IEEE
PEDS
2005
GAtrained
GRNN
for
Intelligent
Ultra
Fast
Charger
for
NiC
atteries
Panom
Petchjatupom,
Noppadol
Khaehintung,
PhinyoWicheanchote
Khamron
Sunat
and
Wiwat
Kiranon
Faculty
of
Engineering,
Mahanakom
University
of
Technology
Test
Engineer
Department
Bangkok,
Thailand.
SanminaSCISystems
co.
ltd.
Thailand
Email:
{Panom,Noppadol,
Khamron,Wiwat
}
mut.ac.th
Email:phinyo.wgsanminasci.com
Abstract
This
paper
presents
an
intelligent
genetic
algorithm
successfully
applied
to
NiCd
batteries
[6].
With
less
GA)
technique
for
training
of
a
neural
network
controller
to
computational
burden,
it
has
beenfound
that
its
data
selectionachieve
a
compactnetworkand
to
decrease
battery
charging
performance
is
greater
than
some
techniques,
for
example
time.
An
ultra
fast
charging
device
for
NickelCadmium
NiCd
ANFIS.
Moreover,
the
computational
afford
reduction
of
batteries
is
designed
through
the
generalized
regression
neural
GRNN
has
beenimproved
by
further
in
[7].
Nonetheless,
the
network
GRNN)
and
implemented
with
the
data
selection
of
construction
GRNN
still
was
ot
thoroughly
MATLAB SIMULINK
fortesting
and
operating
on
real
system.
considered.
The
inputoutput
data
for
training
neural
networks
were
collected
from
GA.
The
suitable
data
were
selected
to
establish
Genetic
algorithms
are
probabilistic
search
techniques
that
GRNN
comprising
only
13
processing
elements.
Each
node
of
emulate
the
mechanics
of
evolution
[8].
They
are
capable
of
RBFs
is
an
extendable
support
functionthat
overcomes
the
globally
exploringa
solutionspace,
pursuing
potentially
drawback
in
the
existing
compact
support
radial
basis
functions
fruitful
paths
while
also
examining
additional
random
points
to
CSRBF).
Experiments
with
real
time
implementation
clearly
reduce
the
likelihood
of
settling
for
a
local
optimum.
The
show
that
the
proposed
technique
notonlyrequires
less
neuralapplied
GAtrained
RBF
network
for
maximum
power
point
processing
units
but
also
yields
less
MSE
than
RBF
technique.
tracking
in
photovoltaic
applications
was
presented
in
[9].
Keywordscomponent;
genetic
algorithm;
fast
charging;
the
This
paper
proposes
an
efficient
approach
for
battery
fast
generalized
regression
neuralnetwork;
radial
basis
functions;
chargingusing
GAtrained
GRNN
controller.
Its
main
advantages
over
other
algorithms
are
the
simple
topological
I.
INTRODUCTION
structure,
precision
in
nonlinear
dynamics
approximation,
and
fast
learning
capability.
The
remainder
of
the
paper
is
In
present
the
secondary
or
rechargeable
batterles
have
organized
as
follows.
Following
this
a
system
configuration
of
been
massivelyused
in
variety
of
electrical
appliances
with
th
prpoe
fs hrigmtoisadeednScinI.
wide
range
of
size
and
capacity.
These
batteries
can
be
the
proposed
fastcharging
method
is
addressed
in
Section
II
rechrge
afer
ingexhustd
b
pasingcurentthrugh
The
GAtrained
GRNN
controller
forthe
fastcharging
method
recharged
after
being
exhausted
by
passing
current
through
is
reviewed
in
Section
III
In
Section
IV,
the
development
of
them
in
the
opposite
direction
tothat
of
discharge
current
[1].
fastcharging
is
discussed
and
experimental
results
with
The
widely
used
secondary
batteries
are
Nickel
Cadmium
Ni
emphasis
on
performance
of
the
proposed
GRNN
controller
are
Cd ,Nickel
Metal
Hydride
NiMH)
andLithium
Ion
LiIon ,
given.
Finally,
conclusion
is
drawn
in
Section
V.
In
principal,
charging
of
a
battery
may
be
a
straightforward
task.
In
practical,
however,
it
could
involvecomplicated
control
algorithm,
especially
when
high
power
demand
needs
II.
PROPOSED
SYSTEM
to
be
supported
by
highperformance
while
cost
effective
battery
charger.
A.
