Anda di halaman 1dari 26

Computers & Fluids 33 (2004) 3156

www.elsevier.com/locate/compuid

Extension of ENO and WENO schemes to


one-dimensional sediment transport equations
 rnjaric-Z
 ic, S. Vukovic *, L. Sopta
N. C
Faculty of Engineering, University of Rijeka, Vukovarska 58, 51000 Rijeka, Croatia
Received 19 June 2002; received in revised form 12 February 2003; accepted 18 February 2003

Abstract
Essentially nonoscillatory and weighted essentially nonoscillatory schemes are high order resolution
schemes constructed for the hyperbolic conservation laws. In this paper we extend these schemes to the onedimensional bed-load sediment transport equations. The diculties that arise in the numerical modelling
come from the fact that a nonconservative product is present in the system. Our specic numerical approximations for the nonconservative product are based on two ideas. First is to include the inuence of
that term in the system upwinding and the second is to dene the numerical approximation in such a way
that the obtained scheme solves the system for the quiescent ow case exactly. As a consequence, the resulting schemes give excellent results, as it can be seen from the numerical tests we present. On the opposite,
the numerical results obtained by applying the pointwise evaluation of nonconservative product on the
same tests present unacceptably large numerical errors.
2003 Elsevier Ltd. All rights reserved.
AMS: 65M06; 76M20; 35L65
Keywords: ENO and WENO schemes; Bed-load sediment transport; Nonconservative product; Exact conservation
property; Conservation law

1. Introduction
In this paper we are concerned with the construction of a numerical approximation for the
bed-load sediment transport equations. The rst step is the choice of the mathematical model

Corresponding author. Tel.: +385-51-651481/651494; fax: +385-51-675818.


 rnjaric-Z
 ic), senka.vukovic@ri.hinet.hr (S. Vukovic), luka.sopta@riteh.hr
E-mail addresses: nelida@riteh.hr (N. C
(L. Sopta).
0045-7930/04/$ - see front matter 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0045-7930(03)00032-X

32

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

governing this physical phenomenon. For this there are several approaches; for example Brufau
et al. [16] modelled this problem as a debris ow. Here, we follow the model proposed by
Hudson [7].
Here we must point out that the considered system contains a nonconservative product, so the
rst goal in this paper is to nd the appropriate way of incorporating the nonconservative product
term in the numerical scheme. In particular, in dealing with the problem of the presence of a
nonconservative product term we follow Castro et al. [13]. Also, a similar approach has been
proposed by Gosse [17]. The fundamental idea is that when in numerical scheme we compute
characteristic elds locally, we use a local approximation of the original system, which is hyperbolic and linear and whose Jacobian matrix also includes the nonconservative product term.
Then we proceed by modifying the well known essentially nonoscillatory (ENO) and weighted
ENO (WENO) schemes. The central idea of ENO and WENO schemes is to compute the
smoothness for several possible stencils which cover a given region of space and then interpolate
this zone with high order of accuracy using the ENO (resp. WENO) reconstruction via primitive
function developed by Harten, Osher, Shu, Enquist and Chakravarthy [35] (resp. Osher, Liu,
Chan [11], Shu [12] and Jiang and Shu [18]). High order of accuracy in time is achieved by using
RungeKutta type integration. These schemes are shock capturing, without spurious oscillations
and achieve high order of accuracy in the regions of smoothness. The framework for here presented new schemes are modied ENO and WENO schemes, which we developed and applied to
the shallow water equations in [14]. The main idea of those schemes is to obtain balancing between the ux gradient and the source term [6,9] based on the concept of the exact conservation
property (C-property), which was introduced by Berm
udez and Vazquez [1,2]. Here we extend this
concept to the bed-load sediment transport equations and create schemes that achieve balancing
between the ux gradient and the nonconservative product term.
In Section 2 we give the mathematical model for the bed-load sediment transport phenomenon
[7,8] that we use in this paper. In Section 3 we present the new schemes and in Section 4 we prove
they validate the exact C-property [2], which is here appropriately dened for the observed
conservation law system. We also make a few observations about the accuracy of the proposed
numerical schemes. Finally, in Section 5 we present numerical results for dierent test problems
that are proposed by Hudson [7].
2. One-dimensional sediment transport equations
In the one-dimensional sediment transport system the mathematical model for the moving
water is taken to be the one-dimensional shallow water system, which consists of the mass conservation law
1
ot h ox hv 0
and the momentum balance law


ot hv ox hv2 12gh2 ghox H :

In the above equations x is the space coordinate along the channel axis, t is the time variable, h is
the water depth, v is the water velocity, g is acceleration due to gravity, and H is the bed level (see
Fig. 1). When the bed level H is xed, i.e., H H x, the system (1) and (2) is strictly hyperbolic.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

33

Fig. 1. The bed-load sediment transport.

When the sediment transport phenomenon occurs, the bed level becomes a function of the time
variable too and the equation that describes the bed-load sediment transport must be added
ot H nox q 0:

Here the parameter n is given with


1
;
4
1
where  is the porosity of the riverbed, and it depends on the type of the sediment (see [7]). The
term q qv; h is the total volumetric sediment transport rate in the x-direction, i.e., the sediment
transport ux. One possible mathematical model for the sediment transport ux is the simple law
n

qv Avjvjm1 ;

discussed by Grass [15]. Here the parameter A depends on the type of the sediment, while the
parameter m is bounded with 1 6 m 6 4. From (5) it is obvious that a smaller value of a parameter
A means that a riverbed interacts slower with water ow. Some of the experimentally obtained
values for A can be found in [7]. There are also other experimentally obtained laws for the sediment transport ux, for example van Rijns law [10], but their application is out of the scope of
this paper.
As we already pointed out in the Introduction, the problem that arises after Eq. (3) is added to
the system (1) and (2), is that the right side is not the standard source term any more. In fact, it
includes the derivations of the variables, so it is a nonconservative product. The system (1) and (3)
can be rewritten in a homogeneous form as
ot h ox hv 0;
ot v ox 12v2 gh H 0;
ot H nox q 0;
but then the variables are not conserved. In order to avoid this we propose the following formulation

