Proceedings of

Turbine 99 - Workshop on Draft Tube Flow

June 20–23, 1999, Porjus, Sweden

APPROXIMATIONS OF NUMERICAL ERRORS AND BOUNDARY CONDITIONS IN A

DRAFT TUBE

John Bergström
Luleå University of Technology
SE-971 87 Luleå
Sweden

E-mail: John.Bergstrom@mt.luth.se

ABSTRACT 125rpm and has a flow rate of 149.0 m3/s. The model turbine
A simulation of the three-dimensional flow field in the has a head of 4.5 m, speed of 595 rpm, flow rate of 0.533 m3/s
Turbine 99 draft tube using a RST closure is reported in this and delivers approximately 32 kW. The geometry of the draft
paper. The best efficiency point measurements were used as tube is a typical design from the 1940’s with a sharp outer
boundary conditions. The report focuses on boundary corner in the bend (figure 1).
conditions, grid quality, numerical error, analysis of the
resulting flow field and comparison to experimental data. The The experiments were made at Vattenfall Utveckling,
incomplete measurements at the inlet made it necessary to Älvkarleby, Sweden (Andersson and Karlsson, 1999a and
estimate boundary conditions for the radial velocity, the 1999b). The data supplied for the workshop were conducted at
Reynolds stresses and the turbulent length scale. A block- 60% load, close to the highest efficiency for the system. Two
structured grid was constructed to achieve a grid of high quality, cases were measured (both at a head of 4.5 m), one at the
which made it necessary to reconstruct the initial surface highest efficiency point (Q=0.533 m3/s) on the propeller curve
topology of the draft tube. The numerical errors, grid and and one at a higher flow rate (Q=0.544 m3/s). Laser Doppler
iterative errors were estimated to 3%-7% and 0.8-0.002%, Velocity measurements at one cross section further downstream
respectively. Comparison with pressure measurements at the of the draft tube were conducted as well as the inlet boundary
upper and lower centerline yielded satisfactory results for the conditions. Pressure measurements at several locations at the
finest grid (700000 grid points). Comparison with velocity walls were also conducted.
measurements at a cross section further downstream yielded
qualitatively good results for the main flow but poor agreement One engineering quantity, the pressure recovery factor, will
for the secondary flow. The flow field had several interesting be used, in conjunction with detailed data, to compare the
flow phenomena including the vortex below the runner cone measurements with the calculations. The present contribution
and a large separated region at the outlet. will focus on the construction of the grid, numerical errors and
the boundary conditions.
INTRODUCTION
The Turbine 99 workshop is aimed to be a follow-up to the It is important that the solution from a CFD calculation is
GAMM Workshop held at EPFL in Switzerland, 1989 (Sottas close to being grid independent. However, it is in most cases
and Ryhming 1993) where the most advanced numerical impossible to achieve a completely grid independent solution.
simulations at that time were used. The goal with the Turbine 99 Therefore it is necessary to estimate the errors. The errors in
workshop is to assess current state-of-the art numerical numerical flow simulations can be subdivided in two parts, grid
simulations of the flow field in a hydraulic turbine draft tube by errors and iterative errors. Grid errors (discretization errors)
comparing numerical calculations with LDV and pressure were estimated by using Richardson extrapolation to estimate
measurements. the solution for a grid having infinitely many grid points. This
estimation can then be used to calculate the error. These
The draft tube is a scale model (1:11) of the real draft tube methods require the use of several grids in a sequence (at least
belonging to the Hölleforsen hydro power plant. It is situated in three), (Bergström and Gebart 1999), (Ferziger and Peric,
the Indals River and was started in 1949. The real plant has a 1996).
head of 24.9 m, delivers 140 MW (three turbines), rotates with

