Professional Documents
Culture Documents
ABSTRACT
1 INTRODUCTION
_______________________________________
© The author(s) and/or their employer(s), 2014
65
However, all above countermeasures need extra cost and complexity of the
turbocharging system.
This paper describes the concepts and the procedure of the design optimization
system based on an ANN and a GA. Experimental and numerical investigations of
the newly designed impellers are presented and discussed.
2 OPTIMIZATION STRATEGY
Start
Geometry
2.1 Optimization procedure
Figure 1 shows schematically the design Requirement
procedure. The optimisation is driven by a GA
GA in which the performance of each Geometry Blade
Generation
geometry is analysed by means of an ANN
ANN
trained on the information contained in a
3D NS
database. Once the GA has found an Analysis
Performance Prediction
optimum it is verified by a Navier Stokes
solver(11) and is added to the database,
Performance Data Learning
resulting in a more accurate ANN. This Base
procedure is repeated for a given number
of iterations. A more complete description Stop
of the optimisation method can be found Figure 1: Optimisation procedure
in published papers (12),(13).
66
2.3 Geometry definition
The impeller is backswept with splitter blades. The blade number is allowed to be
changed in between minimum 4+4 and maximum 6+6. The minimum and
maximum axial length are fixed. The maximum lean angle at the leading edge and
trailing edge, which are not directly controlled by the used parameterisation, are
given as design constraints.
The geometry is defined by the meridional contour and the camberline blade angles
of the full and splitter blades at hub and shroud. The meridional contour definition is
schematically shown in Figure 2, where subscript 0 denotes the leading edge of the
blade and 3 denotes the trailing edge. The inlet contour at the hub and shroud is
defined by a second-order Bézier curve. The hub inlet and leading edge diameters
are fixed. The hub axial position and the shroud leading edge diameter are design
variables. The leading edge axial position of the full blade at the shroud is set equal
to the leading edge at hub. The points denoted by in (Xin,Rin), A and 0 (X0,R0) are
Bézier control points and constitute the Bézier polygon. The leading edge position of
the splitter blade is defined by the design variables Ule at hub and shroud. Ule is the
percentage of meridional length where the leading edge starts.
The blade meridional contour at hub and shroud are defined by third-order Bézier
curves with 4 control points, denoted by 0, 1, 2 and 3. At hub both axial and radial
coordinates of points 1 (X1,R1) and 2 (X2,R2) are design variables. The axial
position of point 1 and both coordinates of point 2 are design variables. The trailing
edge diameter is set equal at hub and shroud. The trailing edge diameter is not a
design variable but it is adjusted to compensate the variation in the blade trailing
edge metal angle, which is a design variable. The diffuser exit diameter is fixed
equal at the hub and at the shroud. The diffuser exit width is also fixed. The diffuser
meridional contour is defined by second-order Bézier curves and the intermediate
Bézier control point B is computed to have a parallel diffuser downstream of that
point and smooth trailing edge curve.
The camberline blade angle definition is achieved using third-order Bézier curves in
the form of Bernstein polynomials as shown in Figure 3.
(u ) 0 (1 u )3 1u (1 u ) 2 2u 2 (1 u ) 3u 3 (1)
Here u is a parameter between 0 at the leading edge and 1 at the trailing edge. The
parameters β0 and β3 are the metal angles at the leading and trailing edges, β1 and
β2 are intermediate parameters in eq. 1 and have no physical meaning. These
67
parameters are the 8 design variables for the full blade camberline at hub and
shroud. The splitter blades are restricted to have the same trailing edge metal angle
as the full blades, hence this results in 6 additional design variables for the splitter
blades. Same pitch at the hub and shroud is imposed at the trailing edge. The
description of the blade is completed with the pre-specified thickness distribution. In
total 27 design variables define the geometry.
The vector of the 27 shape parameters, normalized by their limit values (the ANN
requires input and output values between 0 and 1), constitutes the ANN input. The
performance vector (as well the ANN output) from which the objective is computed,
consists of the two mass flows m1 and m2 (see ∆mass penalty below), the total-to-
total efficiency, distortion and skew of the diffuser exit radial velocity profile and full
blade and splitter blade Mach numbers. The Mach number is parametrized at 20
points each, along respectively the pressure and suction sides of the blades and
splitters. The total performance vector has 5+8*20=165 elements. The total
penalty is the sum of the following items computed using the elements of the
performance vector:
Mass penalty
The difference between the required mass flow and the actual mass flow is
penalized.
