www.seipub.

org/scea Study of Civil Engineering and Architecture (SCEA) Volume 2 Issue 1, March 2013
12
ParameterSelectionsinSimulatingthe
PhysicalDiffusionPhenomenaofSuspended
LoadbyLowOrderDifferentialScheme
NumericalDispersion
ShuaiChen
*1
,JunxingWang
2

StateKeyLaboratoryofWaterResourcesandHydropowerEngineeringScience,WuhanUniversity
Wuhan,Hubei,China
*1
liveangela@whu.edu.cn;
2
jxwang@whu.edu.cn

Abstract
In this paper, the process of the introduction of numerical
diffusionisinvestigatedwhenloworderdifferentialscheme
inconvectiondiffusionequationfordiscretesuspendedload
is in use. According to the results of instances calculation,
theoptimalchoiceforparameterswhenphysicaldiffusionis
simulated by numerical diffusion is obtained. Finally,
comparison between several commonly used loworder
schemesisconductedandthemostviableloworderscheme
in simulating physical diffusion by numerical diffusion is
proposedindealingwithsuspendedload.
Keywords
DifferentialScheme;ConvectionDiffusionEquationofSuspended
Load;NumericalDiffusion
I nt r oduc t i on
In general, using the loworder scheme for numerical
calculations will more or less introduce numerical
diffusion (Xie 1990). Different schemes have varying
degrees of numerical diffusion. In theory this is not
conducive to the exact solution of numerical
calculation,bututilizingthenumericaldiffusion,such
as to simulate the physical diffusion phenomenon, is
also an effective method (Yang 1993). As for how to
use the numerical diffusion to simulate the physical
diffusion, and how to select parameters to achieve an
acceptable level of the simulation results, they are the
issuestobediscussedinthispaper.
Convec t i on-Di f f usi on Equat i on f or One-
Di mensi onal Suspended Load
Equationforsuspendedloadmovementis(Zhengand
Zhao2001):
'
( )
S
x
A AS QS S
AD S r
t x x x t
c c c c c
+ =
c c c c c

Introducingtheflowcontinuityequation,theequation
abovecanberewrittenas:
'
1 ( )
S
x
A S Q S S S r
AD
t A x A x x A t
c c c c c
+ =
c c c c c

where
x
D the longitudinal diffusion coefficient, for
sedimentdiffusion,itcanbeapproximatedas
*
0.25
x
D u h =
Generally, for transportation issue of suspended load,
the longitudinal diffusion of suspended load is much
smaller than the longitudinal conviction of sediment,
which is often overlooked. If the carrier velocity and
diffusion coefficient
x
D are constants without
considering the source item, the convectiondiffusion
equationwillbe
2
2 x
S S S
u D
t x x
c c c
+ =
c c c

For this equation, the main numerical difficulty is to
calculate the convection term because it strictly
demands the conservation of matter, while the
differential method for numerical solution often
cannot achieve this. If the solution to convection
diffusion is considered under the premise of a good
solution to convectiondiffusion term, the probability
of successful numerical solution would be higher. If
ignoring the diffusion term, we can get the pure
convectionequation
0
S S
u
t x
c c
+ =
c c

Study of Civil Engineering and Architecture (SCEA) Volume 2 Issue 1, March 2013 www.seipub.org/scea
13
Est abl i shment of Di f f er ent i al Sc heme
Select Upwind scheme discrete pure convection
equation, when 0 u > , the differential equation is
(Courant1928)
1
1
( )
n n n n
j j r j j
S S C S S
+

=
where
r
C isCroutnumberand

/
r
C C t x = A A
Thedifferentialequationcanbereducedto

1
1
0
n n n n
j j j j
S S S S
C
t x
+

+ =
A A

The first item in the above equation is the forward
difference quotient of / S t c c , and the second term is
backward difference quotient of / S t c c , then

C u =
can be drawn. We will exam the compatibility,
stability and convergence of the scheme as follows.
Expand
1 n
j
S
+
and
1
n
j
S

onpoint(j,n)
2
1 2
2
1
( ) ( ) ...
2
n n n n
j j t j
S S
S S t t
t t
+
c c
= + A + A +
c c