1
 Iterative errors can be calculated in several different ways, NUMERICAL METHOD
e.g. using the residuals or to calculate an estimate of the real The code used for solving the governing equations is the
iterative error, (Ferziger and Peric, 1996). In the present case parallel version of the commercial code CFX (AEA
the residuals were used. Because of the coupling between the Technology, 1997). It is a finite volume based code using a
iterative error and the residual, the reduction of the residuals of structured non-staggered multi-block grid. To avoid
a certain order of magnitude implies that the error has decreased checkerboard oscillations in pressure and velocities, the Rhie-
by a comparable amount, (Ferziger and Peric, 1996). Chow (1983) interpolation method is used. The data transfer
between blocks is done by the introduction of dummy cells
Closely connected to the convergence rate and the outside the boundary of each block. This will make each block
numerical error is the grid quality. If a poor grid is generated, overlap a neighbor block. The interior values in one block will
this might result in poor convergence or divergence. In this become the boundary conditions for the neighbor block and
case, where a block-structured grid is used, it is important how vice versa (AEA Technology, 1997).
the geometry is divided into subdomains (or “blocks”). Keeping
the control volume edges as orthogonal as possible was one of All terms in all equations are discretized using second-
the problems which were encountered when trying to generate order centered differencing apart from the convective terms in
the grid for the Turbine 99 case. the momentum equations. The convective terms are discretized
using higher order upwind, a second-order method. The
One of the challenges in engineering calculations is how to convective term in the remaining equations are discretized using
treat boundary conditions at inlet and outlet. Such a problem, as hybrid differencing.
in the current case, is the incompleteness of the boundary
conditions, i.e. not all the necessary inlet and outlet conditions The PISO algorithm (Jang et. al, 1986) is used for the
are given as input data for the calculations. This missing pressure-velocity coupling. This algorithm was found to give
information must be estimated, guessed or in some way better iterative convergence rate than the SIMPLEC algorithm.
prescribed. This is a common problem when trying to simulate The PISO algorithm (which is originally a time-marching
“real” cases. Usually we do not know all velocity components, procedure) uses one or several correction steps (additional
Reynolds stresses or length scales that must be specified for a pressure correction equations), (Jang et. al, 1986).
simulation. In the workshop case there is probably much more
information available than what is common. More commonly TURBULENCE MODEL
the only available information would be the flow rate and the The turbulence model is a Reynolds stress model, (Daly
runner speed for a draft tube simulation. One way to get all the and Harlow, 1970), (Rotta, 1972), (Naot et. al, 1970). It uses
necessary information at the inlet of the draft tube would be to individual equations for the Reynolds stresses and one equation
simulate the flow field separately in the runner. However, a for the dissipation. The terms that must be modeled in the
runner simulation would also need inlet information that might Reynolds stress equation are the turbulent diffusion and the
not be so easily obtained (but probably easier than for the draft pressure strain. The turbulent diffusion is modeled using a
tube inlet). gradient-type model (Daly and Harlow, 1970). The pressure
strain is written as the sum of three terms. A return to isotropy
Starting by describing the numerical method this report term by (Rotta, 1972), a return to isotropy of production term
continues with the turbulence model, treatment of the boundary by (Naot et. al., 1970) and a wall reflection term (Gibson and
conditions, grid generation, numerical errors, comparison to Launder, 1978), which is omitted in the present turbulence
experimental data and finally some required results for the model (AEA Technology, 1997).
workshop.
For the dissipation tensor it is assumed that the turbulence
is locally isotropic. A modeled ε-equation similar to the one
used in the k-ε model is used. More details about the model can
be found in (Bergström, 1997) and (AEA Technology, 1997).

BOUNDARY CONDITIONS

Inlet and outlet boundary conditions

At the inlet, the measurements from the highest efficiency
point on the propeller curve (case T) at the Ia(1) radius were
used. At this location the axial (U) and tangential (V) velocity
components were measured. Also the root mean square values
of U and V, u’ and v’ and the cross correlation of U and V, u’v’
Fig. 1 The Turbine 99 draft tube geometry.

2
were measured. This is not enough information for a simulation derivative of the velocity at the wall and at the outer edge of the
using the current Reynolds stress model. The missing data were boundary layer. Of these four conditions only the derivative at
the radial velocity, one of the normal Reynolds stresses w’w’, the wall could not be calculated from experiments. Then the
all Reynolds shear stresses and the turbulent dissipation which derivative (dU/dr) at the inner wall was adjusted until the inner
must be specified for the current turbulence model. The and outer boundary layer had approximately the same thickness
velocities were not measured all the way to the walls (the draft (~ 4 mm). The final velocity profile can be seen in figure 3.
tube wall and the rotating runner cone wall) so some estimation
of the velocity variation from the last measurement point up to
the wall must be done.

4
Radial velocity
The radial velocity was computed assuming that the 3.5
velocity vector was parallel to the walls at the runner cone wall
and the outer draft tube wall. The angle of the walls at the inlet 3

measured from a vertical plane was θ = 16.8 degrees at the 2.5

runner cone wall and θ = –6.9 degrees at the outer draft tube
2
wall. This angle was used to calculate the radial velocity as:
1.5
ur = uaxial tan(θ ) (1)
1

where the angle θ was assumed to vary linearly from the runner 0.5

cone to the outer wall (figure 2).

0
0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24

0.6

0.4 Fig. 3 Assumed axial velocity

0.2
The same procedure was used for the tangential (or
0 azimuthal) velocity profile (figure 4).
-0.2

-0.4

7
-0.6

-0.8 6

-1 5

-1.2
0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 4

Fig. 2 Assumed radial velocity. 2

Axial and tangential velocities 1

The axial and tangential velocities were measured from 0

r=103.5 mm to r=235.9 mm so the variation close to the wall 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24

(rinner=98.09 mm router=236.46 mm) had to be estimated. At the

outer draft tube wall the velocity was specified close enough to
the wall to make a linear interpolation to the no-slip condition at Fig. 4 Assumed tangential velocity. The high peak at
the wall possible. the inner radius is the velocity of the runner cone.