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

34

ot u ox f gx; u; ox u;

where the conserved variables vector u, the ux f, and the nonconservative product g are given as
0
1
0
1
0 1
0
h
hv
7
u @ hv A; f @ hv2 12 gh2 A; and g @ ghox H A:
nq
0
H
Naturally, the rst step in our construction of a numerical approximation is the local linearization of the system, and for that we need to compute the Jacobian matrix of the ux
0
1
0
1 0
of @
A
8
gh  v2 2v 0 A:
ou
vd
d 0
Here d Amn=hvm1 for v P 0. The obvious singularity of the matrix A could cause diculties in the numerical approximations. However, this is not a problem here. In fact, since the
nonconservative product g includes the derivatives of the variables, it is quite natural to include
this term in the local linearization. Therefore we propose to write the nonconservative product in
the form
0
1
0 0
0
9
g Box u where B @ 0 0 gh A:
0 0
0
Here we want to point out that if the matrix B was constant, the system (6) and (7) would be a
homogeneous hyperbolic conservation law system with the ux equal to f  Bu. Obviously, this is
not true in our case. However, the following linearization of the system (6) and (7)
ot u Aox u Box u
can be considered locally and
troduce
0
0
e A  B @ gh  v2
A
vd

10
this is enough for the numerical approximation. We can now in1
2v
d

1
0
gh A;
0

11

which is exactly the matrix that corresponds to the linearization (10). Furthermore, we can prove
in the same way as it was done in [7], that under the assumptions
vx; t P 0

12

and
hx; t > 0 for vx; t > 0;
13
e is nonsingular with real and dierent eigenvalues which can be evaluated from the
the matrix A
characteristic polynomial
P k; u k3  2vk2  kghd gh  v2  ghvd:

14

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

For the eigenvalues denoted with kp , the corresponding right eigenvectors are
0
1
1
B
C
kp
C;
rp B
p 2
@ v  k
A
1
gh
while the left eigenvectors can be written as
0 r s
1
2
k
k

v

gh
1
@
A
lp r
kp
p
k  k ks  kp
gh

35

15

16

for p; r; s 1; 2; 3 and r; s 6 p. With this we solved the task of computing local characteristic
elds, and we can move to the construction of our schemes, which are in fact an extension of
standard ENO and WENO schemes.
There is also the possibility of using the formulation in which the momentum balance law (2) is
rewritten in the form
ot v ox hv2 12gh2 ghH gH ox h:
Following [7], we refer to this formulation as the C-formulation. If we apply the approach explained in the previous paragraph to this formulation, we would actually obtain the identical
linearization as for the formulation (6) and (7). Moreover, the proposed extended numerical
schemes will also coincide on both the formulations.

3. Extended ENO and WENO schemes


In order to apply the ENO and WENO schemes to system (6) and (7) we rewrite it in the form
ot u Lu; x; t

17

with L dened as
Lu; x; t ox fu gx; u; ox u:

18

For the time integration of (17) we use the standard approach in ENO and WENO schemes,
i.e., a RungeKutta type method [35,11,12,18], so we have to propose a numerical approximation of (18).
In particular, if an approximation ui ; i 0; . . . ; N for the solution u at any time t is known, we
propose to take
Li 

1
f 1  f i12 gi ;
Dx i2

19

i 0; . . . ; N. Here f i1 stands for the numerical ux on the i 12th cell boundary. gi is numerical
2

approximation of the nonconservative product at the ith cell, and the ith cell is dened as
xi12 ; xi1 with the center at xi .
2

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

36

Furthermore, for the numerical approximation of (19) we need local characteristic elds, i.e.,
eigenvalues k

p
1,
i
2

p
1,
i
2

left eigenvectors l

p
1,
i
2

and right eigenvectors r

i 0; . . . ; N  1, p 1; 2; 3,

must be found. Since they are given with (15) and (16), we just have to chose a state average at
e in that state. Our choice is the van Leer
i 12th cell boundary in order to evaluate the matrix A
average
ui ui1
20
ui12
2
and the reason for that will become clear in Section 4.
Then, we use the ENO or WENO reconstruction algorithm. Generally, ENO and WENO
reconstruction solves the problem of numerically approximating a function v
at position xi1 , if
2

the values v

k ; k 0; . . . ; N; of this function at the cell centers xk ; k 0; . . . ; N ; are known. Here we


present only three essential facts about ENO and WENO reconstruction (Eqs. (21), (23), and
(24)), because we use them in this paper. The complete presentation of these algorithms can be
found in [35,11,12,18].
First fact is that the nal result of both the r 1th order ENO and the 2r 1th WENO
reconstruction on a uniform mesh of cells can be written in the form
v

i12;r

ss
max
X

r;s v

ss

min

r
X

r;s;j virsj ;

s s

min ; . . . ; smax :

21

j0




Here the range of indices is bounded with s


min 0, smax r, or with smin 1, smax r 1, and vk

as stated above is the value of the function v at xk , k 0; . . . ; N. The coecients ar;s;j ; j

0; . . . ; r; s s

min ; . . . ; smax , have known values that for example can be found in [12] and that are

a consequence of the polynomial reconstruction of the function v


. Each weight x

r;s v ;

s smin ; . . . ; smax , is computed from some smoothness indicator that measures smoothness of the
function v
over the sth stencil of points

fxirs ; . . . ; xis g; s s

Sr;s
min ; . . . ; smax :