There is a tolerance value of mref/300.0 with mref=mref for which there is no penalty.
Here wm is the weight of the mass penalty with respecto to the other penalties.
Initially it was 300 and adjusted during the iterations. The mass flow m1 and m2 are
the values for splitter blade pressure side and suction side respectively.
Δmass penalty
The difference between the mass flows of each flow channel (divided by the splitter
blade) is penalized. Here wm2 is the weight of the penalty.
2
m m2 (3)
Pm 2 wm 2 1
m1 m 2
Efficiency penalty
The difference between the required efficiency (always 100%) and the actual
efficiency is penalized. The efficiency used in the penalty function is the total-to-
total adiabatic efficiency. Here weff is the weight of the penalty.
Distortion-skew penalty
The diffuser exit flow is desired to be as uniform as possible. Any distortion or skew
of the radial velocity profile is penalized.
68
Pdis skew wdis | d 1 | wskew | s | (5)
2 Vmid (6)
d
VR VL
2 (V R V L ) (7)
s
VR VL
The radial velocity profile at the diffuser exit is considered. The velocity at the
diffuser exit mid-channel is Vmid, near the hub is VR and near the shroud is VL.
Loading penalty
The loading difference between the splitter and full blade is penalized.
2 ( AB AS ) (8)
PL wL
AB AS
AB and AS are the full blade and splitter blade loading. The loading is calculated as
the area between the suction and pressure side isentropic Mach numbers along the
normalized blade length.
2 ( M P M min ) (10)
Pmach peak wmach peak
M P M min
The peak Mach number Mp is searched near the leading edge. The minimum Mach
number Mmin is also confined to that interval.
2( M i M i 1 )
Pmach acc wmach acc max( ,0) (11)
xi xi 1
The summation in eq. 11 is the first derivative of the Mach number and it is
performed after the location of the peak Mach number up to approximately the
middle of the blade, avoiding high Mach numbers near the trailing edge.
69
( M i M i 1 ) ( M i 1 M i )
Pmach dec wmach dec max ,0 (12)
xi 1 xi xi xi 1
The summation in eq. 12 is performed downstream the location of the peak Mach
number and it is confined to approximately half the blade.
3. OPTIMIZATION RESULTS
In this study two optimized impellers were designed by the same optimization
procedure. The only difference between both sduties is the blade number limitation.
In the first optimization (OPT1) the minimum blade number is 5+5. In the second
optimization (OPT2) the minimum blade number is 4+4. The maximum blade
number is 6+6 in both cases. Table 1 shows the specifications of the optimum
design found in both optimization studies compared to the baseline impeller. Figure
4 show the optimum design impellers and the baseline impeller. Figure 5 show the
meridional contour of the optimum design impellers with the baseline impeller.
Baseline
OPT1
OPT2
As described later, OPT1 has 0.5-1.5% higher efficiency with slightly wider
operating range compared to the baseline impeller. On the other hand, OPT2 has a
twice as wide operating range with very suitable characteristic operating curve with
a sufficient negative pressure gradient, at a 1% lower efficiency compared to the
baseline impeller.
Figure 7 shows the comparison of the shroud Mach number distribution of iteration
17 obtained by ANN and NS calculation. A general good agreement is observed.
Moreover, the Mach number distribution is very smooth. A good split of blade
loading between the full blade and the splitter blade is also obtained. The best
impeller is not chosen only for its high efficiency. The Mach number distribution, off-
design performance and mechanical considerations play as well an important role in
the selection process. Consequently the authors selected the iteration 17 as the
best impeller(OPT1) because it has a good Mach number distribution and equivalent
70
blade loading between the full blade and the splitter blade with a preferable blade
geometry for manufacturing, almost radial leading edge of the splitter blade. The
blade number of OPT1 is 5+5. OPT1 has larger inlet blade height and smaller exit
blade width compared to the baseline impeller.