At the inner wall (the runner cone) the velocity profile was Reynolds Stresses
assumed to have the same boundary layer thickness as at the To calculate the Reynolds stresses from the given rms
outer wall. Then a fourth order polynomial velocity profile was values of the velocities it was assumed that
assumed close to the inner wall. This requires the value and the

3
 ui ' ui ' = ui '⋅ui ' (2) same size as the smallest distance between the runner blades.
The Kaplan runner in the test case has 5 blades and the runner
blade angle was about 25 degrees measured from a horizontal
The Reynolds stresses were then linearly interpolated from
plane (Urban Anderson, Personal communication). The smallest
the experimental data (figure 5 shows the rms values). The u’
distance at the mean radius (167 mm) was then 0.09 m.
and v’ profiles are quite similar so it was assumed that w’=v’,
i.e. that the turbulence is approximately isotropic at the inlet of
the draft tube. The only Reynolds shear stress given was u’v’. Wall boundaries
This stress was about 10% of the normal stresses. Because no In the current Reynolds stress model, a wall function
information about the remaining shear stresses (u’w’ and v’w’) approach is only used for the dissipation but not for the
were given and the fact that the u’v’ stress was small, all Reynolds stresses. The Reynolds stresses are linearly
Reynolds shear stresses were set to zero extrapolated from values in control volumes interior to the flow
(Clarke and Wilkes, 1989).
u' v ' = v ' w' = u' w' = 0 (3) Outlet
At the outlet all quantities were given a zero normal
Both the velocities and the Reynolds stresses were assumed gradient except for velocities, which were given a constant
to be axisymmetric at the inlet. This made it necessary to gradient. The pressure is extrapolated from upstream values.
transform both these quantities depending on the azimuthal This is equivalent to assume that the flow field is fully
positions at the inlet of the draft tube. developed at the outlet. Initially the velocity is first given a zero
normal gradient. The difference from actual mass flow rate is
0.8
then calculated. Then a constant multiplied by the outward
going normal is added to the velocities in order to ensure exact
0.7 global mass conservation.
0.6
GEOMETRY AND GRID GENERATION
0.5 The basis for the 3D geometry construction for this test
case was the original paper drawing of the model draft tube.
0.4 This paper drawing was converted to a 3D solid model using
0.3
the CAD software I-DEAS. This task caused some trouble
because the real draft tube was not exactly constructed by the
0.2 measures in the paper drawing. After consulting Urban
Andersson at Vattenfall Utveckling, Sweden (who also
0.1
conducted the boundary condition measurements for this
0 workshop) the geometry was adjusted to be as close as possible
0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24
to the real geometry. This geometry was then converted to
several CAD formats. In this case the IGES format was used.

Fig. 5 The rms values of the axial (u’, blue, dash- The solver used, CFX 4.2 (AEA Technology, 1997), is a
dotted curve) and tangential velocities (v’, red, upper block-structured code. The geometry must therefore be
curve) subdivided into several inter-connecting solids before creating
the grid. The surfaces of the original model were not suited to
Turbulent dissipation ε create a mesh of high quality, a must if good convergence rate
The turbulent dissipation ε was calculated by assuming that should be possible. The original model was therefore redrawn
quite drastically in order to achieve a well-behaved grid.
k 3/ 2
ε = Cµ (4) The most critical task was to create a high quality mesh at
l
the upper wall of the bend. The grid for this part of the
geometry can be seen in figure 6. This grid is one way to
where k is the turbulent kinetic energy (which can be calculated
achieve a high quality mesh. There are probably several other
from the normal Reynolds stresses) and Cµ is a model constant
possible ways to reconstruct the geometry.
(same as in the k-ε model Cµ=0.09 ). In order to use this
relationship the turbulent length scale or integral scale (l) has to
be specified. (Tennekes and Lumley, 1972) define l as the size
of the largest eddies or the width of the flow. In this case it
seemed reasonable to assume that the largest eddies were of the

4
 where h is the grid cell size, α is a grid refinement factor and a1
is a coefficient that depends on the derivatives but do not
depend on h. Eq. (5) can be used for three grids in a sequence
resulting in an expression for p, the actual order (as opposed to
the formal order) of the numerical scheme.

