output 1 output 2 output 3 output 4
8.2. METHODS F OR THE INTERPOLA TION OF THE REDUCED MODELS
(P1) (P2) (A1) (A2)
referene
subspae
R V =
svd([V (p 1 ), . . . . . . , V (p K )])
, SVDof the projetion
matries
R V =
svd([ω 1 V (p 1 ), . . . . . . , ω K V (p K )])
,weighted SVD
of the projetion
matries
R V = V (p j 0 )
,proje-tionmatrixofhosen
referenemodel
R V = V (p j 0 )
,proje-tionmatrixofhosen
referenemodel
trans-formation
matries
T j = M j =
(R V T V (p j )) −1
T j = M j =
(R V T V (p j )) −1
T j = U j Z j T
is givenby the SVD of
V (p j ) T R V = U j Σ j Z j T
,and the matrix
M j
is obtained as
M j = V (p j ) T E(p j )V (p j ) −1
T j = U j Z j T
is givenby the SVD of
V (p j ) T R V = U j Σ j Z T j
,and the matrix
M j
is obtained as
M j = V (p j ) T E(p j )V (p j ) −1
manifolds
for
inter-polation
no manifoldis
ho-sen
nomanifoldis
ho-sen
the manifold of real
n × n
,n × m
andp × n
matries,dependingonwhihmatrixto
in-terpolate
the manifold of real
n × n
,n × m
andp × n
matries,dependingonwhihmatrixto
in-terpolate - for
A(p ˆ j )
the manifold of the
non-singular matries
ishosen
Table8.3. Two-stepmethodsfortheinterpolationof
re-duedordermodels.
(Af-P1) (Af-P2) (Af-A1) (Af-A2) (Af)
Firststep Calulationofreduedordermodelsfortheane
param-etersinthenewparameterpoint
p new
see8.2.Seond
step
in-terpolation
method
used
(P1),
see
Table
8.2
(P2),
see
Table
8.2
(A1),
see
Table
8.2
(A2),
see
Table
8.2
no
inter-polation
neessary
only
ane
pa-rameters
8.3. Redutionandinterpolationusingreformulationone
Tosimplifythepresentation,wextheparameters
µ, γ, ρ
,soonlyoneaneparameter
θ
remains. After thereformulation(R1) andasalingof thematriesN 1
andN 2
asexplainedinSetion8.1.1,thefollowingsystemisobtained:
Σ bilin (θ) :
(E 0 + θE θ ) ˙ x(t) = A 0 x (t) + X 6 k=1
gN k u g k (t)x(t) + Bu g (t), y (t) = Cx (t ),
(8.5)
with
u g (t ) = 1
g(1+θ) θ
g T 0 T ∞ θT ∞ (1 + θ)S(t) T
, A 0 = A 0 (γ, ρ) + hA h0 (γ, ρ) + 1
1 + µ A 1
1+µ (γ, ρ) + µA µ (γ, ρ), N 1 = A 1
1+θ (γ, ρ),
N 2 = A θ (γ, ρ) + hA hθ (γ, ρ),
N 3 = · · · = N 6 (γ, ρ) = 0,
B =
0 0
.
.
. .
.
.
0 0 1 1 + µ B 1
1+µ (γ, ρ) + µB µ (γ, ρ) + B 0 (γ, ρ) hB h0 (γ, ρ) hB hθ (γ, ρ) B S (γ, ρ)
.
Nowtheresultsforthelargemodelwith
n = 71, 978
degreesoffree-domfrom Setion 4.3 are disussed. As noted before, the
N k
are largeandneedtobesaledbeforearedutionofthesaledsystem(8.5)anbe
performed.
UsingBIRKAas givenin Algorithm3and theKronekerprodut
approxi-mation(f. Setion6.1),wereduethemodelasgiven byequation(8.5)
atvedierentsamplingpoints
θ ∈ {0, 0.5, 1, 1.5, 2}
toareduedorderofr = 700
. Afterthe redution,stablemodelsare obtainedby using aonesidedprojetion
V
inthelastmodel(f. BIRKA-oSinSetion7.2.1.2).Theinterpolation between the redued models at the samplingpoints is
ondutedusingmethods(P2),(A1)and(Af)fromSetion8.2.
