Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
=
(
(
(
A + I
A + I
(
(
(
V
(
(
(
+
(
(
(
V
(
(
(
(
(
E
s y s x s
y r x r
s
r
ss rs
sr rr
s
r rr
I M M
M M
a
a
K K
K K
s
a
a S 0
0 0
0
3
~
0 0
0 0 0
0 0
0
~
3
1
0
|
o
|
(5)
Steady-State Analysis
Under steady-state assumptions, after one period T we wish to achieve
the same primary variables, i.e.
If we define the nonlinear state transition function as , this becomes:
) (- +
(6)
Steady-State Solution Techniques
The state transition function in (6) is determined by (5)
To achieve the steady-state solutions, we used the following numerical
techniques:
o Backward Euler Integration
o Shooting-Newton Algorithm
o Matrix-free GMRES
(5)
(6)
Shooting Newton Method
Modify
NO
Time integration of the differential equation
YES
Initial Value Final Value
Error < Tolerance?
Target Value (=Initial Value) End Iteration
Calculate
Modification
The steady-state solution in the time domain is obtained by
using a shooting Newton method:
Applying the Newton-Raphson method, we get:
Where J
+
is the Jacobian of the nonlinear state transition function
(8)
In our solver we define
(7)
Matrix-free GMRES Method
Generalized minimum residual method (GMRES) :
A Krylov-subspace method
When solving Ax=b, no direct access of matrix A is used, A only
need to be accessible via a subroutine that returns y=Az
The Jacobin
1
1
1
) 0 (
) (
) , 0 , ) 0 ( (
+
c
c
=
j
j
j
x
T x
T x J
can not be explicitly written
(8)
A
x b
Efficiency Improvement
By using GMRES, the computation time of the shooting-Newton
method is dramatically reduced.
Flux Density Distribution
Rotor Loss Distribution
Harmonic analysis
0 20 40 60 80 100
0
0.05
0.1
0.15
0.2
0.25
harmonics
r
o
t
o
r
l
o
s
s
e
s
(
w
)
After we achieve the steady-state solution, we perform a Discrete Fourier
Transform of x(t) and calculate the eddy current rotor losses for each
harmonic.
Rotor Loss Design Study
Design parameters are changed to study the effect
on rotor losses
The stator current peak is adjusted to maintain
constant steady-state mechanical power
Segmenting the PM poles
Other Rotor Design Aspect
Laminated backiron
Different permanent magnet materials
Stator Design: 36 Stator Slots
Stator Design: Open Slot vs. Closed Slot
open slot
closed slot
0 20 40 60 80 100
0
0.05
0.1
0.15
0.2
0.25
harmonics
l
o
s
s
(
w
)
Techniques for Reducing Rotor Losses
Laminating rotor backiron
Segmenting the Permanent magnet poles
Increasing slot number
Closing the stator slots
Summary
Purpose of Finite element analysis for electric machines
The flywheel energy storage system
The steady-state nonlinear finite element solver
Its application
Further research into numerical modeling of electric
machines
Special machine design and analysis
Future Work
2011 Chevrolet Volt Propulsion System
(Credit: GM)
Thank you!
This work has been supported by NASA Grant NAG3-2598
Currents flowing in permanent magnet are eddy currents. In machine
design, magnets can be electrically insulated from each other and the
rotor backiron.
Eddy currents meet large impedance at the
end of the machine, surface charge will
accumulate and electric potential is built
across the magnet
Permanent Magnets
) (
1
2
0
A M A
t
A
V + V = V
c
c
| o
o
Electric Scalar Potential Dynamics
| o o V
c
c
=
t
A
J
eddy
(2)
s
ds J
s
eddy
c
}
= o
c
o
|
c
= V
ds
t
A S
s
V +
c
c
= V
}
) ( | |
o
c
The charge relaxation time constant c/o is extremely small in
permanent magnets, and so our system consists of a set of fast
dynamics (electric scalar potential) and slow dynamics (magnetic
vector potential). Therefore, singular perturbation techniques can
be used to analyze the system.