φα 2 h − φα 3 h α 3p − α 2p
= p (6)
φα 1 h − φ α 2 h α 2 − α1p

The three grids have to be in the asymptotic range for this

analysis to work (h has to be small enough). p can then be used
Fig. 6 The block topology at the upper wall. to calculate an approximation of the exact solution (φextrapolated).
α is a grid refinement factor defined as
If the original topology (figure 7) was used, almost zero-
degree angle elements were generated at the sidewall of the 1/ 3
draft tube.  N1 
α=  (7)
 N2 

where N1 is the number of control volumes (or grid points) for

the finest grid and N2 for a coarser grid. Then the grid error can
be estimated using:

φ extrapolated − φ h
er ,approx = (8)
φ extrapolated

A question remaining to be answered is what value φh one

should use to calculate the grid error. A natural suggestion in
this case would be the pressure recovery factor that is one way
Fig. 7 Result of grid generation when using the to describe the efficiency of the draft tube. The pressure
original surfaces. recovery factor is defined as:

The large errors in these badly shaped elements made it Pout :wall − Pin:wall (9)
impossible to achieve a converged solution during the initial C pr = 2
1  Q
test calculations. This shows how important a good mesh is to ρ 
the solution. 2  Ain 

NUMERICAL ERRORS where Pout:wall and Pin:wall are the mean wall pressures at the
The largest numerical error in Computational Fluid outlet and inlet. This definition is adopted because the pressure
Dynamics is the discretization error or grid error as found out measurements were only conducted at the walls. The number of
by (Bergström and Gebart, 1999). This error will be considered grid points, the grid refinement factor and the pressure recovery
first. The second error, the iterative error, is usually a smaller factor are presented in table 1.
problem than the discretization error.

Grid error
Richardson extrapolation was used to calculate the grid
error as described in (Ferziger and Peric, 1996). This method
expand the discretization error as the leading term in a Taylor
series:

ε h = φ exact − φ αh = a1 (αh) p +H (5)

5
Grid No. of grid points α Cpr Cpr - s φ(α ) = φ extrapolated + ηα p (12)
1 725779 1 0.87174 0.93217
2 515752 1.1206 0.87838 0.94154
3 361279 1.2618 0.86549 0.93217

Table 1. Grid sizes, grid refinement factors and

pressure recovery factor. Cpr is based on wall
pressure. Cpr-s is based on mean values at the inlet
and outlet surfaces.

Using the values of table 1 (the Cpr based on wall pressures) in

eq. (6) yielded a negative left-hand side. This suggests that the
error decreases with alternating sign, i.e. we have oscillatory
grid convergence (Celik and Karatekin, 1997). Celik and
Karatekin suggest an alternative way of calculating the actual
order p for this case:

φα 2 h − φ α 1 h α1p + α 2p
= p (10)
φα 2 h − φα 3 h α 2 + α 3p
Fig. 8 Richardson extrapolation (p=1 and 2) applied
to the result from three grids in a sequence. alpha is
This equation resulted in p=5.668. This is of course not a the grid refinement. The box, circle and plus symbols
reasonable value of p (which should be between 1 and 2 are the results from grid 1,2 and 3, respectively.
because both higher order upwind and hybrid differencing is
used). This can depend on the way in which the grid is refined. Eq. (7), (11) and (12) can also be used to estimate the necessary
The grid is not refined equally in all direction because of the grid size for a grid error of e.g. 1%. This results in a necessary
necessity to have a fine grid close to the walls. If the grid size of 3.9*106 grid points for p=2 and 222*106 for p=1.
Richardson extrapolation method should be meaningful, grids Unfortunately, due to available computer resources it is not
even finer than the finest used here would have to be generated. possible to check if these really are the required grid sizes.
It is possible that the coarsest grid (361279 grid points) is still However, we still do not know for sure that the grids are in the
not fine enough. asymptotic range making it a bit hazardous to conclude
anything about the grid convergence error.
However, it is possible that the two finest grids are in the
asymptotic range and then it is still possible to estimate the grid Iterative error
convergence error. But then the actual order p can not be The iterative convergence was controlled by the residuals
calculated. A natural suggestion would be to calculate the error of the equations. The residuals are coupled to the real iterative
using Richardson extrapolation using both p=1 and 2. In this error. This means that a reduction in the residuals implies a
way we will at least get an upper and lower limit for the grid reduction in the iterative error. If zero initial values are used as
convergence error (but one can not be absolutely sure that the a starting guess then the initial error will be equal to the solution
grids really are in the asymptotic range). Richardson itself. E.g. if the residuals have fallen about 4 orders of
extrapolation can be expressed as (Bergström, 1997): magnitude, then the error should have fallen by a comparable
amount, i.e. the error is about 0.01% of the solution, (Ferziger
α 2p φ h − φ α 2 h and Peric, 1996). (Ferziger and Peric, 1996) also suggest that
φ extrapolated = (11) the iterations can be stopped when the residuals for the inner
α 2p − 1
iterations have decreased 1-2 orders of magnitude (Table 2) and
3-5 orders of magnitude for the outer iterations (Table 3). The
Assuming p=2 yields φextrapolated=0.84578 and er,approx=- convergence criterion for the outer iterations is fulfilled for all
3.06950% and for p=1, φextrapolated=0.81669 and er,approx=- equation except for some of the Reynolds stresses. For the inner
6.74108%. To summarize: the error varies between 3-7% if p is iterations the convergence criterion is fulfilled for all equation
assumed to be between 1 and 2. except for the pressure correction equation. However, it should
be noted that table 2 only shows the average reduction in the
A plot of Cpr for the three grids is shown in figure 8. The curves residuals for all outer iterations.
in this figure are obtained by solving for η in:

