Identification of LPV partial differential equation models
Julien Schorsch, Marion Gilson, Vincent Laurain, Hugues Garnier
To cite this version:
Julien Schorsch, Marion Gilson, Vincent Laurain, Hugues Garnier. Identification of LPV partial
differential equation models. 52nd IEEE Conference on Decision and Control, CDC 2013, Dec 2013,
Florence, Italy. Paper ThA17.3. hal-00861956
HAL Id: hal-00861956
https://hal.archives-ouvertes.fr/hal-00861956
Submitted on 14 Sep 2013
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Identification of LPV partial differential equation models
J. Schorsch, M. Gilson, V. Laurain and H. Garnier
Abstract— This paper deals with the identification of linear
parameter varying (LPV) models described by partial differential equations (PDE). A direct identification of continuous
space-time LPV-EDP systems in an input-output setting is
investigated in the case of an additive output noise. The
continuous space-time LPV-PDE model is firstly proposed to
be rewritten as a multi-input single-output linear time-space
invariant model and an iterative optimization is then developed
to estimate efficiently the model parameters. The performance
of the proposed method is then illustrated via a representative
simulation example based on an Alsace river-flow measurement.
I. I NTRODUCTION
Partial differential equations (PDE) describe a wide range
of physical behaviors with many applications in various
fields from the state-of-health in advanced batteries [1]
to larger scale problems such as simulation of the Arctic
sea ice temperature [2]. The main characteristics of PDE
with respect to the usually considered ordinary differential
equations (ODE) is the introduction of partial derivatives
along different dimensions, mostly depicting the time and
space components of the considered system.
PDE models are most of the time derived from physical laws, and therefore expressed in continuous time-space
(CTS). To solve the identification problem of PDE systems,
two main approaches can be distinguished:
• Indirect approaches which includes two steps. The
first involves the identification of a discrete time-space
(DTS) model. In the second step, the identified DTS
model is converted into CTS (see e.g. [13]).
• Direct approaches: which aim at estimating the coefficients of the CTS model directly from sampled
measured signals (see e.g. [15], [14], [16]).
In the usual ODE linear system case, one of the main
limitation of indirect approaches is the impossibility to deal
with non-regularly sampled data. In the presented PDE case,
while it is fair to assume a regular time-sampling, the
spatial-sampling is directly bounded to a physical sensor
implantation and hence limited by cost and physical possible
positioning. Consequently, in practice, indirect identification
of PDE systems might often not be applicable. In order
to cope with these limitations, direct approaches such as
proposed in [15] aim at directly identifying the physical
CTS parameters, with the inherent difficulty to approximate
J. Schorsch, M. Gilson, V. Laurain and H. Garnier
are with University of Lorraine, CRAN, UMR 7039, 2
rue
Jean
Lamour,
F-54519
Vandœuvre-les-Nancy,
France,
CNRS, CRAN, UMR 7039, France julien.schorsch,
marion.gilson, vincent.laurain,
hugues.garnier@univ-lorraine.fr
partial derivatives from sampled noisy signals. This difficulty
is commonly solved by a low pass filtering of the data.
Nevertheless, it was only recently that the noise modeling
problem with respect to these filters was addressed for direct
identification of PDE systems (see e.g. [7], [19]).
Another statement is that the PDE considered for system identification are mostly linear time and space invariant (LTSI). Nevertheless, in practice most of the physical
behaviors have a nonlinear or time-space varying nature.
Consider for example a pollutant transport along a river.
From a conceptual reasoning, it clearly appears that the
transport speed directly depends on the river flow or the
river profile for example. Consequently, this paper aims at
the identification of linear parameter varying (LPV) PDE
systems. Identification of LPV systems has gained some
strong interest in the past few years, mostly in the discretetime (DT) case [10], [11]. Identification of DTS-LPV-PDE
has been dealt with in [3], using an instrumental variable
(IV) method. The use of LPV models has also been applied
to the identification of LTSI-PDE system by approximating
the PDE at each working point with a DT-LPV model such
as in [4], [5].