22

The only dierence between the ENO and WENO algorithm is in the evaluation of the weights. In
the ENO reconstruction only one stencil is chosen for the reconstruction, i.e., the weight belonging to this stencil is equal 1 and all the other stencils have zero weight. In the WENO reconstruction the weight values are distributed more evenly and normalized, i.e., the sum of
weights is equal 1. Even more, in both cases

ss
max
X

ss

min

r;s v

r
X

r;s;j 1:

23

j0

Eq. (23) is the second fact that we want to emphasize. The last fact regarding ENO and WENO
reconstructions that we need is that values of the weights do not change if any constant value is
added to the function, i.e.,

r;s v const xr;s v ; s smin ; . . . ; smax :

24

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

37

This can be easily veried in Refs. [35,12,18] and it is a logical property of the weights since
smoothness of a function does not change with translation.
Now, we can explain how the ENO or WENO reconstruction algorithms can be applied to the
evaluation of the numerical ux. Actually, we present the locally LaxFriedrichs formulation
(LLF) and the Roe with entropy x formulation (RF) of the ENO and WENO schemes in a form
that is not common. The transition from the classical to the formulations we use can be found in [14].
3.1. Numerical approximations of the ux
There are two main ideas in the ENOLLF and WENOLLF schemes that are used to evaluate
the numerical approximation of the ux component in the pth characteristic eld on the i 12th
cell boundary. First is the ux splitting
p

p;

p;

fi1 fi1 fi1 ;


2

where uxes on the right side of the above equation are dened in such a way that their Jacobian
matrix is positive resp. negative dened. In the considered formulation the LaxFriedrichs ux
splitting is used. The second crucial thing is to evaluate numerically the approximation for terms
p;
p;
f 1 and f 1 using the ENO or WENO reconstruction algorithm for the functions
i
2

i
2

1
p
p
f ai1 u li1
2
2
2
p
1 is a numerical
i
2
p
p
ai1 jki1 j:
2
2

Here a

and

1
p
p
f  ai1 u li1 :
2
2
2

approximation for maxxi1 ;xi1 jkp j. Actually, we just take

As we showed in [14] a reformulation of the stated expressions can be made and we can write

1
p
p
p
p
p
f i f i1  ai1 ui1  ui li1 Pi1; Pi1; ;
fi1
25
2
2
2
2
2
2
p
where the terms P 1 are the high order polynomial corrections which are obtained as values of
i ;

the ENO or WENO reconstruction for functions vp;


dened with

1
p
p
p
vp;

f
ai1 u  f I

ai1 uI
li1 :
2
2
2
2
Here I i, I  i 1. From (21) it follows that
p
Pi1;

ss
max
X

ss

min

p;

r;s v

r
X

26

p;

r;s;j virsj :

27

j0
p p

In the case of the RF formulation, if ki ki1 6 0, the entropy x is adopted and the value f
p p
ki ki1

p
1
2

is

obtained by using the LLF formulation. Otherwise, i.e., if


> 0, the numerical ux component in the pth characteristic eld on the i 12th cell boundary is obtained by Roe formulation
using the ENO or WENO reconstruction algorithm for the function

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

38
p

1
sgnki1
2

f li1 :

2
2
Finally, as in the LLF case, we can write
8
p
< f i lp1 P p1 ;
if ki1 > 0;
i2
i2;
p
2
fi1
: f i1 lp1 P p1 ; otherwise:
2
i
i ;
2

The terms

p
P 1
i ;

28

are again the polynomial corrections that are computed using the ENO or

WENO reconstruction algorithm, i.e., using Eq. (27), but now for functions vp;
dened with
p

p;

1
sgnki1
2

f  f I
li1 :

29

3.2. Numerical approximations of the nonconservative product


In order to complete the denition of our extended ENO and WENO schemes, a numerical
treatment for the nonconservative product g Box u over the ith cell must be dened. The
pointwise evaluation, where the term ox uji is obtained by some numerical dierentiation formula
that approximates it to the desired order of accuracy, produces very poor numerical results, as it
can be seen from the numerical tests we present in Section 5. Because of that, we propose to approximate the nonconservative product following the idea of balancing between the ux gradient
and the source term [2,6,9]. Thus, our goal is to nd the discretisation for the nonconservative
product that will ensure that the appropriate equilibria of the underlying mathematical model are
p
maintained by the numerical scheme. The rst step is to decompose the term gi into two parts
p

gi gi1;L gi1;R :

30

Particularly, in the LLF formulation of ENO or WENO schemes, for the approximation in the
pth characteristic eld we propose formulas

1
p
p
p
p
Bi;i1 ui1  ui li1 Qi1; Qi1; ;
gi1;L
31
2
2
2
2
2Dx

1
p
p
p
p
Bi;i1 ui1  ui li1  Qi1;  Qi1; :
gi1;R
32
2
2
2
2
2Dx
Here we introduce
 0

u u00
0
00
b
b
B u ; u B
33
2
b uj ; uk .
and notation Bj;k B
At rst glance it may seem that the rst term in the expressions (31) and (32)

1
p
Bi;i1 ui1  ui li1
34
2
2Dx

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

39

is simply the central rst order approximation for the nonconservative product. Actually, this is
e.
not true because the upwinding of that term is already included in the matrix A
Term (34) balances with the rst term of the numerical ux in expression (25). In order to
obtain the complete balancing between the ux gradient and the nonconservative product additional terms to the expression (34) obviously must be added, and these are the polynomial corp
rections Q 1 in (31) and (32). For evaluation of that correction terms we use the same algorithm
i ;

p
1
i ;

as for the terms P

only now applied to the function wp;


