Page 376 - Contributed Paper Session (CPS) - Volume 2

CPS1889 Subanar
SRWNN. The forecast value was obtained by aggregating process for all
subseries.
Inspired by those previous study, we proposed a hybrid NN method by
combining NN with singular spectrum analysis (SSA). In the case of forecasting
value, it is obtained from an ensemble of neural network models for several
components that are determined based on SSA decomposition.
The aim of this study is to show that the proposed hybrid approach
improves the forecasting accuracy comparing with results obtained from the
SSA with linear recurrent formula (LRF) and single NN. We select an hourly
electricity load series as the case study to lead that the method is able to solve
such a complex series forecasting problem.

2. Methodology
The methods we use in this work are briefly presented below.

2.1. SSA decomposition
As declared in many references, i.e. Elsner (2002), Golyandina & Zhigljavsky
(2013), Golyandina, Nekrutkin, & Zhigljavsky (2001), and Hassani (2007), SSA
is a decomposition tool that consists of four steps, namely embedding,
singular value decomposition (SVD), grouping and diagonal averaging. The
detail discussion for the four steps can be found in Golyandina & Zhigljavsky
(2013). The two important things in SSA that we need to pay attention are the
window length and the grouping selection. In selecting the window length we
can take a value that is proportional to the seasonal period but no more than
a half of the sample size (Golyandina, 2010). Steps in SSA decomposition are
displayed in Figure 1.

Figure 1: Steps in SSA decomposition

365 | I S I W S C 2 0 1 9