基于XGBoost和数值天气预报的黄淮海平原参考作物蒸散量预测模型研究  

Forecasting of Reference Crop Evapotranspiration in Huang-Huai-Hai Plain Based on XGBoost Machine Learning Model and Numerical Weather Prediction Data

作　　者：朱春霞 秦安振[2] ZHU Chunxia;QIN Anzhen(Center for Soil and Fertilizer and Agricultural Resources Protection of Zhoukou City,Zhoukou,Henan 466000;Institute of Farmland Irrigation,Chinese Academy of Agricultural Sciences,Xinxiang,Henan 453002)

机构地区：[1]周口市土壤肥料和农业资源保护中心,河南周口466000 [2]中国农业科学院农田灌溉研究所,河南新乡453002

出　　处：《中国农学通报》2024年第28期126-133,共8页Chinese Agricultural Science Bulletin

基　　金：河南省重点研发与推广专项(科技攻关)(222102110175);新乡市科技攻关计划项目(GG2021024)。

摘　　要：为提高黄淮海平原参考作物蒸散量(ET_(0))的预测精度,在豫北地区新乡市利用2020—2021年历史气象数据和2022年日数值天气预报数据(最高气温、最低气温、太阳总辐射量、日照时数、相对湿度和2 m风速),建立反向传播(BP)、极限梯度提升(XGBoost)和梯度提升决策树(CatBoost)3种预测ET_(0)的机器学习模型,并与FAO-56 Penman-Monteith模型的结果进行比较。结果显示,气象参数中太阳总辐射量(Ra)、最高气温(T_(max))和最低气温(T_(min))与ET_(0)的相关性最高,可作为模型的输入因子。从预测时间尺度来看,3种机器学习模型对1~16 d的ET_(0)预报效果最佳。其中,XGBoost模型在验证期的R=0.875、RMSE=0.230 mm/d、MAE=0.181 mm/d、MAPE=8.45%。R较CatBoost和BP模型平均提高10.2%,RMSE、MAE和MAPE平均下降39.9%~62.4%。鉴于XGBoost模型预测ET_(0)的精度和稳定性,推荐将其作为黄淮海平原参考作物蒸散量的预测方法。This study seeks to enhance the prediction accuracy of reference crop evapotranspiration(ET_(0))in the Huang-Huai-Hai Plain.Three models,namely BP,XGBoost,and CatBoost,were developed utilizing historical daily meteorological data from 2020-2021.These data included maximum air temperature(T_(max)),minimum air temperature(T_(min)),total solar radiation(Ra),sunshine hours(S),relative humidity(RH),and wind speed at a height of 2 m(U2)for training purposes.The models were tested and forecasted using daily numerical weather prediction data from Xinxiang city,located in northern Henan Province,for the year 2022.The forecasted results were compared with the ET_(0) data calculated using FAO-56 Penman-Monteith model.The results showed that Ra,T_(max),and T_(min) were most correlated with ET_(0) among all factors,and therefore were considered as input factors for model running.The three models generated acceptable ET_(0) accuracy at 1-16 d forecast scale.In model testing,XGBoost model had R=0.875,RMSE=0.230 mm/d,MAE=0.181 mm/d,and MAPE=8.45%,respectively.On average,the R value of XGBoost model was increased by 10.2%,and values of RMSE,MAE,and MAPE were decreased by 39.9%-62.4%,compared to BP and CatBoost models.In view of the accuracy and stability of the XGBoost model,it can be a recommended model for ET_(0) forecasting in the Huang-Huai-Hai Plain.

关 键 词：P-M模型 BP神经网络模型 参考作物蒸散量 机器学习 预测模型 黄淮海平原 极限梯度提升(XGBoost) 梯度提升决策树(CatBoost) 

分 类 号：F301.2[经济管理—产业经济]