Page 375 - Contributed Paper Session (CPS) - Volume 2
P. 375
CPS1889 Subanar
Time series forecasting model by singular
spectrum analysis and neural network
1,2
Subanar
1 Department of Mathematics, Universitas Gadjah Mada, Indonesia
2 Study Program of Statistics, Universitas Sebelas Maret, Indonesia
Abstract
The aim of this study is to present the decomposition based method for
enhancing forecasting accuracy. A hybrid neural network (NN) model is
established based on several aggregate components obtained by SSA
decomposition. We work with an hourly electricity load data series to show
that the proposed method can handle the complex pattern in the data. The
result shows that the proposed method outperforms SSA with linear recurrent
formula (LRF) and single NN model.
Keywords
SSA; hybrid; NN; electricity; complex
1. Introduction
Neural network (NN) is a flexible modeling tool due to its capability in
approximating any type of relationship in data without requiring specifics
assumptions. There were many successful examples of NN implementation in
solving forecasting problem (see (Adhikari & Agrawal, 2012; Di Persio &
Honchar, 2016; Khashei & Bijari, 2010; Yolcu, Egrioglu, & Aladag, 2013; G. P.
Zhang & Qi, 2005).
Various NN based methods are developed to improve forecasting
accuracy. Researchers began to pay attention to improve NN performance by
employing a decomposition-based method. (see Huang & Wang, 2018; Wu,
Chau, & Li, 2009; Wu & Chau, 2011). Wu et al. (2009) used SSA and DWT as a
data processing tool to improve mapping relationship between input and
output of the NN model. Both SSA and discrete wavelet transform (DWT) were
implemented to decompose original series which the results are then used as
a basis for generating the inputs of the NN model. Meanwhile, Wu & Chau
(2011) showed that SSA as data pre-processing can eliminate the lag effect in
the NN model. Both Wu et al. (2009) and Wu & Chau (2011) employed the
decomposition-based method in the field of rainfall-runoff modeling where
the architecture of NN has more than one variable input.
Recently, Huang & Wang (2018) discussed a hybrid DWT and stochastic
recurrent wavelet neural network (SRWNN). DWT is used to decompose the
original series into several subseries and then modeling each subseries by
364 | I S I W S C 2 0 1 9