基于自动LightGBM的贵州局地大气加权平均温度模型构建  

Model Establishment of Atmospheric Weighted Mean Temperature in Guizhou Regin based on Automatic LightGBM

作　　者：方省 张琼莉 张显云[1] FANG Sheng;ZHANG Qiongli;ZHANG Xianyun(Mining College,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学矿业学院,贵州贵阳550025

出　　处：《贵州大学学报（自然科学版）》2024年第5期112-117,共6页Journal of Guizhou University:Natural Sciences

基　　金：贵州大学“SRT计划”资助项目(贵大SRT字[2022]332号)。

摘　　要：针对贵州地形起伏大、探空站数量少,以及现有大气加权平均温度(T_(m))模型不能很好地刻画T_(m)及其垂向变化的空间差异性和日变化特征等问题,结合贝叶斯超参数优化和LightGBM机器学习方法各自的优势,提出了一种顾及T_(m)及其垂向变化的空间差异性、年周期、季节周期和日变化特征的自动机器学习建模方法,并以包围威宁探空站的4个ERA5格网点为例,构建了一种无气象参数依赖的贵州局地T_(m)经验模型(WNTm模型)。实验结果表明:WNTm模型在训练集和验证集上均取得了较高的拟合精度,其不仅可以诊断出T_(m)的日变特征,还能较好地刻画T_(m)的垂向变化趋势;以探空站气象资料计算的T_(m)为参考值,WNTm模型相比于目前较优的GPT3模型取得了更高的预测精度,平均绝对误差和均方根误差分别降低了14.63%和20.14%。该研究方法和思路可为进一步改善T_(m)的精度提供一种新的途径。To address the problems of large topographic fluctuation,small number of radiosonde in Guizhou,and the fact that the existing weighted mean temperature(T_(m))model can not describe the spatial difference and diurnal variation of T_(m) and its vertical variation.This paper combined the advantages of Bayes hyperparameter optimization and LightGBM machine learning method,thus proposed an automatic machine learning modeling method,which took into account the spatial difference,annual cycle,seasonal cycle and diurnal variation of T_(m) and its vertical variation.A local empirical model of T_(m)(WNTm model)in Guizhou province without meteorological parameter dependence is constructed by taking 4 ERA5 grid points surrounding Weining Sounding Station as an example.The experimental results show that the WNTm model achieves high fitting accuracy on both training set and verification set.It can not only diagnose the diurnal characteristics of T_(m),but also describe the vertical variation trend of T_(m) well.Compared with the current better GPT3 model,WNTm model can achieve higher prediction accuracy,and the mean absolute error and root-mean-root error are reduced by 14.63%and 20.14%respectively.The research methods and ideas in this paper can provide a new way to further improve the accuracy of T_(m).

关 键 词：大气加权平均温度 日变化特征 垂向变化 自动LightGBM 

分 类 号：P228.9[天文地球—大地测量学与测量工程]