Journal of The Electrochemical Society,
Год журнала:
2024,
Номер
171(12), С. 120522 - 120522
Опубликована: Дек. 3, 2024
Lithium-ion
batteries
are
widely
used
in
new
energy
vehicles,
but
capacity
regeneration
and
fluctuations
during
aging
affect
the
accuracy
of
remaining
useful
life
(RUL)
prediction.
Complete
charge/discharge
data
often
unavailable
actual
usage.
To
address
these
issues,
this
paper
proposes
a
combined
model
for
RUL
prediction
using
partial
data.
Five
health
indicators
extracted
from
voltage
vs
time
curve
processed
variational
mode
decomposition
to
remove
outliers
noise,
improving
correlation
between
HIs
battery
capacity.
Spearman’s
coefficient
verifies
relationship
The
Kolmogorov-Arnold
Networks-Structured
State
Space
(KAN-S4)
is
then
developed,
capturing
spatial
correlations
long-term
degradation
patterns.
Experimental
validation
our
laboratory
University
Maryland's
CALCE
center
shows
that
KAN-S4
achieves
accurate
predictions,
even
under
complex
conditions
like
rapid
decline.
demonstrates
strong
robustness
generalization
across
varying
usage
scenarios.
Scientific Reports,
Год журнала:
2024,
Номер
14(1)
Опубликована: Окт. 9, 2024
Accurate
runoff
forecasting
is
of
great
significance
for
water
resource
allocation
flood
control
and
disaster
reduction.
However,
due
to
the
inherent
strong
randomness
sequences,
this
task
faces
significant
challenges.
To
address
challenge,
study
proposes
a
new
SMGformer
forecast
model.
The
model
integrates
Seasonal
Trend
decomposition
using
Loess
(STL),
Informer's
Encoder
layer,
Bidirectional
Gated
Recurrent
Unit
(BiGRU),
Multi-head
self-attention
(MHSA).
Firstly,
in
response
nonlinear
non-stationary
characteristics
sequence,
STL
used
extract
sequence's
trend,
period,
residual
terms,
multi-feature
set
based
on
'sequence-sequence'
constructed
as
input
model,
providing
foundation
subsequent
models
capture
evolution
runoff.
key
features
are
then
captured
layer.
Next,
BiGRU
layer
learn
temporal
information
these
features.
further
optimize
output
MHSA
mechanism
introduced
emphasize
impact
important
information.
Finally,
accurate
achieved
by
transforming
through
Fully
connected
verify
effectiveness
proposed
monthly
data
from
two
hydrological
stations
China
selected,
eight
compare
performance
results
show
that
compared
with
Informer
1th
step
MAE
decreases
42.2%
36.6%,
respectively;
RMSE
37.9%
43.6%
NSE
increases
0.936
0.975
0.487
0.837,
respectively.
In
addition,
KGE
at
3th
0.960
0.805,
both
which
can
maintain
above
0.8.
Therefore,
accurately
sequence
extend
effective
period
Lithium-ion
batteries
inevitably
experience
a
decline
in
State
of
Health
(SOH)
due
to
prolonged
use,
and
continued
operation
increases
safety
risks.
Therefore,
it
is
essential
develop
models
that
can
accurately
predict
SOH.
Cyclic
aging
experiments
are
initially
conducted
on
lithium
using
self-built
experimental
platform
collect
data
charging
voltage
temperature
aging.
A
multi-channel
temporal
convolutional
neural
network
employed
perform
feature
extraction
the
multi-source
data,
preserving
dependencies
features
over
time.
The
input
enables
capture
degradation
simultaneously,
enhancing
its
ability
characterize
at
any
moment.
SOH
prediction
then
carried
out
combination
Gated
Recurrent
Unit
(GRU)
Self-Attention
(SA)
mechanism.
SA
ensures
accuracy
by
calculating
weight
distribution
features,
allowing
GRU
focus
most
significant
aspects
data.
Finally,
model
proposed
this
study
compared
with
traditional
Long
Short-Term
Memory
model,
encoder
fusion
model.
results
show
although
similar
some
models,
still
lower
than
study.
Compared
other
mean
absolute
error
reduced
more
29%
average,
root
square
least
20%
average.
Energies,
Год журнала:
2025,
Номер
18(5), С. 1236 - 1236
Опубликована: Март 3, 2025
Estimating
the
state
of
health
lithium-ion
batteries
in
energy
storage
systems
is
a
key
step
their
subsequent
safety
monitoring
and
optimization
management.
This
study
proposes
method
for
estimating
based
on
feature
reconstruction
Transformer-GRU
parallel
architecture
to
solve
problems
noisy
data
poor
applicability
single
model
different
types
operating
conditions
batteries.
First,
incremental
capacity
curve
was
constructed
charging
data,
smoothed
using
Gaussian
filtering,
diverse
features
were
extracted
combination
with
voltage
curve.
Then,
this
used
CEEMDAN
algorithm
reconstruct
IC
features,
which
reduces
due
process
collection
processing.
Lastly,
cross-attention
mechanism
fuse
Transformer
GRU
neural
networks,
improve
its
ability
mine
time-dependent
global
estimation.
conducted
experiments
three
datasets
from
Oxford,
CALCE,
NASA.
The
results
show
that
RMSE
estimation
by
proposed
0.0071,
an
improvement
61.41%
accuracy
baseline
model.
