Professional Documents
Culture Documents
Turbo 136 08 081010
Turbo 136 08 081010
Ernesto Casartelli
Lucerne University of Applied Sciences and Arts,
Analysis
Engineering and Architecture,
This paper describes the development and validation steps of a characteristics-based
Technikumstrasse 21,
Horw 6048, Switzerland
explicit along with a novel fully implicit mixing plane implementation for turbomachi-
nery applications. The framework is an unstructured 3D RANS in-house modified solver,
based on open-source libraries. Particular attention was paid to mass-conservation,
Thomas Mokulys accurate variables interpolation, and algorithm stability in order to improve robustness
MAN Diesel and Turbo Schweiz AG,
and convergence. By introducing a specific interface, allowing the use of algebraic mul-
Hardstrasse 319,
tigrid solvers together with multiprocessor computation, a speed up of the numerical so-
Postfach 2602
lution procedure was achieved. The validation of both mixing plane algorithms is
Zurich 8021, Switzerland
carried out on an industrial radial compressor and a cold air 1.5 stages axial turbine.
[DOI: 10.1115/1.4026323]
Sebastiano Mauri
MAN Diesel and Turbo Schweiz AG,
Hardstrasse 319,
Postfach 2602
Zurich 8021, Switzerland
group at the University of Stuttgart [12] extended this theory to a 2 Concept of the Mixing Plane
fully 3D NRBC.
As already mentioned, mixing planes average the flow quanti-
Today, implementations of mixing planes can be found in most
ties in the circumferential direction before passing them to their
commercial CFD software packages and are widely used in the
adjacent side, as illustrated in Fig. 1. The justification of this pro-
industry. In 2008, Holmes [13] made an attempt to create a list
cess is based on the observation that in real turbomachinery appli-
with all of the criteria a mixing plane should ideally fulfill. A sum-
cations the flow is also largely mixed in the gap between the rotor
mary of the advantages and disadvantages of turbomachinery
and the stator. This mixed-out-state makes the simulation inde-
CFD, including the influences on the final solution of mixing
pendent of the relative position of the rotor and stator and results
planes, was then given by Denton [14] in 2010.
in a circumferentially averaged steady state solution reflecting the
The aim of this work is to introduce a novel implicit approach
mean solution of all rotor-stator positions. Despite the similarity
of the mixing plane, integrated in an in-house modified pressure
R between the mixing process and the mixed-out-state, the differ-
based all-Mach solver [15] based on the OpenFOAMV [16,17]
ence in the way the mixing occurs can cause problems. Even if
framework. A standard explicit approach based on the previously
the flow is largely mixed out in reality, the process is more grad-
published concepts was implemented for comparison and will
ual than the instantaneous mixing at the mixing plane. This instan-
briefly be explained. The underlying concept of these earlier
taneous mixing will result in a higher entropy rise and can be seen
interfaces is the mathematical separation of a physically con-
as a loss of total pressure over the interface. Studies by Fritsch
nected domain. Together with the accompanied iterative solution
and Giles [20] showed an increase in the total loss of about 10%
procedure, this leads to considerable stability problems and
when instantaneous mixing was applied.
demands special treatment of flow variables at the interface. The
superior behavior of the novel implicit approach will be shown in
this paper. 3 Convections and General Requirements
A basic formulation of this kind of boundary condition already To make the mixing plane easy to use and applicable to a vari-
R
exists in an extended version of OpenFOAMV and was published ety of different turbomachinery applications with different flow
by Jasak and Beaudoin [18]. The validation of this mixing plane regimes and numerical schemes, several requirements must be sat-
was only shown for an incompressible Francis turbine [19] and isfied. The mixing plane has to be:
further calculations with this basic formulation showed mass con-
• mesh independent
servation problems. This basic formulation was extended to han-
• applicable for compressible and incompressible flow
dle compressible and incompressible flows, thus increasing the
• applicable to multiprocessor and multigrid calculations
applicability domain. Acceleration of the simulation was achieved
by introducing a framework allowing multigrid techniques and Figure 2 shows the convention of labeling used in the explana-
multiprocessor calculations. tions thereafter. In order to ensure mesh independency, the values
Mixing planes are generally used to represent an average solu- are interpolated at the boundary from the unstructured grid at the
tion of a full rotation of the rotor, which demands a uniform total mixing plane to an intermediate virtual interface, covering the
pressure distribution after this interface. The test case validation same bounding box as the original boundary but built up from
presented in the paper shows that the tight coupling between the radial slices, as shown in Fig. 1.
variables due to the implicit formulation increases the quality of
downstream total pressure uniformity.