UltraFastChargingMethod
When
charging
at
high
rate,
temperature
inside
thebattery
Recently,
fastcharging
algorithms
have
beenproposed
in
could
change
very
rapidly,
as
shown
by
typicalprofiles
of
many
places.
Their
common
purpose
iS
to
speed
up
thebatterybattery
voltage
and
temperature
during
a
charging
period
in
recharging
time.
In
[2],
forinstance,
the
fastcharging
Fig.
1.
lt s
and
tosgicatly
decring
efic
in
alorth
bae
upo
th
reato
beweX
otg
n
Fig.
1.
This
can
lead
to
significantly
decreasing
efficiency
of
temperased
upo
tatiro
e
tw t
voltage
ant
the
charging
process,
and,
hence,
the
degradation
or
even
teprtr
of
tebteyhsenpretd.
Moepacia
serious
damage
of
the
battery.
Therefore,
this
effect
must
be
work
in
[3]
employs
a
nonlinear
controller
to
cope
with
the
sevalusat
eoforeap
atioro.
Tfastcha
i
procss.
problem.
Some
intelligent
control
techniques,
such
as
Neuro
s
p
Fuzzy
[4]
and
Adaptive
NeuroFuzzy
Inference
System
The
maximum
rate
of
charging
currentfor
NiCd
batteries
ANFIS)
[5]
have
also
been
applied
for
fastbattery
chargers.
is
8C
[5],
where
C
is
capacity
in
AmpHour
unit
of
the
battery.
These
algorithms,
though
very
sophisticated
and
adaptable
to
In
general,
the
battery
will
be
charged
at
different
rates
with
the
environment,
require
relatively
high
performance
processor
different
durations,
accordingly
to
internal
pressure
and
and
may
not
be
appropriate
for
certain
applications
where
cost
temperature
o
the
battery.
When
the
battery
reaches
its
full
iS
a
prime
concern.
charge
level,
the
temperature
gradient
will
increase
An
utra
astcharingtechiqueusig
th
Geeralzed
significantly,
and
the
charging
rate
should
be
slowed
down
in
RegressionNeural
Network
GRNN)
has
been
introduced
and
ore oaodteoecag
tt Teeoe
h
atr
0780392965/05/ 20.00
©
2005
IEEE
1
194
Slope
=Temperature
gradient
/
Time
X
6D
Figure
1.
Typical
profiles
of
battery
temperatureand
voltage
during
fast
charging
process.
Figure
2.
Inputoutput
relationship
of
a
GRNNbased
battery
charger
temperature
T)
and
temperaturegradient
dTldt)
can
be
used
in
the
determination
of
the
rate
of
charging
current
C, .
As
an
example,
561
points
data
set
representing
the
inputoutput
relationship
during
a
battery
fastcharging
process
using
GRNN
algorithm
is
shown
in
Fig.
2.
B.
Circuit
Configuration
The
circuit
configuration
of
the
proposed
battery
charger
is
shown
in
Fig.
3.
It
comprises
of
a
voltage
control
current
source
circuit
and
battery
temperature
and
voltage
detector
circuits.
The
charging
current
command
is
sent
via
the
digital
to
analog
converter
of
xPC
Targets
Computer
[12],
and
thevoltage
is
applied
to
the
current
source
circuit
viathe
non
Figure
3.
Circuit
configurationof
the
proposed
ultra
fast
charger.
inverting
input
of
opamp
1/4
LM324).
During
charging,
the
battery
temperature,
T,
is
measured
by
a
temperature
sensor
LM335
and
is
fed
to
a
digital
computer
via
analog
to
digital
converter
xPC
Targetscomputer).
The
temperature
gradient,
iRBFtljx1jj
dTldt,
is,
then,
computed
using
a
unit
delay.
Both
temperature
C

II lII
and
temperature
gradient
are
us
as
theparameters
for
C, x)
n
1
evaluating
the
rate
of
charging
current.
The
battery
voltage
is
ER.BF Ix,x.jj
also
sent
to
digital
computer
in
order
to
checkwhether
the
charging
process
should
be
terminated.
where
x
E
Rn
is
input
vector
of
dimension
n,
RBF .)
is
radial
III.
GAGRNN
FoR
FASTCHARGING
basisfunction,
denotes
the
Euclidean
norm,
Y
arethe
weights
or
parameters
of
network,
xi
e
Rn
are
known
asthe
A.
The
GRNN
centers
of
RBF .),
and
n
is
the
number
of
centers.
