区域气象站日气温极值延长方法研究与误差分析  

Study on Regression Approaches of Extending Daily Temperature Extremes for Local Weather Stations and the Error Analysis

作　　者：殷悦 吴利红[1] 严睿恺 马浩[1] 刘昌杰 Yin Yue;Wu Lihong;Yan Ruikai;Ma Hao;Liu Changjie(Zhejiang Climate Center,Hangzhou 310017,China;x092.Zhejiang Meteorological Observatory,Hangzhou 310017,China)

机构地区：[1]浙江省气候中心,杭州310017 [2]浙江省气象台,杭州310017

出　　处：《气象与环境科学》2023年第4期95-103,共9页Meteorological and Environmental Sciences

基　　金：国家重点研发计划(2019YFD1002201);浙江省气象局一般项目(2021YB06)。

摘　　要：区域气象站由于建站时间短,难以提供具有气候统计学意义的气象资料,因此对区域站短序列气象要素进行延长是部分区域气象服务(如农业保险气象指数产品开发)的需求。为此,采用单元线性回归、逐步回归、神经网络三种建模方案,基于浙江省66个国家站气压、气温、相对湿度、风速、日照等具有天气意义的气象要素,对浙江省100个随机选取的区域站逐日最低/高气温开展延长方法的探究,并对模型的估计误差进行分析。结果表明,相较于传统单元线性回归,利用多因子的逐步回归和神经网络能有效提高延长数据的准确性。神经网络由于能描述非线性统计关系,估计的准确性优于基于线性统计的逐步回归的准确性。三种建模方案的误差分析显示:逐日最低气温的估计误差在68月较小、12月次年2月较大,逐日最高气温的估计误差在12月次年2月较小、35月较大;样本数的增加对模型的估计准确性有一定提升作用;较高海拔区域站逐日最低/高气温的估计误差,高于较低海拔区域站的估计误差。It is difficult for local weather stations(LWSs)to provide a long-time series of meteorological observation data which are of statistical significance in climate studies due to the short history of LWSs.Therefore,it is necessary to extend the short-time series of meteorological variables observed at LWSs for some regional meteorological services(such as agricultural insurance meteorological index product development).Using three regression schemes(the univariate linear regression,stepwise regression,and neural network),and based on the meteorological elements of weather significance such as air pressure,temperature,relative humidity,wind speed and sunshine at 66 national stations in Zhejiang Province,we investigate the daily minimum/maximum temperature extension method for 100 randomly chosen LWSs in Zhejiang Province and analyze the estimation errors of different schemes.The result shows that the multi-factor stepwise regression and neural network can effectively improve the estimation accuracy of the extended data,compared with the traditional univariate linear regression.The neural network is capable of revealing nonlinear statistical relationship,so it performs better than the stepwise regression that only describes linear relationship.The error analysis of the three schemes suggests that the mean absolute error(MAE)of daily minimum temperature is smaller from June to August and larger from December to the next February,while the MAE of daily maximum temperature is smaller from December to the next February and larger from March to May.With the rising number of samples,the estimation accuracy is improved for all models to some extent.The estimation errors of daily minimum/high temperature of LWSs at higher latitudes are larger than those at lower-latitude LWSs.

关 键 词：区域气象站 资料延长 气温极值 逐步回归 神经网络 

分 类 号：P413.21[天文地球—大气科学及气象学]