characteristics of svd in st plasma shape reproduction method based on ccs
DESCRIPTION
CUP2008, EAST, ASIPP, Mar. 2, 2009 (Mon.). Characteristics of SVD in ST Plasma Shape Reproduction Method Based on CCS. Kazuo NAKAMURA 1 , Shinji MATSUFUJI 2 , Masashi TOMODA 2 , Feng WANG 3 , - PowerPoint PPT PresentationTRANSCRIPT
Characteristics of SVD in ST Plasma Shape Reproduction Method Based on CCS
Kazuo NAKAMURA1, Shinji MATSUFUJI2, Masashi TOMODA2, Feng WANG3, Osamu MITARAI4, Kenichi KURIHARA5, Yoichi KAWAMATA5, Michiharu SUEOKA5, Makoto HASEGAWA1, Kazutoshi TOKUNAGA1, Kohnosuke SATO1, Hideki ZUSHI1,
Kazuaki HANADA1, Mizuki SAKAMOTO1, Hiroshi IDEI1, Shoji KAWASAKI1, Hisatoshi NAKASHIMA1 and Aki HIGASHIJIMA1
1Research Institute for Applied Mechanics, Kyushu University, Kasuga, Fukuoka, 816-8580 Japan, 2Interdisciplinary Graduate School of Engineering Sciences, Kyushu University,
Kasuga, Fukuoka, 816-8580 Japan 3Institute of Plasma Physics, Chinese Academy of Sciences, Hefei, 230031 China
4Tokai University, Toroku, Kumamoto, 862-8652 Japan 5Japan Atomic Energy Agency, Mukoyama Naka-shi, Ibaraki-ken, 311-0193 Japan
CUP2008, EAST, ASIPP, Mar. 2, 2009 (Mon.)
Abstract
LSM (Least Square Method) has been used to solve the inverse problem from flux loop measurement to the boundary values on the CCS (Cauchy Condition Surface). When the CCS method is applied to real experimental data, noise superposition is inevitable. By introduction of SVD (Singular Value Decomposition) and truncation of the least SV components, we can expect the shape reconstruction robustness against the noise. [1] K. Kurihara: Fusion Eng. Design, Vol. 51-52 (2000) 1049-1057.
[2] F. Wang, K. Nakamura, O. Mitarai, K. Kurihara, Y. Kawamata, M. Sueoka, K. N. Sato, H. Zushi, K. Hanada, M. Sakamoto, H. Idei, M. Hasegawa, et al.: Engineering Sciences Reports, Kyushu University, Vol. 29, No. 1 (2007) 7-12.
Flux (known)
Boundary values on CCS
(unknown)
Noise
Inverse problem
Normal calculation
Characteristics of SVD in ST Plasma Shape Reproduction Method Based on CCS
• Abstract
• Cauchy Condition Surface method– Observation equation– Least Square Method– Singular Value Decomposition– Magnetic surface and Largest Closed Flux Surface
• Plasma shape reproduction by SVD
• Robustness by least SV truncation
• Summary
bUVWx
bxUWV
bAAAx
bAxAA
bAx
t
t
tt
tt
)(
)(
)()(
)()(
1
1
-1000
-500
0
500
1000
10008006004002000
CCS( Cauchy Condition Surface)Method
・ Unknown: Boundary values on CCS
・ Known: Magnetic sensor (flux loop, magnetic probe), PFC current
・ Solve the observation equation
・ Calculate vacuum magnetic surfaces
・ Identify LCFS as plasma boundary
observation equation
LSM
Error⇒(Condition number)2
SVD
Error ⇒ (Condition number)1
Condition number = max SV / min SV
[1] K. Kurihara: A new shape reproduction method based on the Cauchy-condition surface for real time tokamak reactor control, Fusion Eng. Design 51-52 (2000) 1049.
Vacuum
Plasma
We consider “vacuum region”.
We think “plasma” is outside of the “vacuum region”.
“Flux value” in the “vacuum region” is calculated from the boundary values on “CCS”.
