基于卷积神经网络的区域温度预报订正研究

Research on the Regional Temperature Forecast Correction Based on Convolutional Neural Network

摘要: 为提高ECMWF (European Centre for Medium-Range Weather Forecasts)模式2 m温度预报产品的预报精度，选取2018年~2021年ECMWF模式2 m温度预报产品及CLDAS (CMA Land Data Assimi-lation System)格点融合数据，进行了误差分析并使用卷积神经网络模型进行订正研究，结果表明：1) ECMWF模式2 m温度预报产品夏季预报精度要高于冬季，且在转折性低温天气的情况下误差较大；2) 经过卷积神经网络模型的订正，2 m温度预报产品精度得到提高，预报准确率整体提升约5%，平均绝对误差和均方根误差均降低了0.5℃~0.8℃；3) 对比传统使用历史长期资料的订正方式，模型可以使用更少的数据，达到较好的订正效果，减少统计工作量。

Abstract: In order to improve the forecast accuracy of ECMWF (European Centre for Medium-Range Weather Forecasts) mode 2 m temperature prediction products, the 2 m temperature prediction products in ECMWF mode from 2018 to 2021 and CLDAS (CMA Land Data Assimilation System) lattice data were selected. The error analysis and the convolutional neural network model are used to carry out the revised research, and the results show that: 1) the summer forecast accuracy of ECMWF mode 2 m temperature prediction products is higher than that in winter, and the error is larger in the case of inflectional low temperature weather; 2) After the revision of the convolutional neural network model, the accuracy of the 2 m temperature prediction product has been improved, the overall prediction accuracy has been improved by about 5%, and the average absolute error and root mean square error have been reduced by 0.5˚C~0.8˚C; 3) Compared with the traditional revision method of using historical long-term data, the model can use less data, achieve better correction effect, and reduce statistical workload.