MSSA in the Wind Speed Forecasting Reinaldo Castro Souza PUC-RIO Moisés Lima de Menezes PUC-RIO / UFF José Francisco M. Pessanha UERJ
MSSA in the Wind Speed Forecasting
Reinaldo Castro SouzaPUC-RIO
Moisés Lima de MenezesPUC-RIO / UFF
José Francisco M. PessanhaUERJ
Summary
Objective
SSA Method
MSSA extension
PAR(P) Models
Proposed Methodology
Case Study
Conclusions
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Objective
Investigate the predictive gain obtained when weuse multi-channel singular spectrum analysis(MSSA) and univariate SSA integrated to thePAR(p) model applied to two-dimensional vectortime series.
The methodology is illustrated with an applicationin the modelling of two wind speed time series intwo anemometric station located in braziliannortheast region.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Singular Spectrum Analysis (SSA)
Trajectory MatrixTime Series
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Singular Spectrum Analysis (SSA)
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Noise time seriesHence:
ApproximateTime Series
Original Time Series
Proposed Methodology
Forecasts
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Two time series
SSA 1
(SERIES 1)
(SERIES 2)
SSA 2
MSSAFiltering the two series
simultaneously
Filtering the two series separately
PAR (p)Two SSA
approximatetime series
TwoMSSA
approximatetime series
Case Study• Two monthly wind speed series of Brazil northeast: Petrolina and
Pesqueira. 16 years (jan/96 – dec/11) – T=192.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
456789
10
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11Sp
ee
d (
m/
s)
Time (months)
Wind Speed (Petrolina)
456789
1011
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11Sp
ee
d (
m/
s)
Time (months)
Wind Speed (Pesqueira)
Plots of the 9 First Eigenvectors Associated to Trajectory Matrix from two Monthly Wind Speed Series
(using L optimum equal to 93 in SVD by MSSA)
ISF 2013 - SEOUL KOREA - June 23-26, 2013
80
90
100
110
1 9 17 25 33 41 49 57 65 73 81 89 97
1(98.005%)
-12
-2
8
1 9 17 25 33 41 49 57 65 73 81 89 97
2(0.569%)
-12
-2
8
1 9 17 25 33 41 49 57 65 73 81 89 97
3(0.563%)
-5
-3
-1
1
3
1 9 17 25 33 41 49 57 65 73 81 89 97
4(0.065%)
-3
-1
1
3
1 9 17 25 33 41 49 57 65 73 81 89 97
5(0.050%)
-4
-2
0
2
4
6
1 9 17 25 33 41 49 57 65 73 81 89 97
6(0.048%)
-4
1
1 9 17 25 33 41 49 57 65 73 81 89 97
7(0.042%)
-4
-2
0
2
4
1 9 17 25 33 41 49 57 65 73 81 89 97
8(0.042%)
-4
-2
0
2
4
1 9 17 25 33 41 49 57 65 73 81 89 97
9(0.032%)
Scatter Plots of 2 First Paired Eigenvectors Associated to TrajectoryMatrix from two Monthly Wind Speed Series by MSSA
ISF 2013 - SEOUL KOREA - June 23-26, 2013
-12
-7
-2
3
8
85 90 95 100 105 110
2 (
0.5
69
%)
1 (98.005%)-15
-12
-9
-6
-3
0
3
6
9
12
15
-15 -12 -9 -6 -3 0 3 6 9 12 15
3 (
0.5
63
%)
2 (0.569%)
Scatter Plots of 2 Paired Harmonic Eigenvectors Associated to Trajectory Matrix from two Monthly Wind Speed Series
The number of vertices of each regularpolygon is equal to the harmonic period ofeach eigenvector of the screen plot.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
-15
-12
-9
-6
-3
0
3
6
9
12
15
-15 -12 -9 -6 -3 0 3 6 9 12 15
3 (
0.5
63
%)
2 (0.569%)
-4
-3
-2
-1
0
1
2
3
4
-4 -3 -2 -1 0 1 2 3 4
8 (
0.0
42
%)
7 (0.042%)
Scatter plots of 3 Paired Noise Eigenvectors Associated to Trajectory Matrix from two Monthly Wind Speed Series by MSSA
These eigenvectors were classified statistically(via BDS test) as noise.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
-3
-2,5
-2
-1,5
-1
-0,5
0
0,5
1
1,5
2
-3 -2 -1 0 1 2
38
(0
.00
8%
)
37 (0.008%)-1,5
-1
-0,5
0
0,5
1
1,5
2
-2,5 -2 -1,5 -1 -0,5 0 0,5 1 1,5 2
54
(0
.00
5%
)
53 (0.005%)
-2
-1,5
-1
-0,5
0
0,5
1
1,5
2
2,5
-2 -1 0 1 2
66
(0
.00
4%
)
65 (0.004%)
5
6
7
8
1 22 43 64 85 106 127 148 169 190
-2
-1,5
-1
-0,5
0
0,5
1
1,5
2
1 22 43 64 85 106 127 148 169 190
-2
-1
0
1
2
1 22 43 64 85 106 127 148 169 190
MSSA - Components from Wind Speed Time Series (Petrolina)
MSSA-Component 1 (Trend).
