A Singular Spectrum Analysis Technique to Electricity Consumption Forecasting

Bisher M. Iqelan. Int. Journal of Engineering Research and Application www.ijera.com
ISSN : 2248-9622, Vol. 7, Issue 3, ( Part -3) March 2017, pp.92-100
www.ijera.com DOI: 10.9790/9622- 07030392100 92 | P a g e
S
A Singular Spectrum Analysis Technique to Electricity
Consumption Forecasting
Bisher M. Iqelan
(Department of Mathematics, the Islamic University of Gaza, Gaza Strip, Palestine)
ABSTRACT
Singular Spectrum Analysis (SSA) is a relatively new and powerful nonparametric tool for analyzing and
forecasting economic data. SSA is capable of decomposing the main time series into independent components
like trends, oscillatory manner and noise. This paper focuses on employing the performance of SSA approach to
the monthly electricity consumption of the Middle Province in Gaza StripPalestine. The forecasting results are
compared with the results of exponential smoothing state space (ETS) and ARIMA models. The three techniques
do similarly well in forecasting process. However, SSA outperforms the ETS and ARIMA techniques according
to forecasting error accuracy measures.
Keywords: SSA, ETS state space, ARIMA, Forecasting, Electricity consumption time series
I. INTRODUCTION
Nowadays, life is impossible without
electricity. Electricity provides homes and public
places with lights and heat. Without it, health,
education, finance, technology and other critical
services collapse. Therefore, electricity consumption
is without a doubt an important issue that has been
of interest over the past few years.
It is always more challenging to examine
densely populated places. For instance, Gaza Strip is
a very small Palestinian territory that is home to a
population of more than 2 million people. As a
result, it suffers from a chronic crisis in the
electricity supply for many years mainly because of
the Israeli siege. This paper focuses on the Middle
Province of Gaza Strip or known as Central Gaza
Strip which consists of the refugee camps of Bureij,
al-Maghazi, and al-Nussairat, and the city of Deir al-
Balah. It attempts to provide a coherent electricity
consumption forecasting for the Middle Province of
Gaza Strip using Singular Spectrum Analysis SSA.
Although it is considered as a new
nonparametric tool, there exists an extensive
literature on Singular Spectrum Analysis (SSA). To
begin with, [1] assert that the role of Singular
Spectrum Analysis (SSA) is to provide estimates of
the statistical dimension. SSA also aims to describe
the main physical phenomena reflected by the data.
It gives adaptive spectral filters connected with the
dominant oscillations of the system, and clarifies the
data’s noise characteristics. Moreover, [2] add that
SSA is both a linear analysis and prediction method.
It is superior to the other classical spectral methods
because of the data-adaptive character of the
eigenelements it is based on. It can also use concepts
from nonlinear dynamics. Adding to that, [3] asserts
that SSA is a powerful tool used in time series
analysis. It gives a much more accurate forecast
results than other methods when applied to many
practical problems. Its aim is to decompose original
series into small number of independent and
interpretable components like oscillatory
components, a structureless noise, and a slowly
varying trend. Moreover, [4] also illustrate that SSA
technique performs four steps. First, computing the
trajectory matrix. Second, constructing a matrix for
applying SVD. Third, grouping and corresponding to
splitting the matrices that were computed at the SVD
step. Finally, reconstructing the one-dimensional
series. In addition, [5] add that SSA has attained
successful application in different branches such as
meteorological, biomechanical, hydrological,
physical sciences, economics and others. The aim of
SSA is to look for nonlinear, non–stationary, and
intermittent or transient behavior in an observed
time series. Furthermore, [6] employ SSA to
decompose the original electricity price series into
trend, periodic and noisy components. This approach
is evaluated by analysing and forecasting the day
ahead electricity prices in the Australian and Spanish
electricity markets. The forecasting results assert the
dominance of the SSA approach compared with the
other forecasting techniques. In addition, [7] assert
that SSA is a powerful and well-developed tool of
time series analysis and forecasting. SSA can be
applied to a wide number of time series analysis
problems like exploratory analysis for data-mining
and parameter estimation in signal processing.
Adding to that [8] also assumes that SSA is a model-
free tool that can be applied to all types of series. It
comprises time series analysis tools, multivariate
statistics tools, dynamical systems and signal
processing tools. Finally, [9] investigate the use of
SSA in mid-term forecasting of the monthly
RESEARCH ARTICLE OPEN ACCESS