We examine the temperature distribution at four dierent points in the
model: At the bottom of the housing, on the oil, in the upper bearing
andatthebottomoftherotor. Resultsfortheinterpolatedmodelsattwo
dierentparameterpoints
θ new ∈ {0.45, 1.65}
fortwo{0, 2}
,three{0, 1, 2}
andve
{0, 0.5, 1, 1.5, 2}
samplingpointsanbefoundin theFigures8.1and8.2,fortherstandthefourthoutput,respetively.
Thequalityoftheapproximationimproveswithinreasingthenumber
ofsamplingpoints. Whenusing ve samplingpoints,the interpolated
re-duedmodelsfor
θ new ∈ {0.45, 1.65}
yieldgoodresults fortherst threeoutputs.Itseemshoweverdiulttoapproximatethefourthoutput,whih
evenwithvesamplingpointsonlyleadstogoodmodelsforthe
ap-proahviaaglobalprojetionmatrix(Af),as itanbeseeninFigure8.2.
Thismight berelated to thefat that this output lieson thebottom of
therotorandisnotdiretlyattahedto thestator(asmainheatsoure).
Henetheheatanonlybetransferredviahousingandange.
250 300 350 400 450 500 550
Temperature(K)
Twosamplingpoints
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6
relativeerror
250 300 350 400 450 500 550
Temperature(K)
Threesamplingpoints
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
relativeerror
0 200 400 600
250 300 350 400 450 500 550
Time(s)
Temperature(K)
Fivesamplingpoints
Original
θ = 0.45
Originalθ = 1.65
(P2)
θ = 0.45
(P2)θ = 1.65
(A1)
θ = 0.45
(A1)θ = 1.65
(Af)
θ = 0.45
(Af)θ = 1.65
0 200 400 600
10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
Time(s)
relativeerror
Figure8.1. Firstoutput(bottomofthehousing),
inter-polation ofreduedorder models(r=700) in adierent
numberofsamplingpoints,withresultsindierent
inter-300 350 400
Temperature(K)
Twosamplingpoints
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6
relativeerror
300 350 400
Temperature(K)
Threesamplingpoints
10− 1
10− 2
10− 3
10− 4
10− 5
relativeerror
0 200 400 600
300 350 400
Time(s)
Temperature(K)
Fivesamplingpoints
Original
θ = 0.45
Originalθ = 1.65
(P2,2sp)
θ = 0.45
(P2,2sp)θ = 1.65
(A1,2sp)
θ = 0.45
(A1,2sp)θ = 1.65
(Af, 2sp)
θ = 0.45
(Af,2sp)θ = 1.65
0 200 400 600
10− 1
10− 2
10− 3
10− 4
Time(s)
relativeerror
Figure 8.2. Fourthoutput (bottomoftherotor),
inter-polation of redued order modelsin a dierent number
ofsamplingpoints,withresultsindierentinterpolation
Table8.4. Costs for the redution and interpolationof
themodelwithoneaneparameter.
Method Costs
OineRedutioninoneparameterpoint 1 weekper
sam-plingpoint
OnlineInterpolationwith(A1)or(A2) 20-25min
OnlineInterpolationwith(P1)or(P2) 10-15min
OineCalulationoftheglobalprojetion
matrix
20min
OnlineAssemblingofthemodelinthenew
parameterpoint
<1min
Aswehaveonsideredamodelinoneaneparameter,itwaspossible
to usethemethod viaaglobalprojetionmatrix(Af)andno (additional)
interpolationbetweenthereduedordermodels. Thismethodalwaysleads
togoodresults,andheneitanbereommendedwhenevertheparameter
dependenyis aneand thealulationoftheSVDof allmatries
V (θ j )
does not exeed the omputational apaity. Method (A1) outperforms
(P2) in approximationof the rstoutput (ve samplingpoints), whereas
(P2)performsbetterfortheoutputstwotofour.Hene,oneannotstate
thatoneinterpolationmethodisbetterthantheother.
The redution of the large model for one sampling point required up to
oneweekon
12
CPUswith3
GBRAMeah. Sosamplinginmorethanoneparameterwilleasilyexeedtheavailableresouresorleadtoextremelylong
simulationtimes 2
. Hene,theinterpolationmethodswillnowbetestedon
the smaller model with
n = 2, 969
degreesof freedomfrom Setion4.3.Inaddition, we willhangethe reformulationmethod, anduse theseond
reformulation(f. Setion8.1.2, (R2)),as there willbeno needto sale
themodelspriortotheredution,aswehavenotedthatasalinginthe
N k
inreasesthereduedorder(f. Remark6.3.2).