3 3 31 1
1( ) ( , ) ) ( , ) ) ( )
2
M M
F i i B i i C PFi i
x W x z z W x z t z W x I
2 2 21 1
( ) ( , ) ( ) ( , ) ( ) ( )M M
B F B i i B B i i C B PFi i
Bt x W x z z W x z t z W x I
1 1 11 1
( ) ( , ) ( ) ( , ) ( ) ( )M M
f F f i i B f i i C f PFi i
x W x z z W x z t z W x I
4 4 41 1
( ) ( , ) ( ) ( , ) ( ) ( )M M
F i i B i i C PFi i
x W x z z W x z t z W x I
Flux values at flux loop:
Magnetic field at magnetic probe:
Boundary values on CCS:
Outside of plasma:
Magnetic surface → LCFS → Plasma boundary
ECfC IxW )(8
ECfC IxW )(5
ECfC IxW )(6
ECfC IxW )(7
Boundary values on CCS (unknown)
Eddy current (unknown)
Least Square Method and Singular Value Decomposition
yAAAx
yAxAA
yxA
LSM
~~
~~
1
yUΔVx
yxVΔU
yxA
SVD
~
~
1
⊿: Diagonal component is SV
Observation euation
Normal equation
Multiply the transposed matrix
Multiply the inverse matrix
bUVWx
bxUWV
bAIAAx
bAxIAA
xxbAxbAx
bAAAx
bAxAA
bAxbAx
bAx
t
t
tt
tt
tt
tt
tt
t
)(
)(
)()(
)()(
)()min(
)()(
)()(
)()min(
1
1
1
Observation equation
Minimize the error
LSM
Error⇒(Condition number)2
Minimize the error and the norm
LSM
Error⇒(Condition number)2
SVD
Error ⇒ (Condition number)1
11111
11
1
11
321
3
2
1
321
333222111
00001
0
0
0
0
1
,
00
00
00
,
vVWVUWVuAu
uUUWvUWVAv
vvvVWuuuU
vuvuvuUWVA
ttttt
t
tttt
Singular Value Decomposition
Eigen (Characteristic) Value Decomposition
11111
11
1
11
321
3
2
1
321
333222111
00001
0
0
0
0
1
,
00
00
00
,
0det
0
uUUUUuAu
uUUuUUAu
uuuUuuuU
uuuuuuUUA
IA
uIA
uAu
ttttt
t
tttt
Definition of eigen value and eigen vector
Eigen (Characteristic) equation must hold good.In order that this equation has non-trivial solution,
Larger eigen-value component has larger information.
Larger singular-value component has larger information.
Least Square Method and Singular Value Decomposition
yAAAx
yAxAA
yxA
LSM
~~
~~
1
yUΔVx
yxVΔU
yxA
SVD
~
~
1
⊿: Diagonal component is SV
Matrix is multiplied only one time.Matrices are multiplied two times.
bUVWx
bxUWV
bAIAAx
bAxIAA
xxbAxbAx
bAAAx
bAxAA
bAxbAx
bAx
t
t
tt
tt
tt
tt
tt
t
)(
)(
)()(
)()(
)()min(
)()(
)()(
)()min(
1
1
1
Observation equation
Minimize the error
LSM
Error propagation ⇒ (Condition number)2
Minimize the error and the norm
LSM
Error propagation⇒(Condition number)2
SVD
Error propagation ⇒ (Condition number)1
11111
11
1
11
321
3
2
1
321
333222111
00001
0
0
0
0
1
,
00
00
00
,
vVWVUWVuAu
uUUWvUWVAv
vvvVWuuuU
vuvuvuUWVA
ttttt
t
tttt
Condition number = max SV / min SV
3 3 31 1
1( ) ( , ) ) ( , ) ) ( )
2
M M
F i i B i i C PFi i
x W x z z W x z t z W x I
2 2 21 1
( ) ( , ) ( ) ( , ) ( ) ( )M M
B F B i i B B i i C B PFi i
Bt x W x z z W x z t z W x I
1 1 11 1
( ) ( , ) ( ) ( , ) ( ) ( )M M
f F f i i B f i i C f PFi i
x W x z z W x z t z W x I
4 4 41 1