MSSA - Component 2 (Harmonic).
MSSA - Component 3 (Noise).
ISF 2013 - SEOUL KOREA - June 23-26, 2013
W - correlation between the three components of Wind Speed Time Series (Petrolina)
ISF 2013 - SEOUL KOREA - June 23-26, 2013
The weighted correlation shows that the components
are unrelated. Therefore, they are well separable
Trend Harmonic Noise
Trend 1 0.001 0.002
Harmonic 0.001 1 0.028
Noise 0.002 0.028 1
Extracted Noise from Time Series from Monthly Wind Speed Time Series (Petrolina)
The null hypothesis of BDS test (independence of the noise time series) is not
rejected at 5% of significance level until the dimension 6.
MSSA - Component 3
Dimension BDS
Statistic
Std. Error Z-Statistic Prob.
2 0.006156 0.004463 1.379423 0.1678
3 0.005718 0.007098 0.805545 0.4205
4 0.004280 0.008457 0.506136 0.6128
5 0.004037 0.008817 0.457835 0.6471
6 0.000435 0.008504 0.051183 0.9592
ISF 2013 - SEOUL KOREA - June 23-26, 2013
-1,5
-1
-0,5
0
0,5
1
1,5
1 22 43 64 85 106 127 148 169 190
MSSA - Components from Wind Speed Time Series (Pesqueira)
MSSA-Component 1 (Trend).
MSSA - Component 2 (Harmonic).
MSSA - Component 3 (Noise).
ISF 2013 - SEOUL KOREA - June 23-26, 2013
5
6
7
8
9
10
1 22 43 64 85 106 127 148 169 190
-2
-1
0
1
2
1 22 43 64 85 106 127 148 169 190
-2
-1,5
-1
-0,5
0
0,5
1
1,5
2
1 22 43 64 85 106 127 148 169 190
W - correlation between the three components of Wind Speed Time Series (Pesqueira)
ISF 2013 - SEOUL KOREA - June 23-26, 2013
The weighted correlation shows that the components
are unrelated. Therefore, they are well separable
Trend Harmonic Noise
Trend 1 0.001 0.003
Harmonic 0.001 1 0.029
Noise 0.003 0.029 1
Extracted Noise from Monthly Wind Speed Time Series (Pesqueira)
The null hypothesis of BDS test (independence of the noise time series) is not rejected at 5% of
significance level until the dimension 6.
MSSA – Component 3
Dimension BDS
Statistic
Std. Error Z-Statistic Prob.