Costs for theredution and interpolationan befound in Table8.4.
Exept for the redutionthat hasbeen performed on 12CPUs with 3GB
RAM eah, the alulations have been performed on visualizationnodes
2
AdisussionexplainingthelongsimulationtimesanbefoundinSetion7.2.3.
thatareusedsimultaneouslybydierentusers. Dependingon thememory
demandsandtheloadsoftheotherusers,thealulationtimesandier.
8.4. Redutionandinterpolationusingtheseondreformulation
Forthepresentationoftheresultsobtainedbyusingtheseond
refor-mulation(R2)(f. Setion8.1.2),themodelwith
n = 2, 969
willbeused.IthasbeenpresentedinSetion4.3andisshowninFigures4.7and4.8. For
threedierentpointsthetemperatureproleismonitored: Onthebottom
ofthehousing(output1),onthestator(output2)andontheupperpart
oftherotor(output3).
To obtain stable redued order models the one-sided approah only
V(f. Chapter7.2)ishosen. Thisleads to largerreduedordersas an
originalBIRKAhoweverstabilityispreservedautomatily,whihisruial
fortheinterpolationsteps. Foreverysamplingpoint
p j = (θ j , µ j , γ j , ρ j )
theoriginalmodelwasreduedtoanorderof
r = 100
. Theparameters(θ j , µ j )
are ane,and theparameters
(γ j , ρ j )
arenon-ane, heneourexplainedtwo-stepapproahapplies. Thesamplingpointsaregivenas:
2sp:
θ j , µ j ∈ {0, 2}
andγ j , ρ j ∈ {1, 3}
;2 4
samplingpoints3sp:
θ j , µ j ∈ {0, 1, 2}
andγ j , ρ j ∈ {1, 2, 3}
;3 4
samplingpoints5sp:
θ j , µ j ∈ {1, 0.5, 1, 1.5, 2}
andγ j , ρ j ∈ {1, 1.4122, 2, 2.5878, 3}
;5 4
samplingpointswhere
{1.0489, 1.4122, 2, 2.5878, 2.9511}
aretheChebyhevpointswithin[1, 3]
. Weuse1
and3
insteadof1.0489
and2.9511
aseahoftheparam-etersisinthelosedinterval
[1, 3]
.For the interpolation of the models, we will use four dierent methods.
First,aninterpolationinallfourparameters
(θ j , µ j , γ j , ρ j )
willbeperformeddiretly(one-step approah)by usingthetwo interpolationmethods(A1)
and (P2). In addition, a two-stepapproah will be applied by using the
methods(Af-P2)and(Af-A1)-seeSetion8.2.
0 200 400 600 300
350 400 450 500
Time(s)
Temperature(K)
Firstoutput
Original
(P2,2sp) (A1,2sp) (Af-P2,2sp)
(Af-A1,2sp) (P2,3sp) (A1,3sp) (Af-P2,3sp)
(Af-A1,3sp) (P2,5sp) (A1,5sp) (Af-P2,5sp)
(Af-A1,5sp)
0 200 400 600
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
Time(s)
relativeerror
Figure8.3. Temperatureurvesfromreduedmodels
ob-tainedbyinterpolationwithdierentmethodsand
num-bers ofsamplepointsin point
θ = 1.67, µ = 1.78, γ = 2.36, ρ = 1.22
.InFigure8.3 theresultsfortwo,threeandvesamplingpointsinthe
rstoutputfortheinterpolationpoint
p new 0 = (θ = 1.67, µ = 1.78, γ = 2.36, ρ = 1.22),
and redued order
r = 100
are shown. For two samplingpoints (dottedlines)thetwo-stepmethods(i.e. (Af-P2)and(Af-A1))leadto better
re-sultsthantheone-stepmethods(i.e. (P2)and(A1)). Forthreesampling
points(dashedlines), theone-stepmethodsgetbetter ingeneral,and for
ve samplingpoints(dashdotted lines), the approximation using the
one-stepmethodsissuientlyaurateespeiallyfortheapproah(A1).