A thorough validation was carried out by a comparison with the 4 Explicit Algorithm
measurement data and simulation of several turbomachinery The implementation of an explicitly coupled boundary condi-
applications. The presented test cases are a centrifugal compressor tion generates two computational domains which have to be sepa-
with a vaned diffusor and an axial turbine. The centrifugal com- rately solved. The decomposition of a physically connected
pressor was used to test the improvements of the implicit mixing domain results in the generation of an additional outlet and inlet.
plane with respect to the explicit formulation, especially if oper- As ordinary inlet and outlet boundaries, unphysical flow structures
ated in off-design conditions. The axial 1.5 stage test case proves due to reflections can occur. A brief overview of the NRBC was
the stability of the implemented mixing plane in a multi-row already given in the Introduction and is not discussed further in
calculation. this paper, nor is it implemented in the presented mixing plane.
is based on the Open MPI library [24] and widely used within
R
OpenFOAMV. Special attention was paid to the addressing during
the agglomeration. The size of the agglomerated boundary is no
longer the size of the original boundary-mesh and the initially cal-
culated addressing can no longer be used. An algorithm which
constantly calculates the new addressing for each coarsening level
and processor was, therefore, implemented. Figure 5 shows the
faces and the labels of their addressing for the finest and a coars-
ened level, together with the subdomains where they will be allo-
cated for further calculations.
lp ¼ gttc w (7)
The plotted values are reference values with respect to the inlet
volume-flow coefficient at the maximum gttc of all simulations.
h0 0.00156 0.02634
p0 0.00381 0.22129
s 0.00018 0.04542
m_ 0.00671 0.13542
7.3 Comparison and Validation. The mixing, occurring at Fig. 11 Pressure rise in the rotor and stator
the interface, results in a rise in entropy, which is accompanied by
a loss in total pressure. Therefore, the first comparison deals with
the evolution of the entropy and total pressure over the mixing
plane. Figure 8 shows the averaged rise in entropy through the total pressure from the inlet to the outlet between the two
control surfaces perpendicular to CP2. The continuous line repre- algorithms and the development along streamwise direction is
sents the solution of the implicit mixing plane, while the dashed nearly identical. The comparison with the commercial code results
curve is the solution achieved with the explicit formulation. To shows a comparable total pressure rise, especially in the rotor, de-
arrive at the mixed-out-state, the flow quantities have to adapt to spite the higher entropy rise in the stator. The results shown in
this distribution before the interface. This is clearly visible in Figs. 8 and 9 deal with the overall solution of the simulation. Fur-
Fig. 8 since the rise in entropy is located upstream of the mixing ther validation of the mixing plane was then carried out by a com-
plane, representing the additional losses accompanied with the parison of the up- and downstream flow quantities close to the
mixing process. The implicit formulation leads to results similar interface. Representative quantities for the mass, momentum, and
to the explicit formulation but shows fewer oscillations. The com- energy are listed in Table 2 and the conservative behavior is
parison with a commercial code shows good agreement for the expressed in terms of the percentage of the difference. The calcu-
rotor section but has higher entropy rise over the stator blade. The lated values clearly show the superiority of the implicit formula-
reason for this behavior will be explained later. Since a rise in en- tion for all of the listed quantities. The stage polytropic total-to-
tropy is a loss, which is accompanied by a total-pressure drop, total efficiency, work coefficient, and polytropic head coefficient
Fig. 9 shows the streamwise evolution of the total pressure for the against the inlet volume-flow coefficient are shown in Fig. 10 and
two algorithms. It can be seen that there is no difference in the compared to the measurement data. The work coefficient shows
Fig. 13 Contours of entropy at CP1: (a) CC1 frozen rotor; (b) CC1 mixing plane; (c) implicit mix-
ing plane; (d) explicit mixing plane
seen from the dashed lines, in the rotor section there is very good position of the rotor and stator, resulting in uniform pitchwise
agreement between the implemented implicit algorithm and the enthalpy and entropy for a fixed spanwise height on the down-
solution achieved with the commercial code. In the stator section, stream side of the domain, as stated by Denton [14]. Figures 13
however, the solution achieved by the commercial software has and 14 show the contours of the entropy and enthalpy distribution
less of a pressure drop for a higher volume flow, leading to better along CP1. All of the figures have the same range and fifty uni-
efficiency. A further investigation of the stator section has shown formly distributed isolines in order to assess the quality of the for-
that the solution obtained with the commercial software has sig- mulation. For a better comparison, the position dependent solution