Regarding direct identification of CT-LPV models, only a
few methods are available in the ODE case [6], and to the
best of authors’ knowledge, there is no specific work in the
PDE case. Consequently, this paper aims at identifying CTSLPV-PDE systems from sampled data under some realistic
noise conditions. As pointed out in [6], the identification
of CT parameters with physical meaning in LPV systems
requires the use of a direct identification scheme which
offers challenging theoretical issues. In the DT ODE case,
a convolution type representation of the system dynamics,
i.e. an LPV Impulse Response Representation (IRR) [10]
is used in order to define predictors with respect to some
model class. In the CTS case however, no IRR has been
developed yet and thus the same concept cannot be used
to define the predictors such as usually done [12]. Yet, a
system reformulation was proposed in [6] in order to apply
the linear time invariant (LTI) system theory to the LPV
identification problem. This reformulation is here extended
to the PDE case, allowing the use of a least squares method
which provides efficient estimates for output-error type of
noise in the prediction error minimization (PEM) sense.
The paper is organized as follows: in Section 2, the general
class of CTS-LPV-PDE systems in an input-output (IO)
representation intended for identification is introduced, along
with the main issues related to direct CTS identification. In
section 3, the reformulation of the system which enables the
predictor quality definition is provided. Finally, an iterative
least-squares (LS) algorithm dedicated to this model identification is detailed in section 4 and it is thoroughly analyzed
on a relevant example in Section 5.
II. P ROBLEM FORMULATION
A. System description
Consider the data generating CTS-LPV system described
by the following spatio-temporal PDE:
S
Ao (p, ρ)χ̊(t, x) = Bo (p, ρ)u(t, x)
(1)
B. Model considered
It can be seen that the analysed process is fully characterized by the knowledge of functions {åit ,ix } and {b̊jt ,jx }.
Yet a main difficulty in the estimation of LPV systems is
that in practical cases, these functions are a priori unknown
nonlinear functions. Even if some recent work gives some
preliminary results oriented towards a model structure selection, the usual solution to overcome this problem is to
parameterize these functions using a sum of a priori known
basis functions [6], [10]. Consequently, this paper focuses on
the usual assumption used in the literature where the process
model denoted by Gθ is defined in the following form:
y(t, x) = χ̊(t, x) + vo (t, x)
Gθ :
This equation is subject to initial conditions and/or boundary conditions. u(t, x) denotes the input variable observed
at time t and at space point x, χ̊(t, x) the noise-free output
and vo (t, x) the additive output noise. The parameter varying
nature of the system behavior is caracterized by ρ, the socalled scheduling variable with ρ = ρ(t, x) denoting the
dependency of the system on external deterministic signal
both in time and space. It can be noticed that as a first
step towards more general noise assumptions, throughout this
paper, the noise is supposed to be independent on ρ.
In this PDE context, p = (pt , px ) denotes the partial
differential operators:
pit ,ix u(t, x) = pitt pixx u(t, x) =
∂ it +ix u(t, x)
∂tit ∂xix
(2)
∂
∂x .
∂
∂t
and px =
Ao (p, ρ) and Bo (p, ρ) are
where pt =
bivariate polynomials in p of degrees nt , nx and mt , mx
respectively:
(B(p, ρ, θ), A(p, ρ, θ)) = (Aθ , Bθ )
(4)
where the ρ-dependent polynomials A and B are parameterized as
nt ,nx
(it ,ix )6=(nt ,nx )
X
ait ,ix (ρ)pit ,ix
A(p, ρ, θ) = pnt ,nx +
it ,ix =0
Aθ :
(5)
nα
X
a
0
alit ,ix fl (ρ)
i ,i (ρ) = ait ,ix +
t x
l=1
it = 0, · · · , nt ix = 0, · · · , nx
Bθ :
mX
t ,mx
bjt ,jx (ρ)pjt ,jx
B(p, ρ, θ) =
jt ,jx =0
mβ
X
0
+
bljt ,jx gl (ρ)
(ρ)
=
b
b
jt ,jx
jt ,jx
l=1
jt = 0, · · · , mt jx = 0, · · · , mx
(6)
m
nt ,nx
+
Ao (p, ρ) = p
nt ,nx
(it ,ix )6=(nt ,nx )
X
åit ,ix (ρ)pit ,ix (3a)
it ,ix =0
mX
t ,mx
b̊jt ,jx (ρ)pjt ,jx
Bo (p, ρ) =
(3b)
jt ,jx =0
where åit ,ix (ρ) and b̊jt ,jx (ρ) are analytic functions with
dependence on ρ at time t and at space x, which is called
static dependence, note that ρ = ρ(t, x).