2 0.000525 0.000568 0.925448 0.3547
3 0.001376 0.001234 1.114620 0.2650
4 0.001597 0.002008 0.795097 0.4266
5 0.001581 0.002859 0.553033 0.5802
6 0.005030 0.003764 1.336379 0.1814
ISF 2013 - SEOUL KOREA - June 23-26, 2013
-2
-1,5
-1
-0,5
0
0,5
1
1,5
2
1 22 43 64 85 106 127 148 169 190
Filtering time series (Petrolina and Pesqueira) by MSSA and Original time series
ISF 2013 - SEOUL KOREA - June 23-26, 2013
4
5
6
7
8
9
10
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Petrolina Petrolina (MSSA)
4
5
6
7
8
9
10
11
12
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Pesqueira Pesqueira (MSSA)
Filtering time series (Petrolina and Pesqueira) based on SSA using window
length – L equal to 96.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
4
5
6
7
8
9
10
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Petrolina Petrolina (SSA)
4
6
8
10
12
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Pesqueira Pesqueira (SSA)
Comparison between the filtering by MSSA and SSA (Petrolina)
ISF 2013 - SEOUL KOREA - June 23-26, 2013
4
5
6
7
8
9
10
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Petrolina Petrolina (SSA)
4
5
6
7
8
9
10
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Petrolina Petrolina (MSSA)
Comparison between the filtering by MSSA and SSA (Pesqueira)
ISF 2013 - SEOUL KOREA - June 23-26, 2013
4
6
8
10
12
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Pesqueira Pesqueira (MSSA)
4
6
8
10
12
jan/96 jul/97 jan/99 jul/00 jan/02 jul/03 jan/05 jul/06 jan/08 jul/09 jan/11
Pesqueira Pesqueira (SSA)
Scatter plots with trendline - Petrolina
ISF 2013 - SEOUL KOREA - June 23-26, 2013
4
5
6
7
8
9
10
4 5 6 7 8 9 10
Pe
tro
lin
a (
SSA
)
Petrolina
4
5
6
7
8
9
10
4 5 6 7 8 9 10
Pe
tro
lin
a (
MS
SA)
Petrolina
Scatter plots with trendline - Pesqueira
ISF 2013 - SEOUL KOREA - June 23-26, 2013
4
5
6
7
8
9
10
11
12
4 6 8 10 12
Pe
squ
eir
a (
SSA
)
Pesqueira
4
5
6
7
8
9
10
11
4 6 8 10 12
Pe
squ
eir
a (
MS
SA)
Pesqueira
Comparison between Mean and Standard Deviation in Original time series, SSA and MSSA of Petrolina
ISF 2013 - SEOUL KOREA - June 23-26, 2013
PETROLINA Mean Standard Deviation
Month Original SSA MSSA Original SSA MSSA
January 5.9543 5.8964 5.8769 0.67807 0.53464 0.42529
February 5.9840 5.7253 5.7463 0.78301 0.56048 0.47308
March 5.6773 5.8151 5.8361 0.78543 0.68726 0.51030
April 6.2016 6.2973 6.2043 0.76899 0.65919 0.54709
May 6.8644 6.9539 6.8116 0.68510 0.71811 0.60127
June 7.6413 7.4656 7.4933 0.85914 0.81526 0.67732
July 7.7810 7.8534 7.9906 0.78248 0.79876 0.73271
August 8.2303 8.0778 8.1029 0.83149 0.72173 0.72404
September 7.8183 7.8591 7.7981 0.79368 0.59225 0.64441
October 7.1837 7.2509 7.2285 0.78491 0.62828 0.53599
November 6.5102 6.6287 6.6170 0.56942 0.58426 0.44951
December 6.1359 6.1195 6.1214 0.59486 0.46908 0.41776
Comparison between Mean and Standard Deviation in Original time series, SSA and MSSA of Pesqueira
ISF 2013 - SEOUL KOREA - June 23-26, 2013
PESQUEIRA Mean Standard Deviation
Month Original SSA MSSA Original SSA MSSA
January 7.4461 7.5673 7.5564 1.466812 1.27299 1.18675
February 7.4726 7.3121 7.2867 1.505957 1.35201 1.31279
March 7.1856 7.1079 7.1051 1.311712 1.42076 1.36174
April 6.6023 6.8138 6.8231 1.256665 1.27799 1.25004
May 6.5351 6.4296 6.4366 1.268066 1.00404 1.04172
June 6.0355 6.1661 6.1926 0.950773 0.90588 0.87616
July 6.4069 6.3782 6.3835 0.98693 0.87387 0.91598
August 7.1029 7.0080 7.0410 1.210563 1.04558 1.17135
September 7.7801 7.8292 7.8385 1.616289 1.39401 1.42067
October 8.2268 8.3271 8.3246 1.672698 1.45432 1.48573
November 8.2243 8.2442 8.2755 1.543823 1.34718 1.37672
December 7.9871 7.8648 7.8460 1.409316 1.28417 1.24755
Comparison between autorregressive order in modeling PAR(p), PAR(p) – SSA and PAR(p) - MSSA
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Petrolina
Pesqueira
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