6
 u-velocity residuals 1,79*101
v-velocity residuals 1,72*101
w-velocity residuals 1,64*101
mass source residuals 5,88*100
ε residuals 1,10*101
uu residuals 5,56*101
vv residuals 5,00*101
ww residuals 4,76*101
uv residuals 5,56*101
vw residuals 4,55*101
wu residuals 5,26*101

Table 2. Average reduction in residuals for the inner

iterations (ratio of the initial residual and the final
residual).

Residual Reduction Error

Fig. 9 Pressure at the upper centerline. The pressure
u-velocity residuals 1.4*104 0.007%
has been normalized by dividing the pressure at all
v-velocity residuals 6.2*103 0.016%
locations by the pressure at the first point (at
w-velocity residuals 6.4*103 0.016% Distance=0). Distance is the distance in meters from
mass source residuals 5.3*104 0.002% the inlet of the draft tube.
ε residuals 3.0*103 0.033%
uu residuals 4.9*102 0.204%
vv residuals 4.4*102 0.227%
ww residuals 1.2*102 0.833%
uv residuals 1.3*103 0.077%
vw residuals 4.2*102 0.238%
wu residuals 5.8*102 0.172%

Table 3. Reduction in residuals of the outer iterations

(ratio of residuals from second and last iterations)
and error for each variable.

COMPARISON TO EXPERIMENTAL DATA

Pressure measurements
Pressure measurements were (after the workshop) available
at the upper and lower centerline of the walls and at the walls at
the inlet and outlet of the draft tube. Figures 9 and 10 shows the
pressure for the three grids at upper and lower centerline and
the measured pressure.
Fig. 10 Pressure at the lower centerline. The pressure
has been normalized by dividing the pressure at all
locations by the pressure at the first point (at
Distance=0). Distance is the distance in meters from
the inlet of the draft tube.

Figure 9 shows that grid 1 (725779 grid points) agrees

quite well with experiment except for the pressure drop at
Distance=0.5 which is over predicted and an area of too high
pressure between Distance=1 to 2. Grid 1 shows the same

7
agreement for the lower centerline (fig. 10) but the pressure is
too high between Distance=0.5 to 1.5.

Sorting the pressure recovery factor by increasing values

results in the order: Grid 1, Grid 3 and Grid 2. The pressures at
the upper and lower centerline behave in the same way. This
again indicates that at least the solution from the coarsest grid is
not in the asymptotic range (required for eq. 6 to work).

Velocity measurements
The velocity was measured at the inlet (the boundary
conditions for the workshop) and at a cross section further
downstream (named cross section III in the workshop) (figure
11).

Fig. 12 Velocity perpendicular to cross section III. The

upper part shows the CFD result from Grid 1. The
gray scale is the same for both pictures.

Figure 13 shows the v-velocity component (in positive y-

direction) in cross section III. For this component the agreement
is not good. The experiments indicate that there is one large
Fig. 11 Location of cross section III. A contour plot vortex rotating in the anti-clockwise direction but the simulation
visualizes the velocity component perpendicular to predicts two counter rotating vortices (cf. fig. 14).
the cross section.

In figure 12 the velocity component perpendicular to cross

section III is plotted. The agreement is at least qualitatively
close to the experiments. There is a low speed area to the left
and lower left in both experiment and simulation and a high-
speed area in the lower right corner.

8
 Quantity Inlet Cross section III
αaxial 1.0488 1.1763
αswirl 6.1553*10-2 2.0140*10-2
β 1.1026 1.0362
S 0.1816 6.0363*10-2

Table 4. Engineering quantities required for the

Turbine 99 workshop.
αaxial and αswirl are kinetic energy corrections factors
(Staubli and Deniz, 1994), β is the momentum correction factor
(Daugherty et. al, 1989) and S is the Swirl intensity (Seeno et.
al, 1978).