electricity consumption for the residential class in
Brazil. The results show that the SSA method with
graphical analysis of singular vectors presented the
more accurate forecasts.
The section above contains a summary of
electricity consumption with an application on the
Middle Province of Gaza Strip. Moreover, an
overview of previous literature is also added. The
remainder of the paper is organized as follows.
Section 2 discusses the methodology. Section 3
contains the application to real data. Section 4
consists of comparison between SSA approach and
other well-known models. Section 5 presents the
conclusion of this paper.
II. METHODOLOGY
Consider a real-valued nonzero time series
 1 2
, , ,N N
Y y y y  of length N . The main
point of SSA is to make a decomposition of the
original time series into the sum of independent
fundamental component parts such as slowly varying
trend, oscillatory component and noise.
The SSA technique consists of two stages:
decomposition and reconstruction and both of which
contain two distinct steps. A brief description of
each stage and a discussion on the methodology of
the SAA technique will be presented in the
following sections following [10] and [11].
2.1 Decomposition
This stage includes two steps: embedding
and Singular Value Decomposition.
2.1.1. Step 1: Embedding:
Embedding can be considered as a
procedure that takes a univariate time series
 1 2
, , ,N N
Y y y y  and makes it a
multivariate set of observations 1 2
, , , K
X X X
where the lagged vector , 1, 2, ,i
X i K  is
defined as 1 1
( , , , )
T L
i i i i L
X y y y R  
  .
Let (1 )L L N  be some integer called window
length (the single parameter of the embedding step)
and 1K N L   . As a result, the trajectory
matrix can be constructed as shown below in
equation (1):
1 2
2 3 1
1 2
1
K
K
K
L L N
y y y
y y y
X X X X
y y y


 
 
   
   
 
 



   

(1)
The trajectory matrix X is a Hankel
matrix that means all the elements on the off
diagonals ( constant)i j  are equals [3]. For
example, consider the observations
(10, 20, 30, 40, 50, 60),Y  6N  . Choose
window length 4L  then 6 4 1 3K     ,
and the trajectory matrix is:
1 0 2 0 3 0
2 0 3 0 4 0
3 0 4 0 5 0
4 0 5 0 6 0
X
 
 
 
 
 
 
(2)
2.1.2. Step 2: Singular Value Decomposition (SVD):
From the trajectory matrix X , define the covariance
Matrix
T
X X . Let 1 2 L
     denote the
eigenvalues of
T
X X and the corresponding
eigenvectors 1 2
, , , L
U U U in SSA literature are
donated by (the empirical orthogonal functions). Let
d be the number of nonzero eigenvalues (that is the
rank of
T
X X ) and set the principal components
T
i i i
V X U  , then the SVD of the trajectory
matrix X can be expressed as a sum of matrices:
1 2 d
X E E E    (3)
where , 1, 2, ,
T
i i i i
E U V i d   . The
gathering  , ,i i i
U V is referred to the
th
i
eigentriple (square root of eigenvalue, eigenvector,
factor vector) of the matrix X . The ratio
1
d
i i
i
 

 is the share of the trajectory matrix X
explained by the sum in (1). For more details see
[12].
2.2 Reconstruction
This stage includes two steps: grouping and
averaging.
2.2.1. Grouping:
The grouping procedure divides the set of
elementary matrices indices  1, ,d
into m disjoint subsets 1
, , m
I I . Let
 ,1 ,
, ,
k k k p
I i i  be a group of indices
,1 ,
, , , 1, ,k k p
i i k m  . Then the matrix
kI
X corresponding to the group k
I is defined as
,1 ,2 ,k k k k pI i i i
X E E E    . Then the
corresponding decomposition is written as:
1 mI I
X X X   (4)

For a given group , 1, ,k
I k m  the
share of the component kI
X into the equation (3) is
measured by the contribution of the corresponding
eigenvalues:
1k
d
i i
i I i
 