Consideringthreeotherinterpolationpoints
p new 1 = (θ = 1.67, µ = 1.78, γ = 2.976, ρ = 2.73),
p new 2 = (θ = 1.56, µ = 1.2, γ = 1.47, ρ = 1.634),
and
p new 3 = (θ = 0.34, µ = 0.13, γ = 1.134, ρ = 1.22),
theresultsfortheinterpolatedmodelsanbefoundinFigures8.4 to8.6.
Oneobservesthat one obtainsgood resultsforve samplingpointsin all
fourdierentinterpolationpoints
p new i
. Therearehoweverdierenesinthequalityoftheapproximation. Thepoint
p new 1
isforexamplenotperfetlyapproximatedbytheapproahes(Af-P2)and(Af-A1). Inaddition,onean
observeosillationsin theapproximationsby (Af-P2)and(Af-A1). They
ourwheneverthereisasignianthangeinthedynamisofthemodel.
In general: For few samplingpoints, the two-step methods (Af-A1)
and(Af-P2)(i.e. usingaglobalprojetionmatrixfortheaneparameter
dependenyandtheninterpolatingthenon-aneparameters)leadto
bet-terresultsthanadiretinterpolation. However,asthenumberofsampling
pointsinreases,theapproaheswithdiretinterpolation(i.e. (A1),(P2))
performas goodastheones withaglobalprojetionmatrixfortheane
parameters,orevenbetter. Hene,iftheredutioninonesamplingpointis
timeonsuming(asit isusingBIRKAf. Setion7.2.3),itisdesirable
tosampleasfewpointsaspossible.Ifthealulationofaglobalprojetion
matrixintheaneparametersisnottootimeonsuming,usingfew
sam-plingpointsandoneofthetwo-stepmethods((Af-A1)and(Af-P2))yields
satisfatoryresults.
300 350 400
T e m p e ra tu re (K )
First Output
10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
300 350 400 450
T e m p e ra tu re (K )
Second Output
10− 1 10−2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
0 200 400 600
280 290 300 310 320
Time (s)
T e m p e ra tu re (K )
Third Output
Original θ = 1.67, µ = 1.78, γ = 2.976, ρ = 2.73 (P2, 5sp)
(A1, 5sp) (Af & P2, 5sp) (Af & A1, 5sp)
0 200 400 600
10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
Time (s)
re la ti ve e rr o r
Figure8.4. Interpolationofreduedordermodelsinve
sampling points at
p new 1 = (θ = 1.67, µ = 1.78, γ =
2.976, ρ = 2.73)
.300 350 400 450
T e m p e ra tu re (K )
First Output
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
300 350 400 450
T e m p e ra tu re (K )
Second Output
10− 2
10− 3
10− 4
10− 5
10− 6
re la ti ve e rr o r
0 200 400 600
270 290 310
Time (s)
T e m p e ra tu re (K )
Third Output
Original θ = 1.56, µ = 1.2, γ = 1.47, ρ = 1.634 (P2, 5sp)
(A1, 5sp) (Af & P2, 5sp) (Af & A1, 5sp)
0 200 400 600
10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
Time (s)
re la ti ve e rr o r
Figure8.5. Interpolationofreduedordermodelsinve
sampling points at
p new 2 = (θ = 1.56, µ = 1.2, γ =
1.47, ρ = 1.634)
.300 350 400
T e m p e ra tu re (K )
First Output
10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
300 350 400
T e m p e ra tu re (K )
Second Output
10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
0 200 400 600
270 290 310
Time (s)
T e m p e ra tu re (K )
Third Output
Original θ = 0.34, µ = 0.13, γ = 1.134, ρ = 1.22 (P2, 5sp)
(A1, 5sp) (Af & P2, 5sp) (Af & A1, 5sp)
0 200 400 600
10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
Time (s)
re la ti ve e rr o r
Figure8.6. Interpolationofreduedordermodelsinve
sampling points at
p new 3 = (θ = 0.34, µ = 0.13, γ =
1.134, ρ = 1.22)
.Table 8.5. Costs for the redution and interpolation of
themodelwithtwoaneandtwonon-aneparameters
redutionin
sampling
points(sp)
oine redutioninoneparameterpoint:
≈ 30
min.for2sp:
2 4 · 30
min= 8
h,for3sp:
3 4 · 30
min≈ 1.7
days,for5sp:
5 4 · 30
min≈ 13
days.one-stepmethod
online Interpolationwith (A1)
< 10
min,(A2)
< 15
min,(P1)
< 10
min,(P2)
< 5
min.two-stepmethod
oine one globalprojetion matrixfor xed
non-aneparameters
(θ k , µ l , γ J ˆ , ρ J ˆ )
:≈ 1
min,for2sp:
2 2 · 1
min= 4
min,for3sp:
3 2 · 1
min≈ 6
min,for5sp:
5 2 · 1
min≈ 25
min.online interpolationofnon-aneparameterswith:
(A1)<5s,
(A2)<10s,
(P1)<20s,
(P2)<5s.
