Unsteady Rotor-Stator Interaction in A Low Pressure Centrifugal Compressor

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Unsteady Rotor-Stator Interaction in a

Low Pressure Centrifugal Compressor



MIHAI LEONIDA NICULESCU
&
, STERIAN DNIL
*

&
Ph. D. Student,
*
Professor
Faculty of Aerospace Engineering
Politehnica University of Bucharest
1 Polizu, sect. 1, 011061, Bucharest
ROMANIA
[email protected], [email protected], http://www.aero.pub.ro

Abstract: - The aim of this paper is to study the unsteady rotor-stator interaction in a low-pressure centrifugal
compressor using the finite volume method to solve the Unsteady Reynolds-Averaged Navier-Stokes. In order to
understand better, the rotor-stator interaction, the unsteady results are processed using both Adamczyk decomposition
[8] and Proper Orthogonal Decomposition (POD) [3, 4]. Both decompositions show the behavior of unsteady rotor-
stator interaction but the POD modes also show the numerical errors.

Key-Words: - Unsteady Rotor-Stator Interaction, Adamczyk decomposition, POD, CFD, Compressor, URANS

1 Introduction
The flow in turbomachinery is complex: three-
dimensional, viscous, and unsteady, with time scales
that vary considerably. This behavior makes difficult
the complete flow analysis. Usually, in the industrial
practice, the flow is assumed steady in the reference
frame linked to the studied row. Furthermore, one can
consider that every flow is composed by a main flow
and a secondary flow that contains physical phenomena
with no null velocity rotor. The secondary flows are
characterized by vortexes that lead to three-dimensional
behavior of flow and generate the losses due to the
entropy increase.
There are some sources of unsteady phenomena, in
turbomachinery and C. Dano [1] describes them after
their origins. Because the rotor-stator interaction can
affect dramatically the turbomachinery performance, we
paid it a special attention in this paper. The majority of
researchers that studied this interaction from the
numerical point of view focused their research on
transonic turbomachinery; therefore, there is very few
information about the rotor-stator interaction for low
velocity turbomachinery. Moreover, a recent study [2]
showed important discrepancies between experimental
and numerical results for a low-pressure centrifugal
stage. Unfortunately, this study did not succeed to
identify the effects that caused the major discrepancies
between experimental and numerical results. For this
reason, we have considered that it is useful to study the
rotor-stator interactions in a low-pressure centrifugal
stage, using both Adamczyk and proper orthogonal
decomposition.
Up to now, the Fourier transform is a common tool
for data storage of for the analysis of any periodic
signal. Some recent studies [3, 4] clearly showed that
POD is a more efficient method to extract the dominant
modes involved in unsteady flow field. Unfortunately,
these studies applied POD only for one-dimensional
decompositions. In order to take the full advantage of
POD method, we have applied it for decomposition of
full three-dimensional flow field.


2 Nomenclature
e internal energy (J/kg)
f
e
external acceleration (m/s
2
)
F
x
, F
y
, F
z
vectors of convective components of
flux
G
x
, G
y
, G
z
vectors of diffusive components of flux
h static enthalpy (J/kg)
I rothalpy (m
2
/s
2
)
p static pressure (Pa)
r radius (m)
R gas constant (J/(kgK))
S vector of source term
T static temperature (K)
t time (s)
u, v, w Cartesian components of velocity (m/s)
V absolute velocity (m/s)
W relative velocity (m/s)
thermal conductivity (W/(mK))
dynamic viscosity (kg/(ms))

t
eddy viscosity (kg/(ms))
azimuthal (circumferential) angle (rad)
static density (kg/m
3
)
shear stress tensor (Pa)
angular velocity (rad/s)

Proceedings of the 3rd WSEAS Int. Conference on FINITE DIFFERENCES - FINITE ELEMENTS - FINITE VOLUMES - BOUNDARY ELEMENTS
ISSN: 1790-2769 209 ISBN: 978-960-474-180-9
Subscript
R rotor
t turbulent
Superscript
eff effective (laminar + turbulent)


3 Governing Equations
For a three-dimensional rotating Cartesian coordinate
system, the unsteady Reynolds-averaged Navier-Stokes
equations using the Favre averaging (a mass-weighted
averaging) could be written in the conservative form as
[5, 6]
( ) ( ) ( ) y y
x x z z
F G
F G F G Q
S
t x y z


+ + + =

(1)
where
2 2 2
2 2
u
v
Q
w
W r
e

(
(
(
(
= (
(
(
| |
(
+
|
(
\
(2)
2
2
2
, ,
x y z
u v w
u p vu wu
F F F uv v p wv
uw vw w p
uI vI wI