 
  . For more details see
[13]. Note that in the previous numeric example,
there are only two nonzero components and at most
two groups that can be defined as
   1 2
1 , 2I I  .
2.2.2. Step 4: Diagonal averaging:
The group of m components chosen in the
above step is used to rebuild the deterministic
components of the original time series N
Y . The aim
of diagonal averaging is to transform all matrix
components 1
, ,
mI I
X X in equation (4) above to
Hankel matrices
1
, ,
mI I
X X  where
jI
X is the
Hankelization of
jI
X for 1, ,j m  . The
expression (4) is transformed to the Hankelized
form:
1 mI I
X X X    (5)
If ij
w stands for an element of a matrix W , then
the
th
k element of the produced time series is
obtained by averaging ij
w over all ,i j such that
1i j k   . (For example, the 3rd
element is
obtained by 1,3 2 ,2 3 ,1
( ) 3w w w  and so on).
Let  1 2
, , ,j j jN
y y y   denote the
reconstructed components of the original time series
corresponding to the Hankel matrix
jI
X for
1, ,j m  . Therefore, the original time series
will be decomposed into the sum of m series such
that:
1
, 1, 2 , ,
m
k jk
j
y y k N

    (6)
Forming diagonal averaging on the
components of the previous example, the Hankelized
expansion becomes:
1 2
1 0 2 0 3 0
2 0 3 0 4 0
3 0 4 0 5 0
4 0 5 0 6 0
1 5 .3 8 2 1 .6 3 2 8 .7 0 5 .3 8 1 .6 3 1 .3 0
2 1 .6 3 2 8 .7 0 3 8 .2 7 1 .6 3 1 .3 0 1 .7 3
2 8 .7 0 3 8 .2 7 4 9 .9 1 1 .3 0 1 .7 3 0 .0 8
3 8 .2 7 4 9 .9 1 6 2 .3 9 1 .7 3 0 .0 8 2 .3 9
X X X
 
 
   
 
 
 
    
   

    
   
   
   
 
(7)
Consequently, the reconstructed time series
components are:
1 0 1 5 .3 8 5 .3 8
2 0 2 1 .6 3 1 .6 3
3 0 2 8 .7 0 1 .3 0
4 0 3 8 .2 7 1 .7 3
5 0 4 9 .9 1 0 .0 8
6 0 6 2 .3 9 2 .3 9
Y
     
     

     
     
       
     
     
     
     
(8)
2.3 Forecasting
The basic SSA forecasting method is
known as the recurrent method (R-forecasting). It
uses the concept of Linear Recurrent Formula
(LRF). A time series  1
, ,N N
Y y y  satisfies
the LRF and can be successfully forecasted via the
SSA. N
Y recognizes the LRF of order d if:
1 1
, 1, ,t t d t d
y y y t d N  
      (9)
where 1 2
, , , d
   are the coefficients of the
LRF of order d . Wide class of time series is ruled
by LRF such as harmonic, polynomial and
exponential series. Subsequently, there is a
summarized description of the so called SSA
recurrent forecasting algorithm (for more details see
[10]).
The LRF coefficients 1 2
, , , d
   in the
linear combination (9) can be obtained by using the
eigenvectors acquired in the SVD step of the SSA
algorithm. Let j
U

denote the vector of the 1st
1L  components of the eigenvector j
U and j
 is
the last component of (fo r 1, , , )j
U j r r L  .
The LRF coefficients can be calculated as follows:
1 2 1 2
1
1
( , , , )
1
r
T
L j j
j
U   




  

 (10)
where
2 2
1
r
j
j
 

  .

Considering these notations, the time series
1
ˆ ˆ ˆ( , , )N h N h
Y y y 
  can be defined by the
following formula:
1
1
fo r 1, ,
ˆ
fo r 1, ,
i
L
i
j i j
j
y i N
y
y i N N h





 
  



 

(11)
where ˆ ( 1, , )i
y i N  are the reconstructed
time series. Then 1
ˆ ˆ, ,N N h
y y 
 are the h -step-
ahead recurrent forecast. Regarding the previous
numerical example, let 3r L  . Then the LRF
coefficients can be calculated and have the values:
1 2 3
( , , ) ( 1 .5 0 , 2 .0 0 , 0 .5 0 )
T T
      (12)
and the 1st-step-ahead forecast is then:
7
ˆ 0.50 * 62.39 2.00 * 49.91
( 1.50) * 38.27 73.61.
y  
  