Since
the
Let
us
consider
Fig.
2.
From
our
observation,
when
3D
network
size
directly
depends
on
the
number
of
training
data
surface
of
the
T,
dTldt
and
C,
was
plotted,
it
should
be
[13],
t
therefore,
must
be
carefully
selected
in
order
to
get
a
approximated
by
a
smooth
surface
or
the
data
regression
compact
network,
and,
at
the
same
time,
it
can
represent
the
method.
Accordingly,
the
generalized
regression
neural
net
surface
appropriately.
work
GRNN
comprising
of
threelayers
is
selected
to
perform
this
task.
The
firstlayer
comprises
of
input
nodes.
The
second
B.
The
selection
process
of
training
data
layer
comprises
of
neurons
each
of
which
is
represented
by
a
Since
GRNN
behaves
like
a
functional
approximation
or
a
basisfunction.
The
output
ofnetwork
according
to
the
input
x;
nonlinear
regression,
its
characteristic
highly
depends
on
the
ya,
depends
on
the
weighted
combination
of
thefunctional
characteristic
of
training
data.
Once
the
training
model
is
distance
between
the
input
data
and
thecenter
of
the
basis
completed,
in
order
to
evaluate
the
goodness
the
selected
datafunction
in
the
hidden
layer.
the
charging
current
C,
can
be
set,
the
MeanSquare
Error
MSE)
of
the
validation
set
must
be
computedby
evaluated.
MSE
is
defined
by
1195
Ansioglnput
lnl
L;nOuti
eural
N~t rk
CRgte_Sl
hg
PClUII UpUI
l
~~V
Temp
to
TanpS
Cff
TSi
n
C
PCL S12PG
dTldt
Figure
4.
The
GRNN
controller
implementation
block
diagram.
stringlI
I
500
100
1
l
F4
MSEE
C,d n C
C, n 2
2)
string
521
13
21
.
W
where
C,{n
denotes
the
desiredoutput
of
thevalidation
set,
string
I
I
201
.......
C n
denotes
the
actual
output
of
GRNN
model,
and
N
is
the
Figure
5.
Exampleof
string
for
selectedinputoutput
datapoint
for
GRNN
number
of
datapoints
taken
for
model
validation.
is
shown
in
Fig.
6.
It
is
clearly
seen
that
after
320
generations
C.
The
GAtrained
GRNNfor
Data
Selection
the
GA
gives
the
best
result
of
fitness
value,
4.37x
103.
In
GA,
the
set
of
variables
for
a
given
problem
is
encoded
Inthe
experiment,
we
appliedthe
proposed
data
selection
into
a
string
population
which
is
analogous
to
a
chromosome
criteria
tothe
charger
of
aPanasonic
NiCd
battery
[14].
Table
in
nature.
Each
string
contains
a
number
of
alleles,
and
each
H
shows
a
set
of
data
having
6
out
of
561.
Itis
a
data
set
thatfeature
of
the
system
located
at
a
specific
position
in
the
string
the
GAtrained
GRNN
with
13
processing
units
yields
a
very
is
called
gene.
Each
string,
therefore,contains
possible
solution
low
MSE
as
shown
in
Table
III.
If
the
whole
data
set
was
used,
to
the
problem.
The
GA
performs
with
a
set
of
binarycoded
the
number
of
processing
units
of
the
trained
GRNN
will
be
2x
strings
of
set
parameter,representing
thesuitable
data
for
561+1=1,123
units.
GAtrained
GRNN
of
this
size
has
a
battery
fastcharging
thatwill
be
trainedfor
the
GRNN
These
comparableperformance
butpossessesahigher
cost
than
GA
data
may
be
the
optimum
solution
to
the
problem.
The
trained
GRNN
of
thesmaller
size.
searching
of
selected
inputoutput
data
point
i
start
from
a
randomly
generated
set
of
population.
The
center
of
RBF
xi
can
B.
Neural
Network
setup
be
decoded
as
The
simulations
were
conducted
to
evaluatethe
xi
=
inputoutput
data
point
pop
3
performance
of
GRNN
controllerfor
fastcharging
using
xj=mput outputdatapoultWop
MATLAB/Neural
Network
Toolbox.
A
Simulink
model
ofwhere
pop
is
the
population.
The
fitness
*
of
each
GRNN
for
real
time
control
of
battery
charger
as
shown
in
Fig.
chromosome
is
evaluated
as1
was
implemented.