150 1.5
NS Efficiency penalty NS Full blade
NS Mass penalty NS Splitter blade
NS Loading penalty ANN Full blade
ANN Splitter blade
100 1.0
Mach number
Penalty
50 0.5
0 0.0
0 10 20 30 40 50 60 70 0.0 0.2 0.4 0.6 0.8 1.0
Iterations Nodimensional meridional length
Figure 6: NS and ANN FigureFigure
7: Shroud Mach number
7: Shroud Mach
penalty breakdown distribution
numberof OPT1
distribution
f OPT1
3.2 Optimised impeller 2 (OPT2)
In Figure 8 the convergence history of total to total efficiency is shown. Similar to
the selection strategy of OPT1, iteration 46 was selected as best impeller (OPT2) in
this optimization run. Iteration 46 is one of the best designs and has the third
highest total to total efficiency shown in Figure 8. Figure 9 shows the shroud Mach
number distribution of iteration 46 obtained by ANN and NS calculation. Similar to
OPT1, iteration 46 was selected as best because it has not only high efficiency but
also a good Mach number distribution and an equal loading split between full blades
and splitter blades as shown in Figure 9. The number of blades is 4+4. The inlet
blade height is slightly smaller. Axial length is smaller than that of baseline impeller.
A remarkable feature of OPT2 is the forward inclined leading edge of the splitter
blade.
1.5
NS Full blade
1%
NS Splitter blade
ANN Full blade
Total to total efficiency
0.5
0.0
0 10 20 30 40 50 60 70 0.0 0.2 0.4 0.6 0.8 1.0
Iterations Nondimensional meridional length
Figure 8: Efficiency history Figure
Figure 9: Shroud
9: Shroud MachMach
number
number
distribution ofdistribution
OPT2
71
4. PERFRORMANCE TEST RESULTS AND DISCUSSIONS
The optimal shapes resulting from both optimization processes have been further
analyzed both experimentally and numerically. Performance tests have been
conducted using the compressor inside an automotive turbochager. Both optimal
impellers are each powered by a turbine driven by heated air or exhaust gases,
representing engine like conditions. Each compressor volute has the same geometry,
as well as same exit diameter of the vaneless diffuser. Each diffuser has a different
width but the ratio of diffuser width and impeller exit width is fixed as same value.
The compressor characteristics have been measured and compared with the
baseline impeller.
Figure 11 shows the comparison of the peak efficiency for each rotational speed
shown in Figure 10. Compared to the baseline OPT1 has a 0.5% higher peak
efficiency around the total to total pressure ratio of 1.8 and more than 1.0% higher
efficiency at pressure ratios above 2.2. On the other hand, OPT2 has about 1.0%
lower efficiency compared to the baseline impeller.
In Figure 12 the operating range of the optimum designed impellers and the
baseline impeller are compared. The operating range is defined as the flow range
between maximum flow rate and surge limit at each rotational speed and
normalized with the surge limit. In this study the maximum flow late is defined as
the flow rate at which the compressor efficiency drops below 65%. The horizontal
axis of Figure 12 means the pressure ratio at surge limit at each rotational speed.
The operating range of OPT1 is slightly smaller up to pressure ratio 1.9 compared to
baseline impeller. However, the baseline impeller decreases its operating range
rapidly at pressure ratios above 1.9. OPT1 does not have this rapid decrease of its
operating range and remains a wide operating range up to a pressure ratio 2.5.
72
OPT2 has almost equivalent operating range of the baseline impeller at the pressure
ratio below 1.9 but a very wide operating range at pressure ratios beyond 1.9.
OPT2’s operating range has more than doubled at the pressure ratio above 2.2
compared to the baseline impeller. OPT2 has a very gradual decrease of its
operating range with pressure ratio increase. As a result, OPT2 has a suitable
compressor characteristic for turbocharger applications which require a wide and
stable operating range.
Even though OPT1 and OPT2 have been selected as best impellers which show
higher efficiency compared to the baseline impeller during the optimization
procedure, OPT1 and OPT2 have completely different characteristics as
demonstrated by the performance test. OPT1 has achieved higher efficiency
compare to the baseline impeller as expected. On the other hand OPT2 has lower
efficiency but has a significant increase of its operating range. The difference of the
compressor characteristics of OPT1 and OPT2 shows the diversity and complexity of
the aerodynamic design of a centrifugal compressor impeller. There is a possibility
to get the variety of geometry and characteristic of optimum design such as OPT1
and OPT2 with same design procedure. The discrepancy of OPT2’s efficiency
between the estimation and the test result remains to be solved.
In Figure 13 the limiting stream lines on the blade surface, stream lines (only in
right figures of Figure 13) and the entropy distribution and are compared. Also
shown in Figure 13 are the identified vortex cores which are coloured with
normalized helicity. Normalized helicity is defined as cosine of the angle made
between the vortex vector and the velocity vector, and the domain where its
absolute value is 1 indicates a strong rolling-up of a streamwise vortex. Regarding
73
the entropy distribution the regions which have a relatively high entropy are
mapped in streamwise sections in Figure 13.