, dened with

b uI
; uuI
 ulp1 :
wp;
B
i

35

In the RF formulation, for the case when LLF approach in the numerical ux evaluation is
p p
used, i.e., when ki ki1 6 0, we compute the nonconservative product term using (30), (31), (32),
and (35). Otherwise, we propose to perform the computations using (30) and the expressions


p
1  sgnki1
1 p
p
p
2
Bi;i1 ui1  ui li1
36
gi1;L
Q 1 ;
2
2
2Dx
Dx i2;

p

gi1;R


p
1 sgnki1
2

2Dx

Bi;i1 ui1  ui li1 


2

Here again the polynomial corrections Q

p
1
i ;

1 p
Q 1 :
Dx i2;

37

must be in balance with P

p
1
i ;

and we propose to

evaluate them using the ENO or WENO reconstruction algorithm for the functions wp;
, dened
with


p
1  sgnki1
2
b uI
; uuI
 ulp1 :
B
38
wp;

i2
2
The last very important change in relation to the standard ENO and WENO algorithms is the
way of evaluating the weights of stencils for the polynomial reconstruction. As in [14] we propose
to evaluate weights of stencils for the dierence of the functions vp;
 wp;
, instead of using just
p
p
values for vp;
. This nally leads to the denitions for P 1 and Q 1
i ;

Pi1;

2

p
Qi1;

ss
max
X

p;

 wp;

r;s v

r
X

ss

min

ss
max
X

ss

min

p;

r;s;j virsj ;

i ;

39

j0

p;

 wp;

r;s v

r
X

p;

r;s;j wirsj :

40

j0

We can conclude that in the dened schemes the inuence of the nonconservative product is
accounted not only in the upwinding of the rst component of the ux, but also in the polynomial
correction evaluations for the ux and for the nonconservative product.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

40

4. The exact conservation property for the sediment transport equations


In this section we prove that the extended versions of ENO and WENO schemes for the bedload sediment transport equations have the exact conservation property (C-property). Following
the original idea of the C-property, which has been dened for the shallow water equations [1,2],
we dene here: a numerical scheme for the bed-load sediment transport veries the exact
C-property if it is exactly compatible with the quiescent ow
h H const: and v 0:

41

First, we see that in the case of the quiescent ow, the ux and the nonconservative product term
reduce to
0
1
0
1
0 1
h
0
0
u @ 0 A; f @ 12 gh2 A; g @ ghox H A;
42
0
0
H
while eigenvalues and right and left eigenvectors become
k1 c;

k2 c;

r1

1
1
1
@ c A;
2c
0

r2

k3 0;
0 1
1
1@ A

c ;
2c
0

43
0

r3

1
1
1
@ 0 A;
2c
1

44

1
0 1
0
1
c
c
0
l1 @ 1 A; l2 @ 1 A and l3 @ 0 A;
45
c
c
c
p
where c gh.
What we must prove now is that if we use the schemes presented in the previous section and if
conditions (41) are true then
Li 0;

46

for i 0; . . . ; N , where Li is dened with (19).


From the denition (20) for the numerical approximation of the variable on the i 12th cell
boundary it follows that in numerical approximations for (43)(45) we must take the value of c as
p
ci;i1 ghi;i1 :
Here the term hi;i1 is evaluated from
hj;k

hj hk
:
2

Let us notice that in both, the LLF and the RF formulations, in the expressions (25) and (28) for
the ux and expressions (30), (31), (32), (36), and (37) for the nonconservative product we can
clearly distinguish two parts. The rst part is the result of the simple upwinding, while the second
one includes the ENO or WENO polynomial corrections. In order to prove the exact C-property,

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

41

for both formulations we separately prove the balancing of the upwind parts and of the polynomial corrections. Particularly, the polynomial corrections are balanced if
p

Pi1;
Qi1;
0:

47

Since the stencil weights x

r;s , s smin ; . . . ; smax , and polynomial interpolation coecients ar;s;j ,

j 0; . . . ; r, s s

min ; . . . ; smax , are the same in the ux and in the nonconservative product, the
p;

p;

and wk . It follows that


only dierence in the above polynomial corrections are the values vk
with the proof of
p;

vk

p;

 wk

48

for p 1; 2; 3, k 0; . . . ; N; we can establish balancing of the polynomial corrections for the


proposed schemes.
4.1. The exact C-property of the LLF formulation
p
1
i ;

If we omit the polynomial corrections P

p
1
i ;

and Q

in the expressions (25), (31), and (32) the

rst, upwinded part of the (19) in the pth characteristic eld, p 1; 2 can be written in the form

1
p
p
f i f i1  ai1 ui1  ui  Bi;i1 ui1  ui li1
2
2
2Dx

1
p
p
f i1 f i  ai1 ui  ui1 Bi1;i ui  ui1 li1 :

2
2
2Dx


49

Simple computations lead us to the equalities


p

f i lj1 f i lk1
2

50

and
p

uj  uk li1 ci;i1 hj Hj  hk  Hk 0;

51

for j; k 0; . . . ; N . Here the last relation follows from the quiescent ow condition (41). If we use
(50) and (51), for proving that the expression (49) is equal 0, it remains to prove the equality
 p

52
f i1  f i  Bi;i1 ui1  ui li1 0;
2

for p 1; 2; 3. Actually, we can prove the more general equality


 p

f j  f k  Bj;k uj  uk li1 0;

