Multivariate Time Series Analysis Charles D. Camp MSRI July 18, 2008
PCA: 2 variable example Weakly Correlated VariablesStrongly Correlated Variables
PCA algorithm
PCA algorithm, cont.
An example using Column Ozone Data
Brewer-Dobson circulation and Planetary Waves Upwelling planetary waves break in shaded region Drives the Brewer- Dobson circulation Transports heat to the polar vortex Effect on the strength of the polar night vortex Courtesy of M. Salby
2D CTM Model The Caltech/JPL two-dimensional chemistry and transport model (2D CTM) is used to investigate interannual variability of the total ozone column. Forced by the monthly mean meridional circulation (isentropic circulation) and eddy diffusivity calculated from the NCEP/DOE Reanalysis2 data (NCEP2). Compared to the MOD observations.
Isentropic Mass Stream Function Seasonal Cycle derived from NCEP Units: 10 9 kg/s
IAV of Isentropic Stream Function EOFsPCsPower SpectraCumul. Variance 70% 88% 94%
Part II: TOMS and MOD data sets TOMS: 1°×1.25° lat-lon grid, Nov Apr.1993, monthly means Merged Ozone Data (MOD) combines TOMS and SBUV data: 5°×10° lat-lon grid; Nov Dec.2000, monthly means
MOD decomposition & PCA data is deseasonalized: mean for each month removed then detrended: linear trend removed. Anomaly field is spectrally filtered to remove intra-annual variability. Principal Component Analysis (PCA) is performed to get EOF patterns and PC time series.
MOD EOF patterns and PC time series 42% 75% 90% 93% EOFs PCs Amplitude SpectraCumul. Var.
MOD EOF 1: QBO (and Decadal) Captures 42% of the interannual variance. R=0.80 ( 1% ) <= 0
MOD EOF2: Decadal and QBO Captures 33% of the interannual variance. R= ( 5% ) <= 0
Filtered PCs 1 & 2: Separating the QBO and Decadal signals PC1 PC2 [15, 72] mo.[72, max] mo.
Linear Combinations of EOFs 1 & 2: Patterns for QBO and Decadal Signals Using the standard deviations of the filtered PCs as weights, take weighted sum and difference of EOFs 1 & 2. (Zonal averages shown) EOF 1 EOF2 sum => QBO diff => decadal
EOF 3: interaction between QBO and annual cycles (QBO-annual beat) Captures 15% of the interannual variance. <= 0
QBO-annual beat (analysis with intra-annual variability) Quadratic nonlinearity between the QBO and annual cycles creates signals with periods of 20 and 8.6 months (for a average QBO period of 30 months):
EOF 4: ENSO Captures 3% of the interannual variance. R=0.71 ( < 0.1% ) <= 0
ENSO in TOMS R=0.76 ( 0.1% ) <= 0