机器学习算法在降水和气温多模式集成中的应用  

Application of machine learning algorithms in multimodal integration of precipitation and temperature

作　　者：鞠琴[1,2] 吴金雨 王兴平 刘小妮 王逸夫 段远强 吴可馨 蒋晓蕾 JU Qin;WU Jinyu;WANG Xingping;LIU Xiaoni;WANG Yifu;DUAN Yuanqiang;WU Kexin;JIANG Xiaolei(The National Key Laboratory of Water Disaster Prevention,Hohai University,Nanjing 210098,China;China Meteorological Administration Hydro-Meteorology Key Laboratory,Nanjing 210024,China;Sichuan Province Zipingpu Development Co.,Ltd.,Chengdu 610091,China;College of Hydraulic Science and Engineering,Yangzhou University,Yangzhou 225000,China)

机构地区：[1]河海大学水灾害防御全国重点实验室,江苏南京210098 [2]中国气象局水文气象重点开放实验室,江苏南京210024 [3]四川省紫坪铺开发有限责任公司,四川成都610091 [4]扬州大学水利科学与工程学院,江苏扬州225000

出　　处：《水资源保护》2024年第3期106-115,共10页Water Resources Protection

基　　金：国家重点研发计划项目(2021YFC3201104);中央高校基本科研业务费专项资金资助项目(B240203007);江苏省大学生创新创业项目(202310294012Z,202310294176Y);水灾害防御全国重点实验室“一带一路”水与可持续发展科技基金项目(2022491111,522012232)。

摘　　要：选取CMIP6中5种全球气候模式,利用算术平均、权重平均、多元线性回归、BP神经网络、长短期记忆(LSTM)神经网络和随机森林(RF)等6种多模式集成方法,基于黄河流域水源涵养区历史降水量和气温数据,评估不同集成方法的模拟效果,并选取模拟效果最好的多模式集成方法预估未来SSP1-2.6、SSP2-4.5和SSP5-8.53种情景下黄河流域水源涵养区的降水和气温变化趋势。结果表明:多模式集成能很好地再现基准期降水和气温变化,3种机器学习算法表现相对较好,其中LSTM神经网络最好;在未来3种情景下,多年平均降水量均有所增加,四季降水量变化各有差异;SSP1-2.6情景下年降水量峰值出现在各时段初期,SSP2-4.5和SSP5-8.5情景下的年降水量呈增长趋势,远期下降趋势较明显;3种情景下气温都呈上升趋势,但变化差异较大,增温幅度和速率由小到大为SSP1-2.6、SSP2-4.5、SSP5-8.5,秋季气温增幅最大,冬季最小;多模式集成方法对未来降水量和气温的预估存在较大的不确定性,均表现为中远期大于近期,降水量预估的不确定性比气温大,其中降水量秋冬季不确定性明显大于春夏季。Using six multimodal integration methods,including arithmetic averaging,weighted averaging,multiple linear regression,BP neural network,long-short-term memory(LSTM)neural network,and random forest(RF),this study integrated five global climate models(GCMs)data in CMIP6,and based on historical precipitation and temperature data of the Yellow River Basin water conservation region,the simulation performance of different integration methods were evaluated.The multimodal integration method with the best performance was selected to predict future precipitation and temperature under three scenarios(SSP1-2.6,SSP2-4.5,and SSP5-8.5).The results show that the multimodal integration could well reproduce the variations of historical precipitation and temperature,and the LSTM neural network method has the best performance.In three scenarios,future average annual precipitation all increases,but the change of seasonal precipitation in the future varies.Under the SSP1-2.6 scenario,the annual precipitation peaks occur at the beginning of each period,while annual precipitation increases in the near term and decreases obviously in the long term under the SSP2-4.5 and SSP5-8.5 scenarios.Future temperature in three scenarios shows upward trends of different degrees,and the amplitude and rate of temperature increase from small to large are:SSP1-2.6,SSP2-4.5,SSP5-8.5.Future temperature increases greatest in autumn and least in winter.There are large uncertainties in the future precipitation and temperature predicted by multimodal integration methods,and the uncertainty in the medium to long term is greater than in the short term.The uncertainty in future precipitation projection is relatively greater than that of temperature,and the uncertainty in autumn and winter is significantly greater than that in spring and summer.

关 键 词：CMIP6 全球气候模式 多模式集成 LSTM神经网络 黄河流域 

分 类 号：P468[天文地球—大气科学及气象学]