2
2
1 2
1
( ) ( ) ...
2
n n n n
j j j j
S S
S S x x
x x

c c
= A + A +
c c

Take them into the differential equation, and use the
pureconvectionequationtoget
2 2
2
(1 )
2
r r
x C C S S S
u
t x t x
A c c c
+ =
c c A c

Assume
2
(1 )
2
r r
n
x C C
D
t
A
=
A

Then
2
2 n
S S S
u D
t x x
c c c
+ =
c c c

Comparing to the pure convection equation, only the
right term of the equation tends to be zero when the
space and time step are small enough, which means
theyarecompatible.Taking /
r
t C x C A = A into
n
D ,we
canget
(1 )
2
r
n
xC C
D
A
=

where C u = , 0 1
r
C < < , when 0 x A , 0 t A , and
0
n
D = . So the difference equations and differential
equationsarecompatible,asthestabilityconditionfor
Upwind is / 1 u t x A A s . Using Lax equivalence
theorem again we can get that the solution of
differential equation converges to the solution of
differentialequations.
From the above proof it is clear that, Upwind scheme
dispersesapureconvectionequationandanequation
similar to convective diffusion can be obtained.
Although the differentiated items
2 2
/
n
D S x c c can
make the equation compatible, calculation shows that
the item also makes the solution of differential
equations no longer converge to the solution of the
original differential equation. Instead it converges to
the solution of the convectiondiffusion equation,
which means the implicit numerical diffusion,
resulting from the limitation of x A , t A , which cannot
be infinitely close to zero. However, since it is known
that difference equations converge to the convection
diffusion equation, then it is used to identify the
required parameter values, so that the differential
equations can converge precisely to differential
equations, namely numerical diffusion can simulate
thephysicaldiffusionwell.
The I nst anc e Anal ysi s of Par amet er s Ef f ec t
As differential equation is similar in form with
convectiondiffusion equation,
n
D might be called
numerical diffusion coefficient. It is clear that the
closeness between numerical diffusion coefficient
n
D
and physical diffusion coefficient
x
D can determine
the accuracy of the numerical diffusion in simulating
physicaldiffusion,thustheproblemistofindtheright
parameters,so that
n
D can be close to
x
D (Papadakis
and Metaxas 2011). On the other hand, three
parameters can determine
n
D which are velocity u ,
space interval x A and time step t A , where flow rate
should be in accordance with the actual engineering
value, and x A is subject to the restrictions on the
boundarysimulationaccuracy(itwilltakeondifferent
values in accordance with the different size of the
project and the complexity of the border. Usually, for
example, the smaller the size of x A is, the more
precise the boundary simulation results will be. But it
willincreasetheamountofcalculation),onlythetime
step t A has wider choice. By selecting different t A
we can get the corresponding numerical diffusion
coefficient
n
D . Then comparing
n
D with different
longitudinal diffusion coefficient
x
D , the impact of
t A on calculated results can be examined
(Gasiorowski2013).
Buildthefollowingmodel:
www.seipub.org/scea Study of Civil Engineering and Architecture (SCEA) Volume 2 Issue 1, March 2013
14
Model1
Length is 10km, width is 2m, water depth is 4m,
Chezy coefficient C is 25, and flow rate is 0.5m/s,
x A =200m. Take respectively t A =400s, 399.5s, 399s,
360s,300s,200s
Thecalculationresultsareshownasfollows(Table1).
TABLE1CALCULATIONRESULTS(U=0.5M/S)
t A
(s)
Parameters
r
C
n
D
x
D
400 1 0 0.063246
399.5 0.99875 0.0625 0.063246
399 0.9975 0.125 0.063246
360 0.9 5 0.063246
300 0.75 12.5 0.063246
200 0.5 25 0.063246
100 0.25 37.5 0.063246
50 0.125 43.75 0.063246
0.5 0.00125 49.9375 0.063246
In the above table, only when t A =399.5s, the
difference between numerical diffusion coefficient
n
D
and longitudinal diffusion coefficient
x
D is the
smallest at 0.000746. At the same time, the difference
between
r
C and 1 is 0.00125, when t A decreases by
0.5sto399s.Thetwodiffusioncoefficientsdiffersinan
order of magnitude, indicating that only when
r
C is
verycloseto1thephysicaldiffusioncanbesimulated
more accurately. On the other hand, the numerical
diffusion coefficient
n
D will increase with the
decrease of
r
C . When
r
C tends 0, a solution with
largererrorwillbegot.
Here the time step is accurate to the extent of 0.5,
which is not conducive to calculation. Since the time
step may be influenced by the flow velocity, the flow
rateischangedandthemodelismodifiedasfollows:
Model2
Flow velocity u =1m/s, taking t A =400s, 300 s 200s,
199.5s, 199s, 198s other parameters are unchanged.
In the results (Table 2) when t A is greater than x A ,
numerical diffusion coefficient is negative. It will
repeatthelastcalculationlawwhentheyareequal.In
factiftheformulaischangedfor
n
D ,wecanfindthe
contradictions.
2 2
(1 )
2 2
r r
n
x C C u x u t
D
t
A A A
= =
A