(13)
2.4 Parameters Selection
As explained in the previous sections, SSA
approach requires two important parameters: the
window length L (in the embedding step) and the
grouping effect parameter r (in the reconstruction
stage). Certainly, there are great efforts and
suggestions of various methods of selecting the
suitable value of L and r (see [10], [14], [15] and
[16]). Taking into account theoretical results for
structure of the trajectory matrix and seprabelity, it
seems better to select window length less than
2N or 3N as other recommended values (see
[17]). Moreover, values for L and r be selected
depend on both information provided by time series
under study and analysis that need to be performed.
In this paper, SSA is used as a technique of
forecasting. Therefore, the implemented criteria are
based on the forecasting errors. However, different
error measures are used to find the error forecast.
These include Mean absolute error (MAE), Mean
square error (MSE), Root mean square error
(RMSE) and Mean absolute percentage error
(MAPE). These accuracy measures are defined in
Table 1 below.
Table 1 Accuracy Measures
Acronyms Definition Formula
MAE Mean absolute
error 1
| |
n
ti
e n

MSE Mean square error 2
1
n
ii
e n

RMSE Root mean square
error M S E
MAPE Mean absolute
percentage error
1
| |
100
n
i i
i
e y
n
where n is the number of observations in the
sample. Note that the selection of L and r is
obtained by minimizing the forecasting error using
these accuracy measures (see [16] and [18]).
Figure 1. Monthly electricity consumption in the
middle area of Gaza Strip (2005:11-2016:12)
III. APPLICATION TO REAL DATA
SSA approach is used to decompose and
forecast the monthly electricity consumption series
(in GWh) in the Middle Province of Gaza Strip. The
time series contains 134 observations from
November 2005 to December 2016. We have
transformed the data by dividing it by 1000000.
Table 2 shows some descriptive statistics for this
data set. On the other hand, Fig. 1 displays the
electricity consumption over the period (2005:11 -
2016:12).
Table 2. Descriptive Statistics for Monthly
Electricity Consumption (in GWh) for the Middle
Area of Gaza 2005:11-2016:12
n Min. 1st
Qu
Median Mean 3rd
Qu
Max.
134 5.26 10.22 13.45 12.81 15.56 20.02
Visual analysis of the drawn time series in
Fig. 1 specifies that it has a trend and this trend can
be resembled either by exponentially increasing
function or linear function. Moreover, it seems that
the seasonal component has a complicated and
changeable attitude. Fig. 2 displays the periodogram
of electricity consumption series which emphasizes
this assumption. The periodogram is established for
purifying and revising the representation of the time
series.

Figure 2. Periodogram of series electricity
consumption (the first 70 points)
Moreover, the first 122 observations from
November 2005 to December 2015 are used as a
training sample for model prediction and the
remaining 12 observations from January 2016 to
December 2016 are used as a testing data set to
assess the electricity consumption forecasts. Results
from SSA technique are compared with other
models’ results, like Box-Jenkins (ARIMA) model
and Exponential Smoothing (ETS) model. SSA
computations are implemented by using the RSSA
package in R (see [19]), while the other two models
performed using the forecast package [20]. For each
approach, the following performance indexes related
to the forecasting errors are offered: MAE, MSE,
RMSE and MAPE, and residuals test statistics.
3.1. Electricity consumption by SSA
As explained in the previous section, it is
required as a first step to select the proper
parameters, i.e., the window length L and the group
indices r . As mentioned before, the suitable values
of L and r are that allowing the minimal values of
the accuracy measures of Table 1.
The process of selection L and r carried
out through two steps. The first step is fixing
, 2 2L L N  and changing the value of
, 1 1r r L   until reaching the choice of the
pair ( , )L r which minimizes the accuracy
measures. Repeat this step many times with different
choices of L . The second step is choosing the
optimal pair ( , )L r , throughout all pairs of the first
step, which provide minimum accuracy measures.
Table 3 shows different choices of ( , )L r and the
corresponding values of the accuracy measures (all
values are rounded).
Table 3 Different selections of ( , )L r with
corresponding values of accuracy measures
( , )L r MAE MSE RMSE MAPE
(37,3) 1.568 3.952 1.988 0.102
(38,9) 1.516 3.914 1.978 0.101
(39,7) 1.501 3.622 1.903 0.097
(40,7) 1.416 3.560 1.887 0.094
(41,7) 1.454 4.014 2.004 0.097
(42,12) 1.445 3.417 1.848 0.095
(55,4) 1.528 4.134 2.033 0.102
(56,4) 1.488 3.886 1.971 0.099
Note that many other selections and their
corresponding values of accuracy measures are also
computed but they are improper so ignored and not
reported. Comparing the values in Table 3 concludes
that the best selection of SSA parameters of
electricity consumption time series is that 40L 
and 7r  . Fig. 3 illustrates eigenvectors from the
first 12 eigentriples. It shows the participation of
each eigenvalue after the SVD stage using 40L 
as window length required for the embedding step.
Figure 3. The first 12 principal components plotted
as time series.
Fig. 4 displays the plot of the reconstructed
components. It describes the 12 most significant
initial reconstructed components of the original
electricity consumption time series. After taking a
quick look, the first two reconstructed components
are related to slow motion components (i.e., the
trend behavior), while the remainder of the
reconstructed components are connected to
fluctuating components.