Like in any identification problem, it is assumed that
a data sequence DN,L {y(tk , xℓ ), u(tk , xℓ ), ρ(tk , xℓ )}N,L
k,ℓ=1
generated by S is available. Let h be the spatial sampling
distance between two observation points, and Ts the temporal
sampling interval. L denotes the number of spatial observation locations and N the number of temporal sampling
instants. The data-generating system in Eq (1) is then identified by defining a model set and some information criterion
characterizing the quality of a given model based on DN,L .
As a next step, a model class suitable for the identification
of such systems is given.
β
α
In this parametrization, {fl }nl=1
and {gl }l=1
are meromorphic functions of ρ, with static dependence. They can be
chosen for examples as linearly independent functions. The
associated model parameters θ are stacked columnwise:
T
θ = [a0,0 · · · ant ,nx −1 b0,0 · · · bmt ,mx ] ∈ Rnθ
(7)
where
(8)
ait ,ix = a0it ,ix a1it ,ix · · · anitα,ix ∈ Rnα +1
mβ
∈ Rmβ +1
bjt ,jx = b0jt ,jx b1jt ,jx · · · bjt ,j
x
(9)
and nθ = ((nt + 1)(nx + 1) − 1)(nα + 1) + (mt + 1)(mx +
1)(mβ + 1).
Let introduce also G = {Gθ | θ ∈ Rnθ }, as a collection of
all process models in the form of (4).
Furthermore, it has to be noticed that the system presented
in Eq (1) displays a CTS representation of the noise. Nevertheless, in order to avoid the rather difficult mathematical
problem of treating sampled CTS random process in terms of
a filtered piece-wise constant CTS noise source (see [9], [8]),
it is here proposed to consider only the sampling instances
of the noise and that consequently, the model denoted as Mθ
is then defined as the following hybrid structure [18]:
A(p, ρ, θ)χ(t, x) = B(p, ρ, θ)u(t, x)
Mθ :
(10)
y(tk , xℓ ) = χ(tk , xℓ ) + e(tk , xℓ )
Furthermore, for clarity’s sake, it is considered that the
additive noise e(xℓ , tk ) is a zero mean, random spatial array,
with no temporal or spatial correlation and that vo (tk , xℓ )
also fulfills these assumptions. In order to handle the more
complicated case of a filtered DT noise, some hints can be
found in [6].
Even though the system behavior is completely characterized by the LTSI vector parameter θ in the model structure
defined in (4), the quality assessment of a given estimator can
not be straight-forwardly defined in the CTS-LPV setting.
A solution to this problem is developed in the next section
where a system reformulation based on [6] is proposed,
allowing the definition of a prediction error minimization
framework in the CTS-LVP-PDE case.
III. R EFORMULATION
now expressed as
(i ,i )
χ̊l t x (t, x) = fl (ρ)χ̊(it ,ix ) (t, x)
(13a)
{it , ix , l} ∈ {0 · · · nt , 0 · · · nx , 1 · · · nα }
u(jt ,jx ) (t, x) = gl (ρ)u(jt ,jx ) (t, x)
(13b)
l
{jt , jx , l} ∈ {0 · · · mt , 0 · · · mx , 0 · · · mβ }
Therefore, the process part of the LPV model is rewritten
as a MISO system with ((nt + 1)(nx + 1) − 1)nα +
,nx ,nα
(mt + 1)(mx + 1)(mβ + 1) inputs {χlit ,ix }intt=0,i
and
x =0,l=1
mt ,mx ,mβ
l
{ujt ,jx }jt =0,jx =0,l=0 . By using (11), (1) can be rewritten in
terms of the sampled output signal y(tk , xℓ ) as
nt ,nx
(it ,ix )6=(nt ,nx ) nα
X
X åli ,i (i ,i )
t x
χ̊l t x (tk , xℓ )
y(tk , xℓ ) = −
F
(p)
o
it ,ix =0
l=1
(14)
mβ l
mt ,mx X
X
b̊jt ,jx (jt ,jx )
u
(tk , xℓ ) + eo (tk , xℓ ).
+
Fo (p) l
j ,j =0
t
x
l=0
At this point, it must be noticed that despite the apparent
writing complexity, equations (14) and (1) are equivalent.
Therefore the model to be identified is still given by equation
(10). The gain here is the LTSI formulation of the LPVPDE system which consequently enables the use of a widely
spread theory on linear systems for the expression of the optimization in the prediction error sense. The obvious drawback
is the lack of a priori knowledge on noise-free outputs χ̊it ,ix .
