梅雨降水季节预测的多方法比较

基于1961—2000年逐月降水观测资料和全球大气再分析资料,分析了6—7月长江中下游(108°～123°E,27°～33°N)梅雨的时空分布特征。通过观测诊断和数值试验确定了影响梅雨异常偏多的3个前期因子:4—5月平均的西北太平洋海平面气压正异常;3月至5月北大西洋海平面气压负变压倾向;1月至4月西伯利亚的2 m温度负倾向。利用这3个具有物理意义的影响因子构建了梅雨季节预测模型,该模型在训练期(1961—2000年)和独立预测期(2001—2022年)均具有显著的预测技巧(相关系数分别为0.79和0.77,均方根误差分别为0.59和0.68)。同时,基于相似的潜在预测因子,对比了利用偏最小二乘回归方法和5种机器学习方法(随机森林、轻量级梯度提升机、自适应提升、类别型特征提升、极端梯度提升)建立的预测模型的技巧。虽然训练期(1961—2000年)偏最小二乘回归和机器学习建模拟合效果更高,但在独立预测期(2001—2022年)上述模型的预测技巧显著降低(相关系数均低于0.44,均方根误差均大于0.93),出现了明显的过拟合问题。本研究强调梅雨的短期气候预测应建立在物理机制基础之上,而使用机器学习方法需谨慎。

A comparative study of multiple methods for seasonal prediction of Meiyu rainfall

This study elucidates the spatiotemporal characteristic of June—July mean Meiyu rainfall over the middle and lower reaches of the Yangtze River basin(27°—33°N,108°—123°E) using Chinese monthly gauge precipitation data and global atmospheric reanalysis datasets from 1961 to 2000.Three physically meaningful precursors play pivotal roles in enhancing Meiyu rainfall during June and July.First,positive sea level pressure anomalies over the subtropical western Pacific (SWP) during April—May strengthened the western North Pacific subtropical high by exciting Kelvin wave responses and enhancing Walker circulation.This phenomenon facilitates moisture transport from the tropics to the Yangtze River via southerly winds.The mechanism underlying SWP’s impact on Meiyu highlights the persistent influence of atmosphere-ocean interaction over the Indo-Pacific basin from spring to summer.Second,the negative tendency of sea level pressure over the North Atlantic from March to May (NAP) reflects the influence of North Atlantic Oscillation (NAO)-related mid-latitude wave trains on Meiyu.From spring to early summer,the evolution of NAO-related wave trains across Eurasia strengthens the Northeast Asian cyclone and enhances Meiyu rainfall.Third,the cooling tendency of surface temperature over East Siberian from January to April (EST) is closely associated with the extratropical westerly jet by amplifying the temperature gradient between the tropics and polar regions.This condition favors the maintenance of meridional circulation over East Asia and enhances Meiyu rainfall.The aforementioned mechanisms have been verified in corresponding numerical experiments based on a linear baroclinic model.Consequently,a physically-based empirical (PE) model based on these three predictors exhibited significant prediction skills,with a temporal correlation coefficient (TCC) of 0.79 and 0.77 and a mean square skill score (RMSE) of 0.59 and 0.68 during the training period (1961—2000) and independent forecast period (2001—2022),respectively.For comparison,the partial least squares (PLS) regression method and five machine learning methods (Random Forest,LightGBM,Adaboost,Catboost,and XGboost) are employed to conduct seasonal predication of Meiyu based on the same potential precursors.Although the PLS model and five machine learning models exhibit prefect hindcast skills (TCCs of LightGBM,Catboost,and XGboost all being 1.00) during the training period,their skills diminish dramatically in the independent forecast period of 2001—2022 (with the maximum TCC being 0.43 and the minimum RMSE being 0.94),indicating a significant overfitting problem.Hence,the PE model based on physically meaningful precursors demonstrates superior and stable independent prediction skills in Meiyu rainfall forecasts.The findings of this study underscore the advantages of the PE model and emphasize caution in the use of machine learning methods in climate prediction.Additionally,the comparison of multiple methods for seasonal prediction of Meiyu in this study provides practical scientific references for operational departments engaged in seasonal climate prediction.