Figure 4. Reconstructed components related to the
first 12 eigentriples
Furthermore, for the sake of evaluating the
separability of the various eigentriples, Fig. 5 displays
a graphical representation of the w-correlation matrix.
The correlation between components m and n can be
represented as a cell ( , )m n
F F . It has a color scale
from black to white equivalent to 1 to 0, respectively.
Notice that cell 1 2
( , )F F is white (w-correlation=0), so
components 1 and 2 are clearly separable whereas
cell 3 4
( , )F F are grey color (w-correlation=0.25), so
components 3 and 4 are almost separable. Components
5-6 and 7-8 have high w-correlations so they could be
grouped. Also, the first two eigentriples in Fig. 5
correspond to the trend components while the next
eigentriples produce harmonic (oscillatory)
components. The large shining cells on the top-right
shows noise components. Therefore, the first 7
eigentriples will be used in the reconstruction of the
electricity consumption series.
IV. COMPARISON
In this section, the SSA approach is
compared with several well-known models, the Box-
Jenkins model namely, Autoregressive Integrated
Moving Average (ARIMA) models and the
Exponential Smoothing (ETS) models.
Figure 5. A graphical representation of the w-
correlation matrix.
4.1. ARIMA Model
A non-stationary time series  t
X is said
to follow a non-stationary autoregressive integrated
moving average (ARIMA) denoted by
A R IM A ( , , )p d q if it is expressed as:
( ) ( )
d
p t q t
Y      B B (14)
where t
 are identically and independently
distributed as  2
0, , 1, 2, ,N t N   and
N is the number of observations, d is the order of
non-seasonal differences and  is the non-seasonal
differencing operator, 1   B . µ is the mean of
a series assuming that after differencing it is
stationary.
As mentioned above, B is the backshift
operator, used to simplify the representation of lag
values, by 1t t
X X 
B . Also, ( )p
 B and
( )q
 B are the autoregressive polynomial of B for
order p and the moving average polynomial of B
for order q respectively where:
2
1 2
) 1(
p
p p
       BB B B (15)
and
2
1 2
1( )
q
q q
       B B B B (16)
More details can be found in ([21], [22], [23], [24],
and [25]).
4.1.1. Parameters Estimation
The estimation of parameters for ARIMA
model is a nonlinear problem that requires
some special processes such as the maximum
likelihood method or nonlinear least-squares
estimation. At this stage of model building, the
estimated parameter values should minimize
the sum of squared residuals. For this purpose, many
software packages are applicable for
fitting ARIMA models. In this current study, the
forecast package in R software will be used.
To choose the best ARIMA model based on
observation data, we use the corrected
Akaike Information Criterion (AICc). For more
details see [21].
Based on the methodology summarized
above, to calculate the point forecasts for electricity
consumption in the Middle Province of Gaza Strip
for the 12 months from January 2016 to December
2016, the results showed that the best ARIMA
model was A R IM A (1 ,1 ,4 ) :
4
2
(1 1 .2 0 1 0 .2 0 1 ) 0 .0 2 8 ( )t t
Y     B B B (17)
where
2
4
3 4
( ) (1 0.202 0.102 0.116 0.100 )    B B B B B (18)

