Research Square (Research Square),
Journal Year:
2024,
Volume and Issue:
unknown
Published: Oct. 21, 2024
Abstract
Predicting
the
price
and
volatility
of
international
crude
oil
futures
is
a
complex
task.
This
paper
presents
novel
hybrid
prediction
model,
VMD-BiTCN-BiGRU-Attention,
which
integrates
variational
mode
decomposition
(VMD)
advanced
deep
learning
techniques
to
forecast
nonlinear,
non-stationary,
time-varying
characteristics
sequences.
Initially,
sequence
decomposed
into
multiple
modes
using
VMD,
enabling
capture
different
frequency
components.
Each
independently
predicted
bidirectional
time
convolutional
network
(BiTCN),
captures
temporal
information
enhances
long-term
dependencies
through
dilated
convolution.
Subsequently,
gated
recurrent
unit
(BiGRU)
models
more
effectively,
while
an
attention
mechanism
adjusts
weights
BiGRU
outputs
emphasize
critical
information.
The
model’s
predictions
are
optimized
with
Adam
algorithm.
Empirical
results
demonstrate
that
model
adept
at
forecasting
non-stationary
nonlinear
prices.
Furthermore,
Diebold-Mariano
(DM)
test
confirms
this
surpasses
15
other
regarding
accuracy
performance,
achieving
optimal
key
metrics:
R²
=
0.9953,
RMSE
1.4417,
MAE
0.7973,
MAPE
1.5213%.
These
findings
underscore
its
potential
for
enhancing
prediction.
Sensors,
Journal Year:
2024,
Volume and Issue:
24(23), P. 7492 - 7492
Published: Nov. 24, 2024
Addressing
the
issues
of
low
prediction
accuracy
and
poor
interpretability
in
traditional
matte
grade
models,
which
rely
on
pre-smelting
input
assay
data
for
regression,
we
incorporate
process
sensors'
propose
a
temporal
network
based
Time
to
Vector
(Time2Vec)
convolutional
combined
with
multi-head
attention
(TCN-TMHA)
tackle
weak
characteristics
uncertain
periodic
information
copper
smelting
process.
Firstly,
employed
maximum
coefficient
(MIC)
criterion
select
strongly
correlated
grade.
Secondly,
used
Time2Vec
module
extract
from
variables,
incorporates
time
series
processing
directly
into
model.
Finally,
implemented
TCN-TMHA
specific
weighting
mechanisms
assign
weights
features
prioritize
relevant
key
step
features.
Experimental
results
indicate
that
proposed
model
yields
more
accurate
predictions
content,
determination
(