The predicted pressure recovery factor was Cpr=0.87174

and the energy loss coefficient was ζ=0.1135 (for the finest
grid). The measured pressure recovery factor was
Cpr=1.122767. The error is 22%; i.e. the simulated pressure
recovery factor is 22% lower than the measured value.

Flow field and pressure distribution

The most striking result from the simulation is the large
vortex that starts just below the runner cone and extends all the
way down to the bottom wall (figure 15). This was expected,
e.g. (Batchelor, 1967) shows analytically that in the case of a
Fig. 13 V-velocity (in positive y-direction to the right in rotating flow inside a pipe having a inner boundary shaped as a
the figure) at cross section III. The upper part shows cone, the azimuthal velocity becomes infinitely large
the CFD result from Grid 1. The gray scale is the same downstream the apex of the cone.
for both pictures.

Fig. 14 Secondary velocity in cross section III.

Engineering quantities
A number of engineering quantities were requested for the Fig. 15 The vortex below the runner cone visualized
workshop. These are presented in table 4. All values are for grid by vortexlines. The vortexlines correspond to the
1. vorticity vector as streamlines correspond to the
velocity vector.

9
 The vortex is also an area of recirculation as shown in
figure 16. The velocity in the area below the cone is directed
upwards towards the cone. This is because of the high rotational
speed in the vortex inducing a low pressure just below the cone
(figure 17). Flow visualizations with fluorescent dye showed
qualitatively the same results.

Fig. 18 Separated regions visualized by iso-surfaces

of the x-components of the velocity (ux=0).

Fig. 16 The vortex visualized using two planes (x=0

and y=0) representing the velocity (m/s) in the z-
direction (w). The w-velocity is positive in the vortex
indicating an upward flow in this area.

Fig. 19 The separated region at the upper wall (y=-

0.33). The colors represent the pressure distribution
and shows the adverse pressure gradient inducing
the separation.

The separation zones are also visible at the outlet.

Visualizing the flow field (figure 20) also shows that large
velocities at only a small part of the outlet dominate the flow.

Fig 17 The low pressure (Pa) in the area below the

apex of the cone forcing the flow field upwards.

Separated regions were one of the phenomena of interest

for the workshop. This simulation results in three regions close
to the outlet where separation occurs, the largest one
downstream the edge at the upper wall. Figure 18 shows these
three regions. A closer view of the largest region is shown in
figure 19.

10
 to some extent, how the geometry is divided into subdomains).
In this case it was decided that the grid should have as equal
resolution as possible in all directions in order to be sure that all
flow phenomena were captured. It is of course possible to guess
in advance where interesting areas in the computational domain
exist and create a finer grid there. However, it is possible that
some unexpected phenomena arise.

The prediction of the pressure recovery factor is dependent

on the pressure at the walls, which in turn is dependent on the
prediction of the remaining variables close to the wall. One of
the reasons for the poorly predicted pressure recovery factor
can be that the wall boundary conditions for the Reynolds
stresses are not defined by wall functions in the present code.
Instead, the Reynolds stresses are linearly extrapolated from
values in control volumes interior to the flow (Clarke and
Wilkes, 1989). Using wall functions for the Reynolds stresses
might improve the solution, e.g. like the work by (Perzon,
Fig. 20 The outlet flow field. The flow field is non-
1997). A wall treatment similar to the work by (Perzon, 1997)
uniform with large velocities at the y>0 lower corner
will be included in a future project.
of the outlet. There are also low magnitude inward
velocities in two regions indicating two separation
The creation of the geometry and grid was a major part of
zones. (Cf. figure 18).
the work. The block-structured grid used required drastic re-
construction of the original CAD geometry. This was necessary
There were no further major separation or recirculation
in order to achieve a high quality grid (which is necessary to
zones found in the domain in addition to the ones already
achieve good iterative convergence), especially at the upper
mentioned.
wall where the cross section changes from being circular to
rectangular. If the original CAD surfaces were used to create the
COMPUTER REQUIREMENTS
subdomains, large errors in badly shaped elements at the upper
CFD simulations of this type undoubtedly require large
wall made it impossible to achieve a converged solution during
computers. In this case a parallel version of the code was used
the initial test calculations. This shows how important a good
on a 10 processor Silicon Graphics Onyx2 at High Performance
mesh is to the solution.
Computing Center North (HPC2N) in Umeå, Sweden. CPU
time, number of iterations and number of processors for all
The problem of achieving a high quality grid always
grids are reported in table 5.
appears when a block-structured grid is used in combination
with a complicated geometry. There are of course some rules of
No. of grid points No. of iter. CPU time processes thumb one can follow when creating the grid, but each time a
725779 2000 9.5 hours 6 new geometry is encountered completely new problems arise. It
515752 2000 6.8 hours 6 is possible that the use of an unstructured grid makes grid
361279 2000 6.2 hours 4 generation easier in this case. However, if a block-structured
grid is the only choice, a way of improving the quality of the
Table 5. Computational times for all grids. grid can be the use of more advanced grid generation software
available (e.g. ICEM). This can help the user to create block-
DISCUSSION structured grids of higher quality, but in the end it is still up to
Starting with the boundary conditions at the inlet it was the user to decide the topology of the grid. This can be a
necessary to make additional assumptions about the velocity problem when comparing results from calculations where the
variation in the boundary layer. However, the question is how same original geometry has been used, but different topologies.
important this really is. Even if the grids were large, the radial One at least can hope that grid refinement studies show that the
resolution of the grid was only 11 control volumes at the inlet topology is of secondary importance. However, this is yet to be
(with smaller control volumes at the walls but not fine enough proven.
to completely resolve the boundary layer). This means that the
variation of the velocity in the boundary layer in the simulation The grid error was 3%-7% depending on the assumed order
really is no better than the original measurements. The reason of the differencing schemes and the iterative error was
for using such a “small” number of control volumes at the inlet approximately 0.8-0.002% depending on the variable. The
is that one have to decide were the grid should be refined (and errors were generally larger for the Reynolds stresses. This