A solution to cope with this issue is proposed in Section IV.
However, by momentarily assuming the knowledge of χ̊it ,ix ,
the prediction error can be expressed as
A PEM cannot be directly applied to identify a model
described by (10) because of the non-commutativity of the
LPV polynomials. This problem is well-known when dealing
with LPV models. In [6], the solution proposed to identify
a LPV-ODE model is therefore to rewrite the LPV system
as a multi-input single-output (MISO) LTI model for which
the PEM can be applied.
εθ (tk , xℓ ) = y(tk , xℓ ) − ŷθ (tk , xℓ )
(15)
The same idea is used here for LPV-EDP model. Indeed,
if the system belongs to the model set defined in (10), the where ŷθ (tk , xℓ ) is the one-step-ahead predictor [12]:
LPV-PDE model can be expressed as a CTS-MISO model
nt ,nx
(it ,ix )6=(nt ,nx ) nα
X
X ali ,i
by rewriting the data-generating system (1) as the following
(i ,i )
t x
χ̊l t x (tk , xℓ )
ŷθ (tk , xℓ ) = −
LTSI representation
F
(p,
θ)
it ,ix =0
l=1
(16)
nt ,nx
mβ
mX
t ,mx X
l
(it ,ix )6=(nt ,nx )
b
j
,j
(j
,j
)
X
t
x
t
x
u
(tk , xℓ ) .
+
(nt ,nx )
(t, x) +
å0it ,ix χ̊(it ,ix ) (t, x)
F (p, θ) l
χ̊
j
,j
=0
l=0
t x
it ,ix =0
|
{z
}
Based
on these considerations, the following section
Fo (p)χ̊(t,x)
presents
a
LS-based identification method which provides
nt ,nx
(it ,ix )6=(nt ,nx ) nα
X
X
efficient
estimates
i.e:
+
ålit ,ix fl (ρ)χ̊(it ,ix ) (t, x) (11)
N,L
|
{z
}
1 X 2
it ,ix =0
l=1
(it ,ix )
θ̂
=
arg
min
εθ (tk , xℓ ),
(17)
χ̊
(t,x)
l
θ∈Rnθ N L
mβ
mX
t ,mx X
k,ℓ=1
=
b̊ljt ,jx gl (ρ)u(jt ,jx )(t,x)
for data generating systems S written as in (1) and such that
{z
}
|
jt ,jx =0 l=0
(j ,j )
S ∈ Mθ is informative with respect to Mθ .
u t x (t,x)
l
where g0 (ρ) = 1 and the superscript (jt , jx ) for a signal,
like u(jt ,jx ) , denotes the jtth partial derivative in time and
the jxth partial derivative in space and
y (nt ,nx ) (tk , xℓ ) = ϕT (tk , xℓ )θo + Fo (p)eo (tk , xℓ )
nt ,nx
(it ,ix )6=(nt ,nx )
Fo (p) = pnt ,nx +
X
IV. I TERATIVE LEAST SQUARES FOR LPV-EDP SYSTEMS
Using the LTSI system reformulation (14), y(tk , xℓ ) can
be rewritten in the following regression form:
å0it ,ix pit ,ix
(12)
it ,ix =0
Note that in this way, the main advantage is to transpose
the time-space variation of the coefficients onto the signals
(18)
where,
ϕ(tk , xℓ ) = −y(tk , xℓ ) · · · − y (nt ,nx −1) (tk , xℓ )
(0,0)
(n ,n −1)
−χ̊1 (tk , xℓ ) · · · − χ̊nαt x (tk , xℓ )
(0,0)
(mt ,mx )
T
u0 (tk , xℓ ) · · · umβ
(tk , xℓ ) ]
(19)
h
iT
mβ
0
α
θo = å00,0 · · ·å0nt ,nx −1 å10,0 · · ·ån
(20)
nt ,nx −1 b̊0,0 · · · b̊mt ,mx
In the given context, contrary to the LPV formulation
where the LPV dependency appears into the signal definitions, the filter Fo (p) commutes and allows the rewritten of
(18) as follows
(nt ,nx )
yf
(tk , xℓ ) = ϕT
f (tk , xℓ )θo + eo (tk , xℓ )
(i ,i )
(21)
(j ,j )
where yf (tk , xℓ ), χ̊l,ft x (tk , xℓ ) and ul,ft x (tk , xℓ ) represent the outputs of a prefiltering operation.
