LSTM模型在嘉陵江流域参考作物蒸散量计算中的适用性研究  

Applicability of LSTM Model for Calculating Reference Crop Evapotranspiration in the Jialing River Basin

作　　者：葛杰 康秀 白钰 张晓鹏 董超超 王晶 GE Jie;KANG Xiu;BAI Yu;ZHANG Xiao-peng;DONG Chao-chao;WANG Jing(PowerChina Northwest Engineering Corporation Limited,Xiʹan 710065,Shaanxi Province,China;Hebei Tianhe Consulting Co.,Ltd,Shijiazhuang 050000,Hebei Province,China;Xianyang Water Conservancy and Hydropower Planning Survey and Design Institute,Xianyang 712000,Shaanxi Province,China)

机构地区：[1]中国电建集团西北勘测设计研究院有限公司,陕西西安710065 [2]河北天和咨询有限公司,河北石家庄050000 [3]咸阳市水利水电规划勘测设计研究院,陕西咸阳712000

出　　处：《中国农村水利水电》2024年第11期62-70,共9页China Rural Water and Hydropower

基　　金：陕西省科技统筹创新计划项目(2016KTZDNY-01-01)。

摘　　要：为有效提高缺乏气象数据时嘉陵江流域参考作物蒸散量(ET0)计算精度,选取流域及周边20个气象站58 a(1960-2017年)逐日气象数据,基于不同气象要素组合,构建16种基于长短期记忆网络(Long Short-Term Memory,LSTM)的ET0计算模型,并将其与Hargreaves-Samani、Makkink和Irmark-Allend等3种在嘉陵江流域计算精度较高的模型进行对比,研究LSTM模型在嘉陵江流域的适用性。结果表明:①LSTM模型能精准地捕捉输入参数与ET0之间复杂的非线性关系,基于最高气温(T_(max))、最低气温(T_(min))和地球外辐射(R_(a))建立的LSTM2模型计算精度能达到应用要求(MAPE平均为14.6%,RMSE平均为0.476 mm/d,NSE平均为0.891,R2平均为0.903),模型计算精度随输入气象要素数量的增加而升高。②模型输入参数中增加R_(a),可有效提高模型计算精度(MAPE平均降低17.3%,RMSE平均降低11.1%,NSE平均升高0.779%,R2平均升高0.715%)。③基于最高气温(T_(max))、最低气温(T_(min))、地球外辐射(R_(a))和日照时间(n)构建的LSTM12模型是嘉陵江流域缺乏气象数据时最适宜的ET0计算模型。④LSTM模型的计算精度均优于相同输入参数依赖下的Hargreaves-Samani、Makkink和Irmark-Allen模型。⑤LSTM模型在嘉陵江流域具有很强的可移植性,不同气象站点建立的LSTM相互移植时,能保持较高精度(MAPE低于7.42%,RMSE低于0.242 mm/d,NSE高于0.972,R2高于0.980)。因此,基于长短期记忆网络(LSTM)建立的ET0模型在嘉陵江流域具有很好的适用性,可作为气象数据缺乏时嘉陵江流域ET0计算的推荐模型。In order to effectively improve the calculation accuracy of reference crop evapotranspiration(ET0)in the Jialing River basin when meteorological data is lacking,daily meteorological data in the past 58 years(1960-2017)were collected from twenty meteorological stations in Jialing River basin and it′s vicinity,16 ET0 calculation models were built based on Long Short-Term Memory(LSTM)according to different combinations of meteorological elements.The calculation accuracy of these models was compared with the Hargreaves-Samani,Makkink,and Irmark-Allen models to investigate the applicability of LSTM models in the Jialing River basin.The results showed that:①LSTM models could accurately capture the complex nonlinear relationship between the input parameters and ET0.The LSTM2 model,which was based on the maximum temperature(T_(max)),the minimum temperature(T_(min)),and the extraterrestrial radiation(R_(a)),achieved a high calculation accuracy(on average,MAPE=14.6%,RMSE=0.476 mm/d,NSE=0.891,R2=0.903),meeting the requirements for practical applications.The calculation accuracy of the model increased with the number of input meteorological elements.②If R_(a) is added to the input parameters,it is a significant improvement in model accuracy(on average,MAPE is reduced by 17.3%,RMSE is reduced by 11.1%,NSE is increased by 0.779%,and R2 is increased by 0.715%).③The LSTM12 model based on T_(max),T_(min),R_(a),and sunshine time(n)is the most suitable ET0 calculation model when there is a lack of meteorological data in the Jialing River Basin.④Under the same input parameters,the calculation accuracy of the LSTM model was better than that of the Hargreaves-Samani,Makkink,and Irmark-Allen models.⑤The LSTM model demonstrated strong portability in the Jialing River basin.When LSTM models established at different meteorological stations were transplanted to each other,they still maintained high accuracy(MAPE<7.42%,RMSE<0.242 mm/d,NSE>0.972,R2>0.980).Therefore,the ET0 model based on Long Short Term Memory Network(LSTM)has go

关 键 词：嘉陵江 参考作物蒸散量 缺乏资料 LSTM 适用性 

分 类 号：TV213.9[水利工程—水文学及水资源] S274.3[农业科学—农业水土工程]