基于循环神经网络的欧亚中高纬夏季极端高温年代际预测模型研究

近几十年来频繁发生的极端高温事件严重威胁着自然生态系统、社会经济发展和人类生命安全。针对生态环境脆弱的欧亚中高纬地区,首先评估了当前主流动力模式(CMIP6 DCPP)对于该地区夏季极端高温的年代际预测水平,并构建了基于循环神经网络(Recurrent Neural Networks,RNN)的年代际预测模型。多模式集合平均(Multi-Model Ensemble,MME)的评估结果显示,得益于大样本和初始化的贡献,当前动力模式对于60°N以南区域(South Eurasia,SEA)展现了预测技巧,准确预测出了其线性增长趋势和1968—2008年间主要的年代际变率,然而模式对于60°N以北区域(North Eurasia,NEA)极端高温的年代际变率几乎没有任何预测技巧,仅预测出比观测低的线性增长趋势。基于86个初始场的动力模式大样本预测结果,RNN将2008—2020年间NEA和SEA极端高温的年代际变率预测技巧显著提高,距平相关系数技巧从MME中的-0.61和-0.03,提升至0.86和0.83,均方差技巧评分从MME中的-1.10和-0.94,提升至0.37和0.52。RNN的实时预测结果表明,在2021—2026年,SEA区域的极端高温将持续增加,2026年很可能发生突破历史极值的极端高温事件,NEA区域在2022年异常偏低,而后将呈现波动上升。

Decadal prediction of summer extreme high temperatures in Eurasian mid-high latitudes using on Recurrent Neural Networks

The frequency of extreme high temperature events has increased against the backdrop of global warming,posing serious risks to natural ecosystems,socio-economic development,and human safety.The Eurasian mid-high latitudes,or core regions of the Belt and Road area,feature fragile ecological environments highly susceptible to climate change,with limited adaptive capacities to extreme weather events.In recent decades,the frequent occurrence of extreme high-temperature events in these latitudes has resulted in tens of thousands of fatalities and billions of dollars in economic losses.Accurate prediction of extreme high temperatures in this region,especially on a decadal scale,is urgently needed by governmental decision-makers to effectively address climate change and promote sustainable development.This paper assesses the decadal predictive skill of current state-of-the-art dynamical models (CMIP6 DCPP) for summer extreme high temperatures in the Eurasian mid-high latitude region.We utilize the anomaly correlation coefficient (ACC) to assess the model''s skill in capturing the observed variability phase and the mean-square skill score (MSSS) as a deterministic verification metric sensitive to amplitude errors.By comparing DCPP hindcasts (initialization) with historical simulations (external forcing),we examine the sources of predictive skill.The evaluation results show that multi-model ensemble average (MME) exhibits high predictive skill for the region south of 60°N (South Eurasia,SEA),accurately forecasting its linear growth trend and prominent decadal variability during 1968—2008.However,MME shows almost no predictive skill for the decadal variability of extreme high temperatures in the North Eurasia (NEA) region,only forecasting a linear growth trend lower than observed.To improve decadal predictive skills,we developed a three-layer recurrent neural network (RNN).This model utilizes the large-sample model predictions of 86 initial fields as input,with training and testing periods of 1968—2007 and 2008—2022,respectively.Significant improvements in extreme high temperature skills in NEA and SEA during test period of 2008—2020 were observed in the RNN model.The ACC skills of NEA and SEA in RNN are 0.86 and 0.83,respectively,compared to-0.61 and-0.03 in MME.Meanwhile,the MSSSs of NEA and SEA in RNN are 0.37 and 0.52,whereas they are-1.1 and-0.94 in MME,respectively.Real-time forecasts from RNN indicate that extreme high temperatures in the SEA region will continue to rise from 2021 to 2026,with a record-breaking event in 2026.Meanwhile,the NEA region is predicted to experience anomalously fewer events in 2022,followed by fluctuating increases.A comparison of the performance of various input sizes in RNN reveals that large sample sizes are necessary for the RNN model.Additionally,incorporating additional predictors with significant physical mechanisms for extreme high-temperature events may further enhance decadal prediction skills,warranting further investigation.Nevertheless,this study provides new insights into current decadal prediction of extreme climate,offering promising scientific support for governmental decision-makers in addressing climate change.