1
y(tk , xℓ )
(22)
yf (tk , xℓ ) =
Fo (p)
and
h
(n ,n −1)
ϕf (tk , xℓ ) = −yf (tk , xℓ ) · · · − yf t x (tk , xℓ )
(0,0)
(n ,n −1)
−χ̊1,f (tk , xℓ ) · · · − χ̊nαt,f x (tk , xℓ )
(0,0)
(m ,m )
T
u0,f (tk , xℓ ) · · · umβt,f x (tk , xℓ ) ]
(23)
It has to be noticed that the interest of the filter Fo (p) is
twofold. Firstly it is used to filter the input/output signals
used for the identification to come up with the regression
equation (21) written in terms of white noise eo (tk , xℓ ). In
other words, Fo (p) is used to filter the data such that a simple
LS algorithm applied to these pre-filtered data leads to the
statistically optimal estimate under the condition S ∈ Mθ .
Secondly, another important particularity of this method is
to provide directly a CTS model, therefore the regressor
ϕf (tk , xℓ ) (23) is built up form the time-space derivatives of
the data (and not from the delayed version of these data as in
the DTS case). The main problem to be solved is that these
derivatives are unknown and has therefore to be estimated.
This well-known problem is usually solved by using low
pass filtering on the input and output signals, as e.g. statevariable-filters. However these filters present the drawback
of requiring the critical choice of a design variable. On the
contrary, the particular advantage of the presented method is
to be free of this drawback since the filter Fo (p) is not only
used for the prediction error minimization but also for the
approximation of the unknown derivatives.
As a result, the filtered regressor ϕf (tk , xℓ ) in (21) is
directly numerically approximated instead of the regressor
ϕ(tk , xℓ ) in (18) which makes use of unknown time-space
derivatives and still requires a pre-filtered operation to access
the space-time derivatives, with a filter which has to be
chosen by the user.
Consequently, under the assumption of a priori known
Fo (p) and χ̊(it ,ix ) , the optimal estimator in the prediction
error minimization sense is the LS estimator
−1
N,L
X
·
θ̂LS =
ϕf (tk , xℓ )ϕT
f (tk , xℓ )
k,ℓ=1
N,L
X
k,ℓ=1
(24)
(n ,n )
ϕf (tk , xℓ )yf t x (tk , xℓ )
(it ,ix )
Naturally, the filter Fo (p) and the signals χ̊
are
unknown and therefore and iterative procedure is proposed
in order to cope with this issue, similarly as it is achieved
in the approaches developed for the ODE case in [21], [22].
From Eq. (24) two comments may be formulated. Indeed
the identification problem defined by Eq. (21) has been
chosen to be solved by an iterative LS method. Alternatives
could have also been chosen, as e.g. more advanced PEM
method [12]. However, these latter need a robust initialization
which seems not easy to handle in this CTS-PDE-LPV model
identification case. Moreover, the iterative solution appears
as a natural choice since the proposed method involves an
iterative solution (or relaxation) algorithm in which, at each
iteration, an auxiliary model is used to generate a part of the
regressor ϕf (tk , xℓ ) as well as the associated pre-filters.
As a result the regressor used for the identification is
not only based on reconstructed signals but on real input,
measured output, and reconstructed output.
LPV-LS-PDE algorithm:
The proposed LPV-LS-PDE model identification algorithm
can be summarized as shown below. The algorithm consists
of three first steps for the initialization and five iterative
steps used to estimate the filter, the derivatives and finally
the model parameters.
•
•
Step 1 Initialization of θ0 for l = 0 only:
iT
h
θ̂0 = â00,0 · · · â0nt ,nx −1 0 · · · 0b̂00,0 · · · b̂0mt ,mx 0 · · · 0
using for instance the LS-PDE method proposed in [19].
In this first step the system is considered as a LTSI
system.
Step 2 Compute a first estimated continuous-filter on
the basis of the estimates obtained in Step 1
1
Q(p, θ̂0 ) =
F (p, θ̂0 )
where F (p, θ̂0 ) is given as (12).
