Based on an intermediate coupled model (ICM), a probabilistic ensemble prediction system (EPS) has been developed. The ensemble Kalman filter (EnKF) data assimilation approach is used for generating the initial ensemble conditions, and a linear, first-order Markov-Chain SST anomaly error model is embedded into the EPS to provide model-error perturbations. In this study, we perform ENSO retrospective forecasts over the 120 year period 1886–2005 using the EPS with 100 ensemble members and with initial conditions obtained by only assimilating historic SST anomaly observations. By examining the retrospective ensemble forecasts and available observations, the verification results show that the skill of the ensemble mean of the EPS is greater than that of a single deterministic forecast using the same ICM, with a distinct improvement of both the correlation and root mean square (RMS) error between the ensemble-mean hindcast and the deterministic scheme over the 12-month prediction period. The RMS error of the ensemble mean is almost 0.2°C smaller than that of the deterministic forecast at a lead time of 12 months. The probabilistic skill of the EPS is also high with the predicted ensemble following the SST observations well, and the areas under the relative operating characteristic (ROC) curves for three different ENSO states (warm events, cold events, and neutral events) are all above 0.55 out to 12 months lead time. However, both deterministic and probabilistic prediction skills of the EPS show an interdecadal variation. For the deterministic skill, there is high skill in the late 19th century and in the middle-late 20th century (which includes some artificial skill due to the model training period), and low skill during the period from 1906 to 1961. For probabilistic skill, for the three different ENSO states, there is still a similar interdecadal variation of ENSO probabilistic predictability during the period 1886–2005. There is high skill in the late 19th century from 1886 to 1905, and a decline to a minimum
The initial ensemble perturbations for an ensemble data assimilation system are expected to reasonably sample model uncertainty at the time of analysis to further reduce analysis uncertainty. Therefore, the careful choice of an initial ensemble perturbation method that dynamically cycles ensemble perturbations is required for the optimal performance of the system. Based on the multivariate empirical orthogonal function (MEOF) method, a new ensemble initialization scheme is developed to generate balanced initial perturbations for the ensemble Kalman filter (EnKF) data assimilation, with a reasonable consideration of the physical relationships between different model variables. The scheme is applied in assimilation experiments with a global spectral atmospheric model and with real observations. The proposed perturbation method is compared to the commonly used method of spatially-correlated random perturbations. The comparisons show that the model uncertainties prior to the first analysis time, which are forecasted from the balanced ensemble initial fields, maintain a much more reasonable spread and a more accurate forecast error covariance than those from the randomly perturbed initial fields. The analysis results are further improved by the balanced ensemble initialization scheme due to more accurate background information. Also, a 20-day continuous assimilation experiment shows that the ensemble spreads for each model variable are still retained in reasonable ranges without considering additional perturbations or inflations during the assimilation cycles, while the ensemble spreads from the randomly perturbed initialization scheme decrease and collapse rapidly.