When u =1m/s, ( ) / 2
n
D x t = A A , obviously x A needs
to be larger than t A to make
n
D positive, in fact, this
iscausedbythefailuretosatisfythestabilitycondition
/ 1 u t x A A s . And, from another perspective, the
accuracy of
n
D shows direct linear relationship with
the precision of x A as well as t A , which demands
highly for x A and t A , in the actual calculation, it is
oftenimpossibletomeetsucharequest,sootherflow
ratesshouldbetakentocalculateagain.
TABLE2CALCULATIONRESULTS(U=1M/S)
t A
(s)
Parameters
r
C
n
D
x
D
400 2 100 0.126491
300 1.5 50 0.126491
200 1 0 0.126491
199.5 0.9975 0.25 0.126491
199 0.995 0.5 0.126491
198 0.99 1 0.126491
100 0.5 50 0.126491
50 0.25 75 0.126491
0.5 0.0025 99.75 0.126491
Model3
Flow velocity u =0.8m/s, taking t A =250s, 249.5s, 249s,
200s, 100s, 50s The remaining parameters are
unchanged.
Model4
Flow rate u =0.2m/s, taking t A =1000s, 999.5s, 999s,
998s, 900s, 800s The remaining parameters are
unchanged.
CalculationresultsareshowninTable3andTable4
respectively.
Study of Civil Engineering and Architecture (SCEA) Volume 2 Issue 1, March 2013 www.seipub.org/scea
15
TABLE3CALCULATIONRESULTS(U=0.8M/S)
t A
(s)
Parameters
r
C
n
D
x
D
250 1 0 0.101193
249.5 0.998 0.16 0.101193
249 0.996 0.32 0.101193
200 0.8 16 0.101193
100 0.4 48 0.101193
50 0.2 64 0.101193
0.5 0.002 79.84 0.101193
TABLE4CALCULATIONRESULTS(U=0.2M/S)
t A
(s)
Parameters
r
C
n
D
x
D
1000 1 0 0.025298
999.5 0.9995 0.01 0.025298
999 0.999 0.02 0.025298
998 0.998 0.04 0.025298
900 0.9 2 0.025298
800 0.8 4 0.025298
500 0.5 10 0.025298
100 0.1 18 0.025298
0.5 0.0005 19.99 0.025298
Combining the above four models; we can draw the
followingrules:
When u isintheinterval(0,1],thelongitudinal
diffusion coefficient
x
D increases as u
increase, and the maximum value is close to
0.126 (this is the maximum value when the
waterdepthis4mandChezycoefficientis25,
dependingonthecircumstances,thevaluewill
bedifferent);
When u is in the interval (0,1], if u is small,
the numerical diffusion simulation of physical
diffusion is better, which is the closest when
u =0.5m/s;
When u is in the interval (0,1], the simulation
is better when the time step value is 0.5s
smaller than the t A that makes
r
C =1, but
thedifferenceof0.5sisnotideal;
When u is in the interval (0,1], although the
closer
r
C is to 1, the better the simulation is,
however within a small range close to 1, it is
not in line with this rule. As shown in the
followingtable.
TABLE5VARIATIONS IN PARAMETERS WITH DIFFERENT U
u
(m/s)
t A
(s)
Parameters
r
C
n
D
x
D
x n
D D %
0.2 999.5 0.9995 0.01 0.025298 0.0153 60.48
0.5 399.5 0.99875 0.0625 0.063246 0.0007 1.11
0.8 249.5 0.998 0.16 0.101193 0.0588 58.11
1.0 199.5 0.9975 0.25 0.126491 0.1235 97.64
0
0.5
1
1.5
2
2.5
10 100 1000
lgt(s)
C
r
u=0.2m/s
u=0.5m/s
u=0.8m/s
u=1.0m/s