( ) ( , ) ( ) ( , ) ( ) ( )M M
F i i B i i C PFi i
x W x z z W x z t z W x I
At flux loop:
At magnetic probe:
On CCS:
Outside of plasma:
Magnetic surface → LCFS → Plasma boundary
ECfC IxW )(8
ECfC IxW )(5
ECfC IxW )(6
ECfC IxW )(7
Cauchy condition (unknown)
Eddy current (unknown)
(Largest Closed Flux Surface)
bUVWx
bxUWV
bAAAx
bAxAA
bAx
t
t
tt
tt
)(
)(
)()(
)()(
1
1
-1000
-500
0
500
1000
10008006004002000
CCS( Cauchy Condition Surface)Method
・ Unknown: CC on CCS
・ Known: Magnetic sensor (flux loop, magnetic probe), PFC current
・ Solve the observation equation
・ Calculate vacuum magnetic surfaces
・ Identify LCFS as plasma boundary
observation equation
LSM
Error⇒(Condition number)2
SVD
Error ⇒ (Condition number)1
Condition number = max SV / min SV
[1] K. Kurihara: A new shape reproduction method based on the Cauchy-condition surface for real time tokamak reactor control, Fusion Eng. Design 51-52 (2000) 1049.
(Largest Closed Flux Surface)
-0.4
-0.2
0.0
0.2
0.4
Flux
loop
val
ue (
mW
b)
403020100
Outer---Upper---Inner Flux Loop---Lower---Outer
Flux value measured Singular Vector Singular Vector1 Singular Vector2
-0.4
-0.2
0.0
0.2
0.4
Flux
loop
val
ue (
mW
b)
403020100
Outer---Upper---Inner Flux Loop---Lower---Outer
Flux value measured Singular Vector3 Singular Vector4 Singular Vector5
SV of coefficient matrix in observation equation
Flux loop value and lower modes
Flux loop value and higher modes
8
6
4
2
0
Sin
gula
r Val
ue
6543210-1Singular Value No
Even Even Even
Odd Odd Odd
Larger singular-value component has larger information.
-0.4
-0.2
0.0
0.2
0.4
Flu
x lo
op v
alue
(m
Wb)
403020100
Outer---Upper---Inner Flux Loop---Lower---Outer
Flux value measured Singular Vector Singular Vector2 Singular Vector4
Flux loop value and even modes
SV of coefficient matrix in observation equation
-0.4
-0.2
0.0
0.2
0.4
Flu
x lo
op v
alue
(m
Wb)
403020100
Outer---Upper---Inner Flux Loop---Lower---Outer
Flux value measured Singular Vector1 Singular Vector3 Singular Vector5
Flux loop value and odd modes
8
6
4
2
0
Sin
gula
r Val
ue
6543210-1Singular Value No
Even Even Even
Odd Odd Odd
Larger singular-value component has larger information.
Error and contribution fraction (singular value)^2
Flux loop value and even-mode accumulation
-0.4
-0.2
0.0
0.2
0.4
Flu
x lo
op v
alue
(m
Wb)
403020100
Outer---Upper---Inner Flux Loop---Lower---Outer
Flux value measured Singular Vector0 Singular Vector2 Singular Vector4
-0.4
-0.2
0.0
0.2
0.4
Flu
x lo
op v
alue
(m
Wb)
403020100
Outer---Upper---Inner Flux Loop---Lower---Outer
Flux value measured Singular Vector0 Singular Vector02 Singular Vector024
bUVWx
bUxVW
bxUWV
bAx
t
tt
t
)(
)()(
)(
1
-1000
-500
0
500
1000
10008006004002000
W: Diagonal component is SV.
