Adaptive Filtering and Prediction of Intraseasonal Oscillations

Abstract

A statistical model based on the combination of singular spectrum analysis (SSA) and the maximum entropy method (MEM) is applied to monitor and forecast outgoing longwave radiation anomalies (OLRAs) in the intraseasonal band over the Indian?Pacific sector and in the pan-American region. SSA is related to empirical orthogonal function analysis (EOF) but is applied to time series. The leading SSA modes (T-EOFs) are orthogonal and they are determined from the training period before filtering. The OLRA time series can be projected onto T-EOFs to obtain the principal components (T-PCs). To obtain fluctuations in any frequency band, one can partially sum up a chosen subset of T-EOFs and the related T-PCs in that band. The filter based on the SSA modes is data adaptive and there is no loss of end points. It is well suited for real-time monitoring of intraseasonal oscillations. In the Pacific and the pan-American region, there are three leading modes (T-EOFs) of oscillations with periods near 40, 22, and 18 days. The T-PCs associated with these modes are quasiperiodic and they can be modeled by an autoregressive process. To perform forecasts, the MEM is used to determine the autoregressive coefficients from the training period. These coefficients are used to advance T-PCs. The summation of T-EOFs and T-PCs related to three preferred modes gives the predicted OLRAs. For 5-day mean OLRAs, the averaged correlation between the predicted and the observed anomalies is 0.65 at the lead times of four pentads (20 days). The SSA?MEM method is effective for any time series containing large oscillatory components. The deficiency of this method is that the forecasted magnitudes of anomalies are usually weaker than observations.