Use the filter in order to generate the estimates of the
derivatives
(j ,jx ) mt ,mx ,mβ
}jt =0,jx =0,l=0
{ul,ft
(it ,ix ) nt ,nx
}it =0,ix =0
{yf
(jt ,jx ) mt ,mx ,mβ
}jt =0,jx =0,l=0
= Q(p, θ̂0 ){ul
,nx
= Q(p, θ̂0 ){y (it ,ix ) }intt=0,i
x =0
and the regressor ϕf (tk , xℓ ) built up from the filtered
noisy output only and the input for this initialization
step
h
(n ,n −1)
ϕf (tk , xℓ ) = −yf (tk , xℓ ) · · · − yf t x (tk , xℓ )
(0,0)
(n ,n −1)
−y1,f (tk , xℓ ) · · · − ynαt,f x (tk , xℓ )
(0,0)
(m ,m )
T
u0,f (tk , xℓ ) · · · umβt,f x (tk , xℓ ) ]
•
step 3 Compute an estimate θ̂1 :
θ̂1 =
−1
N,L
X
k,ℓ=1
ϕf (tk , xℓ )ϕT
f (tk , xℓ )
·
N,L
X
k,ℓ=1
(nt ,nx )
ϕf (tk , xℓ )yf
(tk , xℓ )
Consequently, the initial estimate proposed for the
LPV-LS-EDP algorithm is an LTSI-LS-EDP estimate
of Mθ . Set τ = 1. [End of initialization part.]
analysis, the boundary and initial conditions are assumed to
be zeros.
V. E XAMPLE
•
Step 4 At any iteration τ +1, compute an estimate of the
ˆ k , xℓ ) via numerical approximation
noise-free output χ̊(t
of
ˆ x) = B(p, ρ, θ̂τ )u(t, x)
A(p, ρ, θ̂τ )χ̊(t,
based on the estimates θ̂τ at the previous iteration.
Step 5 Compute the estimated continuous filter
Q(p, θ̂τ ) =
1
F (p, θ̂τ )
ˆ k , xℓ ) in order to generate
and use the filter as well as χ̊(t
the estimates of the derivatives which are needed:
(it ,ix ) nt ,nx ,nα
}it =0,ix =0,l=1
ˆ (it ,ix ) }nt ,nx ,nα
= Q(p, θ̂τ ){χ̊
l
it =0,ix =0,l=1
(j ,jx ) mt ,mx ,mβ
}jt =0,jx =0,l=0
= Q(p, θ̂τ ){ul
ˆ
{χ̊
l,f
{ul,ft
(it ,ix ) nt ,nx
}it =0,ix =0
{yf
•
(jt ,jx ) mt ,mx ,mβ
}jt =0,jx =0,l=0
,nx
= Q(p, θ̂τ ){y (it ,ix ) }intt=0,i
x =0
Step 6 Build the filtered estimated regressor ϕ̂f (tk , xℓ )
as:
h
(n ,n −1)
ϕ̂f (tk , xℓ ) = −yf (tk , xℓ ) · · · − yf t x (tk , xℓ )
ˆ (0,0) (tk , xℓ ) · · · − χ̊
ˆ (nt ,nx −1) (tk , xℓ )
−χ̊
1,f
nα ,f
(0,0)
(m ,m )
T
u0,f (tk , xℓ ) · · · umβt,f x (tk , xℓ ) ]
•
ϕ̂f (tk , xℓ ) vector is composed of the filtered noisy
output, the filtered noise-free estimated output and the
input
Step 7 Compute the LPV-LS-PDE estimate θ̂τ +1 at the
(τ + 1)th iteration:
θ̂
τ +1
=
−1
N,L
X
k,ℓ=1
ϕ̂f (tk , xℓ )ϕ̂T
f (tk , xℓ )
•
N,L
X
k,ℓ=1
·
(n ,n )
ϕ̂f (tk , xℓ )yf t x (tk , xℓ )
Step 8 If θ̂τ +1 has converged or the maximum number
of iterations is reached, then stop, else increase τ by 1
and go to Step 4.