SVD
2.0
1.5
1.0
0.5Sum
of Fl
ux L
opp̂
2
543210Singular Value No
1.00
0.95
0.90
0.85
0.80
Contribution from
SV
Sum of Flux Loop Diff^2 Contribution from SV
0.20
0.15
0.10
0.05
0.00
-0.05
CC
S fl
ux
6543210-1
CCSM
Flux value on CCS Singular Vector0 Singular Vector2 Singular Vector4
Flux value and eigen vectors of even modes on CCS
out ← up ← in in ← down ← out
-1000
-500
0
500
1000
10008006004002000
-1000
-500
0
500
1000
10008006004002000
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Z(m
)
0.60.40.20.0
R(m)
CCS ECODE
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Z(m
)
0.60.40.20.0
R(m)
CCS ECODE
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Z(m
)
0.60.40.20.0
R(m)
CCS ECODE
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Z(m
)
0.60.40.20.0
R(m)
CCS ECODE
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Z(m
)
0.60.40.20.0
R(m)
CCS ECODE
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Z(m
)
0.60.40.20.0
R(m)
CCS ECODE
1 only 1 - 2 1 - 3
1 - 4 1 - 5 1 - 6
bUVWx
bxUWV
bAIAAx
bAxIAA
xxbAxbAx
bAAAx
bAxAA
bAxbAx
bAx
t
t
tt
tt
tt
tt
tt
t
)(
)(
)()(
)()(
)()min(
)()(
)()(
)()min(
1
1
1
Observation equation
Minimize the error
LSM
Error⇒(Condition number)2
Minimize the error and the norm
LSM
Error⇒(Condition number)2
SVD
Error ⇒ (Condition number)1
11111
11
1
11
321
3
2
1
321
333222111
00001
0
0
0
0
1
,
00
00
00
,
vVWVUWVuAu
uUUWvUWVAv
vvvVWuuuU
vuvuvuUWVA
ttttt
t
tttt
Least Square Method and Singular Value Decomposition
yAAAx
yAxAA
yxA
LSM
~~
~~
1
yUΔVx
yxVΔU
yxA
SVD
~
~
1
⊿: Diagonal component is SV
Shape difference increases in proportional to noise of flux value
Noise of flux value is simulated by random noise from ±1% to ±5% and shape difference is defined as the maximum difference of plasma shape.
RLRL ZZRRdiff ,,,max
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
Flux Loop Diff
Max/FL
20151050Measure Noise [%]
SV0 to SV1
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
Flux Loop Diff
Max/FL
20151050Measure Noise [%]
SV0 to SV1 SV0 to SV3 SV0 to SV5
oddevenoddevenoddeven
SVSVSVSVSVSV 543210
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
Flux Loop Diff
Max/FL
20151050Measure Noise [%]
SV0 to SV5
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
Flux Loop Diff
Max/FL
20151050Measure Noise [%]
SV0 to SV3
oddevenoddevenoddeven
SVSVSVSVSVSV 543210
25
20
15
10
5
0Sum
of Fl
ux L
opp
Diff
2̂
20151050Measure Noise [%]
SV0 to SV1 SV0 to SV3 SV0 to SV5
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
Flux Loop Diff
Max/FL
20151050Measure Noise [%]
SV0 to SV1 SV0 to SV3 SV0 to SV5
oddevenoddevenoddeven
SVSVSVSVSVSV 543210
Summary: Characteristics of SVD in ST Plasma Shape Reproduction Method Based on CCS
• Ideal equilibrium magnetic surface by equilibrium code– CCS (Cauchy-Condition Surface) method is a numerical approach
to reproduce plasma shape, which has good precision in conventional tokamak and also in spherical tokamak.
– When the CCS method is applied to real experimental data, noise superposition is inevitable.
– SVD may be effective, since error is proportional to condition number, though the error is proportional to the square in LSM
– By introduction of SVD (Singular Value Decomposition) and truncation of the least SV components, we can expect the shape reconstruction robustness against the noise.
– It is expected that noise effect decreases though detailed magnetic surface information is lost, if smaller SV is neglected.
– The truncation simulation of the less SV component shows shape reproduction error less than no truncation, when the measurement error exceeds a certain value.
– In this configuration, the ST plasma is symmetrical with respect to the equatorial plane and the number of free parameters on CCS is effectively half of 6.
– When unsymmetrical-shape plasma is considered or the degree of freedom is increased, the certain value for inversion may decrease.