基于机器学习的CMA-MESO模式气温预报订正方法研究

doi:10.3969/j.issn.1673-503X.2025.03.003

气象与环境学报 ›› 2025, Vol. 41 ›› Issue (3): 18-28.doi: 10.3969/j.issn.1673-503X.2025.03.003

• 论文 • 上一篇

基于机器学习的CMA-MESO模式气温预报订正方法研究

张会^1,2,3,4,5, 陈军明^2,6, 王亚强^2,6, 马凤莲^1,7, 周煜^1,7, 卢宇坤^4,5, 刘通^4,5, 张良玉⁵

1. 中国气象局雄安大气边界层重点开放实验室, 河北雄安新区 071700;
2. 中国气象科学研究院灾害天气科学与技术全国重点实验室, 北京 100081;
3. 河北省气象与生态环境重点实验室, 河北石家庄 050021;
4. 保定市生态气象智能监测与服务重点实验室, 河北保定 071000;
5. 保定市气象局, 河北保定 071000;
6. 雄安气象人工智能创新研究院, 河北雄安新区 070001;
7. 雄安新区气象局, 河北雄安新区 071700

收稿日期:2024-01-02 修回日期:2024-04-18 发布日期:2025-09-29
通讯作者: 陈军明,男,研究员,E-mail:chenjm@cma.gov.cn。 E-mail:chenjm@cma.gov.cn
作者简介:张会,女,1976年生,高级工程师,从事人工智能气象应用技术工作,E-mail:zhhqx2000@163.com。
基金资助:
中国气象局能力提升联合研究专项(24NLTSZD01)和河北省保定市科技局(2211ZG001)共同资助。

A study on the temperature forecast correction method of the CMA-MESO model based on machine learning

ZHANG Hui^1,2,3,4,5, CHEN Junming^2,6, WANG Yaqiang^2,6, MA Fenglian^1,7, ZHOU Yu^1,7, LU Yukun^4,5, LIU Tong^4,5, ZHANG Liangyu⁵

1. China Meteorological Administration Xiong'an Atmospheric Boundary Layer Key Laboratory, Xiong'an New Area 071700, China;
2. State Key Laboratory of Severe Weather Meteorological Science and Technology, Beijing 100081, China;
3. Hebei Key Laboratory of Meteorology and Ecological Environment, Shijiazhuang 050021, China;
4. Baoding Key Laboratory of Intelligent Monitoring and Service on Ecological Meteorology, 071000, China;
5. Baoding Meteorological Service, Baoding 071000, China;
6. Xiong'an Institute of Meteorological Artificial Intelligence, Xiong'an New Area 070001, China;
7. Xiong'an New Area Meteorological Service, Xiong'an New Area 071700, China

Received:2024-01-02 Revised:2024-04-18 Published:2025-09-29

摘要/Abstract

摘要： 为提高雄安新区及上游保定地区气温预报准确性,利用CMA-MESO中尺度天气模式预报产品和地面观测数据,采用线性回归(Linear Regression)、长短期记忆全卷积网络(LSTM-FCN)和轻量级梯度提升机(LightGBM)等三种机器学习方法,围绕站点划分和特征因子选择设计4种订正方案开展气温预报研究。结果显示:采用分区站点模型优于全站点模型,LightGBM在全部方案中表现最优。特别是结合起报时刻前48 h的实况数据以及预报时刻前4·k的预报或实况要素(在预报时效0～36 h 内:0～12 h预报时效采用预报时刻前0～12 h的实况数据,k取0～12;13～36 h预报时效采用预报时刻前12h的预报数据,k固定为12)构建复合特征因子,LightGBM预测性能进一步提升。37个预报时效均在CMA-MESO模式预报结果基础上提高了精度,尤其在海拔超过1000 m的高原地区,均方根误差订正改善率超过30%。此外,在转折性天气背景下,这些方法依旧展现出较强的适应能力。从整体预报性能来看,LightGBM最优,均方根误差、平均绝对误差、准确率分别为1.86 ℃、1.42 ℃、75%,相比CMA-MESO预报,分别提高了36.5%、38.9% 和44.4%。