53

for p 1; 2; 3, since by taking into account the expressions for the terms on the left side in (53) we
obtain



g 2
hj hk
g1p
p
2
h  hk  g
Hj  Hk 1
hj hk hj  hk Hj  Hk 0;
2 j
2
2
54

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

42

for p 1; 2. The last relation follows again from (41). Furthermore, in the third characteristic eld
the eigenvalue is zero and the left eigenvector is orthogonal to the numerical ux and to the
numerical nonconservative product term, so (53) is proved. We can nish the proof of the exact
C-property for the LLF formulation by proving (48). Using (26) and (35) we obtain


p;

p;

p
p
f k  f I

ai1 uk  uI
 Bk;I
uk  uI
li1
vk  wk
2
2
 p

p
p
f k  f I
 Bk;I
uk  uI
li1
ai1 uk  uI
li1 0
55
2

for p 1; 2; 3, k 0; . . . ; N. The rst term in the above expression is equal zero because of (53)
and the second one because of (51).
4.2. The exact C-property of the RF formulation
It is obvious that we have to consider only the case when the pure Roe formulation is used since
we solved the LLF case. Again, as in the LLF case we distinguish two parts of the proof. First, in
order to prove the balancing of the upwinded parts, from (46), (28), (36), and (37) it follows that
we have to prove


p
p
p
 f I
li1  f I
1 li1 BI
1;I
uI
 uI
1 lI
1 0;
56
2

for p 1; 2: In fact, from (56), after applying (50) we obtain



 p
 f I
 f I
1  BI
1;I
uI
 uI
1 lI
1 0;

57

for p 1; 2 and this is just a special case of (53). For the balancing of the polynomial corrections
similarly to the previous calculations we can nd that
 p

p;

p;

f k  f I
 Bk;I
uk  uI
li1 0;
58
vk  wk
2

for p 1; 2. Finally, in the third characteristic eld, because of ki1 0, all the terms in (19) are
2

identically equal zero. With this the Eq. (46) is established and we nished the proof of the exact
C-property for both formulations of the extended ENO and WENO schemes.
4.3. Order of accuracy
At this point we would like to make a few observations about the order of accuracy of the
proposed schemes. As it is usual for the nite dierence ENO and WENO schemes for the hyperbolic conservation laws systems we consider separately the order of accuracy for the time
operator and the order of accuracy that refers to the space operator. The order of accuracy for the
time operator is determined with order of the RungeKutta method that is used.
Since the approximation of the ux derivative in the space operator follows the conservation
form, i.e.,
ox fuji 

1
f 1  f i12 ;
Dx i2

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

43

accuracy of that part relate to the accuracy of the numerical evaluation of the numerical uxes f i1
2

and f i1 . From the theoretical background of the standard ENO and WENO schemes we know
2

that when the rth order reconstruction algorithms are used, in the regions when the solution is
smooth, the approximation of the ux is r 1th order accurate in ENO schemes and 2r 1th
order accurate in WENO schemes [35,12,18]. Therefore, we have
p

Pi1;
vp;
xi12 ODxR ;

59

where the polynomial corrections P

p
1
i ;

are dened with (39), while R r 1 in the ENO and R

2r 1 in the WENO case. Incorporation of (59) and the denition (26) in (25) for the LLF case
leads to the relation
p

fi1 fxi12 li1 ODxR :

60

Similarly, in the RF formulation, by applying the relation (59) and using a denition (29) for
evaluating (28), we obtain (60) once again. Thus, the evaluation of the ux on the cell boundaries
is Rth order accurate in both cases. Furthermore, we must point out that the Jacobian matrix
proposed in the paper does not change the order of accuracy of the ux. Namely, it only help us to
evaluate correctly the local characteristic elds and as consequence to obtain the appropriate ux
splitting in the LLF formulations and to use the appropriate upwinding in the RF formulations of
the schemes.
As the approximation of the ux derivative in the numerical scheme must be done in the
conservative form, in our schemes there is an absolute priority to evaluate a nonconservative
product in such a way that the balancing between the ux gradient and the nonconservative
product [2,14] would be obtained. Actually, only some algorithm that respects this balancing can
improve the numerical accuracy and remove numerical errors otherwise present. Therefore we
consider a noncentered contribution for the source term (30). The higher order polynomial corrections that are dened with (40) and that appears in the numerical examination of terms (31)
and (32) in the LLF case, resp. (36) and (37) in the RF case, are just approximations obtained by
using ENO or WENO reconstruction algorithm for functions wp;
. Since a given denitions for
the functions wp;
are actually a consequence of the demand for the balancing we want to obtain,
p
p
the order of accuracy of terms g 1 and g 1 just relate to the accuracy of evaluating wp;
xi1 .
i ;L
2

i ;R
2

From the theory of the ENO and WENO reconstruction algorithm we know
p

Qi1;
wp;
xi12 ODxR ;

61

so the evaluation is Rth order accurate.


In [14] we presented an extension of ENO and WENO schemes for one-dimensional shallow
water ow, that is to some extent similar to the algorithm we propose here. There we also give a
discussion on accuracy and a convergency test that is very incouriging in regard to the experimentally established order of accuracy of new schemes.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

44

5. Numerical results
In this section we present results obtained with dierent versions of extended ENO and WENO
schemes for several test problems proposed in [7,8].
All the problems we discuss are initial problems with x 2 0; L , L 1000 m and moving riverbed as an initial condition. Since they are all initial problems, following [7] we treat boundary
points numerically using a free ow boundary conditions
ui u0

and

uNi uN ;

i 1; . . . ; r:

62

In our numerical investigation we focus on three types of test problems. In all of them, the
sediment transport ux is supposed to be
qv Av3 :