( ( (
( ( (
+
( ( (
( ( ( = = = +
( ( (
+
( ( (
( ( (

(3)
where rothalpy I is defined by
( )
2
2
2 2
r W
I h

= + (4)
If we assume that the fluid is Newtonian, the
diffusive flux G may be written as
0
eff
xx
eff
xy
x
eff
xz
eff eff eff eff
xx xy xz
G
T
u v w k
x


(
(
(
(
=
(
(
(

+ + +
(

(5)
0
eff
xy
eff
yy
y
eff
yz
eff eff eff eff
xy yy yz
G
T
u v w k
y


(
(
(
(
=
(
(
(

+ + +
(


(6)
0
eff
xz
eff
yz
z
eff
zz
eff eff eff eff
xz yz zz
G
T
u v w
z


(
(
(
(
=
(
(
(

+ + +
(

(7)
According to the Boussinesq hypothesis, the shear
stresses
eff
may be written as
( )
2
2
3
eff
xx t
u v w
x y z

| |
= +
|

\
(8)
( )
2
2
3
eff
yy t
v u w
y x z

| |
= +
|

\
(9)
( )
2
2
3
eff
zz t
w v u
z y x

| |
= +
|

\
(10)
( )
eff eff
xy yx t
u v
y x

| |
= = + +
|

\
(11)
( )
eff eff
xz zx t
u w
z x

| |
= = + +
|

\
(12)
( )
eff eff
yz zy t
w v
y z

| |
= = + +
|

\
(13)
The Sutherlands formula could be used to determine
the dynamic viscosity as function of temperature,
while the eddy viscosity
t
is computed with a
turbulence model.
For gases, the external force f
e
due to the
gravitational acceleration is very small, therefore it can
be neglected. Moreover, we can assume that the thermal
conductivity is the single heat source. If the Cartesian
coordinate system is rotating about z axis with constant
angular velocity , source term S could be written as
( )
( )
2
2
0
2
2
0
0
x v
S
y u

(
(
+
(
(
=

(
(
(
(

(14)
The pressure is obtained from the equation of state,
p RT = (15)


4 Numerical Simulation
The numerical simulations of the three-dimensional
viscous flow were carried out on a centrifugal
compressor designed, manufactured and tested by
COMOTI, with commercial CFD code FLUENT that is
based on finite volume method where each unknown
takes an average value on each discretization cell. The
Proceedings of the 3rd WSEAS Int. Conference on FINITE DIFFERENCES - FINITE ELEMENTS - FINITE VOLUMES - BOUNDARY ELEMENTS
ISSN: 1790-2769 210 ISBN: 978-960-474-180-9
computational domain generated in Gambit was split
into eight blocks to facilitate the building of a fully
structured mesh as shown in Fig. 1. The mesh for
which, the results are given, has about 253 000
hexahedral cells for the impeller passage and 127 000
hexahedral cells for the vaned diffuser passage.
In order to decrease the computational time,
impressively, the time discretization is made with a
backward implicit first order scheme and multigrid
technique is used. To take into account the physical
properties of flow, the convective fluxes are discretized
with the Roe scheme, which is a Godunov-type scheme
[5, 6]. Because the turbulence is not a critical issue of
this study, we used the Spalart-Allmaras model, which
is a one-equation model [7].
At the inlet, a uniform stagnation pressure (96 310
Pa) and temperature (300 K) are imposed, turbulent
viscosity ratio
t
/ is 10 and the flow is normal to inlet.
At the outlet, a uniform static pressure (156 000 Pa) is
imposed. At the left and right sides of computational
domain, the rotational periodic boundary conditions are
imposed. All the walls have been assumed adiabatic.
The shaft speed of impeller is 14 915 rpm.

Fig. 1 Computational domain of centrifugal compressor


5 Adamczyk Decomposition
Non-uniformities and unsteadiness due to the rotor-
stator interaction introduce major complexity in the
analysis of the turbomachinery flow field. This problem
can be considerably simplified if we apply the method
of Adamczyk [4, 8] that proposed the decomposition of
an arbitrary field u associated to a turbomachinery in
four contributions through the successive application of
averaging operators:
( ) ( ) ( ) ( ) ( )
*
1 2 3 4
, , , , , , , , , , ,
R
R
u z r t u z r u z r u z r u z r t
t