Secondary flow
LE tip leakage vortex
Secondary flow
Accumulation of low momentum fluids
by secondary flow
(a) Baseline
Hn [-] Tip leakage flow Tip leakage flow
(b) OPT1
Hn [-] Tip leakage flow Tip leakage flow
Secondary flow
LE tip leakage vortex
Accumulation of low momentum fluids
by secondary flow
(c) OPT2
From Figure 13(a), it is clear that the baseline impeller has a strong secondary flow
rolling up from hub to shroud on the suction surface at the inducer. Because of this
strong secondary flow, low momentum fluids start to accumulate on the suction
surface of the full blade near the end of the inducer. This low momentum fluids
combined with the tip leakage flow more downstream the full blade accumulate and
compose the high entropy region near the tip corner of the splitter blade pressure
74
surface at the impeller exit. As shown in Figure 13(a) the LE(leading edge) tip
leakage vortex is generated at just downstream the full blade leading edge. This LE
tip leakage vortex moves into the passage of the splitter blade suction surface and
composes a low flow region near the shroud. This LE tip leakage vortex travels
further downstream while accumulating low momentum fluid and composes the
high loss region near the shroud of the splitter blade suction surface at the impeller
exit. At the tip corner of the full blade pressure surface at the impeller exit mostly
tip leakage flow downstream the inducer accumulates and highest loss region is
composed. As demonstrated above the strong secondary flow motion and tip
leakage flow are the main causes of loss generation in the baseline impeller.
In contrast to the baseline impeller OPT1 does not have a remarkable secondary
flow at the inducer as shown in Figure 13(b). Owing to this, the accumulation of low
momentum fluid is suppressed near the tip corner of the splitter blade pressure
surface at the impeller exit compared to the baseline impeller. The LE tip leakage
vortex is remarkable different between the baseline impeller and the OPT1 impeller,
as it only develops more downstream the main blade and postpones its movement
into the passage of splitter blade pressure surface. This is due to a wider blade pitch
distance as a result of the smaller blade number, even though LE tip leakage vortex
is much stronger due to the increased blade loading resulting from the smaller
blade count. Because this stronger LE tip leakage vortex travels into the passage of
the splitter blade pressure surface, the maximum loss of this passage is higher than
that of the baseline impeller. However the area of highest loss region is much
smaller than that of the baseline impeller because there is almost no accumulation
of low momentum fluid by the secondary flow. Regarding the passage of the splitter
blade suction surface, it is obvious that the high loss region is suppressed clearly
compared to the baseline impeller because the LE tip leakage vortex of the full
blade leading edge does not come in and the related loss is diminished. OPT1 has a
very smooth flow without pronounced secondary flow at the inducer and thus
achieves a higher efficiency compared to the baseline impeller. It seems that the
spanwise and streamwise blade loading distribution between the full blade and
splitter blade are optimized to succeed suppressing the secondary flow by the
aerodynamic design optimization system composed of an ANN and a GA in this
study.
The internal flow phenomena and loss generation mechanisms of OPT2 are almost
identical to those of OPT1. In contrast to OPT1, OPT2 has a secondary flow rolling
up from mid-span to shroud on the suction surface at the inducer (similar to the
baseline) as shown in Figure 13(c). Because of this secondary flow, the
accumulation of the low momentum fluid on the full blade suction surface near the
end of the inducer is observed. As a result, OPT2 has a relatively larger area of low
momentum fluid at the exit of the passage of the splitter blade pressure surface and
has lower efficiency compared to OPT1. On the other hand, compared to the
baseline impeller, the secondary flow and the accumulation of the low momentum
fluid are not as severe. It means that the design optimization system has succeeded
to improve the internal flow similar to the design of OPT1. According to the CFD
result, OPT2 has higher efficiency compared to the baseline impeller in spite of
lower efficiency confirmed by the performance test. This discrepancy between CFD
and test results remains to be solved. Significant increase of the operating range
has been confirmed by the performance test of OPT2. The investigations to find the
mechanisms of map width enhancement of OPT2 are beyond the scope of this paper.
The authors have found out the breakdown of the tip leakage vortex plays an
important role in stabilizing the unsteadiness of the flow near the stall condition in a
separate study(14).
75
5. CONCLUSIONS
(1) Two optimum design impellers, OPT1 and OPT2, were designed by an advanced
optimization system. OPT1 has a higher efficiency with slightly wider operating
range compared to the baseline impeller. OPT2 has a twice as wide operating
range compared to the baseline impeller with a minor decrease in efficiency.