Comments on the implementation issue:
In this paper, we considered the practical feasible situation such that only sampled measurements of the spacetime continuous signals (y, u, ρ) are available. In order to
apply a continuous filter on sampled data one can either
interpolate the samples to obtain a continuous signal and
apply the continuous filter on this reconstructed signal or use
a numerical approximation. This is a common problem for
simulation of continuous systems. For simulation purposes,
discrete approximation of the system can efficiently be dealt
with by using available powerful numerical algorithms. Note
in the following example, an Euler method is used to simulate
the true system and the filtering operations. In the above
The performance of the described algorithm is presented
on a representative simulation example. As an example, the
advection-diffusion equation (ADE) is used. This equation is
often used in the water resource quality analysis to describe
the transport and dispersion of a solute (pesticide, pollutant,
...) into a river channel (see e.g. [20]). The ADE can be
written as following:
∂ 2 χ̊(t, x)
∂χ̊(t, x)
= å02 (ρ(t))
∂x2
∂t
∂χ̊(t, x)
(25)
−å01 (ρ(t))
+ b̊00 (ρ(t))u(t, x)
∂x
χ̊(t = 0, x) = 0
y(tk , xℓ ) = χ̊(tk , xℓ ) + eo (tk , xℓ )
where the noise-free signal χ̊(xℓ , tk ) (Fig. 1) represents
the solute propagation in the river and is obtained by
discretization of the partial differential equation by a finite
difference method (see e.g. [17]). the temporal sampling is
Ts = 1day and N = 3650 temporal points are measured.
The spatial sampling is h = 20m for L = 50 space points.
These 50 points can represent 50 sensors distributed in the
river. u(t, x) describes the source of the pollutant, and is
represented by a PRBS signal at the space point x = 20m.
In realistic conditions of simulation the a02 (ρ(t)) diffusion
0
18
100
16
200
14
300
distance in meter
•
12
400
10
500
8
600
700
6
800
4
900
2
1000
0
0
500
1000
1500
2000
2500
3000
3500
Time in day
Fig. 1.
Noise-free output χ̊(t, x)
and the a01 (ρ(t)) advection coefficients are dependent of
the river flow Q(t) (Fig. 2). The scheduling variable ρ is
then taken as ρ(t) = Q(t) and the coefficients a02 (ρ(t)) ,
a01 (ρ(t)) and b00 (ρ(t)) are given by
(26a)
å02 (ρ(t)) = 2 + 0.9ρ(t)
å01 (ρ(t)) = 0.5 + 0.2ρ(t)
(26b)
b̊00 (ρ(t)) = 1
(26c)
In this example, realistic data are used since the river flow
considered here comes from measurements of a river located
in Alsace (France). 10 years of measurements are used in this
example. For model identification purposes, the output signal
is supposed to be corrupted by a two dimensional, zero-mean
R EFERENCES
120
100
80
60
40
20
0
0
500
1000
1500
2000
2500
3000
3500
Time in day
Fig. 2.
Flow river Q(t) in m3 /s
and normally distributed, discrete-time noise signal, with a
Signal-to-Noise Ratio (SNR) defined as
Pχ̊
(27)
SN R = 10log10
Pe
where Pe and Pχ̊ are the average power of the noise
and deterministic output.To provide representative results,
a Monte Carlo simulation analysis for 100 runs is used to
illustrate the performance of the approach. In Table I and II,
it can be noted that the estimates of the proposed method
are unbiased with a relatively small standard deviation when
SN R = 30dB. Moreover, in the case where SN R = 15dB,
the estimates are again unbiased with a higher standard
deviation.
TABLE I
MC SIMULATION RESULTS FOR 100 RUNS ( DIFFUSION COEFFICIENTS )
True value
SNR
mean
std
â002
2
30dB
2.0218
0.1071
â102
0.9
15dB
2.0149
0.43073
30dB
0.8922
0.0446
15dB
0.8695
0.34926
TABLE II
MC SIMULATION RESULTS FOR 100 RUNS ( ADVECTION COEFFICIENTS )
True value
SNR
mean
std
â001
0.5
30dB
0.4993
0.0037
â101
0.2
15dB
0.4998
0.0165
30dB
0.2002
0.0014
b̂000
1
15dB
0.2002
0.0059
30dB
1.0003
0.0016
15dB
0.9981
0.0195
VI. C ONCLUSION
A method has been proposed in this paper to identify
a continuous space-time LPV partial differential equation
model. The approach is based on a MISO linear time/space
invariant reformulation of the data generating system along
with an iterative procedure develop to estimate efficiently
the model parameters into the error-prediction minimization
framework. A realistic example has been used to illustrate
the efficiency of the proposed method. Future research investigates the use of an instrumental variable method to handle
the situation where the additive noise is colored.
[1] S. J. Moura, N. A. Chaturvedi, and M. Krstić, ”PDE estimation
techniques for advanced battery management systems - Part I: SOC estimation”, Proceedings of the American Control Conference, Montreal
(QC), pp: 559 - 565, 2012.