4.2. Exponential Smoothing State Space Model
Another popular type of forecasting models
are exponential smoothing techniques. It’s simple
but very helpful of adjusting time series forecasting.
[26], [27], and [28] have initially introduced these
methods in their early works. The idea behind
forecasting using exponential smoothing is to assign
exponentially declining weights to observations as
they go back in age, i.e., recent observations have a
larger weight than the old ones.
Pegels [29] was the first to classify
exponential smoothing techniques and propose a
taxonomy of the trend component and seasonal
component. Pegels’ taxonomy was later extended by
[30], who added damped trend to the classification.
This extension is then modified by [31], before the
final extension is proposed by [32] who extended the
classification to include damped multiplicative
trends. For more details, see [33].
When fitting an exponential smoothing
model with a state space approach, the ets() function
in forecast package [20] and [34] was used. It is
utilized to choose the suitable model automatically
on the basis of maximum likelihood method (MLE)
and then to calculate the point forecasts for
electricity consumption in the Middle Area of Gaza
Strip for the 12 months from January 2016 to
December 2016. When applying the function, the
results showed that the best performing model
was E T S(M , A , N ) , that is a model with
multiplicative error, additive trend and no
seasonality. The exponential smoothing state space
E T S(M , A , N ) model is:
 
 
4.689
1 1 (1 ),
0.118
1 1 4.689 4.689 0.908
1 1
0 1 0.118 0.118 0.0001
t t
t t
y
x


 
  
 
       
        
       
(19)
where the state vector ( , )
T
t t t
x b  , includes the
level and growth components respectively. For more
details about the structures and notations of
exponential smoothing state space models, see [33].
4.3. Forecasting Results
Table 4 presents the results for several
techniques for the forecasting of the 12-step-ahead
data points for electricity consumption in the Middle
Area of Gaza Strip. The scores show that the SSA
forecasts are comparable with the forecasts acquired
from the Box-Jenkins A R IM A (1 ,1 ,4 ) and the
exponential smoothing state space E T S(M , A , N )
models. Moreover, the SSA forecasts outperform the
forecasts produced by the ARIMA and exponential
smoothing models and consequently the forecasted
values for electricity consumption of the Middle
Province in Gaza Strip are very close to the original
data for the SSA technique. Adding to that, SSA is
able to provide further details about the
decomposition of the time series.
Comparing the results of the measures of
forecast accuracy in Table 5, the SSA technique still
performs best based on all the measures values. The
predicted values using SSA are highly accurate as
MSE=3.5604 is approximately 3 times less than the
other models. RMSE = 1.8869 is 1.5 times less than
the other models and MAPE = 0.0938 is
approximately 2 times less than the other models.
Table 4. Actual and Prediction for Electricity
Consumption (in GWh) for the Middle Province of
Gaza Strip in 2.16:1-2016:12
Prediction
Month Actual
Data
SSA ARIMA
(1,1,4)
ETS(M,
A,N)
01/2016 14.1886 17.2325 18.4250 18.4873
02/2016 12.5564 16.9441 19.1469 18.6053
03/2016 16.4811 16.8670 18.1730 18.7232
04/2016 15.1924 16.9982 18.7956 18.8412
05/2016 16.0113 17.2629 18.5982 18.9592
06/2016 17.4678 17.5599 18.8822 19.0771
07/2016 19.0960 17.8049 18.8836 19.1951
08/2016 18.1212 17.9669 19.0509 19.3131
09/2016 17.6333 18.0759 19.1208 19.4310
10/2016 16.0417 18.2045 19.2478 19.5490
11/2016 20.0189 18.4360 19.3413 19.6670
12/2016 19.2120 18.8230 19.4545 19.7849
By all the odds, the SSA technique has the lowest
forecasting accuracy measures.
Table 5 Forecast Accuracy for the SSA, ARIMA
and Exponential Smoothing Techniques
MAE MSE RMSE MAPE
SSA 1.4158 3.5604 1.8869 0.0938
ARIMA(1,1,4) 2.2399 8.3199 2.8844 0.1499
ETS(M,A,N) 2.3597 8.5087 2.9170 0.1563
Fig. 5 exhibits the actual and forecasted
electricity consumption in the Middle Province of
Gaza Strip, as well as 80% and 95% prediction
intervals. Therefore, a more detailed picture of the
SSA technique can be drawn by diving into the
numbers of the previous tables.

