基于深度学习模型集成的日最高和最低气温订正预报研究  

Daily Maximum and Minimum Temperature Forecasts Correction Based on Deep Learning Model Ensemble

作　　者：卢姝 郭可萌 周悦[4] 傅承浩 许霖 顾雪 LU Shu;GUO Kemeng;ZHOU Yue;FU Chenghao;XU Lin;GU Xue(Hunan Key Laboratory of Meteorological Disaster Prevention and Reduction,Changsha 410118,China;Hunan Meteorological Observatory,Changsha 410118,China;College of Artificial Intelligence,China University of Petroleum,Beijing 102249,China;Institute of Heavy Rain of China Meteorological Administration/Hubei Key Laboratory for Heavy Rain Monitoring and Warning Research/China Meteorological Administration Basin Heavy Rainfall Key Laboratory,Wuhan 430205,China;Xiangxi Prefecture Meteorological Bureau,Xiangxi,Hunan 416000,China)

机构地区：[1]气象防灾减灾湖南省重点实验室,湖南长沙410118 [2]湖南省气象台,湖南长沙410118 [3]中国石油大学人工智能学院,北京102299 [4]中国气象局武汉暴雨研究所/暴雨监测预警湖北省重点实验室/中国气象局流域强降水重点开放实验室,湖北武汉430205 [5]湘西州气象局,湖南湘西416000

出　　处：《热带气象学报》2024年第6期1018-1029,共12页Journal of Tropical Meteorology

基　　金：湖南省气象局2024年创新发展专项(青年专项,CXFZ2024-QNZX23);湖北省自然科学基金气象联合基金(2023AFD096);武汉市自然科学基金项目(2024020901030454)共同资助。

摘　　要：采用2018—2023年中国气象局陆面数据同化系统的气温资料以及欧洲中期天气预报中心的高分辨率模式预报产品(ECMWF-IFS),分别建立基于时空堆叠的残差网络(Res-STS)以及基于自注意力(Self-Attention)机制的长短期记忆网络(Attention-LSTM),并将两个模型进行集成,构建集成神经网络模型(Ensemble),得到涵盖湖南地区的0.05°×0.05°气温网格日最高、最低气温预报产品。结果表明:深度学习模型均有效改善了ECMWF-IFS预报效果,0—24 h预报时效日最高气温的平均绝误差(MAE)相比ECMWF-IFS和中央气象台指导报(SCMOC)分别降低了25.76%~40.40%和15.03%~31.79%,日最低气温的MAE分别降低了10.53%~31.58%和5.31%~19.47%,其中Ensemble模型在绝大多数月份的预报效果均是最优。同时,Ensemble模型有效弥补了ECMWF-IFS对地形复杂区域预报效果弱的缺陷,日最高气温预报准确率(F2)达85%的面积占比为17.31%,而其余模型低于6%;日最低气温F2达90%的面积占比为68.63%,高出单一模型21.08%~63.09%。由此可见,多模型集成能够显著提高气温预报的准确性和可靠性。Using temperature data from the China Meteorological Administration(CMA)Land Data Assimilation System and high-resolution forecast products from the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System(ECMWF-IFS)during 2018—2023,we developed two deep learning frameworks:a residual spatiotemporal stacking network(Res-STS)and a self-attention long short-term memory network(Attention-LSTM).An ensemble model was subsequently developed from these two models to produce 0.05°×0.05°gridded temperature forecasts specific to the Hunan region.Validation results for the year of 2023 indicate that the deep learning models effectively improved the accuracy of ECMWF-IFS forecasts.For the 0~24 h forecasts,the mean absolute error(MAE)of daily maximum temperature was reduced by 25.76%~40.40%compared to the ECMWF-IFS products and by 15.03%~31.79%compared to the products from the System of Central Meteorological Observatory for Correction(SCMOC),CMA.The MAE of daily minimum temperature was reduced by 10.53%~31.58%compared to the ECMWF-IFS products and by 5.31%~19.47%compared to products from the SCMOC,with the ensemble model performing the best.Furthermore,the ensemble model effectively mitigated the limitation of the ECMWF-IFS in forecasting within complex terrain.The proportion of areas achieving an F2 score of 85%for daily maximum temperature was 17.31%,mainly in the Dongting Lake plain area,whereas it was below 6%in other models.For daily minimum temperature,the area with an F2 score of 90%reached 68.63%,which was 21.08%~63.09%higher than those of other models.Overall,the ensemble model exhibited superior forecasting performance in most months.The integration of multiple models and deep learning can significantly enhance the reliability and accuracy of temperature forecasts.

关 键 词：Res-STS Attention-LSTM 多模型集成 ECMWF-IFS 气温预报 

分 类 号：P45[天文地球—大气科学及气象学]