[2] H. Fang, J. Wang, E. Feng, and Z. Li, ”Parameter identification and
application of a distributed parameter coupled system with a movable
inner boundary”, Computers and Mathematics with Applications, vol.
62, pp: 4015-4020, 2011.
[3] M. Ali, A. Ali, H. Abbas, and H. Werner, ”Identification of BoxJenkins models for parameter-varying spatially interconnected systems”, American Control Conference, San Francisco (USA), pp: 145150, 2011.
[4] G. Belforte, F. Dabbene, and P. Gay, ”LPV approximation of distributed parameter systems in environmental modelling”, Environmental Modelling & Software, vol. 20, pp: 1063-1070, 2005.
[5] V. Puig, J. Quevedo, T. Escobet, P. Charbonnaud, and E. Duviella,
”Identification and control of an open-flow canal using LPV models”,
Proceedings of the 44th IEEE Conference on Decision and Control,
and the European Control Conference, Seville (Spain), pp: 1893-1898,
2005.
[6] V. Laurain, R. Tóth, M. Gilson, and H. Garnier, ” Direct identification
of continuous-time linear parameter-varying input/output models”, IET
Control Theory & Applications, vol. 5(7), pp: 878-888, 2011.
[7] J. Schorsch, H. Garnier, and M. Gilson, ”Instrumental variable methods for identifying partial differential equation models of distributed
parameter systems”, In IFAC Symposium on System Identification, pp:
840–845, Brussels (Belgium), 2012.
[8] R. Pintelon, J. Schoukens, and Y. Rolain, ”Box-Jenkins continuoustime modeling”, Automatica, vol. 36(7), pp: 983-991, 2000.
[9] R. Johansson, ’Identification of continuous-time models”, IEEE Transactions on Signal Processing, vol. 42(4), pp: 3887-3897, 1994.
[10] R. Tóth. Modeling and Identification of Linear Parameter-Varying
Systems. Lecture Notes in Control and Information Sciences, Vol.
403. Springer-Germany, 2010.
[11] V. Laurain and M. Gilson and R. Tóth and H. Garnier, ”Refined
Instrumental Variable Methods for Identification of LPV Box-Jenkins
Models”, Automatica, vol. 46(6), pp: 959-967, 2010.
[12] L. Ljung. System Identification, theory for the user. Prentice Hall,
1999.
[13] M. Ali and S. S. Chughtai, and H. Werner, ”Consistent identification of
two-dimensional systems”, American Control Conference, Baltimore
USA, pp: 3464-3469, 2010.
[14] M. S. Sadabadi, M. Shafiee, and M. Karrari, ”System identification of
two dimensional continuous-time systems using wavelets as modulating functions”, ISA Transactions, vol. 47, pp: 256-266, 2008.
[15] C. Chochol, S. Chesne, and D. Remond, ”An original differentiation
tool for identification on continuous structures”, Journal of Sound and
Vibration, 2013.
[16] S. Sagara, Z. J. Yang, and K. Wada, ”Parameter identification of
distributed-parameter systems in the presence of measurement noise”,
Int. Journal of Systems Science, vol. 22(8), pp: 13911401, 1991.
[17] K. W. Morton and D. F. Mayers, Numerical Solution of Partial
Differential Equations, Cambridge University Press, 2005.
[18] H. Garnier and L. Wang (Eds.), editor, Identification of continuoustime models from sampled data, Springer-Verlag, London, 2008.
[19] J. Schorsch, H. Garnier, M. Gilson, and P. C. Young, ”Instrumental
variable methods for identifying partial differential equation models”,
International Journal of Control, in press, 2013.
[20] W. Tych and P.C. Young, ”A matlab software framework for dynamic
model emulation”, Environmental Modelling and Software, vol. 34,
pp: 19-29, 2012.
[21] P. Lohnberg and G. H. J. Wisselink, ”Iterative Least-Squares Parameter
Estimation for ARMA Pulse Response and Output Disturbance”, IEEE
Transactions on Automatique Control, 27(6), pp: 1252-1255, 1982.
[22] F. Dinga, Y. Shi, and T. Chen, ”Auxiliary model-based least-squares
identification methods for Hammerstein output-error systems”, Systems & Control Letters, vol. 56, pp: 373 380, 2007.