利用卷积神经网络提高天气短期气温预报效果  

Using convolutional neural network to improve the effect of short-term weather temperature prediction

作　　者：王莹 杨晓君[1] 王迪 张庆 张楠[1] WANG Ying;YANG Xiaojun;WANG Di;ZHANG Qing;ZHANG Nan(Tianjin Meteorological Observatory,Tianjin 300074,China;School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津市气象台,天津300074 [2]天津大学电气自动化与信息工程学院,天津300072

出　　处：《气象科学》2024年第4期793-800,共8页Journal of the Meteorological Sciences

基　　金：天津市气象局一般科研资助项目(202120ybxm08);福建省灾害天气重点实验室开放课题(2021KFKT05)。

摘　　要：为提高天津地区温度精细化预报的准确性,本文基于欧洲中期天气预报中心综合预报系统ECMWF-IFS、中国气象局全球同化预报系统CMA-GFS模式数据及天津259个区域自动站的逐小时气温实况数据,构建了一种基于U-Net编解码器结构的3D卷积神经网络气温预报模型,能实现天津地区24 h内逐小时气温的网格预报。采用二分搜索的方式对模型众多超参数进行调节,通过148组试验训练得到最优模型,测试集误差为1.226℃。采用多种指标对模型进行检验,结果表明,模型的预报误差整体低于原数值模式,特别是对天津市中南部(含中心城区)和东部沿海有较好的订正效果;其温度日变化预报特征更接近实况,能有效改善原数值模式的日变化预报误差,且模型表现出更强的误差稳定性。In order to improve the accuracy of fine temperature forecast in Tianjin,based on ECMWF-IFS,CMA-GFS model data of China Meteorological Administration and hourly temperature data from 259 regional automatic stations in Tianjin,a 3D convolutional neural network based on U-Net code was proposed to model hourly temperature.Many hyper-parameters were adjusted by dichotomous search method,and the optimal model was obtained by 148 groups of experimental training,and the test set error was 1.226℃.Results show that the prediction error of the model is lower than that of the original numerical model,especially for the central and southern Tianjin(including the central urban area)and the eastern coastal area.The prediction characteristics of diurnal temperature variation are closer to the actual temperature,which can effectively improve the prediction error of the original numerical model,and the model shows stronger error stability.

关 键 词：精细化温度预报 卷积神经网络 日变化预报特征 订正效果 

分 类 号：P457.3[天文地球—大气科学及气象学]