Figure 5. Actual and forecasted (red line) electricity
consumption in the Middle Province of Gaza Strip,
80% and 95% (blue) prediction intervals.
However, to make prediction intervals
using exponential smoothing techniques, the
prediction intervals require that the forecast errors
should be uncorrelated and normally distributed with
mean zero and constant variance.
The sample correlation in Fig. 6 displays
that most of sample autocorrelation coefficients of
the residuals are within the confidence limits,
consequently the residuals are white noise reflecting
that SSA technique is adequate.
Figure 6. Residual diagnostics for SSA technique.
To confirm the evidence of autocorrelations,
the Box-Ljung test p-value result in Table 6 displays
that there is an evidence of no autocorrelations in the
forecasts errors. Furthermore, the Jarque Bera test p-
value shows that there is a strong evidence for
normality of forecasts errors. Moreover, the test of
homoscedasticity p-value proves the evidence that the
variance is constant.
Table 6. Residual Diagnostics Tests
Residual Test p-values
Residual Autocorrelation test
Box-Ljung test
08
8.243 * 10

Residual Normality test
Jarque Bera Test
0.01426
Homoscedasticity test
Box-Ljung test (Squared
Residuals)
0.04635
V. CONCLUSION
The main objective of this research is to
clarify the methodology of Singular Spectrum
Analysis and explain that SSA can be successfully
utilized to analyze and predict the monthly
electricity consumption in the Middle Province of
Gaza StripPalestine. This paper has explained that
the SSA technique is a very powerful appliance for
decomposing and forecasting a nonlinear and/or a
non-stationary time series into a collection of
independent components. In the given example,
regarding the monthly electricity consumption, the
performance of SSA is the leading, compared with
other well-known forecasting models, namely Box-
Jenkins ARIMA model and Exponential Smoothing
State Space ETS model. The numerical results
obtained using R packages declared that the error
came by the SSA technique were smaller than those
obtained by the ARIMA and ETS state space models
according to mean absolute error (MAE), mean
square error (MSE), root mean square error (RMSE)
and mean absolute percentage error (MAPE).
However, the SSA outperforms the ARIMA and
ETS state space models. Adding to that, the obtained
numerical results assert the potentiality of the SSA
technique for electricity consumption predicting
implementations.
REFERENCES
[1] R. Vautard and M. Ghil, Singular Spectrum
Analysis in Nonlinear Dynamics, With
Applications to Paleoclimatic Time Series,
Physica D, 35, 1989, 395-424.
[2] R. Vautard, P. Yiou and M. Ghil, Singular-
Spectrum Analysis: A Toolkit for Short,
Noisy Chaotic Signals, Physica D, 58, 1992,
95-126.
[3] H. Hassani, Singular Spectrum Analysis:
Methodology and Comparison, Journal of
Data Science, 5(2), 2007, 239–257.
[4] N. Simões, L. Wang, S. Ochoa1, J. P. Leitão,
R. Pina, C. Onof, A. Sá Marques, Č.
Maksimović, R. Carvalho, L. David, A
Coupled SSA-SVM Technique for Stochastic
Short-Term Rainfall Forecasting, 12nd
International Conference on Urban Drainage,
Porto Alegre, 2011, 1-8.
[5] M. Zokaei, R. Mahmoudvand, N. Najari,
Comparison of Singular Spectrum Analysis
and ARIMA, Proc. 58th World Statistical
Congress, Dublin, 2011, 3991- 3996.
[6] A. Miranian1, M. Abdollahzade, H. Hassani,
Day-Ahead Electricity Price Analysis and
Forecasting by Singular Spectrum Analysis,
IET Generation, Transmission &
Distribution, 2012, 7(4), 2012, 337–346.
[7] N. Golyandina and A. Korobeynikov, Basic
Singular Spectrum Analysis and forecasting