关键词: CMA-MESO, 预报订正, 机器学习, 气温预报

Abstract: To improve the accuracy of temperature forecasts in Xiong'an New Area and the upstream Baoding region,this study utilizes forecast products from the CMA-MESO mesoscale weather model and surface observation data.Three machine learning methods-Linear Regression,Long Short-Term Memory Fully Convolutional Network(LSTM-FCN),and Light Gradient Boosting Machine(LightGBM) are applied.Four forecast correction schemes are designed,focusing on station classification and feature selection.The results show that models using regionally divided stations outperform those using all stations collectively,and LightGBM delivers the best performance among all schemes.Specifically,when composite feature factors are constructed by combining observed data from 48 hours prior to the forecast start time and forecast or observed variables from 4·k hours before the forecast time(within the 0-36 h lead time:for lead times 0-12 h,actual observations from the 0-12 h period before the forecast time are used,with k ranging from 0-12; for lead times 13-36 h,forecast data from 12 h before the forecast time are used,with k fixed at 12),the predictive performance of LightGBM is further improved.For all 37 forecast lead times,the accuracy is improved over the original CMA-MESO model forecasts.Particularly in plateau regions with elevations above 1000 meters,the RMSE improvement exceeds 30%.Moreover,these methods continue to demonstrate strong adaptability under transitional weather conditions.In terms of overall forecasting performance,LightGBM proves to be the best,achieving a root mean square error(RMSE)of 1.86 ℃,a mean absolute error(MAE)of 1.42 ℃,and an accuracy of 75%,representing improvements of 36.5%,38.9%,and 44.4% respectively compared to the CMA-MESO forecast.

Key words: CMA-MESO, Forecast correction, Machine Learning, Temperature forecast

中图分类号:

P457.3

张会, 陈军明, 王亚强, 马凤莲, 周煜, 卢宇坤, 刘通, 张良玉. 基于机器学习的CMA-MESO模式气温预报订正方法研究[J]. 气象与环境学报, 2025, 41(3): 18-28.

ZHANG Hui, CHEN Junming, WANG Yaqiang, MA Fenglian, ZHOU Yu, LU Yukun, LIU Tong, ZHANG Liangyu. A study on the temperature forecast correction method of the CMA-MESO model based on machine learning[J]. Journal of Meteorology and Environment, 2025, 41(3): 18-28.

参考文献

[1] 代刊,曹勇,钱奇峰,等.中短期数字化天气预报技术现状及趋势[J].气象,2016,42(12):1445-1455.
[2] 曾晓青,薛峰,赵瑞霞,等.几种格点化温度滚动订正预报方案对比研究[J].气象,2019,45(7):1009-1018.
[3] Cui B,Toth Z,Zhu Y J,et al.Bias correction for global ensemble forecast [J].Weather and Forecasting,2012,27(2):396-410.
[4] 尹姗,李勇,马杰,等.延伸期温度预报误差订正技术初探[J].气象,2020,46(3):412-419.
[5] Taylor A A,Leslie L M.A single-station approach to model output statistics temperature forecast error assessment [J].Weather and Forecasting,2005,.20(6):1006-1020.
[6] 王焕毅,谭政华,杨萌,等.三种数值模式气温预报产品的检验及误差订正方法研究[J].气象与环境学报,,2018,34(1):22-29.
[7] 佟华,胡江林,张玉涛.GRAPES模式后处理技术改进应用研究[J].气象科技,2020,48(4):511-517.
[8] 蔡怡,徐枝芳,朱克云,等.CMA-MESO 3 km系统2 m温度预报诊断[J].气象,2023,49(4):400-414.
[9] 罗菊英,周建山,闫永财.基于数值预报及上级指导产品的本地气温MOS预报方法[J].气象科技,2014,42(3):443-450.
[10] 王婧,徐枝芳,范广洲,等.GRAPESRAFS系统2 m温度偏差订正方法研究[J].气象,2015,41(6):719-726.
[11] 何珊珊,蓝盈,戚云枫.GRAPES-GFS模式2 m温度预报的最优时窗滑动订正方法[J].气象科技,2021,49(5):746-753.
[12] 贺圣平,王会军,李华,等.机器学习的原理及其在气候预测中的潜在应用[J].大气科学学报,2021,44(1):26-38.
[13] 孙健,曹卓,李恒,等.人工智能技术在数值天气预报中的应用[J].应用气象学报,2021,32(1):1-11.
[14] 王涛,王乙舒,赵春雨,等.基于机器学习方法的辽宁省初霜冻日期预测模型研究[J].气象与环境学报,2022,38(4):47-56.
[15] 邱贵强,时少英,王洪霞,等.2 m气温集成订正方法及在冬奥延庆赛区的应用[J].应用气象学报,2023,34(4):400-412.
[16] 蒋维东,张荣智,陈博,等.基于BP神经网络的地基微波辐射计雷暴预报试验研究[J].气象与环境学报,2023,39(5):106-112.
[17] 陈昱文,黄小猛,李熠,等.基于ECMWF产品的站点气温预报集成学习误差订正[J].应用气象学报,2020,31(4):494-503.
[18] 方楠,谢国权,阮小建,等.长短期记忆神经网络(LSTM)模型在低能见度预报中的应用[J].气象与环境学报,2022,38(5):34-41.
[19] 李浙华,肖安,郑丽君.基于高分辨率数值预报和深度学习的地面气温预报研究[J].高原气象,2024,43(2):464-477.
[20] 熊世为,郁凌华,胡姗姗,等.基于ECMWF细网格产品的一种优化BP-MOS气温预报方法[J].干旱气象,2017,35(4):668-673.
[21] 陈鹤,蔡荣辉,陈静静,等.基于深度学习方法的气温预报技术应用与评估[J].气象,2022,48(11):1373-1383.
[22] Zhao L,MO C Y,Ma J J,et al.LSTM-MFCN:a time series classifier based on multi-scale spatial-temporal features [J].Computer Communications,2022,182:52-59.
[23] Ke G L.,Meng Q.,Finley T.,et al.LightGBM:a highly efficient gradient boosting decision tree[C].//31st Conference on Neural Information Processing Systems.Long Beach:Curran Associates Inc.,2017:1-9.
[24] 黄丽萍,邓莲堂,王瑞春,等.CMA-MESO 关键技术集成及应用[J].应用气象学报,2022,33(6):641-654.
[25] Karim F,Majumdar S,Darabi H,et al.LSTM Fully Convolutional Networks for Time Series Classification [J].IEEE Access,2018,6:1662-1669.
[26] Zhang H,Wang Y Q,Chen D D,et al.Temperature Forecasting Correction Based on Operational GRAPES-3km Model Using Machine Learning Methods [J].Atmosphere,2022,13,362.
[27] 中国气象局.全国智能网格天气预报业务规定:气发〔2019〕54号[Z].2017,7.