The
training
data
set
taken
from
Table
II
is
chromosome
is
evaluated
asr
applied
to
set
up
the
GAtrained
GRNN
networks.
After
f
=
MSE
4
training
process,
the
surface
of
output
Ct
and
input
battery
temperature
T
with
associatedbattery
temperature
gradient
Theoptimum
solution
can
be
obtained
by
minimizing
the
dTldt
was
obtained,
shown
in
Fig.
2.
fitness
function.
The
lower
fitness
value
string
at
each
searching
step
will
be
chosen
to
be
the
parent
of
the
next
C.
Charging
Results
generation,
while
the
higher
fitness
one
will
be
rejected.
In
In
order
to
determine
the
characteristic
ofnormal
charging
creatingthe
new
offspring,theselected
parentstrings
undergo
this
battery,
the
NiCd
battery
was
tested
intensively
at
charg
reproduction
process,
such
as
crossover
and
mutation
as
ing
current
rate
at
4C
and
8C
at
the
environment
temperature
of
described
in
[8].
The
process
continues
updating
newer
and30
OC.
Inthese
experiments,
we
investigate
the
controller
fitter
chromosome,
until
a
predefinedstopping
condition
is
performance
by
comparing
MSE
of
GAtrained
GRNN
with
satisfied.
MSE
of
ANFIS
[5].
Itis
noticed
that
the
GAtrained
GRNN
In
our
work,
the
number
of
alleles
is
determined
from
the
network
yields
not
only
alower
number
of
processing
units
butinputoutputdata
points
as
same
as
that
used
in
[11].
It
is
a
data
also
a
lower
MSE
than
ANFIS
[5]
controller
and
GRNN
[7]
set
that
the
trained
GRNN
as
shown
in
Fig.
5.
see
Table
III .
Fig.
7
depicts
the
results
of
battery
temperature
and
voltage
during
charging
with
constant
current
C
It
indicates
that
charging
processshould
be
stopped
with
in
1,400
IV.
EXPERIMENTAL
RESULTS
seconds,
becauseovercharge
can
be
detected
by
positive
dTldt
and
negative
dV/dt
at
this
point.Fig.
8
shows
the
result
of
A.
Data
Selection
Results
constant
current
charging
at
8C.
It
indicates
that
charging
This
experiment
aims
to
evaluate
our
training
procedure.
process
should
be
stopped
within
415
seconds.
Although
the
The
population
size
used
in
each
trial
is
20.
Each
system
is
results
from
Fig.
7
and
Fig.
8
indicatethat
charging
at
8C
is
randomly
assigned
between
1
and
561
represented
the
order
of
faster
than
charging
4C,
the
former
yields
a
hightemperature
the
inputoutput
data
points
taken
from
[
11].
Table
I
shows
the
gradient
which
may
cause
batteries
deterioration.Fig.
9
depicts
parameters
used
in
the
GA
process.
The
result
of
GA
selected
1196
TABLE
1.
GENETIc
PARAMErER
USED
2
r
i I ~
50
Crossover
0.8
45
FMutation
0.1
T7
..
....
i
40
I
35
TABLE
II.
INPuT ouTPUT
DATA
FOR
Ni Cd
BAT rERY
CHARGER
30
No.
Inpu
Output
2
__
~~
~~~T
T/dt
Ct
_
_i
4
0.2
8
20
.9
8
I



15
___3
38
0.9
8
10
440
0.58
43
0.4
2
Tempaatie
I I I
I
5
6
5
0
L
5
0
a
0100150200
250
300350
400
45i
TABLE
III
Totne
sen
COMPARISONBETWEEN
ANFIS
RBF
AND
GRNN
BATTERY
CHARGER
Figure
8.
The
battery
voltage
and
temperature
in
a
Ni Cd
battery
ANFIS
[5
GRNN
71
GAtrained
GRNN
1
2VI600mH
at
a
8C
charging
rate.
Number
of
Layers
6
33
Number
of
Nodes
27
1
3
1350
MSE
0.1321
0.0905
0.0046
4.5
4
40
40
Best:
0.
0043719
Mexi:
0.54306
10
3

~
S
0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~.5
Voltage
25
..........~~~~~0.
10
0
1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2
00
400
50
600700

2.0
>
..........Tim
iec
Fiure . Th
ate...oltae .c
ren.andtempratue.ina.NiCd.btte.
