基于神经网络奇异谱分析的ENSO指数预测

A Study of ENSO Index Prediction Based on Neural NetworkSingular Spectrum Analysis

采用奇异谱分析方法研究了ENSO指数及相关序列,结果表明奇异谱分析能很好的对原始序列进行信噪分离,增大了ENSO指数的可预报性。在此基础上,提出了人工神经网络和奇异谱分析相结合的ENSO指数预测方法,进行了不同因子组合的预报试验,预报效果明显优于持续性预报,超前4季的Nino3区、Nino4区预报相关系数仍高于0.5。

ENSO indices and the related time series are studied by singular spectrum analysis (SSA). The results show that SSA can detect the signal and noise from the original time series, enhance the predictability of ENSO indices. And by that, a model based on neural network and singular spectrum analysis (NNSSA) is built to forecast ENSO indices. The model is applied with different combinations of predictors from the time series. It is shown that NNSSA model performance is higher than persistent forecasts evidently, the best in the zonal wind index at 850 hPa and Nio region indices as inputs. The correlation coefficient on Nio 3 and Nio 4 SSTA forecast is still above 0.5 at lead time of 4 quarters. There is also seasonal dependence on SSTA forecast skills in the NNSSA model experiments. Compared with other statistic models (LR, CCA and Persistence etc.), NNSSA model shows comparable or predominant skills on Nio 3 SSTA forecast, and the correlation coefficient decreases very slowly as the lead time increases.