= + + +



(16)
Starting from an arbitrary field u expressed in an
inertial reference frame attached to the stationary row,
the first averaging has as objective to extract the
axisymmetric field independent by time and azimuthal
coordinate. The second averaging is a time averaging in
the inertial reference frame and it extracts from the
remained field, the flow structures attached to the
stationary row while the third averaging also is a time
averaging but in the rotating reference frame and it
extracts from the remained field, the flow structures
attached to the rotating row. Therefore, the third
contribution is steady in the rotating reference frame.
Finally, after three averaging, the residual field (fourth
contribution) represents the unsteady part of initial field
u in the inertial and rotating reference frame associated
to stationary and rotating row, respectively. This
contribution characterizes purely unsteady phenomena
of turbomachinery flow. In order to understand better,
the unsteady rotor-stator interaction, the fourth
contribution was decomposed with POD technique as
shown in the next section.
As it follows, we will give some results for some
control points placed in a section at mid height of blade
of vaned diffuser, at the middle distance between the
blade and the right periodic as shown in Fig. 2. The
numbering of these control points is from upstream to
downstream.

Fig. 2 Placement of control points


Fig. 3 First component of Adamczyk decomposition for
static pressure normalized by inlet static pressure and
absolute velocity normalized by inlet absolute velocity

The first component of Adamczyk decomposition for
static pressure and absolute velocity, at considered
control points is shown in Fig. 3. One sees that the
compression process is smooth while the absolute
velocity has big variations especially in the first part of
Proceedings of the 3rd WSEAS Int. Conference on FINITE DIFFERENCES - FINITE ELEMENTS - FINITE VOLUMES - BOUNDARY ELEMENTS
ISSN: 1790-2769 211 ISBN: 978-960-474-180-9
vaned diffuser where the strong deceleration triggers a
huge jet-wake region accompanied by boundary layer
separation on suction side of vaned diffuser blade. These
phenomena generate huge nonuniformities in the
absolute velocity field as shown in Figs. 4 and 5, which
induce important total pressure losses. For this reason,
the compression process is very slow in the last part of
vaned diffuser. Furthermore, the rectangular trailing
edge of vaned diffuser blade generates additional
nonuniformities, which are shown in Fig. 5 and losses.
The homogenization process of flow begins after the
trailing edge of vaned diffuser blade and it is
accompanied by significant total pressure losses. For this
reason, the air compression is very weak downstream of
the trailing edge.
The Adamczyk decomposition clearly shows that this
classical vaned diffuser with circular arc blades
generates a huge jet-wake zone and important pressure
losses because the channel is extremely divergent in the
first part of vaned diffuser. In order to obtain better
compressor performance, it is necessary to renounce
single circular arc vaned diffuser.


Fig. 4 Second and third component of Adamczyk
decomposition for static pressure normalized
by inlet static pressure and absolute velocity
normalized by inlet absolute velocity


Fig. 5 Isolines of second and third component of
Adamczyk decomposition for absolute velocity in the
section from the middle height of vaned diffuser



6 Proper Orthogonal Decomposition
In the field of fluid mechanics, two approaches have
been used for the POD. Historically the method of
Continuous POD (or the classical method) of Lumley
[9] proceeded by the Snapshot POD of Sirovich [10].
More information regarding the application of the
proper orthogonal decomposition in the analysis of
turbulent flows together with a detailed bibliography is
given in [11]. In this paper, we used the Snapshot POD
because it is much more efficient from the numerical
point of view
The POD is a method that reconstructs a data set
from its projection onto an optimal base. Besides using
an optimal base for reconstructing the data, the POD
does not use any prior knowledge of the data set. It is
because of this that the basis is only data dependent and
this is reason that the POD is used also in analyzing the
natural patterns of the flow field.
For the reconstruction of the dynamic behavior of a
system the POD decomposes the data set in two parts: a
time dependent part, a
k
(t), that forms the orthonormal
amplitude coefficients and a space dependent part,

k
(x), that forms the orthonormal basis. The
reconstructed data set is:
( ) ( ) ( )
1
,
M
k k
k
u x t a t x
=
=

(17)
where M is the number of time instant observations in
the data set.
We denote the error of the reconstructed data set as:
( ) ( ) ( ) ( )
1
, ,
m
k k
k
x t u x t a t x
=
=

(18)
The base from which the data set is reconstructed is
said to be optimal in the sense that the average least
squares truncation error is minimized for any given
number ( m M ) of basis functions over all possible
sets of orthogonal functions:
( ) ,
m
= (19)
where the . is the ensemble average and
( ) .,.
is the
standard Euclidian inner product.
It was shown that the minimization condition for
error (x,t) translates into maximum condition for:
( )
2
,
,
u