63

With variation of the physical parameters, such as parameter A in (63), porosity of the riverbed ,
and the discharge Q we obtain dierent problems. In order to obtain realistic physical initial
conditions with moving riverbed, in all the problems we rst suppose dummy initial conditions
with xed riverbed. Precisely, the initial velocity and initial water level are of the form
hx; 0 10  H x; 0

and vx; 0

Q
:
hx; 0

64

To ensure that the initial conditions of the water ow and riverbed are consistent we proceed as
follows. In the tests where A 1 we rst compute the ow until the time t 75 s using ux limited
Roes scheme [6], while in the case where A 0:001 the water ow is iterated to an equilibrium
state whilst keeping the riverbed xed. The obtained states are taken as the real initial conditions.
5.1. Test Problem 1
Here the initial riverbed is dened with
8


< 2 px  300
;
if 300 6 x 6 500
sin
H x; 0
200
:
0;
otherwise:

65

5.1.1. Test Problem 1.1


In the rst test case we simulate a small pulse that is interacting quickly with the water ow. To
simulate this we take values A 1, Q 10 and  0:2.
We use this test problem for a numerical study of the convergence rate of the presented
schemes. In order to do that and because the analytical solution is not known, we rst compute an
exact solution on a very ne grid with N 10; 240 grid points. Then we compute errors in the
riverbed obtained with the ENOLLF and WENOLLF formulations of the extended schemes
for r 1; . . . ; 5, on much coarser grids. While for r 1; 2 we apply a two-step RungeKutta
method, for r 3; 4; 5 we use a three step RungeKutta method. Depending of the order of the
scheme we also change the CFL coecient. For the case r 1; 2; 3 we use cCFL 0:75, while for
r 4; 5 we reduce it to 0.5 in order to compensate for the lower order of time approximation. The

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

45

Table 1
The convergency test results for Test Problem 1.1
r
1

ENOLLF extended

WENOLLF extended

L1 error

L1 order

L1 error

L1 order

40
80
160
320
640

5.36407619
2.03535310
0.69720372
0.21136733
0.05764226

1.40
1.55
1.72
1.87

5.10662432
1.90082133
0.60924830
0.15069022
0.03061319

1.43
1.64
2.02
2.30

40
80
160
320
640

2.63338661
0.55432786
0.12408960
0.03051834
0.00663355

2.25
2.16
2.02
2.20

1.90105413
0.34576976
0.05891488
0.00843662
0.00110149

2.46
2.55
2.80
2.94

40
80
160
320
640

2.01225866
0.34879049
0.06140986
0.01312447
0.00270923

2.53
2.51
2.23
2.28

1.47673720
0.21561184
0.02863229
0.00477258
0.00089837

2.78
2.91
2.58
2.41

40
80
160
320
640

1.72024649
0.26615983
0.05395468
0.00924458
0.00238462

2.69
2.30
2.55
1.95

1.17635189
0.10971679
0.01405844
0.00294646
0.00066326

3.42
2.96
2.25
2.15

40
80
160
320
640

1.55308743
0.25507816
0.04799502
0.00876928
0.00365260

2.61
2.41
2.45
1.26

0.91373525
0.06599067
0.00857888
0.00230442
0.00058742

3.79
2.94
1.90
1.97

experimentally established errors and orders at time t 200 are presented in Table 1. From the
presented results we can conclude that both ENOLLF and WENOLLF schemes do not achieve
the expected formal order of accuracy but show a signicant increase of the obtained accuracy
with the formal order R of the scheme. It can be noted a decrease of the order of accuracy for the
case r 4; 5 when N 640 grid points in the computations are used. Besides the greater round-o
errors that appear when the high order polynomial reconstruction is used and that certainly reduce the order of accuracy, the main reason for that comes from the fact that we do not compute
the order of accuracy comparing the obtained results with an analytical solution, but with the
solution obtained by using the ENOLLF scheme, r 2. Namely, the results obtained with the
higher order schemes on a much coarser grid approach faster to such an obtained exact solution.
Furthermore, on the same test problem we take Dx 1 m and cCFL 0:85. The real initial
conditions can be seen in Figs. 24. In Fig. 5 we present evolution in the riverbed obtained with
the extended version of the WENOLLF scheme, r 2. The computations are performed for a
much longer computational time and on the extended domain as can be seen in Fig. 6. The
presented evolution in the water level is also computed by using the extended WENO-RF scheme,

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

46
1,0

riverbed

0,8

0,6

0,4

0,2

0,0
0

100

200

300

400

500

600

700

800

900

1000

Fig. 2. The initial riverbed for Test Problem 1.


10,000
9,999

water level

9,998
9,997
9,996
9,995
9,994
9,993
0

100

200

300

400

500

600

700

800

900

1000

Fig. 3. The initial water level for Test Problem 1.

r 2. We can conclude that with the extended schemes a very sharp and clear discontinuity forms
and it can be evaluated for a long computational time. Obtained results verify the stability and the
shock capturing properties of the dened schemes. On the contrary, the results obtained with the
schemes that include a pointwise evaluation of the nonconservative product are very poor. In
Figs. 7 and 8 we zoom in on the details of the riverbed and of the water level in order to emphasize the improvement of the extended schemes. The presented results clearly indicate the superiority of the nonconservative product decomposition over the schemes with the pointwise
evaluation of the nonconservative product.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

47

velocity

1,06

1,04

1,02

1,00
0

100

200

300

400

500

600

700

800

900

1000

Fig. 4. The initial velocity for Test Problem 1.

1,0

riverbed

0,8

0,6

0,4

0,2

0,0
300

350

400

450

500

550
x
t=200 s

600

650

700

750

800

t=700 s

Fig. 5. The evolution in the riverbed for Test Problem 1.1 obtained with the extended WENOLLF scheme, r 2.