11
shows that the iterative error is smaller than the grid desirable. It is also possible to conduct a transient simulation of
convergence error. A result also found out by (Bergström and this case with the presently available boundary conditions. This
Gebart, 1999). Following the recommendations by (Ferziger can probably shed some light on why is it difficult to reach a
and Peric, 1996) the solution has not converged for some of the steady state solution. However, one might ask if turbulence
variables. This can indicate that trying to reach a steady state models based on RANS are suited for transient simulations.
solution in this case is not possible. A transient simulation Information about the behavior of turbulence models (RANS) in
should be included in a future study of this case. unsteady flows can be found in (Leschziner, 1998). Generally
one can say that turbulence models (RANS) can be used only if
In this study the asymptotic range was not reached for the the unsteadiness is quasi-steady or slow but not if it is moderate
present grid sequence. The most obvious reason is that the grids or fast (Leschziner, 1998).
are too coarse. It is also possible that only the coarsest grid is
outside the asymptotic range. The only way to answer this The second phenomenon is the large separation region at
question would be to create one or several larger meshes. the upper wall. This region stretches out outside the outlet.
However, it is also possible that the present grids are sufficient Because an assumption of a fully developed flow at the outlet is
and that the poor iterative convergence (towards a steady state assumed the region outside the draft tube should be included in
solution) for the Reynolds stresses is the reason for not reaching the simulation. This however, will make the grid even larger
the asymptotic range. But as mentioned above, it might not be and the equalization wall in the tank downstream the draft tube
possible to reach a steady state solution in this case. has to be modeled in some way. Also, the small fraction of
incoming fluid into the outlet of the draft tube is probably not a
The comparison with pressure measurements at the upper large error source in the calculation.
and lower centerlines shows quite good agreement for the finest
grid except for the pressure drop at the upper centerline and the ACKNOWLEDGMENTS
area of maximum pressure at the lower centerline. The pressure The financial support of the Swedish National Board for
drop at the upper centerline occurs at the end of the first conical Industrial and Technical Development, Elforsk, and Kvaerner
section of the draft tube (this pressure drop is visible to the right Turbine AB is gratefully acknowledged.
in figure 17). The reason for the disagreement in this area is
probably that the grid is not fine enough in this area to resolve This research was conducted using the resources of High
this sharp drop. The area of high pressure at the lower Performance Computing Center North (HPC2N).
centerline consists of two maximums, the first one also visible
in figure 17 to the lower left and a second, more widespread REFERENCES
area, at the lower back corner of the draft tube. Visualizations AEA Technology, (1997), CFX 4.2: Solver, CFX-
(Andersson and Karlsson, 1999a and 1999b) show that there is International, 8.19 Harwell, Didcot, Oxfordshire OX11 0RA,
a quite large vortex in this corner. This vortex was not predicted United Kingdom.
in the simulation, explaining why the pressure is wrongly
predicted in this area. Andersson, U. and Karlsson, R., (1999a), Quality Aspects
of the Turbine 99 Draft Tube Experiment, to be published in the
The pressure recovery factor was found to be 22% lower Proceedings of Turbine 99 - Workshop on Draft Tube Flow
than the measured value. The reason for this (except for the June 20–23, 1999, Porjus, Sweden
reasons mentioned above) might be that the outer boundary
layer at the inlet of the draft tube is not resolved enough. Andersson, U., (1999b), Turbine 99 – Experiments on
Applying the law of the wall and wall functions too far from the Draft Tube Flows (Test Case T), to be published in the
wall can result in a wrongly predicted pressure. The assumption Proceedings of Turbine 99 - Workshop on Draft Tube Flow
of a radial velocity (fig. 2) which was not measured at the inlet June 20–23, 1999, Porjus, Sweden
can of course also influence the pressure distribution at the
inlet. Batchelor, G.K., (1967), An introduction to Fluid
Dynamics, Cambridge University Press, Cambridge.
The result from the simulation shows two main phenomena.
The first is the vortex downstream the runner cone. This vortex Bergström, J. and Gebart, B.R., (1999), Estimation of
is probably unsteady, with the core of the vortex rotating itself Numerical Accuracy for the Flow Field in a Draft tube,
and parts of this vortex coming off, affecting the flow field all International Journal of Numerical Methods for Heat & Fluid
the way to the outlet. This makes it questionable if a steady state Flow, Vol. 9, No. 4.
solution of this problem is possible to reach. However, a pure
transient simulation of this case would require transient Bergström, J., (1997), Turbulence modelling and numerical
boundary conditions at the inlet and the computational time accuracy for the simulation of the flow field in a curved
would probably be very long if high resolution in time is