with R, Computational Statistics and Data
Analysis, 71, 2013, 934–954.
[8] C. Deng, Time Series Decomposition Using
Singular Spectrum Analysis, Master Thesis,
the Department of Mathematics,
East Tennessee State University, 2014.
[9] M. Lima de Menezes, R. Castro Souza and J.
Francisco Moreira Pessanha, Electricity
Consumption Forecasting Using Singular
Spectrum Analysis, Universidad Nacional de
Colombia, 82 (190), 2015, 138-146.
[10] N. Golyandina, V. Nekrutkin, and A.
Zhigljavski, Analysis of Time Series
Structure: SSA and Related Techniques (Boca
Raton: CRC Press, 2001).
[11] N. Golyandina, A. Zhigljavsky, Singular
Spectrum Analysis for Time Series (Springer
Science & Business Media, 2013).
[12] H. Hassani, S. Heravi, and A. Zhigljavsky,
Forecasting European Industrial Production
with Singular Spectrum Analysis,
International Journal of Forecasting, 25,
2009, 103–118.
[13] S. Rocco, Singular Spectrum Analysis and
Forecasting of Failure Time Series, Reliability
Engineering and System Safety, 114, 2013,
126–136.
[14] H. Hassani, R. Mahmoudvand and M. Zokaei,
Separability and Window Length in Singular
Spectrum Analysis, C. R. Acad. Sci. Paris,
Ser. I, 349, 2011, 987–990.
[15] R. Mahmoudvand and M. Zokaei, On the
Singular Values of the Hankel Matrix with
Application in Singular Spectrum Analysis,
Chilean Journal of Statistics 3, 2012, 43-56.
[16] R. Mahmoudvand, F. Alehosseini, M. Zokaei,
Feasibility of Singular Spectrum Analysis in
the Field of Forecasting Mortality Rate,
Journal of Data Science, 11, 2013, 851-866.
[17] N. Golyandina, A. Korobeynikov, Basic
Singular Spectrum Analysis and forecasting
with R, Computational Statistics & Data
Analysis, 71, 2014, 934-954.
[18] M. Ghodsi and M. Yarmohammadi, Exchange
Rate Forecasting With Optimum Singular
Spectrum Analysis, Journal of Systems
Science and Complexity – Springer, 27, 2014,
47–55.
[19] A. Korobeynikov, Computation- and space-
efficient implementation of SSA, Statistics
and Its Interface, 3(3), 2010, 257-368.
[20] Hyndman RJ. Forecast: Forecasting Functions
for Time Series and Linear Models. R
package version 8.0. 2017.
[21] B. M. Iqelan, Time Series Modeling of
Monthly Temperature Data of
Jerusalem/Palestine, MATEMATIKA, 31(2),
2015, 159–176.
[22] J. D. Cryer and K. Chan, Time Series Analysis
with Applications in R. (Springer, Second
edition, 2008).
[23] D.S. Shumway and D. S. Stoffer, Time Series
Analysis and Its Applications: With R
Examples (Springer, third edition, 2011).
[24] G. Box and G. Jenkins, Time Series Analysis:
Forecasting and Control (Holden-Day, the
University of Michigan, 1976).
[25] P. Brockwell and R. Davis, Time Series:
Theory and Methods (Springer; 2nd ed.
1991).
[26] R. G. Brown, Statistical Forecasting for
Inventory Control (McGraw-Hill, New York.
1959).
[27] C. C. Holt, Forecasting Trends and Seasonals
by Exponentially Weighted Averages (O.N.R.
Memorandum 52/1957, Carnegie Institute of
Technology. 1957).
[28] P. R. Winters, Forecasting Sales by
Exponentially Weighted Moving Averages,
Management Science, 6, 1960, 324–342.
[29] C. C. Pegels, Exponential Forecasting: Some
New Variations. Management Science 15(5),
1969, 311–315.
[30] Jr. Gardner, Exponential Smoothing: The
State of the Art, Journal of Forecasting 4,
1985, 1–28.
[31] R. J. Hyndman, A. B. Koehler, R. D. Snyder
and S. Grose, A State Space Framework for
Automatic Forecasting Using Exponential
Smoothing Methods, International Journal of
Forecasting, 18(3), 2002, 439–454.
[32] J. W. Taylor, Exponential Smoothing with a
Damped Multiplicative Rend, International
Journal of Forecasting, 19, 2003, 715–725.
[33] B. M. Iqelan, Comparison of Parametric and
Nonparametric Techniques for Water
Consumption Forecasting, International
Journal of Scientific & Engineering Research,
8(1), 2017, 1530-1536.
[34] R.J. Hyndman and Y. Khandakar, Automatic
Time Series Forecasting: The Forecast
Package for R, Journal of Statistical
Software, 26(3), 2008, 1–22

A Singular Spectrum Analysis Technique to Electricity Consumption Forecasting

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to A Singular Spectrum Analysis Technique to Electricity Consumption Forecasting (20)

Recently uploaded (20)

A Singular Spectrum Analysis Technique to Electricity Consumption Forecasting