__________________
I........I.....of
RBF
cotolr
o
5t br
100200300400500600
700000900
loot
controller
dCureaeh
hrigcretwe
epratur
2.0
30
increases.
Fr~~~~~~~~~~~~~~~~~~~~~~~~~om
hes reslts
bydetecting
the
oItive
dI
Voltage
~
~~~
40
and
negative
dV/dt
the
Ni Cd
battery
was
reached
the
ull~~~~
Tera
............I40
shourt
ie
wi
h
oubtt
oery
v
limtae,
retad
temperature.
i
iC
atr
construcion
of
highly
compatrle
GorN
oshrgfatharin
200
400
6
400
00
1000
12
8014900
newr
ee
esta
AFSadRF.Frhror tera
eneration
~
sythem
reslso
valu
enation
ftepopsdcitireral
wasimplementedo
Figreg.uTe6
atryoltgresandftmratireingfreealtNiCl
atesn
oulnfirmedtel
reslts.rtdi
his.woul
lea
stoarlow
costsstemt
l.2VI600H
t
4
chargingratetrim
lem
entationei
the
future.
gcfen
hnteprtr
2 050
inrese.
ro
thsereuls bydeectngth
ps1197dId
REFERENCES
[7]
P.
Petchjatuporn,
P.
Wicheanchote,
N.Khaehintung,K.
Sunat,
W.
Kiranonand
S.
Chiewchanwattana,
Intelligent
ultra
fast
charger
ofNi
[1]
D.Linden,
HandbookofBatteries,
McGraw
Hill
Inc.,1995.
Cd
batteries ,
Proc
of
ISCAS05
Conf.,
pp.
51625165,
2005.
[2]
M.
Gonzalez,
F.
.J.
Ferrero,
J.
C.
Anton
and
M.
A.
Perez,
[8]
D.
E.
Goldberg,
Genetic
Algorithm
on
Search
Optimization
and
Considerations
to
improve
thepractical
design
of
universal
and
full
Machine
Learning,
Reading,
MA
AddisionWesley,1989
effective
NiCd/NiMH
battery
fastchargers,
Proc.
of
14th
APEC
99
[9]
L.
Zhang,
B.
Yunfeiand
A.
AlAmoudi,
GARBF
neural
network
Conf.,
Vol
l
pp.
167173.,1999.
[]L
hn B
ufiad
lAod ARFnua
ewr
based
maximum
power
point
tracking
for
gridconnected
photovoltaic
[3]
J.
C.Viera,
M.
Gonzalez,C.Blanco,
and
J.
C.
Alvarez,
Application
systems,
Proc
ofInt.
PEDS
Conf,
pp.
18
23,2002.
range
of
fastcharging
in
NiCd
batteries,
Proc.
of
24th
INTELEC
Conf,
[10]
J.
Binfet
and
B.
M.
Wilamowski,
Microprocessorimplementation
of
pp.
259264,
2002.
fuzzysystems
and
neural
networks,
Proc.
of
IJCNN
01
Conf:,Vol.
1
[4]
Z.
Ullah,B.
Burford,
and
S.
Dillip,
Fast
intelligent
battery
charging:
pp.234239
2001.
neuralfuzzy
approach,
IEEE
Aerospace
and
Electronic
Systems
Maga
[11]
http://www.research.4t.com.
zine,
Vol.
1
pp.
26
34.,
June,
1996,
[5]
A.
Khosla,
S.
Kumar,
andK.K.Aggarwal,
Fuzzy
controllerforrapid
[12]
http://www.mathworks.com.
nickelcadmium
batteries
charger
through
adaptive
neurofuzzy
[13]
S.
Chen,
C.
F.
N.
Covan,
and
P.
M.
Grant,
Orthogonal
least
squareinference
system
ANFIS)
architecture,
Proc.
Of
NAFIPS
2003Conf,
learning
algorithm
forradial
basis
function
networks,
IEEE
Trans.
pp.
540
544,
2003.
Neural
Networks,
Vol.2,
No 2
March,
1991,pp.302309.
[6]P.
Petchjatupom,
P.
Wicheanchote,
N.Khaehintung,
K.
Sunat,
P.
[14]
Panasonic,
Nickel
Cadmium
Handbook,
August,
1998,
p.51.
Sookavatanaand
W.
Kiranon,
Data
Selection
if
a
Compact
GRNN
for
NiCd
Batteries
fast
charging ,Proc.
of
TENCON
2004
Conf,
pp.
213
216,2004.
1198