FIG.1 DISTRIBUTION OF C
r
WITH GROWTH OF THE TIME STEP At (IN LOG
FORM)
0
20
40
60
80
100
120
10 100 1000
lgt(s)
D
n
u=0.2m/s
u=0.5m/s
u=0.8m/s
u=1.0m/s

FIG.2 DISTRIBUTION OF D
n
WITH GROWTH OF THE TIME STEP At (IN LOG
FORM)
Obviously, in the descending process of u ,
r
C is
gettingcloserto1,but
x n
D D isthenminimumwhen
u =0.5m/s.Theexistenceofasmallerdifferencecanbe
www.seipub.org/scea Study of Civil Engineering and Architecture (SCEA) Volume 2 Issue 1, March 2013
16
tested by changing the value of u , however it is no
longerstudiedhere.
In fact, in order to get the parameters conditions of
numerical diffusion that accurately simulate physical
diffusion, formulas are used to reason reversely,
assume
x n
D D =
Weget:
2
*
0.25 0.25
2
n x
u x u t u
D D u h g h
C
A A
= = = =

Simplifyittobe:
0.5 gh
x u t
C
A = A +

Byassuming
0.5 gh
k
C
=
,wecanobtain:
x k
t
u
A
A =

Obviously, t A isalinearfunctionof x A ,where k isa
constant related to the parameters of the physical
model. Take u =0.5, h =4, C =25, x A =200m into the
formula then the calculation shows t A =399.499 which
is very close to the above 399.5, so the difference
between the two diffusion coefficients is insignificant.
Butthisstepsizeisveryunfavorabletocalculation;as
it is tendency to take an integer as much as possible.
Although in different engineering cases, different
parameters of the physical model can improve the
value of t A , its value depends completely on the
differential scheme itself. In order to keep the
generality, the following will horizontally contrast a
variety of commonly used differential schemes under
the current model parameter conditions to examine
whether a scheme can achieve a relatively satisfying
stepvalue.
Compar i son among Var i ous Di f f er enc e
Sc hemes
Although various differential schemes have different
compatibility and stability conditions, as long as
extracting numerical diffusion coefficient after
dispersing pure convection equations respectively,
then equivalent unify with the physical diffusion
coefficient, we can obtain the calculation formula for
stepsize(Table6).
Obviously, these three schemes of Dobbins, Upwind
and implicit have the same method to select the step
size, and they also have the same characteristics
represented by the Upwind scheme. The loworder
Preissmann scheme is more suitable for the large
stride length, and the possibility of an integral step
length has been improved in this way, where =0.6,
u =0.4, t A =19900s, and
n
D =0.05, which shows a
difference of 20.94% with
x
D . When t A =19870s and
n
D =0.065,
n
D show a difference of 2.77% with
x
D
andprecisioncanmeettherequirements.Instead,Lac
Friedrichsschemeismoresuitableforsmalltimestep.
Though the step is smaller, the simulation accuracy is
not improved, such as when t A =247s and
n
D =0.097,
t A is only 0.1s smaller than 247.1s, while the relative
error of diffusion coefficient reaches an unsatisfactory
level,52.83%.
TABLE6RESULTS OF DIFFERENT CALCULATION FORMULATIONS
Schemename
n
D expression t A expression Calculationresults
Dobbins
2
(1 )
2
r r
x C C
t
A
A

