气候变化研究进展 ›› 2020, Vol. 16 ›› Issue (3): 263-275.doi: 10.12006/j.issn.1673-1719.2019.067

• 气候系统变化 •    下一篇

基于LSTM网络的中国夏季降水预测研究

沈皓俊1,罗勇1,2(),赵宗慈1,3,王汉杰1   

  1. 1 清华大学地球系统科学系地球系统数值模拟教育部重点试验室,北京 100084
    2 全球变化与中国绿色发展协同创新中心,北京 100875
    3 中国气象局国家气候中心,北京 100081
  • 收稿日期:2019-03-29 修回日期:2019-04-29 出版日期:2020-05-30 发布日期:2020-06-15
  • 通讯作者: 罗勇 E-mail:Yongluo@mail.tsinghua.edu.cn
  • 作者简介:沈皓俊,男,硕士研究生, shhj410@126.com
  • 基金资助:
    国家重点研发计划项目(2016YFA0602100);国家重点研发计划项目(2017YFA0603700)

Prediction of summer precipitation in China based on LSTM network

Hao-Jun SHEN1,Yong LUO1,2(),Zong-Ci ZHAO1,3,Han-Jie WANG1   

  1. 1 Ministry of Education Key Laboratory for Earth System Modeling, Department of Earth System Science,Tsinghua University, Beijing 100084, China
    2 Joint Center for Global Change Studies, Beijing 100875, China
    3 National Climate Center, China Meteorological Administration, Beijing 100081, China
  • Received:2019-03-29 Revised:2019-04-29 Online:2020-05-30 Published:2020-06-15
  • Contact: Yong LUO E-mail:Yongluo@mail.tsinghua.edu.cn

摘要:

基于BCC-CSM季节气候预测模式系统历史回报数据和国家气象信息中心提供的中国地面降水月值数据,通过多方法对比并讨论了影响预测结果的因素,利用长短期记忆(Long Short-Term Memory,LSTM)网络预测2014年和2015年中国夏季降水。结果表明:LSTM网络的预测效果较逐步回归、BP神经网络及模式输出结果有一定优势。参数调优对于LSTM网络预测效果影响较大,重要参数有隐含层节点数、训练次数和学习率。选择合适的起报月份数据有助于提升季节预测的准确性,利用4月起报的数据预测夏季降水效果较好。海冰分量因子对降水季节预测有正贡献。在2014年、2015年夏季降水回报试验中,LSTM网络对降水整体形势有一定的预测能力,Ps评分分别为74分、71分,距平符号一致率分别为55.63%、55.25%,Ps评分的均值高于同期全国会商及业务模式。

关键词: LSTM网络, 机器学习, 汛期降水, 季节预测

Abstract:

Based on the historical return data of the BCC-CSM seasonal climate prediction model and the monthly data of the surface precipitation in China provided by the National Meteorological Information Center, the factors affecting the forecast results were compared and discussed in this study by multiple methods. The summer precipitation of China in 2014 and 2015 is predicted by using LSTM network. The results show that the prediction ability of the LSTM network is better than that of the stepwise regression, Back Propagation neural network and BCC models. Parameter optimization has a great influence on the prediction effect of LSTM network. The important parameters include the number of hidden layer nodes, training times and learning rate. Selecting suitable starting months is helpful to improve the accuracy of seasonal forecast, and the forecast effect of summer precipitation is better by using the data reported from April. The sea ice component factors have made a positive contribution to seasonal precipitation forecast. In the summer precipitation return experiment in 2014 and 2015, the LSTM network has the ability to predict the overall precipitation situation. Ps score are 74 and 71, anomaly sign consistency rates are 55.63% and 55.25%. The average Ps score is higher than the national consultation and business model in the same period.

Key words: LSTM network, Machine learning, Flood season precipitation, Seasonal prediction

京ICP备11008704号-4
版权所有 © 《气候变化研究进展》编辑部
地址:北京市海淀区中关村南大街46号 邮编:100081 电话/传真:(010)58995171 E-mail:accr@cma.gov.cn
本系统由北京玛格泰克科技发展有限公司设计开发 技术支持:support@magtech.com.cn