基于机器学习技术的蒸发皿蒸发量估算模型  

Estimation Model of Pan Evaporation Based on Machine Learning Technology

作　　者：龙亚星[1,2] 黄勤 李成伟[1] LONG Yaxing;HUANG Qin;LI Chengwei(Shanxi Provincial Atmospheric Detection Technical Support Center,Xi’an 710015;Key Laboratory of Eco-Environment and Meteorology for the Qinling Mountains and LoesPlateau,Xi’an 710016;Sha nxi Provincial Meteorological Information Center,Xi’an 710015)

机构地区：[1]陕西省大气探测技术保障中心,西安710015 [2]秦岭和黄土高原生态环境气象重点实验室,西安710016 [3]陕西省气象信息中心,西安710015

出　　处：《气象科技》2021年第2期166-173,共8页Meteorological Science and Technology

基　　金：秦岭和黄土高原生态环境气象重点实验室开放研究基金课题(2020G-11)资助。

摘　　要：为了弥补国家级气象观测站小型蒸发皿停止观测后蒸发量观测资料的空缺,建立了陕北、关中和陕南3个区域数据集以及榆林、泾河和汉中3个单站数据集,通过建立和优化KNN、MLP模型及其参数,分别建立蒸发量区域估算模型、单站估算模型并对其进行检验。结果表明:(1)进行区域蒸发量估算时,KNN模型表现出良好的泛化性能,均方误差、总相对误差和准确率指标值平均分别为0.42、2.1%、57.0%;陕北MLP模型的泛化性能较差;(2)进行单站蒸发量估算时,基于k近邻法的单站估算模型性能优于区域估算模型,均方误差、准确率指标值平均分别为0.48、55.0%,榆林与泾河总相对误差指标绝对值平均为1.6%,汉中总相对误差指标值相对偏高,达到10.3%。本研究为不同气候区域及单站日、月、季和年蒸发皿蒸发量估算以及日蒸发量数据质量控制提供了一种基于机器学习的方法。In order to make up for the lack of evaporation data after the stop of evaporation pan manual observations at the National Meteorological Observatory,three regional datasets of the northern Shaanxi,Guanzhong and southern Shaanxi and three single station datasets of Yulin,Jinghe and Hanzhong are established.By establishing and optimizing the KNN(K-Nearest Neighbor method)and MLP(MultiLayer Perceptron)models and its parameters,the regional estimation model of evaporation and the single station estimation model are constructed and verified respectively.The results show that:(1)While estimating the regional evaporation,the KNN model shows good generalization performance,and the average Mean Square Error,Total Relative Error and Correct Rate values are 0.42 and 2.1%,57.0%,respectively;the generalization performance of the MLP model in the northern Shaanxi is poor.(2)While estimating the evaporation of a single station,the performance of the single station estimation model based on the K-nearest neighbor method is superior to the regional estimation model,and the average Mean Square Error and Correct Rate index values are 0.48 and 55.0%,the absolute average value of Total Relative Error at Yulin and Jinghe 1.6%,and that at Hanzhong is relatively high,reaching 10.3%.This research provides a tool based on the Machine Learning for the estimation of daily,monthly,seasonal and annual evaporation in different climate regions and single stations and the quality control of daily evaporation data.

关 键 词：K近邻法 神经网络 蒸发量 估算 检验 

分 类 号：P412[天文地球—大气科学及气象学]