12
channel, 1997 ASME Fluids Engineering Division Summer Rotta, J.C, (1972), Turbulente Stromungen, B.G. Teubner,
Meeting, FEDSM’97, Vancouver, Canada, June 22-26. Stuttgart.

Clarke D.S., and Wilkes N.S., (1989), The calculation of Tennekes, H. And Lumley, J.L., (1972), A First Course in
turbulent flows in complex geometries using a differential stress Turbulence, The MIT press, Massachuttes.
model, AERE-R 13428, Engineering Sciences Division,
Harwell Laboratory, Oxfordshire OX11 0RA. Wilcox, D.C., (1993), Turbulence Modeling for CFD,
DCW Industries, La Cañada, California.
Celik I. and Karatekin O., (1997), Numerical Experiments
on Application of Richardson Extrapolation With Nonuniform Seeno, Y., Kawaguchi, N. and Tetsuzou, T., (1978), Swirl
Grids, Journal of Fluids Engineering, Vol. 119, pp. 584-590. flow in conical diffusers, Bulletin of the JSME, Vol. 21., No.
151., January, pp. 112-119.
Daly, B.J. and Harlow, F.H., (1970), Transport equations of
turbulence, Phys. Fluids, Vol. 13, pp. 2634-2649. Sottas, G. and Ryhming, I.L., (1993), 3D-Computation of
Incompressible Internal Flows, Notes on Numerical Fluid
Daugherty, R.L, Franzini, J.B. and Finnemore, E.J., (1989), Mechanics, Vol. 39, Friedr. Vieweg & Sohn Verlagsgesellschaft
Fluid Mechanics with engineering applications, McGraw-Hill mBH, Braunschweig/Weisbaden, Germany.
Book Company, Singapore.
Staubli, T. and Deniz, S., (1994), Analysing draft tube
Ferziger, J.H. and Peric, M., (1996), Computational kinetics, Int. Water Power & Dam Construction, December, pp.
methods for Fluid Dynamics, Springer-Verlag, Berlin 38-42.
Heidelberg.

Gibson, M.M. and Launder, B.E., (1978), Ground effects

on pressure fluctuations in the atmospheric boundary layer,
Journal of Fluid Mechanics, Vol. 86, pp. 491-511.

Jang, D.S., Jetli, R., and Acharya, S., (1986), Comparison

of the PISO, SIMPLER, and SIMPLEC Algorithms for the
Treatment of the Pressure-Velocity Coupling in Steady Flow
Problems, Numerical Heat Transfer, vol. 10, pp. 209-228.

Leschziner, M. A., (1998), The Computation of Turbulent

Engineering Flows with Turbulent-Transport Closures, will be
published in Peyret, R. and Krause, E., eds., Advanced
Turbulent Flow Computations, CISM Publication 395, lecture
notes from the 9-th IUTAM Summer School: Advanced
Turbulent Flow Computations, September 7-11, 1998,
International Centre for Mechanical Sciences, Udine, Italy
(URL: http://www.uniud.it/www/cism/).

Naot, D., Shavit, A. and Wolfshtein, M., (1970),

Interactions between components of the turbulent velocity
correlation tensor, Israel J. Techn., Vol. 8, pp. 259-269.

Perzon, S., (1997), Reynolds Stress Modeling of Flow

Separation on Curved Surfaces, Thesis for the degree of
Licentiate of Engineering, Chalmers University of Technology,
Göteborg.

Rhie, C.M. and Chow, W.L., (1983), Numerical study of

the turbulent flow past an airfoil with trailing edge separation,
AIAA Journal, Vol. 21, No. 11, pp. 1525-1532.

13