5.1.2. Test Problem 1.2


Beginning with the same initial riverbed (65) as in the previous example a slow interaction of
the riverbed with the water ow is obtained by using A 0:001. Furthermore we take Q 10 and
 0:4. Following [7] the computations are performed on a much coarser grid than in the previous example, i.e., Dx 10 m. In [7], the approximate solution that is valid until the characteristics rst cross is evaluated analytically. In Fig. 9 the riverbed obtained with the extended
ENO-RF scheme, r 4 is presented. Again, for the case when the nonconservative product is
evaluated pointwise, a numerical instability occurs as can be seen in Fig. 10.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

48
10,001
10,000
9,999

water level

9,998
9,997
9,996
9,995
9,994
9,993
500

600

700

800
t=1000 s

900

1000

1100

t=1500 s

Fig. 6. The evolution in the water level for Test Problem 1.1 obtained with the extended WENO-RF scheme, r 2.

1,00

riverbed

0,98

0,96

0,94

0,92

0,90
640

670

700

x
WENO-LLF, r=2 pointwise

WENO-LLF, r=2 extended

Fig. 7. Comparison of the schemes for Test Problem 1.1, detail of the riverbed at t 700 s.

5.1.3. Test Problem 1.3


Finally, the last version of the rst problem presents how the riverbed is interacting with a fast
moving water ow. Thus we take Q 50; A 0:001, and  0:4. We present results obtained
with the extended ENOLLF formulation, r 2 and r 5, a space step Dx 10 m, and

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

49

cCFL 0:7, (Figs. 11 and 12). When compared with the results presented in [7] we can conclude
that our approach gives very good results.

water level

10,000

9,998

9,996

9,994

9,992
450

500

550

600

650

700

750

800

x
WENO-LLF, r=2 pointwise

WENO-LLF, r=2 extended

Fig. 8. Comparison of the schemes for Test Problem 1.1, detail of the water level at t 700 s.

1,0

riverbed

0,8

0,6

0,4

0,2

0,0
300

350

400

450

500

550

600

650

700

750

800

x
ENO-RF, r=2 extended

ENO-RF, r=4 extended

Fig. 9. The riverbed for Test Problem 1.2 obtained with the extended ENORF scheme, r 2 and r 4 at t
238; 000 s.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

50
10,005

water level

10,000

9,995

9,990

9,985
0

100

200

300

400

500

600

700

800

900

1000

x
ENO-RF, r=2 pointwise

ENO-RF, r=2 extended

Fig. 10. Comparison of the schemes for Test Problem 1.2, water level at t 1000 s.

5.2. Test Problem 2


The initial riverbed is here dened with
H x; 0

1


x  400
1 exp
5p

66

and we consider the case of a sediment bore that interacts quickly with the water ow. Such a
situation is obtained with Q 10, A 1, and  0:2. The real initial conditions are given in Figs.
1315. The moving of the sediment bore caused by the water ow can be seen in Fig. 16, while in
Fig. 17 we can observe how the water level changes with time. The results are obtained by using
Dx 1 m and cCFL 0:8. Figs. 16 and 17 illustrate the stability of the extended ENOLLF
schemes. From the presented results we can also conclude that our schemes behave very well near
discontinuities because the results obtained are very sharp and clear. With Fig. 18, where the
comparison of the results obtained by using a decomposed evaluation of the nonconservative
product against the pointwise approach is presented, the improvement of the presented schemes is
emphasized again.
For the same test problem numerical errors in the water level at t 200 s obtained with the Roe
formulation of the extended WENO schemes are presented in Table 2. The results are computed
by using a relatively coarse grid with Dx 5 m. The approximation for the analytical solution
which is not known is evaluated as in Test Problem 1.1, i.e., by computations on a very ne grid

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

51

1,0

riverbed

0,8

0,6

0,4

0,2

0,0
300

350

400

450

500

550

600

650

700

750

800

Fig. 11. The riverbed for Test Problem 1.3 obtained with the extended ENOLLF scheme, r 2 at t 1904 s.

10,05
10,00
9,95

water level

9,90
9,85
9,80
9,75
9,70
9,65
9,60
0

100

200

300

400

500

600

700

800

900

1000

Fig. 12. The water level for Test Problem 1.3 obtained with the extended ENOLLF scheme, r 2 at t 1904 s.

with N 10; 240 grid points. Obtained values clearly illustrate the increase of the accuracy with
the formal order R of the scheme.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

52
1,0

riverbed

0,8

0,6

0,4

0,2

0,0
0

100

200

300

400

500

600

700

800

900

1000

Fig. 13. The initial riverbed for Test Problem 2.

10,023

10,022

water level

10,021

10,020

10,019

10,018

10,017
0

100

200

300

400

500

600

700

800

900

1000

Fig. 14. The initial water level for Test Problem 2.

5.3. Test Problem 3


The last test problem presents the behavior of our numerical schemes for the case when the
riverbed is discontinuous. The initial conditions are dened with

1; if 300 6 x 6 500
H x; 0
67
0; otherwise:
A slow interaction of the riverbed with the water ow is studied, i.e., here we take Q 10,
A 0:001 and  0:4. In Fig. 19 results performed using ENOLLF schemes with r 2, resp.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

53

Fig. 15. The initial velocity for Test Problem 2.