x k
t
u
A
A =

399.5
Upwind
2
(1 )
2
r r
x C C
t
A
A

x k
t
u
A
A =

399.5
Implicit
2
(1 )
2
r r
x C C
t
A +
A

k x
t
u
A
A =

399.5
Preissmann
2
(2 1)
(2 1)
2
r
r
x C
C
t

u
A
+
A
2
(1 2 )
2(2 1)
k x x
t

u
A + A
A =

19874.8

20125.2

LacFriedrichs
2 2
(1 )
2
r r
x C C
t
A
A

2 2
5 2
2
x k x k k x
t
u
A + A A
A =
247.1
Note:
1. Inthetable 0.5 ( ) / k gh C = ,where h =4m,Chezycoefficient C =25,flowrate u =0.5m/s, x A =200m;
2. Preissmann scheme only consider the loworder condition, i.e. , u is not equal to 0.5, , , are respectively corresponding to the
differentvaluesof , u ,whereiscorrespondingto =0.6,and u =0.4,iscorrespondingto =0.4,and u =0.6.
Study of Civil Engineering and Architecture (SCEA) Volume 2 Issue 1, March 2013 www.seipub.org/scea
17
Conc l usi ons
The paper starts from Upwind scheme, stating the
process that generates numerical diffusion when
dispersing suspended load convectiondiffusion
equation with a lowlevel difference scheme, and it
tries to simulate the physical diffusion phenomenon
using numerical diffusion, simulation accuracy under
different parameters, such as t A , x A and u , are
calculated. And by horizontally comparing different
differential schemes we can get the following
conclusions:
1) Numericaldiffusionofdifferentialschemecanbe
used to simulate the actual physical diffusion
phenomenon, but it has certain requirements on
the parameter selection. For suspended load
convectiondiffusion equation, the simulation
method is to disperse the pure convection
equation first, and then propose diffusion term.
Since the coefficient in front of diffusion term is
the numerical diffusion coefficient, by assuming
whichequalstothephysicaldiffusioncoefficient,
wecangettherelationshipequationcomposedof
some parameters that meet the simulation
accuracy.
2) Ingeneral,inthevariousparameters,onlyspatial
interval x A and time step t A which are related
to the calculation can be chosen randomly, and
the selection of x A needs to be compatible with
the size of the project and the complexity of the
boundary, so the key is the time step t A . By
calculation, this paper has found the relationship
of t A that meets the accurate simulation under
various differential schemes, but usually the
calculated step sizes are with decimals, which is
not conducive to calculation, so an instance is
used to calculate the relative error between the
numerical diffusion coefficient and the physical
diffusion coefficient when t A takes a similar
integer value. It is found that as the step of low
level Preissmann scheme is larger, the possibility
of an integral step is greatly increased, thereby
increasing the calculation accuracy, which is a
recommendedmethod.
3) Thoughthisarticleonlyanalyzedsuspendedload
convection diffusion equation, it can be
speculated that bedload convectiondiffusion
equation, as well as the convectiondiffusion
phenomena of pollutants in the water, can be
analyzed using the same method. Similarly, the
method can be employed to solve the general
problems by means of its unique superiority of
differential equations clear mathematical
foundation, simple calculation and easy
programming,whichisstillfavourablycompared
toamoreprecisenumericalmethod.
REFERENCES
Gasiorowski D. Balance Errors Generated by Numerical
Diffusion in the Solution of Nonlinear Open Channel
Flow Equations. Journal of Hydrology 476 (2013): 384
394.
Papadakis Antonis P., and Metaxas Andrew C. Optimum
Mesh Dependent Diffusion Coefficient Proof from High
to Low Order Upwind Schemes Utilized in Plasma
Discharges. 10th WSEAS International Conference on
EHAC11 and ISPRA11, 3rd WSEAS Int. Conf. on
Nanotechnology, Nanotechnology11, 6th WSEAS Int.
Conf. on ICOAA11, 2nd WSEAS Int.Conf. on
IPLAFUN11,2011.
RichardCourant,KurtOttoFriedrichs,andHansLewy.On
the Partial Difference Equations of Mathematical
Physics.Math.Ann.100(1928):32.
XieJianheng.RiversSimulation.Beijing:WaterPowerPress,
1990.
Yang Guolu. Rivers Mathematical Model. Beijing: Ocean
Press,1993.
Zheng Bangmin, and Zhao Xin. Calculation of Water
Dynamics.Wuhan:WuhanUniversityPress,2001.