Singular Perturbation Analysis Techniques
Under singular perturbation theory, when analyzing the fast
dynamics, the slow variables can be assumed to be essentially
constant
If the fast dynamics are stable under this condition, the fast
variables will converge to a quasi-steady-state value, which is a
function of the slow variables
When analyzing the slow dynamics, it can be assumed that the
fast variables have converged to their quasi-steady-state value
discussed above.
0 = V
c
c
}
| S ds
t
A
S
m
n
R z z t z t z x g z
R x x t x t z x f x
e = =
e = =
, ) ( ), , , , (
, ) ( ), , , , (
0
0
0
0
c c
c
c
small
) , 0 ,
~
,
~
( 0 t z x g =
(3)
ds
t
A S
s
V +
c
c
= V
}
) ( | |
o
c
Applying Finite Element Method
To solve the Partial Differential Equations
use 2D finite element method
Primary variable : A and V|
Element type : triangular
Shape functions : linear
Procedure used : weighted residuals
Error distribution principle :
Galerkins method
Boundary conditions:
0 =
I z
A
Newton-Raphson Method
Applying the Newton-Raphson method, we get:
Where J
+
is the Jacobian of the nonlinear state transition function
) ( ) (
1 1
1
=
c
c
j j j
x
x f x x
x
f
j
Define : and its called the Jacobian of
j
x
f
x
f
J
c
c
=
) (x f
(5)
In our solver we define
For a nonlinear equation, the Newton-Raphson Method provides the
following iterative procedure to solver for x:
0 ) ( = x f
Equation solving
Now we look closely at the steady state iterative equation :
1
1
1
) 0 (
) (
) , 0 , ) 0 ( (
+
c
c
=
j
j
j
x
T x
T x J
In (5), is not known explicitly but is determined
through Backward Euler Integration. Thus:
1 1
) ( ) , 0 , ) 0 ( (
= +
j j
T x T x
(5)
can not be explicitly written either and would have to be calculated using a
form of numerical integration
Equation solving (2)
One way of solving for the Jacobian is to differentiate both sides of (4)
with respect to
+
J
Where and are matrices derived from (4). can then be computed by
repeatedly solving (6) starting from the initial condition
f
C
p
C
+
J
Computation load analysis of this method:
Matrix-Matrix multiplication on the right side and LU-factorization on the left
side. If the vector has components, the computation work per time step is
of at least order
N
2
N
The Backward Euler Integration Method calculate y at step
n+1 implicitly by :
Backward Euler Integration
Consider the problem:
) ( ) (
2
1 1
h O y hf y y
n n n
+ =
+ +
In the solver we separate the period T into a number of time
steps h. Provided a solution at time t1, x(t1), we then can
compute the solution x(t1+h) at time t1+h, by applying the
backward Euler integration to (5) over the time interval h:
(5)
(7)
Application of the GMRES
Computation load analysis of this method:
Now we consider solving the problem using GMRES method. By using
GMRES method, we dont need to provide explicitly. Instead, we only need
to provide . We apply the procedure (6) to a vector
+
J
Matrix-vector multiplication on the right side and LU-factorization on the left
side. Typically the number of iterations required by GMRES to achieve a
sufficiently low relative error is substantially smaller than problem size. So by
using GMRES, the numerical efficiency is improved.
The eddy currents in rotor are :
Post-processing:
rotor loss calculation
After we achieve the steady-state solution at time 0, the entire response
x(t) can be calculated by integrating (5) for one period. We perform a
Discrete Fourier Transform of x(t) and calculate the eddy current rotor
losses for each harmonic in each element.
}
}
V =
=
Se
i i i s
Se
i e
s
ei
dS A j l
dS J
l
P
2
2
) ( ) (
) (
e | e e o
e
o
u V
c
c
= o o
t
A
J