基于机器学习的Angström-Prescott公式系数的估算  

Estimation of the Coefficients of the Angström-Prescott Formula Based on Machine Learning Methods

作　　者：冯文哲 和志豪 陈上 董文彪 李若彤 于强 冯浩[2,4] 何建强 FENG Wen-zhe;HE Zhi-hao;CHEN Shang;DONG Wen-biao;LI Ruo-tong;YU Qiang;FENG Hao;HE Jian-qiang(Key Laboratory for Agricultural Soil and Water Engineering in Arid Area of Ministry of Education,Northwest A&F University,Yangling 712100,Shaanxi Province,China;Institute of Water-Saving Agriculture in Arid Areas of China,Northwest A&F University,Yangling 712100,Shaanxi Province,China;Jiangsu Key Laboratory of Agricultural Meteorology,Nanjing University of Information Science and Technology,Nanjing 210044,China;State Key Laboratory of Soil Erosion and Dryland Agriculture on the Loess Plateau,Institute of Water and Soil Conservation,Chinese Academy of Science and Ministry of Water Resource,Yangling 712100,Shaanxi Province,China)

机构地区：[1]西北农林科技大学旱区农业水土工程教育部重点实验室,陕西杨凌712100 [2]西北农林科技大学中国旱区节水农业研究院,陕西杨凌712100 [3]南京信息工程大学江苏省农业气象重点实验室,南京210044 [4]中国科学院水利部水土保持研究所黄土高原土壤侵蚀与旱地农业国家重点实验室,陕西杨凌712100

出　　处：《节水灌溉》2024年第7期108-118,共11页Water Saving Irrigation

摘　　要：地表太阳辐射(R_(s))数据在水文、农业和生态等领域具有重要的应用价值。由于目前仅有少数国家气象站点具备直接观测条件,因此Angström-Prescott(A-P)公式被广泛应用于逐日R_(s)的估算。尽管使用A-P公式需提供的两个经验系数a和b已经有FAO(Food and Agriculture Organization)推荐值(a=0.25;b=0.5),但是越来越多的研究指出这两个参数的本地化有助于提高R_(s)的估算精度。利用1967-2017年全国80个具有太阳辐射观测数据气象站的逐日地表太阳辐射(R_(s))及其他常规气象数据,来获取中国内地地区A-P公式的a、b系数。首先,整个中国内地地区被划分为高原山地气候区(Mountain Plateau Zone,MPZ)、亚热带季风气候区(Subtropical Monsoon Zone,SMZ)、温带季风气候区(Temperate Monsoon Zone,TMZ)、温带大陆性气候区(Temperate Continental Zone,TCZ)等4个不同气候区。其次,基于最小二乘法回归得到各气候区不同站点A-P公式系数值,可视为A-P公式系数的观测值。然后,利用4种机器学习算法分别估算全国80个具有太阳辐射观测数据气象站的A-P公式系数,各算法分别结合不同输入因子组合构建不同的A-P公式系数估算模型。最后,评估机器学习算法估算得到的A-P公式系数自身的精度,及其在R_(s)估算中的精度。研究发现在估算系数a时,机器学习模型中基于五因子输入组合的SVM模型的估算精度最高(R^(2)=0.661,RMSE=0.022,nRMSE=0.120)。在估算系数b时,机器学习模型中基于四因子输入组合的ELM模型的估算精度最高(R^(2)=0.550,RMSE=0.031,nRMSE=0.055)。基于所选最优机器学习模型各自估算的a和b系数值来驱动A-P公式进一步估算R_(s),结果表明机器学习模型在MPZ、SMZ、TMZ、TCZ气候区R_(s)估算中的nRMSE分别为0.168、0.225、0.138、0.180。因此,推荐使用五因子输入组合的SVM模型来估算系数a,使用四因子输入组合的ELM模型来估算系数b,可以得到更为准确的�Surface solar radiation(R_(s))data are important in hydrology,agriculture and ecology.Since only a few national meteorological stations have direct observation conditions,the Angström-Prescott(A-P)formula is widely used to estimate daily R_(s).While the two empirical coefficients a and b required by the A-P formula have been recommended by the FAO(Food and Agriculture Organization)(a=0.25;b=0.5),recent studies have emphasized that the localization of the formula parameters could help to improve the estimation accuracy.This study used daily surface solar radiation(R_(s))and other conventional meteorological data from 80 national weather stations with solar radiation observation data from 1967 to 2017 to derive reliable A-P formula coefficients in China mainland.First,the entire Chinese mainland was divided into four climatic zones:the Mountain Plateau Zone(MPZ),Subtropical Monsoon Zone(SMZ),Temperate Monsoon Zone(TMZ),and Temperate Continental Zone(TCZ).Next,the calibrated a and b values of the A-P formula were obtained at each weather station in different climate zones through linear regression,which were regarded as the observed values of the A-P formula coefficients.Four machine learning algorithms were applied to estimate the A-P formula coefficients.Each algorithm combined different input factor combinations to construct different estimation models for A-P formula coefficients.The accuracy of the estimated A-P formula coefficients and their impact on R_(s) estimation were evaluated.Some main conclusions were drawn as follows,when estimating the coefficient a,the SVM machine learning model with the five-factor input combination had the highest estimation accuracy,with R^(2)=0.661,RMSE=0.022,and nRMSE=0.120.When estimating the coefficient b,the ELM machine learning model with the four-factor input combination had the highest estimation accuracy,with R^(2)=0.550,RMSE=0.031,and nRMSE=0.055.Based on the A-P formula and the relevant coefficients a and b estimated with the selected optimal machine learning model to

关 键 词：太阳辐射 Angström–Prescott a、b系数 机器学习 

分 类 号：S161.1[农业科学—农业气象学]