nificantly less separation. Figure 12 illustrates the backflow region achieved with a frozen-rotor approximation is plotted in
at mid span. The dark gray parts are regions of streamwise nega- Figs. 13(a) and 14(a). Since a dense distribution of isolines indi-
tive flow. While the incidence and mass flow are approximately cates high gradients, the nonuniformity in entropy and enthalpy
the same, the part of the blade which has recirculation is on the downstream side of the interface is clearly visible in the
considerably larger for the simulation with the implicit mixing case of the frozen-rotor simulation. Compared to this result, all
plane. The secondary flow structures accompanied by the separa- other implementations of mixing planes provide much more uni-
tion on the stator blade are an additional loss, leading to a higher form fields. With respect to the uniformity criteria, the implicit
drop in total pressure and, thus, lower efficiency. By comparing algorithm performs the best. From an engineering point of view,
the different simulation results with the measurement in Fig. 10, it the rotor and the stator can be considered as two separate domains
seems that the commercial software underestimates the separation with their inlet and outlet conditions coupled to the adjacent do-
in the stator. Compared to frozen-rotor simulations, a mixing main. Applying the total conditions at the inlet, the downstream
plane should yield a solution which is independent of the relative side of the mixing plane interface would also require imposed
total conditions. With respect to the idea behind the mixing plane, 7.4 Axial Turbine Test Case. The axial turbine presented in
this would demand pitchwise uniform total pressure. Figure 15 this paper is a 1.5 stage cold air turbine built at the Institute of Jet
shows the isolines of the total pressure on CP1. It is important to Propulsion and Turbomachinery at Aachen Technical University
point out that despite the small reflections in entropy and enthalpy (IST RWTH Aachen, Germany). This turbine is a well-known test
due to the strong coupling of the boundary condition, a uniform case investigated by many authors. See, for example, Refs.
total pressure distribution has been achieved with the novel [27–29] for further information. At the inlet, total conditions are
formulation. imposed while at the outlet a prescribed static pressure was used.
Validation of the code is carried out by a comparison of the abso- the spanwise positions of the extrema. The solutions from the
lute flow angles and the total pressure distribution. Figure 16 commercial code and the in-house code with the Spalart–Allmaras
shows the axial turbine, including the cutting planes used for the turbulence model are not able to reproduce the correct spanwise
comparison. At cutting planes one to three, the absolute flow position of the extrema of the turning angles. For the first blade,
angles were computed and compared with the measurement data. however, it has to be noted that the position of the extrema is
strongly dependent on the inlet boundary condition, which is
always accompanied with uncertainties about the real flow
7.5 Comparison and Validation. Figures 17, 18, and 19 distribution. The prediction of the absolute flow angle behind the
show the flow angles computed with the implemented implicit rotor shows even higher discrepancies between the different simu-
mixing plane and those obtained with a commercial code (CC1) at lations; see Fig. 18. Again, the in-house solver, together with the
the three cutting planes. Several different turbulence models have k-x SST model, shows superior behavior. The small wave close
been used together with the in-house solver wherein the solutions to the shroud is, however, smeared out. A reason for this could be
of a Spalart–Allmaras [30] and a modified k-x shear stress trans- a too low mesh resolution at the end-walls. Calculation of the flow
port (SST) [31] model with automatic wall treatment are angles using the Spalart–Allmaras model showed discrepancies in
presented. The turbulence model of the commercial code is, again, predicting secondary flow effects close to the hub. The absolute
a k-x SST model with automatic wall treatment. Figure 17 flow angles after the stator in the outlet domain are shown in
presents the solutions after the first stator row. The flow angles Fig. 19. As can be seen, the flow angles are well predicted by the
around midspan are well-predicted by all simulations. Differences k-x model of the in-house code, while the k-x model of the com-
appear in the prediction at the upper and lower end wall. Com- mercial software was unable to correctly predict the secondary
pared to measurement results the k-x SST model with automatic flow effects. The one-equation Spalart–Allmaras model gives rea-
wall treatment of the in-house solver performs best in predicting sonable results in most parts of the computed section but shows
discrepancies close to the hub. A global validation is carried out
by comparing the total pressure distribution. Since the most prom-
ising turbulence model is the k-x SST model with automatic wall
treatment, these results will be used for the comparison and are
presented in Figs. 20 and 21.
The implemented mixing plane shows the expected uniform
total pressure distribution after the interface. This uniformity
respects the underlying concept of a mixed-out-state, reproducing
a solution for a complete revolution.
