基于深度学习及传统机器学习模型估算山东省参考作物蒸散量  

Estimation of Reference Crop Evapotranspiration in Shandong Province Based on Deep Learning and Traditional Machine Learning Model

作　　者：任传栋[1] 王志真 马钊[1] 张敏[1] REN Chuan-dong;WANG Zhi-zhen;MA Zhao;ZHANG Min(Shandong Survey and Design Institute of Water Conservancy,Jinan 250013,China)

机构地区：[1]山东省水利勘测设计院,济南250013

出　　处：《节水灌溉》2022年第3期67-74,共8页Water Saving Irrigation

基　　金：国家科技支撑项目(31410005321403)。

摘　　要：为探寻深度学习模型在区域参考作物蒸散量(Reference crop evapotranspiration,ET_(0))估算中的适用性,以山东省为研究区域,选取了深度神经网络(DNN)、时间卷积神经网络(TCN)和长短期记忆神经网络(LSTM)3种深度学习模型,极限学习机模型(ELM)、广义回归神经网络模型(GRNN)和随机森林模型(RF)3种传统机器学习模型,Hargreaves-Samani模型(HS)、Droogres-Allen模型(DA)、Priestley-Tayor模型(PT)、Marrink模型(MK)、WMO模型(WMO)、Trabert模型(TRA)6种经验模型,以均方根误差(RMSE)、决定系数(R^(2))、平均绝对误差(MAE)和效率系数(E_(ns))为精度评价体系,找出了适用于山东省ET_(0)估算的最优模型,结果表明: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,而3种深度学习模型精度最优,TCN模型在所有模型中精度最高;输入辐射资料的模型精度普遍高于温度模型和质量传输模型,TCN2模型GPI为1.036,在所有模型中排名第1。因此,TCN模型为山东省ET_(0)的最优估算模型使用。In order to explore the applicability of deep learning models in the estimation of regional reference crop evapotranspiration(ET_(0)),this paper took Shandong Province as the research area and selected 3 deep learning models including deep neural network(DNN),time convolutional neural network(TCN)and long short-term memory neural network(LSTM),3 traditional machine learning models including extreme learning machine model(ELM),generalized regression Neural network model(GRNN)and random forest model(RF),and 6 empirical models including Hargreaves-Samani model(HS),Droogrs-Allen model(DA),Priestley-Tayor model(PT),Marrink model(MK),WMO model(WMO),Trabert model(TRA).Taking Root mean square error(RMSE),coefficient of determination(R^(2)),average absolute error(MAE)and efficiency coefficient(E_(ns))as the accuracy evaluation system,the optimal model for ET_(0) estimation in Shandong province was found.The results showed that the accuracy of the machine learning model was generally better than that of the empirical model under the same meteorological parameter input conditions,while the three deep learning models had the best accuracy,and the TCN model had the highest accuracy among all models.The accuracy of the model with radiation data input was generally higher than that of the temperature model and the mass transfer model.The GPI of the TCN2 model was 1.036,ranking first among all models.So the TCN model can be used as a recommended model for ET_(0) estimation in Shandong Province.

关 键 词：山东省 参考作物蒸散量 深度学习 机器学习 辐射资料 估算模型 

分 类 号：S271[农业科学—农业水土工程] TV642.4[农业科学—农业工程]