不同核函数高斯过程回归算法与不同因子输入情况下对长江流域蒸散发量应用研究  

Application of Gaussian Process Regression Algorithm with Different Kernel Function and Different Factor Inputs on Evapotranspiration in the Yangtze River Basin

作　　者：杨梓涵 崔峥铮 张鹏程 YANG Zi-han;CUI Zheng-zheng;ZHANG Peng-cheng(School of Hydrology and Water Resources,Hohai University,Nanjing 210024,China)

机构地区：[1]河海大学水文水资源学院,南京210024

出　　处：《水利科技与经济》2023年第9期19-25,共7页Water Conservancy Science and Technology and Economy

摘　　要：为探明不同核函数高斯过程回归算法在不同使用条件下对参考作物腾发量(ET 0)模拟精度,在长江流域内选择10个代表性气象站点,以PM公式的计算结果作为参考值,以最高气温、最低气温、平均气温、相对湿度、平均本站大气压、日照时数和风速作为主要气象因子,使用灰色关联分析得到因子输入组合,使用二次有理、平方指数、Matern 5/2等3种不同核函数的高斯过程回归算法对ET 0进行模拟,并与Priestley-Taylor、Hargreaves-Samani、Irmak-Allen等3种经典算法计算结果进行对比。结果显示:①在同一站点同一参考公式计算结果下,3种不同核函数高斯过程回归算法和3种经典算法的模拟精度大小排序为:Matern 5/2>二次有理>平方指数>PT>IA>HS,其中Matern 5/2的模拟效果最好,其R 2范围为0.970~0.988。表明在相同气象参数输入条件下,机器学习模型精度普遍优于经验模型。②针对灰色关联分析得到的结果,日最高温度对参考作物腾发量影响较大,其平均关联度为0.8969;日照时数对参考作物腾发量影响较小,其平均关联度为0.8105;其余气象因子对参考作物腾发量的影响适中。③针对不同因子组合输入下同种核函数的高斯过程回归算法,3种不同核函数高斯过程回归算法的模拟ET 0表现效果均为:六因子>五因子>四因子,其中六因子输入的模拟效果最好,其R 2范围为0.908~0.977。In order to explore the simulation accuracy of different kernel function Gaussian process regression algorithms on the emission rate(ET 0)of reference crops under different conditions,10 representative meteorological stations were selected in the Yangtze River Basin,and the calculation results calculated by the PM formula were used as reference values,the maximum temperature,minimum temperature,average temperature,relative humidity,average atmospheric pressure,sunshine hours and wind speed of the station were used as the main meteorological factors,and the factor input combination was obtained by gray correlation analysis,and the quadratic rational,squared index and Matern 5/2 were used.The Gaussian process regression algorithm with three different kernel functions simulated ET 0 and compared the calculation results of Priestley-Taylor,Hargreaves-Samani and Irmak-Allen.The results show that:①under the calculation results of the same reference formula at the same site,the simulation accuracy of the Gaussian process regression algorithm of three different kernel functions and the three classical algorithms are:Matern 5/2>quadratic rational>squared exponent>PT>IA>HS,Matern 5/2 works best with an R 2 range of 0.970~0.988.The results show that the accuracy of the machine learning model is generally better than that of the empirical model under the input conditions of the same meteorological parameters.②According to the results obtained by gray correlation analysis,the maximum daily temperature had a greater effect on the emission of the reference crop,and the average correlation degree was 0.8969,the average correlation degree of sunshine hours had a small effect on the emission of the reference crop,and the average correlation degree was 0.8105,and the influence of other meteorological factors on the emission of the reference crop was moderate.③For the Gaussian process regression algorithm of the same kernel function under different factor combination inputs,the simulated ET 0 performance effect of the Gaussian

关 键 词：参考作物腾发量 灰色关联分析 高斯过程回归 核函数 

分 类 号：S152.7[农业科学—土壤学]