InTable8.5 approximateostsfortheredutionandinterpolationare
summarized. Againthealulationshave beenperformedon visualization
nodes used simultaneously by dierent users. The alulation times are
thereforeonlyapproximationsdependingon load andavailablememory on
thenodes. Itis notsurprising, that the interpolationusing all parameter
pointsisslower that theone, where onlythe non-aneparameters need
tobeinterpolated. Ingeneral,(A1)isfasterthan(A2)and(P2)isfaster
than(P1). Thisisduetothefollowingbehavior: Theredutionin(P2)is
performedusingaweightedSVD.Weusetheweightsthatwillbeusedfor
thelinearinterpolationofthemodelsafterwards.Asonlythenearestmodels
withrespettothenewparameterpointareusedintheinterpolation,only
theprojetionmatries
V (p j )
fromthesemodelsareusedforthealulationof the referene subspae
R V
. In ontrast, all matriesV (p j )
are usedforthe SVD in (P1). This explains longeralulation times. Duringthe
exeutionof (A1)and(A2),theinterpolationisdoneontangentialspaes
ofmatrixmanifolds. Thematriesneedto bemappedto thesespaesby
usingdierentlogarithms(seeTable5.1). Whereasthemanifoldof
n × m
matriesonlyinvolvesasubtration, themanifoldofnonsingularmatries
requiresaninversionandamatrixlogarithm.Thisleadstolongeralulation
times.
8.4.0.1. Interpolationmethods(A2)and(P1). Sofar,onlyresultsfor
theinterpolationmethods(A1)and(P2)havebeenpresented. Thisisdue
to the fat that the obtainedresults for the approahes (A2) and (P1)
arein mostases notasgood asthe resultsforthe otherapproahes. A
omparisonoftheapproahes(P1)and(P2)fortheinterpolationpoint
p new 0 = (θ = 1.67, µ = 1.78, γ = 2.36, ρ = 1.22),
anbefound in Figure8.7, andresults forthe approah(A2) forthe
in-terpolationpoint
p new 3
areshowninFigure8.8. Whereasthemethod(P1)usuallygivesreasonableresults,themethod (A2)hassigniantproblems
intheapproximationofthethirdoutputofthemodel.
Method(A2)fails toprovideareasonableapproximation. Thismight
berelatedtotheinterpolationproedure.First,allmatriesinthesampling
points
A(p j )
(whihbelongtothemanifoldM
ofthenon-singularmatries) needtobetransferredtothetangentialspaeregardingthereferenemodelT A(p j 0 ) M
,thenalassi"interpolationinouraselinearinterpolation isperformedontheseelementsofT A(p j 0 ) M
.Itisnotlear,thatthelassi"interpolationstaysinthetangentialspae,andhenetheinterpolatedmatrix
A(p new )
mightleadto inaurateresults.300 350 400 450 500
T e m p e ra tu re (K )
First Output
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7
re la ti ve e rr o r
250 300 350 400 450 500
T e m p e ra tu re (K )
Second Output
10− 2 10− 3 10− 4 10− 5 10−6 10− 7 10− 8
re la ti ve e rr o r
0 200 400 600
270 290 310 330
Time (s)
T e m p e ra tu re (K )
Third Output
Original (P2, 5sp) (P1, 5sp)
0 200 400 600
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7
Time (s)
re la ti ve e rr o r
Figure8.7. Interpolationusingtheapproahes(P1)and
(P2) in ve sampling pointsat interpolation point
θ =
1.67, µ = 1.78, γ = 2.36, ρ = 1.22
.300 350 400 450
T e m p e ra tu re (K )
First Output
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
300 350 400
T e m p e ra tu re (K )
Second Output
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
re la ti ve e rr o r
0 200 400 600
270 290 310 330
Time (s)
T e m p e ra tu re (K )
Third Output
Original Two sampling points
Three sampling points Five sampling points
0 200 400 600
10− 1 10− 2 10− 3 10− 4 10− 5 10− 6 10− 7 10− 8
Time (s)
re la ti ve e rr o r
Figure 8.8. Interpolation of redued order models in
two,threeandvesamplingpointsatinterpolationpoint
θ = 0.34, µ = 0.13, γ = 1.134, ρ = 1.22
for(A2).8.4.1. Disussionofresults. Inthishapter,resultsfortheredution
and interpolation of thermal modelswith geometri variations have been
presented. Linear parametri models have been reformulated as bilinear
modelsintwodierentways(f.Setion8.1.1and8.1.2)andthenredued
usingBIRKAwith one-sided projetions(f. Setions6.2.2 and7.2.1.2).