= (20)
This maximization can be proven to take place if the
time independent base functions (x) are obtained from
the Fredholm integral equation:
( ) ( ) ( )
1
, ' ' '
M
ij j i
j
R x x x dx x
=
=

(21)
where R
ij
is the correlation kernel. In this way, we
transform this into an eigenvalue problem and
k
is the
Proceedings of the 3rd WSEAS Int. Conference on FINITE DIFFERENCES - FINITE ELEMENTS - FINITE VOLUMES - BOUNDARY ELEMENTS
ISSN: 1790-2769 212 ISBN: 978-960-474-180-9
eigenvalue corresponding of the eigenvector
k
.
Because we can consider the inner product as being the
equivalent of an energy, the value of
k
is linked to
the energy contained in mode
k
and the optimization
process involved can be summarized as: the data set is
projected onto a basis that maximizes the energy
content. While in the classical approach of Lumley [9],
the correlation matrix is constructed as a space
correlation matrix and solving the eigenvalue problem,
we obtain directly the eigenvectors as the spatial modes
and then use them in order to obtain the time-dependent
coefficients
( ) ( ) ( ) ( )
, ,
k k
a t u x t x = (22)
in the Snapshot POD of Sirovich [10], the correlation
matrix is a time correlation matrix:
( ) ( )
1
, , '
V
C u x t u x t dV
V
=

(23)
which is of the size of the square of the number of
snapshots. From the time correlation matrix, we get the
eigenvalues
k
and time dependent eigenvectors
k
(t).
The spatial eigenmodes that are time independent, are
computed according to the formula:
( ) ( ) ( )
1
,
k k
k t
x t u x t dt

=

(24)
where
k k
= (25)
The processed data are the variations of absolute
velocity magnitude and static pressure fields, which
represent the fourth term of Adamczyk decomposition
according to Eq. 16. These variations were obtained
from numerical simulations using the commercial CFD
code Fluent. For each period, we took 20 snapshots and
the time between adjacent snapshots is of t =
9.5781s; therefore, the Snapshot POD of Sirovich
yields 20 eigenmodes for each considered field.
The very high efficiency of the proper orthogonal
decomposition is clearly underlined by Table 1. The
sum of the first two modes represents 90.5% and 95.5%
of the total energy, respectively for the variations of
static pressure and absolute velocity magnitude fields
while the sum of modes 6 to the last mode represents
only 0.163% and 0.236% of the total energy,
respectively for the variations of static pressure and
absolute velocity magnitude fields. Therefore, both
variations of static pressure and absolute velocity
magnitude fields can be very accurately reconstructed
using only the first five modes. Furthermore, these
results confirm that the base from which the data set is
reconstructed is indeed optimal.



Table 1 Fraction of total energy
for the most energetic modes
Mode Fraction of total
energy for variation
of static pressure
Fraction of total
energy for variation
of absolute velocity
1 6.40E-01 5.49E-01
2 2.65E-01 4.06E-01
3 6.63E-02 2.13E-02
4 1.72E-02 1.58E-02
5 8.97E-03 5.73E-03
6 1.41E-03 7.57E-04
7 6.29E-04 4.50E-04
8 1.60E-04 8.52E-05
9 4.58E-05 7.30E-05
10 4.02E-05 5.25E-05


Fig. 6 The first four most energetic modes of
variation of absolute velocity magnitude field

The first two most energetic modes of variation of
absolute velocity magnitude field contain as much as
95.5% of the total energy. The first mode has 54.9% of
the total energy and it represents the interaction
between wakes due to the blunt trailing edge of impeller
blade and potential effects. According to theory of
characteristics, this interaction affects especially the
vaned diffuser region and its peak is located near the
middle distance between impeller and vaned diffuser as
shown in Fig. 6. The second mode contains 40.6% of
total energy and it represents the interaction between
wakes and potential effects in the vaned diffuser region
as well as the propagation of potential effects in the
impeller region. The third and fourth modes have 3.7%
of the total energy and they contain both physical and
numerical information. From the physical point of view,
they contain the information regarding the interaction
between and potential effects as well as the influence of
potential effects in the impeller region. From the
numerical point of view, they represent the numerical
errors that occur at the interface between rotating region
and stationary region and due the rotational periodicity
condition. Furthermore, one sees that the value of the
third mode is not close to zero at the outlet boundary of
Proceedings of the 3rd WSEAS Int. Conference on FINITE DIFFERENCES - FINITE ELEMENTS - FINITE VOLUMES - BOUNDARY ELEMENTS
ISSN: 1790-2769 213 ISBN: 978-960-474-180-9
computational domain because we imposed a uniform
static pressure on this frontier and this is not too correct
according to the theory of characteristics [5, 6].


