MEPM模型:基于深度学习的多变量厄尔尼诺-南方涛动预测模型

MEPM:Multivariate ENSO Prediction Model Based on Deep Learning

厄尔尼诺-南方涛动(ENSO)是发生在热带太平洋年际时间尺度的海-气相互作用的异常现象,并由Nino3.4指数表征其发生情况;除此之外,ENSO与众多极端气候事件密切相关。因此,有效的ENSO预测对于预防极端气候事件和深入研究全球气候变化具有重要意义。然而,目前基于深度学习的ENSO预测大多数是预测一个指数或者单一变量,对于模拟多气候要素下的ENSO预测研究较少。通过提出一种利用多气候变量的ENSO预测模型——MEPM模型,其中包括多变量信息提取模块(MIEM)和时空融合模块(STFM),捕获不同气候变量在时空上的相互依赖性,进而提高ENSO预测的准确性。选取了纬向风应力异常(τ_(x))、经向风应力异常(τ_(y))、海表温度异常(SSTA)和海表下150 m温度异常(SSTA150)4个变量的距平值进行ENSO预测。结果表明:MEPM模型在提前11个月的Nino3.4指数相关技巧上分别比北美多模型集合中的动力预报系统CanCM4、CCSM3和GFDL-aer04高10%、20%和14%。此外,MEPM模型在中期Nino3.4指数相关技巧上显著优于其他深度学习模型,并可提供长达17个月的有效预测。

The El Nino-Southern Oscillation(ENSO)is an anomaly of air-sea interaction on an interannual time scale in the tropical Pacific Ocean,and its occurrence is characterized by the Ni o3.4 index.In addition,ENSO is closely related to many extreme climatic events.Therefore,effective ENSO prediction is of great significance for preventing extreme climate events and in-depth study of global climate change.However,most ENSO predictions based on deep learning predict an index or a single variable,and there are few research on the space-time evolution of ENSO under the simulation of multi-climate factors.MEPM,a multivariate ENSO prediction model,was presented.These include multivariate information extraction module(MIEM)and spatial-temporal fusion module(STFM),to capture the interdependencies of different climate elements in time and space,thereby improving the accuracy of ENSO prediction.The anomalies of latitude wind stress anomaly(τ_(x)),longitude wind stress anomaly(τ_(y)),sea surface temperature anomaly(SSTA)and 150 m below sea surface temperature anomaly(SSTA150)were selected,and sufficient experiments were carried out.The results show that MEPM is 10%,20%and 14%higher,respectively,than dynamic forecasting systems CanCM4,CCSM3,and GFDL-aer04 in North American multi-model ensemble on the average of Nino3.4 index-related techniques 11 months in advance.In addition,MEPM significantly outperforms other deep learning models on Nino3.4 index-related techniques over the medium term and provides valid predictions up to 17 months.