First, results for the rst reformulation (R1) (f. Setion 8.1.1), for a
modelwith
n = 71, 978
andone geometrialparameterhavebeen shown(f. Figures8.1and 8.2). Anadditionalpreproessingstepwas neessary
to avoid problemsresulting from the fat, that the norms of
N k
andA
areof thesamemagnitude. A salingwas introduedandleadto alarge
reduedorder
r = 700
. Theseondreformulation(R2),Setion8.1.2,does notrequire this preproessing. Due to high omputational demands (f.Setion7.2.3),allresultsfortheseondreformulationandfourparameters
have been presented for a smaller model with
n = 2, 969
. Interpolation ofthismodel usingdierentnumbersofsamplingpointsandinterpolationmethods(f.Setion8.2)havebeenperformed.Ingeneral,allmethodsgive
reasonableresults. However, themethod (P2) usinga weightedSVD
toobtainthereferenesubspaeusuallyoutperformsthemethod (P1)
thenon-weightedSVD.Inaddition,it wasnotpossibletoobtaingood
results for the interpolation method on tangentialspaes of non-singular
matries(A2),whereastheinterpolationontangentialspaesof
R k×l
leads to good results (A1). The two approahes (A1) and (P2) usually giveomparableresults,heneit isnotpossibleto favoronemethod overthe
other.Havingtwoaneandtwonon-aneparameters,itisreommended
touse atwo-stepmethod rst alulatea globalprojetion matrixfor
theaneparametersandtheninterpolatethereduedordermodelsinthe
non-aneparameters. Forfewsamplepointsthesemethodsyield usually
betterresultsthantheonestepmethods.
Conlusions and Outlook
9.1. SummaryandConlusions 171
9.2. Futureresearh 173
9.1. SummaryandConlusions
Themain objetiveofthisworkwasto investigatetheuseofbilinear
H 2
-optimal methodsin parametriModel Order Redution. Asshownby BennerandBreiten[11℄,itispossibletoreformulateaertainlassoflinearparametrisystems as bilinearsystems(f. Setion5.3.2). The
parame-tersanthenbeonsideredasinputsandtheredutionanbeperformed
withoutanysamplingandinterpolationintheparameterspae,as mostof
theothermethodsforpMORdo[53,3,37,13℄. Afterobtainingabilinear
model,oneanmakeuseofbilinearModelOrderRedution. Inthiswork,
we fousedon two methods for bilinear
H 2
-optimal Model Order Redu-tion,whiharedesribedinChapter5. BIRKA(f. Algorithm3),originallyobtainedbyBennerandBreiten[12℄,isstatedandnewalgorithmsforthe
bilinear
H 2
-optimalredutionhavebeen developed. Thesealgorithmsuse optimizationon Grassmannmanifoldsandasamainadvantageanpreservestability. Wehaveproventhestabilitypreservationforsymmetri,
bilinearsystemsandanalyzedtheonvergenebehaviorofthealgorithms.
In addition to these theoretial results, several models for the
ther-mal analysisofeletrialmotors havebeen builtusing Comsol
R3.5a (f.
