., Huijsmans, R., and Ger-
5.3 Conclusions ritsma, M. (2014). High order geometric methods with
exact conservation properties. Journal of Computational
This paper presents the PFEM, a new method for Physics, 257, 1444–1471.
the power-preserving discretization of 2D wave equation. Kotyczka, P. (2016). Finite Volume Structure-
Rewriting the wave equation in a weak-form where only Preserving Discretization of Distributed-Parameter
some of the equations are integrated by parts, a skew- Port-Hamiltonian Systems. In 2nd IFAC Workshop
symmetric representation naturally arises with the bound- on Control of Systems Governed by Partial Differential
ary control and observation. After discretization, a finite- Equations (CPDE), volume 49, 298–303. Elsevier B.V.,
dimensional port-Hamiltonian system is obtained in a Bertinoro, Italy.
straightforward way. Kotyczka, P., Maschke, B., and Lefèvre, L. (2018). Weak
form of Stokes-Dirac structures and geometric dis-
The PFEM method can be easily implemented using cretization of port-Hamiltonian systems. Journal of
available Finite Element software (for instance, opensource Computational Physics, 361, 442–476.
software as FreeFem++ (Hecht, 2012)), since it suffices to Moulla, R., Lefevre, L., and Maschke, B. (2012). Pseudo-
properly write the weak-form equations. spectral methods for the spatial symplectic reduction of
An even greater flexibility comes from the choice of the open systems of conservation laws. Journal of Compu-
boundary inputs and outputs, either by a conservative tational Physics, 231(4), 1272–1292.
change of boundary variables through an orthogonal ma- Trenchant, V., Ramirez, H., Gorrec, Y.L., and Kotyczka,
trix, or by the choice of the partition of the variables in P. (2017). Structure preserving spatial discretization
the process of integration by parts. of 2D hyperbolic systems using staggered grids finite
difference. In Proceedings of the 2017 American Control
The extension of the method to other coordinate systems, Conference. Seattle, USA.
and to nonlinear equations will be better explored in Vu, N.M.T., Lefèvre, L., and Maschke, B. (2016). A
further work, (Cardoso-Ribeiro et al., 2018). structured control model for the thermo-magneto-
hydrodynamics of plasmas in tokamaks. Mathematical
REFERENCES and Computer Modelling of Dynamical Systems, 22(3),
181–206.
Boyd, J.P. (2001). Chebyshev and Fourier spectral meth- Vu, N., Lefèvre, L., Nouailletas, R., and Brémond, S.
ods. Dover. (2017). Symplectic spatial integration schemes for sys-
Cardoso-Ribeiro, F.L., Matignon, D., and Pommier- tems of balance equations. Journal of Process Control,
Budinger, V. (2016). Piezoelectric beam with dis- 51, 1–17.
tributed control ports : a power-preserving discretiza- Wu, Y., Hamroun, B., Le Gorrec, Y., and Maschke, B.
tion using weak formulation. In 2nd IFAC Workshop (2015). Power preserving model reduction of 2D vibro-
on Control of Systems Governed by Partial Differential acoustic system: A port Hamiltonian approach. Pro-
Equations (CPDE), 290–297. Bertinoro, Italy. ceedings of the 5th IFAC Workshop on Lagrangian and
Cardoso-Ribeiro, F.L., Matignon, D., and Lefèvre, L. Hamiltonian Methods for Nonlinear Control (LHMNC),
(2018). A Partitioned Finite-Element Method for power- 48(13), 206–211.
preserving discretization of open systems of conservation
laws. Preprint, to be submitted.

© 2018, IFAC (International Federation of Automatic Control) 124