Measurement Science and Technology,
Journal Year:
2024,
Volume and Issue:
36(1), P. 016233 - 016233
Published: Dec. 11, 2024
Abstract
In
recent
years,
thermal
runaway
during
charging
of
lithium-ion
batteries
has
become
a
critical
issue.
This
problem
emerged
as
significant
barrier
to
the
development
power
for
electric
vehicles
(EVs).
paper
addresses
this
challenge
from
data-driven
perspective
by
proposing
temperature
prediction
model
EV
batteries.
The
leverages
both
long
short-term
memory
and
Transformer
algorithms
account
time-series
characteristics
charging.
data
under
varying
capacities
ambient
temperatures
are
extracted
using
Newman–Tiedemann–Gaines–Kim
batteries,
which
is
then
used
optimize
accuracy
hybrid
algorithm
through
training.
Additionally,
real-world
collected
further
validate
model.
Experimental
results
demonstrate
that
proposed
achieves
superior
compared
single
models
convolutional
neural
network
models.
Based
on
model,
residual-based
early
warning
method
incorporating
sliding
window
approach
proposed.
experimental
findings
indicate
when
residual
predicted
EVs
exceeds
threshold,
preemptive
termination
effectively
prevents
runaway.
Natural hazards and earth system sciences,
Journal Year:
2025,
Volume and Issue:
25(1), P. 335 - 351
Published: Jan. 23, 2025
Abstract.
Deep-learning-based
surrogate
models
represent
a
powerful
alternative
to
numerical
for
speeding
up
flood
mapping
while
preserving
accuracy.
In
particular,
solutions
based
on
hydraulic-based
graph
neural
networks
(SWE-GNNs)
enable
transferability
domains
not
used
training
and
allow
the
inclusion
of
physical
constraints.
However,
these
are
limited
due
four
main
aspects.
First,
they
cannot
model
rapid
differences
in
flow
propagation
speeds;
secondly,
can
face
instabilities
during
when
using
large
number
layers,
needed
effective
modelling;
third,
accommodate
time-varying
boundary
conditions;
fourth,
require
initial
conditions
from
solver.
To
address
issues,
we
propose
multi-scale
network
(mSWE-GNN)
that
at
different
resolutions
speeds.
We
include
via
ghost
cells,
which
enforce
solution
domain’s
drop
need
solver
conditions.
improve
generalization
over
unseen
meshes
reduce
data
demand,
use
invariance
principles
make
inputs
independent
coordinates'
rotations.
Numerical
results
applied
dike-breach
floods
show
predicts
full
spatio-temporal
simulation
irregular
meshes,
topographies,
conditions,
with
mean
absolute
errors
time
0.05
m
water
depths
0.003
m2
s−1
unit
discharges.
further
corroborate
mSWE-GNN
realistic
case
study
Netherlands
capabilities
only
one
fine-tuning
sample,
0.12
depth,
critical
success
index
depth
threshold
87.68
%,
speed-ups
700
times.
Overall,
approach
opens
several
avenues
probabilistic
analyses
configurations
scenarios.
Abstract.
Floods
are
among
the
most
hazardous
natural
disasters
worldwide.
Accurate
and
rapid
flood
predictions
critical
for
effective
early
warning
systems
management
strategies.
The
high
computational
cost
of
hydrodynamic
models
often
limits
their
application
in
real-time
simulations.
Conversely,
data-driven
gaining
attention
due
to
efficiency.
In
this
study,
we
aim
at
assessing
effectiveness
transformer-based
forecasting
spatiotemporal
evolution
fluvial
floods
real-time.
To
end,
model
FloodSformer
(FS)
has
been
adapted
predict
river
inundations
with
negligible
time.
FS
leverages
an
autoencoder
framework
analyze
reduce
dimensionality
spatial
information
input
water
depth
maps,
while
a
transformer
architecture
captures
correlations
between
inundation
maps
inflow
discharges
using
cross-attention
mechanism.
trained
can
long-lasting
events
autoregressive
procedure.
model's
performance
was
evaluated
two
case
studies:
urban
flash
scenario
laboratory
scale
along
segment
Po
River
(Italy).
Datasets
were
numerically
generated
two-dimensional
model.
Special
given
analyzing
how
accuracy
is
influenced
by
type
severity
used
create
training
dataset.
results
show
that
prediction
errors
generally
align
uncertainty
observed
physically
based
models,
larger
more
diverse
datasets
help
improving
accuracy.
Additionally,
time
procedure
compared
physical
simulated
event.
also
benchmarked
against
state-of-the-art
convolutional
neural
network
showed
better
These
findings
highlight
potential
enhancing
responsiveness,
contributing
improve
resilience.