Chapter3). LinearparametrisystemshavebeenexportedfromComsol
R by an analysisof theunderlyingequations (f. Chapter 4). Forindustri-allyrelevantproblems,bothphysialandgeometriparametersneedtobe
onsideredandtheparameterdependenyaftertheredutionmustbe
pre-served. Astheresultingmodelsareusuallylarge(inourase
n = 41, 199
,n = 71, 978,
andn = 2, 969
),thebilinearH 2
-optimalredutionmethods havetobeapableofdealingwiththeselargesystems.Thenewlydevelopedmethodsfortheredutionusingoptimizationon
Grassmannmanifoldsare,however,notyetready(f. Setion7.1)forthe
usewiththeselargesystems,butresultsfortheredutionofaheat
equa-tiononasquarehavebeenstated. BIKRA(f.Setion5.5,[12℄)isapable
ofreduingthelargemodels,butseveralproblemshavebeenidentied. In
some ases,the stinessmatrix
A
issingular,themagnitudeof theN k
istoolargeandasalingneedstobeintrodued. Alsounstablemodelshave
been obtainedafter the redution. All theseissueshaven been examined
andsolutionshavebeenproposed(f.Chapter6).
Numerial results for the redution of two dierent types of models
have been obtained. On one hand, a part of an eletrialmotor model,
inorporatingphysialparameters, hasbeen onsidered. Thesemodelsare
parametrizedwithphysialparametersandhaveastruturethat easily
al-lowstoreformulatethemasabilinearmodel. RedutionwithBIRKAyields
good results, notonly in a ertain parameter interval, butglobally in the
wholeparameterrange(f. Chapter7.2,Figure7.8). Theseondtypeof
modelsareeletrialmotormodels,thatinadditiontothephysial
parame-tersuseparametersthatdesribehangesingeometry.Thisleadstomodels
withastruturethatannoteasilyberewrittenasabilinearsystem. Hene
one anreformulatethe model as abilinearmodel for ertainparameters
andinterpolatetheotherparameters(f. Chapter8).Fortheinterpolation,
severalwellknownmethodsfrompMORhavebeenused(f. [53,3,37℄),
whihgenerallyleadtogoodresults. Thereare,however,dierenesinthe
qualityoftheapproximation. For modelswith ananeparameter
depen-deneinertainparameters,usingaglobalprojetionmatrixfortheane
parameter dependeneleads to goodresults and anoutperforma diret
interpolation,espeiallyforfewsamplingpoints.
9.2. Futureresearh
Basedon theworkthathasbeen presented inthis thesis,several
op-portunitiesforfutureresearhhavebeenidentied:
•
The new methods for the bilinearH 2
-optimal MOR using op-timizationmethodson theGrassmannmanifoldas developedinSetions5.5.4and7.1 stillrequiresomeinvestigation:
TheAlgorithmsbilGFA, bilFGFAand bilSQA havenotyet
beentestedonlargeproblems,duetothefatthatoneneeds
tosolvelargebilinearSylvesterequations. Inthefuture,
low-rankapproximationstothesolutionsshouldbeappliedsuh
as theADI iteration (f. [57, 14℄),to allowtreatmentof
largesystems.
ConvergeneandthestabilitypreservationfortheAlgorithms
bilGFA,bilFGFAandbilSQAhavenotyetbeenestablishedfor
bilinearsystemswithnon-symmetri
A
andN k
.Fortheoptimization,oneneedsto orretlysetseveral
pa-rametersto ensurea desentin the objetivefuntion. It
wouldbeanadvantageto identify robustriteriabased on
whihtheseparametersanbehosen.
•
Theredutionofthelargeparametri thermalmodelshasbeendoneusingBIRKA[12℄. Theredutiontimesforourlargemodels
arewithintherangeofseveralhoursto afewdaysfor12 CPUs
with 3GB RAM(see Setion for adisussion7.2.3). However,
thestrutureofBIRKAwouldallowaparallelization,whihould
signiantlyreduetheredutiontime.
•
OneinterpolationapproahbyAmsallem[3℄shows weak perfor-mane for some models (f. Setion 8.4.0.1). This ould beausedby thefatthatour usedinterpolationmethod doesnot
preservethe membershipin thetangentialspae. Thisbehavior
requiresa developmentofinterpolation proedures that dostay
ontheorrespondingmanifold.
•
Theinterpolationmethodsusedfortheredutionofthe paramet-rimodelsrequiretheredutionatseveralsamplingpoints. Thenumber ofsamplingpoints hasa strongimpat on the
ompu-tationaldemands,soit isworthwhiletoexploremethodsto
sys-tematiallyandoptimallysampletheparameterspae,e.g. using
sparsegrids[10℄orlatinhyperubesampling[4,20℄.