7 Conclusions
Both Adamczyk and proper orthogonal decomposition
have been successfully applied to the decomposition of
fully three-dimensional static pressure and absolute
velocity magnitude fields obtained from numerical
simulations using the commercial CFD code Fluent.
The Adamczyk decomposition clearly shows that the
single circular arc vaned diffuser generates a huge jet-
wake region and important pressure losses because the
channel is highly divergent in the first part of vaned
diffuser. In order to obtain better compressor
performance, it is necessary to renounce circular arc
vaned diffuser.
Both variations of static pressure and absolute
velocity magnitude fields can be very accurately
reconstructed using only the first five modes; therefore,
the proper orthogonal decomposition method is a very
efficient method for the data storage of unsteady flows.
Moreover, POD technique is able to capture the
relevant features of the unsteady rotor-stator interaction,
especially, the potential effects and the interaction
between wakes due to the blunt trailing edge of impeller
blade and potential effects. Furthermore, the POD
method clearly show the numerical errors such as those
errors that occur at the interface between rotating region
and stationary region because the information exchange
does not use the characteristic variables, the reflection
of numerical waves at rotational periodic and outlet
boundaries as well as their magnitude. In order to obtain
more accurate results [3, 12], we should impose the
phase-lagged condition, which is not yet available in
Fluent, on the left and right sides of computational
sides, instead of the rotational periodicity condition.


8 Acknowledgements
The present work has been jointly supported by the
CNCSIS UEFISCSU, project number PNII IDEI
1030/2007 (contract number 109/2007) and Contract
PNII CNMP no. 81-027.


References:
[1] C. Dano, valuation de modles de turbulence pour
la simulation dcoulements tridimensionnels
instationnaires en turbomachines, Thse de doctorat,
cole Centrale de Lyon, 2003
[2] P. E. Smirnov, T. Hansen and F. R Menter,
Numerical Simulation of Turbulent Flows in
Centrifugal Compressor Stages with Different Radial
Gaps, GT2007-27376, ASME Turbo Expo 2007:
Power for Land, Sea and Air, Montreal, Canada,
May 2007
[3] N. Rochuon, I. Trbinjac and G. Billonnet, An
Extraction of the Dominant Rotor-Stator Interaction
Modes by the Use of Proper Orthogonal
Decomposition (POD), Journal of Thermal Science,
Vol. 15, No. 2, 2006, pp. 109-114
[4] N. Rochuon, Analyse de l coulement
tridimensionnel et instationnaire dans un
compresseur centrifuge fort taux de pression,
Thse de doctorat, cole Centrale de Lyon, 2007
[5] S. Dnil and C. Berbente, Numerical Methods in
Fluid Dynamics (only in Romanian), Publishing
House of Romanian Academy, 2003
[6] C. Hirsch, Numerical Computation of Internal and
External Flow, Volume 2: Computational Methods
for Inviscid and Viscous Flows, John Wiley and
Sons, New York, 1990
[7] P. R. Spalart and S. R. Allmaras, A One-Equation
Turbulence Model for Aerodynamic Flows, AIAA
Paper 92-0439, 1992
[8] J. J. Adamczyk, Model Equation for Simulating
Flows in Multistage Turbomachinery, NASA
Technical Memorandum 86869, 1984
[9] J. Lumley, Stochastic Tools in Turbulence,
Academic Press, 1970
[10] L. Sirovich, Turbulence and the dynamics of
coherent structures, parts I-III, Q. Appl. Math., Vol.
XLV, No. 3, 1987
[11] G. Berkooz, P. Holmes and J. Lumley, The Proper
Orthogonal Decomposition in the Analysis of
Turbulent Flows, Annual Review of Fluid
Mechanics, Vol. 25, 1993, pp. 539-575
[12] I. Trbinjac, P. Kulisa, N. Bulot and N. Rochuon,
Effect of Unsteadiness on the Performance of a
Transonic Centrifugal Compressor Stage, GT2008-
50260, Proceedings of ASME Turbo Expo 2008:
Power for Land, Sea and Air, Berlin, Germany, 2008
Proceedings of the 3rd WSEAS Int. Conference on FINITE DIFFERENCES - FINITE ELEMENTS - FINITE VOLUMES - BOUNDARY ELEMENTS
ISSN: 1790-2769 214 ISBN: 978-960-474-180-9

You might also like