PAR(p) 1 1 1 1 3 1 1 1 1 1 1 2
PAR(p)- SSA 1 4 1 1 1 1 1 1 1 1 1 3
PAR(p)- MSSA 1 1 1 1 1 1 1 1 1 1 1 1
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
PAR(p) 6 1 2 1 1 5 1 1 1 1 1 2
PAR(p)- SSA 1 1 2 1 1 2 3 5 1 2 1 2
PAR(p)- MSSA 1 1 1 1 1 1 1 1 1 1 1 1
mouth PAR(p) PAR(p) - SSA PAR(p) - MSSA
January 6.120 3.3072 1.48270
February 8.7405 3.1330 0.87904
March 6.1584 4.1720 0.81247
April 5.9539 2.4138 0.63071
May 4.4594 3.3287 0.51478
June 4.5123 2.1882 0.51592
July 2.9755 1.9037 0.54739
August 3.5905 2.3111 0.61316
September 5.0779 2.5236 0.70418
October 5.635 2.4985 0.73237
November 4.1773 1.9260 0.62408
December 4.3738 2.5884 0.68437
MAPE Statistics (In sample) - Petrolina
We can see that the PAR(p) – MSSA was better in all periods.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
MAPE Statistics (In sample) - Petrolina
We can see that the PAR(p) – MSSA was better in all periods.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
0123456789
10
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
PAR(p) PAR(p) - MSSA PAR(p) - SSA
mouth PAR(p) PAR(p) - SSA PAR(p) - MSSA
January 6.1182 4.7278 3.5413
February 5.1540 2.6814 1.1211
March 7.1102 1.9755 1.4781
April 7.4127 3.1508 1.8216
May 5.7543 2.7266 1.7512
June 4.3649 2.4026 1.3590
July 3.8340 1.9483 1.4165
August 4.5714 2.2650 1.7567
September 5.3439 1.7724 1.4193
October 5.9651 2.1032 1.4086
November 7.0626 2.3336 1.2869
December 4.4704 0.8727 1.0379
MAPE Statistics (In sample) - Pesqueira
We can see that the PAR(p) – MSSA was better in almost all periods.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
MAPE Statistics (In sample) - Pesqueira
We can see that the PAR(p) – MSSA was better in almost all periods.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
0
1
2
3
4
5
6
7
8
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
PAR(p) PAR(p) - MSSA PAR(p) - SSA
Conclusions
• In this work we compared the performances of threeforecasting methods: PAR(p), PAR(p) - SSA and PAR(p) -MSSA.
• The SSA method with analyzed approaches showedefficient to extraction of noises from the original timeseries such that generating an approximate time series(less noisy regarding to the original time series).
• The MSSA method in association to PAR(p) is better thanSSA and the correlation structure remain when there are atleast two time series.
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Bibliography• CAO, L. Y. & SOOFI, A., Nonlinear deterministic forecasting of daily dollar exchange rates,
International Journal of Forecasting, (1999), 15(4): 421 - 430.
• ELSNER, J. B. & TSONIS, A. A. Singular Spectrum Analysis: A New Tool in Time Series Analysis. Plenum Press, New York – London, (2010).
• GOLYANDINA, N., NEKRUTKIN, V., & ZHIGLJAVSKY, A., Analysis of Time Series Structure: SSA and Related Techniques, Chapman & Hall/CRC, New York - London, (2001).
• HASSANI, H. Singular Spectrum Analysis: Methodology and Comparison. Journal of Data Science, 5 (2007), 239 – 257.
• HASSANI, H. & ZHIGLJAVSKY, A. Singular Spectrum Analysis: Methodology andApplication to Economics Data. Journal of Systems Science & Complexity (2009) 22: 372-394.
• HSIEH, D. A., Chaos and nonlinear dynamics: Application to financial markets, Journal of Finance, (1991), 46: 1839 -1877.
• KRANE, S. D., An evaluation of real GDP forecasts: Economic Perspectives, (2003): 1996 -2001. Http://www.chicagofed.org/publications/economicperspectives/2003/1qeppart1.pdf.
• ZHIGLJAVSCKY, A., Hassani, H., Heravi, S. Forecasting European Industrial Productionwith Multivariate Singular Sperctum Analysis (MSSA). 31º. International Symposium onForecasting. Prague. (2011).
ISF 2013 - SEOUL KOREA - June 23-26, 2013
Thank you!
Reinaldo C. Souza (PUC-Rio)
Moisés L. Menezes (UFF/PUC-Rio)
José F. M. Pessanha (UERJ)
ISF 2013 - SEOUL KOREA - June 23-26, 2013