机器学习算法对黄河上游人工增雨效果检验的应用  

Application of machine learning algorithm in the effect verification of artificial rainfall on the upper reaches of the Yellow River

作　　者：张玉欣 康晓燕 侯永慧 薛丽梅 田建兵 ZHANG Yuxin;KANG Xiaoyan;HOU Yonghui;XUE Limei;TIAN Jianbing(Meteorological Disaster Prevention Technology Center in Qinghai Province,Xining 810000,China;Key Laboratory for Disaster Prevention and Mitigation in Qinghai Province,Xining 810000,China)

机构地区：[1]青海省气象灾害防御技术中心,西宁810000 [2]青海省防灾减灾重点实验室,西宁810000

出　　处：《气象科学》2024年第6期1154-1162,共9页Journal of the Meteorological Sciences

基　　金：第二次青藏高原综合科学考察研究项目(2019QZKK0104);国家自然科学基金资助项目(42165008);西北区域人影建设研究项目(RYSY201903);青海省自然科学基金资助项目(2021-ZJ-745)。

摘　　要：利用1976—2020年黄河上游8个国家站实测降水数据结合74项环流指数,应用RF和BP神经网络算法,建立了分降水等级的月降水等级和日降水等级预测模型,并检验两种模型对实际降水的预测效果和模型的稳定程度,最终选取合适的预测模型估计自然降水量并结合实际增雨作业和实际降水量检验作业效果。结果表明:两种算法的月、日降水等级模型模拟结果均较好,样本占总样本的75%以内时模型准确率达80%,但对偏极端的月降水量预测结果较差。RF算法建立的模型更为稳定,在非线性作用和缺失上均有较好的包容性,而BP神经网络建立的日降水模型中预测期的误差大于训练期,模型的稳定性相对较差。对2019—2020年甘德县10次火箭增雨作业的效果进行检验,2020年9月18日作业两种算法均预测该日降水为1级,但实际降水等级为2级,认为作业效果有效。Based on the precipitation data from 8 national stations in the upper Yellow River from 1976 to 2020 combined with 74 circulation indexes,the random forest algorithm and the BP neural network algorithm were used to establish the monthly precipitation grade and daily precipitation grade prediction models,and then the prediction effect of the two models on actual precipitation as well as the stability of the model were tested.The influence of precipitation was then tested by combining the actual precipitation and the predicted precipitation,and finally an appropriate prediction model was chosen to estimate natural precipitation.Results show that the two algorithms perform better in simulating monthly and daily precipitation,with model accuracy reaching 80%when samples account for less than 75%of the total samples.However,the results for the prediction of extreme monthly precipitation are subpar.The daily precipitation model created by the BP neural network has an error in the prediction period that is higher than that in the training period,and the model's stability is relatively poor.In contrast,the model created by the RF algorithm was more stable and has good tolerance for nonlinear effects and omissions.In Gande County,from 2019 to 2020,the impact of ten rocket operations to increase rainfall1was examined.The two algorithms both projected level 1 precipitation for September 18,2020,however level 2 precipitation actually fell on that day.The effectiveness of the artificial rainfall enhancement operation is thought to be good.

关 键 词：机器学习 环流指数 降水 随机森林 神经网络 

分 类 号：P48[天文地球—大气科学及气象学]