This successful result clearly demonstrates the benefits of advanced optimization
systems composed of an ANN and a GA, and shows that innovative
aerodynamic design can be found while speeding up the design time.
(2) OPT1 has a 0.5% higher peak efficiency at pressure ratio 1.8 and more than
1.0% higher efficiency above pressure ratio 2.2 compared to the baseline
impeller. The operating range above the pressure ratio 1.9 is wider than that of
the baseline impeller. There is almost no secondary flow on the suction surface
of the full blade. It seems the spanwise and streamwise blade loading, and
loading split between the full blade and splitter blade are optimized to suppress
the secondary flow and achieve a higher efficiency.
(3) OPT2’s operating range has increased more than double at the pressure ratio
above 2.2 with almost equivalent operating range at low pressure ratio and 1%
lower efficiency compared to the baseline impeller. The internal flow phenomena
and loss generation mechanisms of OPT2 are almost identical to those of OPT1.
Owing to a slightly stronger secondary flow compared to OPT1, OPT2 has lower
efficiency. Significant increase of the operating range was investigated in a
separate study. The breakdown of the tip leakage vortex plays an important role
in this.
6. REFERENCES
(1) Osako, K., Jinnai, Y. Samata, A., Suzuki, H., Ibaraki, S., Hayashi, N., 2006,
“Development of the High Performance and High Reliability VG Turbocharger
for Automotive Applications”, MHI Technical Review, Vol. 43, No. 3.
(2) Fisher, F. B., 1988, “Application of Map Width Enhancement Devices to
Turbocharger Compressor Stages”, SAE Paper 880794.
(3) Tomita, I., An, B. And Nanbu, T., 2014, “A New Operationg Range
Enhancement Device Combined with a Casing Treatment and Inlet Guide
Vanes for Centrifugal Compressor”, IMechE, 11th International Conference on
turbochargers and turbocharging.
(4) Ibaraki, S., Ogita, H. and Yamada, T., 2007, “Development of a Wide
Operating Range Turbocharger Compressor with a Low Solidity Vaned Diffuser”,
CIMAC No. 166.
(5) An, B., Shiraishi, T., 2010, “Development of Variable Two-stage Turbocharger
for Passenger Car Diesel Engines”, MHI Technical Review, Vol. 47, No. 4.
(6) Ibaraki, S., Higashimori, H. and Mikogami, T., 1998, “Flow Investigation of a
Centrifugal Compressor for Automotive Turbochargers”, SAE Paper 98-P94.
(7) Ibaraki, S., Higashimori, H. and Matsuo, T., 2001, “Flow Investigation of a
Transonic Centrifugal Compressor for Turbocharger”, 23rd CIMAC.
(8) Ibaraki, S., Matsuo, T., Kuma, H., Sumida, K and Suita, T., 2003,
“Aerodynamics of a Transonic Centrifugal Compressor Impeller”, ASME Journal
of Turbomachinery, Vol.125, No.2,pp. 346-351.
76
(9) Ibaraki, S., Furukawa, M., Iwakiri, K. and Takahashi, K., 2001, “Vortical Flow
Structure and Loss Generation Process in a Transonic Centrifugal Compressor
Impeller”, ASME Paper No. GT2007-27791.
(10) Zangeneh, M., Roduner, D. V. C., 2002, “Improving a Vaned Diffuser for a
Given Centrifugal Impeller”, ASME Paper No. GT-2002-30621.
(11) Arnone A. and Pacciani R., 1995, “Rotor-Stator Interaction Analysis Using the
Navier-Stokes Equations and a Multigrid Method”, ASME Paper 95-GT-177.
(12) Pierret S. and Van den Braembussche R.A., 1998, “Turbomachinery blade
design using a Navier-Stokes solver and Artificial Neural Network”, ASME
Trans. Journal of Turbomachinery, Vol. 121, No. 9, (pp. 326-332).
(13) Alsalihi Z. and Van den Braembussche R.A., 2002, “Evaluation of a Design
Method for Radial Impellers Based on Artificial Neural Network and Genetic
Algorithm”, ASME ESDA 2002/ATF-069, Istanbul.
(14) Tomita, I., Ibaraki, S., Furukawa, M., Yamada, K., 2012, “The Effect of Tip
Leakage Vortex for Operating Range Enhancement of Centrifugal Compressor”,
ASME Paper, GT2012-68947.
77