基于机器学习的CMA-MESO模式气温预报订正方法研究

A study on the temperature forecast correction method of the CMA-MESO model based on machine learning

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

参考文献

相关文章 10

Metrics

本文评价

推荐阅读 0

[1]	张艳晴, 金晨曦, 闵晶晶, 韩超, 董颜, 齐晨. 基于机器学习的北京供暖季气温预报误差订正[J]. 气象与环境学报, 2025, 41(1): 58-65.
[2]	达瓦次仁,落追,洪一航,次仁卓嘎. 基于机器学习方法的气溶胶对西藏高原地区雨季降水的影响[J]. 气象与环境学报, 2024, 40(3): 138-144.
[3]	吴宇恒,白景昌. 基于遥感和机器学习的中国典型城市碳排放驱动因子及预测分析[J]. 气象与环境学报, 2024, 40(1): 105-112.
[4]	王佳津, 王彬雁, 肖递祥, 龙柯吉. 基于目标对象检验法的四川省两种数值模式暴雨预报对比分析[J]. 气象与环境学报, 2023, 39(6): 44-50.
[5]	张煦庭, 刘慧, 刘瑞芳, 巨菲, 刘嘉慧敏, 高星星, 黄少妮, 王楠. 基于极端梯度提升算法的西安市逐小时PM_2.5浓度预报研究[J]. 气象与环境学报, 2023, 39(1): 44-54.
[6]	王涛, 王乙舒, 赵春雨, 王小桃, 秦美欧, 沈玉敏, 侯依玲, 赵建云. 基于机器学习方法的辽宁省初霜冻日期预测模型研究[J]. 气象与环境学报, 2022, 38(4): 47-56.
[7]	孟莹莹, 李树岭, 赵玲, 白雪梅, 闫敏慧. 三种气温预报产品黑龙江省预报效果检验[J]. 气象与环境学报, 2022, 38(2): 21-30.
[8]	张博,于超,黄威,陈传雷. 欧亚持续性阻塞对6月大兴安岭地区气温和降水的影响及气温的前兆信号研究[J]. 气象与环境学报, 2021, 37(4): 63-69.
[9]	王远谋,李家启,陈施吉,唐家萍,夏佰成,韩世刚. 基于机器学习的重庆长江航道雾图像特征识别研究[J]. 气象与环境学报, 2021, 37(1): 106-112.
[10]	孟莹莹, 曹殿斌, 吴岩, 王子洋. 大小兴安岭林区风场动力降尺度对比研究[J]. 气象与环境学报, 2019, 35(4): 8-15.