1,0

riverbed

0,8

0,6

0,4

0,2

0,0
0

100

200

300

400

500

600

700

800

900

1000

x
t=200 s

t=700 s

Fig. 16. The evolution in the riverbed for Test Problem 2 obtained with the extended ENOLLF scheme, r 5.

r 3, with Dx 10 m, cCFL 0:7, and at time t 238; 000 s are presented. From Fig. 19 we can
conclude that the proposed schemes give good results for a long computational time even when
the initial state is discontinuous, and that the shock capturing property of the schemes improves
with the order of the scheme.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

54
10,023

water level

10,022
10,021
10,020
10,019
10,018
10,017
0

100

200

300

400

500
x

600

t=200 s

700

800

900

1000

t=700 s

Fig. 17. The evolution in the water level for Test Problem 2 obtained with the extended ENOLLF scheme, r 5.

10,024
10,023

water level

10,022
10,021
10,020
10,019
10,018
10,017
10,016
570

590

610

630

650

670

690

710

x
ENO-LLF, r=5 pointwise

ENO-LLF, r=5 extended

Fig. 18. Comparison of the schemes for Test Problem 2, water level at t 200 s.

6. Concluding remarks
In this paper we extend the standard ENO and WENO schemes, which have been developed for
the hyperbolic conservation laws, to the one-dimensional bed-load sediment transport equations.

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

55

Table 2
Errors in water level for Test Problem 2
Method

L1 error

WENO-RF

1
2
3
4
5

2.23  105
1.05  105
2.88  106
1.51  106
1.02  106

1,1
1,0
0,9
0,8

riverbed

0,7
0,6
0,5
0,4
0,3
0,2
0,1
0,0
250

300

350

400

450

500

550

600

ENO-LLF extended, r=2

ENO-LLF extended, r=3

Fig. 19. Comparison of the riverbed for Test Problem 3, water level at t 238; 000 s.

The diculty with this conservation law system is that it contains a nonconservative product
term. Our approach is based on including this term in the global upwinding of the system, on the
application of ENO and WENO reconstruction, and on the balancing between the ux gradient
and the nonconservative product term. The new schemes are exactly consistent with the quiescent
ow and inherit shock capturing and nonoscillatory properties of ENO and WENO schemes. This
work can be extended to the numerical approximation of the two-dimensional bed-load sediment
transport system and to the systems with a dierent denition of the sediment ux. A similar
treatment of the nonconservative product can also be used in combination with other numerical
schemes.
References
[1] Berm
udez A, Dervieux A, Desideri JA, Vazquez ME. Upwind schemes for the two-dimensional shallow water
equations with variable depth using unstructured meshes. Comput Meth Appl Mech Eng 1998;155:49.

56

 rnjaric-Z ic et al. / Computers & Fluids 33 (2004) 3156


N. C

[2] Berm
udez A, Vazquez ME. Upwind methods for hyperbolic conservation laws with source terms. Comput Fluids
1994;23(8):1049.
[3] Harten A, Osher S. Uniformly high-order accurate non-oscillatory schemes I. SIAM J Numer Anal 1987;24:279.
[4] Harten A, Engquist B, Osher S, Chakravarthy SR. Uniformly high order accurate essentially non-oscillatory
schemes III. J Comput Phys 1987;71:231303.
[5] Shu C-W, Osher S. Ecient implementation of essentially non-oscillatory shock capturing schemes II. J Comput
Phys 1989;83:32.
[6] Hubbard ME, Garcia-Navarro P. Flux dierence splitting and the balancing of source terms and ux gradients.
Numerical Analysis Report, University of Reading, Department of Mathematics, 1999.
[7] Hudson J. Numerical techniques for morphodynamic modelling. Doctor Thesis, University of Reading,
Department of Mathematics, 2001. A colour postscript version can be obtained at http://www.rdg.ac.uk/
AcaDepts/sm/wsm1/research/theses/jh.html.
[8] Hudson J, Sweby PK. Numerical formulations for approximating the equations governing the bed-load sediment
transport in channels and rivers. Numerical Analysis Report, University of Reading, Department of Mathematics,
2, 2000.
[9] LeVeque RJ. Balancing source terms and ux gradients in high-resolution Godunov methods: the quasi-steady
wave propagation algorithm. Numerical Analysis Report, Department of Applied Mathematics and Department of
Mathematics, University of Washington, 1998.
[10] van Rijn LC. Sediment transport, Part I: bed-load transport. Proc ASCE J Hydraul Div 1984;110(HY 10):1613.
[11] Liu X-D, Osher S, Chan T. Weighted essentially non-oscillatory schemes. J Comput Phys 1994;115:200.
[12] Shu C-W. Essentially non-oscillatory and weighted essentially non-oscillatory shock-capturing schemes for
hyperbolic conservation laws. In: Cockburn B, Johnson C, Shu C-W, Tadmor E, editors. Advanced Numerical
Approximation of Nonlinear Hyperbolic Equations. Lecture Notes in Mathematics, vol. 1697. Springer; 1998.
[13] Castro M, Macas J, Pares C. A Q-scheme for class of systems of coupled conservation laws with source term.
Application to a two-layer 1-d shallow water system. Math Modell Numer Anal 2001;35:107.
[14] Vukovic S, Sopta L. Essentially non-oscillatory and weighted essentially non-oscillatory scheme with the exact
C-property for one-dimensional shallow water equations. J Comput Phys 2002;179:593.
[15] Grass AJ. Sediment transport by waves and currents. SERC London Cent Mar Technol, Report No: FL29, 1981.
[16] Brufau P, Garcia-Navarro P, Ghilardi P, Natale L, Savi F. 1d Mathematical modelling of debris ow. J Hydraul
Res 2000;38:435.
[17] Gosse L. A well-balanced scheme using non-conservative products designed for hyperbolic systems for
conservation laws with source terms. Math Mod Meth Appl Sci 2001;11:33965.
[18] Jiang G, Shu C-W. Ecient implementation of weighted ENO schemes. J Comput Phys 1996;126:202.

Anda mungkin juga menyukai