资源环境科技发展态势分析平台

In this study, a new statistical strategy to improve the long-term prediction skill of a numerical model was developed. This new strategy begins by finding the major principal time series (PTs) in the observations using the self-organizing map (SOM) method. Next, values at the model grid points that are highly correlated with the observational PTs for each ensemble member (EM) are combined to yield a modelled PT. Finally, the model prediction is corrected using the model PTs from the previous step. As the predictors for correction are objectively selected from among the signals found in model prediction, automatically considering their statistical correlation with predictands, the correction strategy is relatively free from the problem of selecting the proper predictor compared to conventional statistical correction methods. In addition, SOM shows a better performance in classifying non-linear complex patterns than conventional data analysis methods, while both SOM and conventional methods such as the empirical orthogonal function show a comparable performance when classifying linear patterns. The new strategy is applied to the 12-month-lead sea surface temperatures hindcasted by the Pusan National University coupled general circulation model. After correction using the new strategy, temporal correlation coefficients and the hit rate are increased while normalized root mean square errors and the false alarm rate are decreased for each season and each lead time. The correction becomes more effective as the lead time increases. In particular, this correction effect is large over the region where the prediction skill without correction is apparently low, which implies that the biases leading to poor prediction skills are effectively reduced by the new strategy. Additionally, the prediction skill is steadily improved for all lead times as the number of EMs is increased, whereas it reaches a plateau when the number of neurons in the output layer of the SOM method exceeds a certain threshold.