基于时间序列LSTM模型的ENSO指数预报试验

The Prediction of ENSO Indexes Based on Time Series LSTM Model

摘要: 为了提高ENSO事件的预报能力，本文利用1951~2008年Niño 3.4和南方涛动指数(Southern Oscilla-tion index，简称SOI)时间序列数据，建立了关于ENSO指数预报的时间序列LSTM模型，并对2009~2018年期间Niño 3.4或SOI指数进行预测。结果表明，对于Niño 3.4指数，时间序列LSTM模型在1~12个月预报时效内都具有很好的预报能力，尽管模型对15/16年间超强厄尔尼诺事件的振幅预测仍有一定的误差。而对于SOI指数，模型仅在1~3个月的短预报时效内有一定的预报能力，原因是SOI中高频的季节内信号使SOI指数不具备长期的可预测性。实验结果也证明，在经过7个月的滑动平均后，模型在不同预报时效下对SOI指数的预报能力都有显著的提升，与Niño 3.4指数接近。而相对其他方法，时间序列LSTM模型具有一定预报能力优势，且具有建模方便，可同时得到不同时间预测结果等优点。

Abstract: To improve the prediction ability of El Niño-Southern Oscillation (ENSO), we use the Niño 3.4 and Southern Oscillation indexes during 1951-2008 to build time series long short-term memory (LSTM) model, and forecast that during 2009-2018. The analysis shows the time series LSTM model has a good forecasting ability for Niño 3.4 index in advance of 1 - 12 months, though underestimates the amplitude of the super El Niño event during 15/16 year. While time series LSTM model can only predict the SOI in advance of 1 - 3 months to a certain extent, because of the high frequency seasonal variability in SOI index. Further experiment proves that the 7-month run-mean SOI is also can be predicted well by time series LSTM model, similar with the Niño 3.4 index. Comparing with other statistical forecasting models, the time series LSTM model is efficient and convenience with ad-vantages in ENSO prediction.