A Forecasting Method Based on Extrema Mean
Empirical Mode Decomposition and Wavelet Neural
Network
JianJia Pan*, Xianwei Zheng, Lina Yang, Yulong Wang, Haoliang Yuan, Yuan Yan Tang
Departmen t of Computer and Information Science,
University of Macau, Macau
*Email: jianjiapan@gm a il. com
Abstract—Time series forecasting i s a widely and important
research area in signal processing and machine learning. With
the development of the artificial intelligence (AI), more and more
AI technologies are used in time series forecasting. Multi-layer
network structure has been widely used for forecasting problems.
In this paper, based on a data-driven and adap t ive method,
extrema mean emp irical mode decomposition, we proposed
a decomposition-forecasting-ensemble approach to time series
forecasting. Experimental result shows the prediction result by
proposed models are better than original signal and EMD based
models.
Index Terms—forecasting, empirical mode decomposition,
wavelet neural network
I. INTRODUCTION
Time series is a set of observations taken sequentially in
time. An important characteristic of time series is stationary.
In gen eral, stationary time series has average properties and
characteristics that do not change with time. Time series anal-
ysis is to measure and estimate the properties or ba sic features
of the potential stoc hastic process from the information given
by the observed series [11]. There have been a number of
techniques developed in statistics to mo del time series.
Widely used established time series m odels include: (1)
linear models, e.g., m oving average, exponential smoothing
and the autoregressive integrated moving average (ARIMA);
(2) nonlinear mod els, for examples, neural network models,
fuzzy system models etc.; and (3) the combination of linear
and nonlinear models.
The three-layer network structure can carry out almost any
arbitrary input-output m apping to describe the relationship in
efficient time series forecasting [8]. When trained on examples
of observation data, the network s can learn th e characteristic
features ’hidden’ in the examp le s of the collected data and
even generalize the knowledge learnt.
Recently, Empirica l mode deco mposition, which decom-
poses signal based on the local characteristics scale and present
time-frequency-energy distribution of da ta , has b een proposed
and widely used in signal processing[1]. Due to the data-driven
property, EMD has been found many successful applications in
variant areas, such as biological, medical sciences, astronomy,
engineer ing and others. EMD has been used for foreca sting
problems, such as a decomposition-and-ensemble ANN learn-
ing paradigm [7], forecasting task for air traffic by E MD and
ANN[9], EEMD and SVM for the stock price time series[10]
In this paper, based on wavelet neural network (WNN) and
an improved EMD (Extrema Mean Empirical mode decom-
position), w e build a EME MD-WNN e nsemble forecasting
method. The EMD method decompose time series into its
basic components, and more accura te forecasts can be obtained
from basic components, lastly the foretasted components are
combined to achieve the lastly forecast result. Experiments
show the prediction result by EMEMD based models are better
than original signal and EMD based models.
II. WAVELET NEURAL NETWORK (WNN)
In wavelet theory, wavelets are a family of f unctions gener-
ated from one function Ψ (x) (called mother wavelet) by the
operation of dilation and translation as follows:
Ψ = {Ψ
i
= |a
i
|
−1/2
Ψ(
x − b
i
a
i
) : a
i
, b
i
∈ R
n
, i ∈ Z}
x = (x
1
, x
2
, · · · , x
n
),
a
i
= (a
i1
, a
i2
, · · · , a
in
),
b
i
= (b
i1
, b
i2
, · · · , b
in
)
(1)
where x is th e input vector, a
i
is the scale parameter and
b
i
is translation parameter. The moth e r wavelet is orthogonal
to all functions which are obtained by dilating (stretching) the
mother by a factor of 2
j
(2 to the j
th
power) and shifting by
multiples of 2
j
units.
In wavelet neural network (WNN), the output is given by
[12]:
f(x) =
M
X
i=1
ω
i
Ψ
i
(x) =
M
X
i=1
ω
i
|a
i
|
−1/2
Ψ(
x − b
i
a
i
) (2)
where Ψ
i
is the wavelet activation fun ction of i
th
neuron
of th e hidden layer. ω
i
is the weight connecting the i
th
neuron of the hidden layer to the output layer neuron. For
the n-dimensional input space, the multivariate wavelet basis
function can be calculated by the n single wavelet basis
functions as follows:978-1-4799-8322-3/15/$31.00
c
2015 IEEE