AbstractWe assess the predictability limits of the large-scale cloud patterns in the boreal summer intraseasonal variability (BSISO), which are measured by the infrared brightness temperature, a proxy for convective activity. A recent developed nonlinear data analysis technique, nonlinear Laplacian spectrum analysis (NLSA), is applied to the brightness temperature data, defining two spatial modes with high intermittency associated with the BSISO time series. Then a recent developed data-driven physics-constrained low-ordermodeling strategy is applied to these time series. The result is a four dimensional system with two observed BSISO variables and two hidden variables involving correlated multiplicative noise through the nonlinear energyconserving interaction. With the optimal parameters calibrated by information theory, the non-Gaussian fat tailed probability distribution functions (PDFs), the autocorrelations and the power spectrum of the model signals almost perfectly match those of the observed data. An ensemble prediction scheme incorporating an effective on-line data assimilation algorithm for determining the initial ensemble of the hidden variables shows the useful prediction skill in the non-El Niño years is at least 30 days and even reaches 55 days in those years with regular oscillations and the skillful prediction lasts for 18 days in the strong El Niño year (year 1998). Furthermore, the ensemble spread succeeds in indicating the forecast uncertainty. Although the reduced linear model with time-periodic stable-unstable damping is able to capture the non-Gaussian fat tailed PDFs, it is less skillful in forecasting the BSISO in the years with irregular oscillations. The failure of the ensemble spread to include the truth also indicates failure in quantification of the uncertainty. In addition, without the energy-conserving nonlinear interactions, the linear model is sensitive with parameter variations. mcwfnally, the twin experiment with nonlinear stochastic model has comparable skill as the observed data, suggesting the nonlinear stochastic model has significant skill for determining the predictability limits of the large-scale cloud patterns of the BSISO.
Expansion for the Moments of a Nonlinear Stochastic Model