8 Conclusions
Two mixing plane approaches have been implemented. One
uses a characteristics-based explicit approach and one uses a novel
implicit formulation. Validation of both implemented mixing
planes showed good agreement with the measurement data. The
implicit mixing plane proved the expected superior behavior com-
pared to the explicit characteristics-based approach. The coupled
treatment of the interface greatly improved stability during start-
up, in the case of backflow or close to choking conditions,
extending the applicability of this approach for turbomachinery
applications. The formerly stated possible inaccuracy due to an
inaccurate averaging technique could be solved with an explicit
Fig. 18 Absolute flow angle at CP2 correction. This allowed the use of a better implementation of the
Fig. 21 Contours of total pressure at CP4 for the second interface: (a) mixing plane 2, CC1; (b)
mixing plane 2, implicit
interpolation and resulted in a massive reduction of computational affects all mixing planes. In fact, the so-called nonreflective
costs. Numerical experiments also showed that the explicit correc- mixing plane used in the past to solve this problem is damping the
tion does not influence the convergence behavior. As already effect of the mixing process without so much efficacy, in terms of
mentioned in the paper, the NRBC treatment is not implemented downstream profile uniformity. This paper points out that the
in the code. It is, therefore, evident to point out that the small reflection is not the direct cause of the nonuniformity in the down-
reflections, due to the use of a strong imposition of the mixing av- stream profile. Instead, the tight coupling between the variables
erage, are not directly responsible for the phenomena of the non- has a major effect in achieving the target of downstream uniform-
uniform profile in the downstream domain. This is a problem that ity. Despite the small reflections in entropy and enthalpy due to
I_n;i ¼ qbi ubn;i ubn;i þ pbi (A1) B.1 Explicit Algorithm. As mentioned earlier, an explicit
X coupling leads to additional inlets and outlets, as visualized in
ðqi;j ui;j ni;j ui;j ni;j þ pi;j ÞSi;j Fig. 23. The subdomains are represented by two systems of linear-
j ized equations and are given in Eqs. (B4) and (B5):
¼ (A2)
Si "CC# ! !
ð U1 b1
1 22
¼ ðqðu nÞðu nÞ þ pÞdS (A3) ¼ (B4)
S CC
22 U2 b2
|fflffl{zfflffl} |fflfflffl{zfflfflffl} |fflfflffl{zfflfflffl}
A1 x1 b1
To create a more general solution, Dzung proposed the introduction
of dimension-free quantities. Therefore a Laval speed is defined, "CC# ! !
dependent only on the stagnation enthalpy and the mass flux 22 U3 b3
CC
¼ (B5)
U4 b4
j 1 E_ 22
|fflffl{zfflffl} |fflfflffl{zfflfflffl} |fflfflffl{zfflfflffl}
a2 ¼ 2 (A4) A2 x2 b2
j þ 1 m_
The normal velocity component can be computed with Eq. (A5) The values from the adjacent side of the interface are used as inlet
and outlet conditions, according to the theory of characteristics.
ubn ¼ la (A5) The influence of the adjacent domain fully acts as an explicit con-
tribution to the source. Therefore, the coupling is only carried out
where l is the solution of the quadratic Eq. (A6), which was cre- before the outer iterations.
ated with a complete set of conservation equations similar to B.2 Implicit Algorithm. An implicit treatment of a coupled
Eq. (A3) interface can be understood as using two separate matrix of coeffi-
cients to build up a system of linearized equations for the com-
l2 b l þ c ¼ 0 (A6) plete solution domain, as shown in Eq. (B7) (see Fig. 24).
It can be seen that the system shown in Eq. (B8) can be built up
from the equation systems of the explicit approach. However, Nomenclature
instead of considering the influence of the coupling in the source A¼ matrix of coefficients
term, these fluxes are implicitly discretized. A system of linear- a ¼ Dzung’s Laval speed
ized equations for the interface fluxes, with respect to the preced- b¼ right hand side
ing example is shown in Eq. (B9): CP ¼ cutting plane
2 3 0 1 0 1 d¼ diameter, m
0 0 0 0 U1 b1 DS ¼ downstream side
6 C C 7 B C B C E_ ¼ total energy flux, kg s3
60 0 7 B U2 C B b2 C
6 2 2 7B C ¼ B C (B9) GGI ¼ general grid interface
60 07 B C B C
4 C
2
C
2 5 @ U3 A @ b3 A h¼ enthalpy, kJ kg1
I_ ¼ impulse, kg m1 s2
0 0 0 0 U4 b4
|fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflffl{zfflfflffl
ffl} |fflfflffl{zfflfflffl} LE ¼ leading edge
AC x b m¼ mass, kg
MRF ¼ multi-reference frame
Adding AI þ AC will lead to a system of linearized equations for NRBC ¼ nonreflecting boundary condition
the complete domain p¼ pressure, Pa