nav emailalert searchbtn searchbox tablepage yinyongbenwen piczone journalimg journalInfo journalinfonormal searchdiv searchzone qikanlogo popupnotification paper paperNew
基于多头注意力机制的蒙古黄芪需水量预测研究
基金项目(Foundation): 财政部和农业农村部国家现代农业产业技术体系资助项目(CARS-21); 巴彦淖尔市科技计划项目(K2024014); 高等教育改革发展项目(DC2500000813)
邮箱(Email): ;
DOI:
发布时间: 2026-11-05
出版时间: 2026-11-05
网络发布时间: 2026-11-05
移动端阅读
摘要:

为精准预测蒙古黄芪需水量,基于内蒙古农业大学海流图科技园区试验田实测气象数据,构建了融合卷积神经网络(CNN)、长短期记忆网络(LSTM)与多头注意力机制(MHA)的CNN-LSTM-MHA预测模型。以Penman-Monteith公式计算的作物需水量为基准,以8种气象因子为输入。结果表明:模型的决定系数(R2 )为0.963,均方根误差(RMSE)为0.312 mm,相对分析误差(RPD)为5.208;与GRU、DLinear、Informer、Transformer等主流模型相比,该模型在各项评价指标上均表现最优,预测值与实测值时序曲线高度一致。消融试验验证了CNN空间特征提取、LSTM时序建模和多头注意力机制融合的有效性与必要性。该模型可为蒙古黄芪规模化种植的智慧灌溉决策提供可靠支持。

Abstract:

To accurately predict the water requirement of Astragalus membranaceus var.mongholicus, this study constructed a CNN-LSTMMHApredi ction model-integrating convolutional neural network(CNN), long short-term memory network(LSTM), and mul ti-head attention mechanism(MHA)-based on measured meteorological data from the experimental field at the Ocean Current Map Scienceand Technology Park of Inner Mongolia Agricultural University. The model uses crop water requirements calculated by the Penman-Monteith equation as the target v ariable and eight meteorol ogical factors as input features. Results show that the model achieves a coefficient of determination(R2) of 0. 963, a root mean square error(RMSE) of 0. 312 mm, and a relative prediction deviation(RPD) of 5. 208. Compared with mainstream mode lssuch as GRU, DLinear, Informer, and Transformer, this model achieves the best performance across all evaluation metrics, and the times eries of predicted values exhibits high consistency with the observed values. Ablation experiments verify the effectiveness and necessity of fusingC NN-based spatial feature extraction, LSTM-based temporal modeling, and the multi-head attention mechanism. This model can provide relia blesupport for intelligent irrigation decision-making in large-scale cultivation of A. membranaceus.

参考文献

[1] Fong B N,Reba M L,Teague T G,et al. Eddy covariance measurements of carbon dioxide and water fluxes in US mid-south cotton production[J]. Agric Ecosyst Environ,2020,292:106813.

[2]王妍,张晓龙,石嘉丽,等.中国冬小麦主产区气候变化及其对小麦产量影响研究[J].中国生态农业学报(中英文),2022,30(5):723-734.

[3] Sidhu R K,Kumar R,Rana P S. Long short-term memory neural network-based multi-level model for smart irrigation[J]. Mod Phys Lett B,2020,34(36):2050418.

[4]刘婧然,刘心,武海霞,等.基于GA优化的支持向量机模型在青椒作物需水量预测中的应用[J].节水灌溉,2021(1):70-76.

[5]邓皓,李文竹,刘婧然,等.基于MIV-MEA-Elman神经网络的核桃果实膨大期需水量预测[J].节水灌溉,2020(4):68-72.

[6]孟玮,孙西欢,郭向红,等.基于人工蜂群径向基神经网络预测参考作物需水量[J].节水灌溉,2020(1):79-83.

[7]李志新,赖志琴,龙云墨.基于GA-Elman神经网络的参考作物需水量预测[J].节水灌溉,2019(2):117-120.

[8]汪强,李清阳,席磊,等.融合多头注意力机制的LSTM冬小麦需水量预测[J].中国农业大学学报,2025,30(4):38-50.

[9]李洁松,刘韬,伍星.基于皮尔逊相关系数统计特征滚动轴承RUL预测[J].航空动力学报,2025,40(7):92-101.

[10]何立新,牛浩然,张峥,等.基于STL分解的灌区典型作物需水预测模型研究[J].中国农村水利水电,2026(6):179-184,194.

[11]胡程达,方文松,王红振,等.河南省冬小麦农田蒸散和作物系数[J].生态学杂志,2020,39(9):3004-3010.

[12]杜云,张婧婧,雷嘉诚,等.冬小麦需水量的预测模型对比分析[J].新疆农业科学,2024,61(7):1590-1596.

[13] Sun B K,Wang W G,Liu G S,et al. Projecting the impact of climate change and elevated CO2 concentration on rice irrigation water requirement in China[J]. Sci Total Environ,2024,912:168489.

[14] Ohana-Levi N,Ben-Gal A,Munitz S,et al. Grapevine crop evapotranspiration and crop coefficient forecasting using linear and non-linear multiple regression models[J]. Agric Water Manag,2022,262:107317.

[15]李丰琇,刘泽龙,左顺利. SIMDual_Kc蒸散模型在南疆滴灌夏玉米的适用性[J].农业工程学报,2025(5):68-78.

[16]姬丽君,席旭东,晋小军.蒙古黄芪物候期研究[J].草业学报,2013,22(1):60-67.

[17]戴小枫,田洪岭.中国道地药材黄芪[M].北京:人民卫生出版社,2025.

[18]杨晓娟,刘园,白薇,等.黑龙江省大豆作物系数的确定[J].干旱地区农业研究,2014(6):53-56.

[19]符太东,韩秋香.基于大数据的智能灌溉决策系统设计研究[J].农机化研究,2024,46(11):86-89,95.

基本信息:

中图分类号:S274;S567.239

引用信息:

[1]汪栋,张圆梦,张雄杰,等.基于多头注意力机制的蒙古黄芪需水量预测研究[J].安徽农业科学().

基金信息:

财政部和农业农村部国家现代农业产业技术体系资助项目(CARS-21); 巴彦淖尔市科技计划项目(K2024014); 高等教育改革发展项目(DC2500000813)

发布时间:

2026-11-05

出版时间:

2026-11-05

网络发布时间:

2026-11-05

检 索 高级检索

引用

GB/T 7714-2015 格式引文
MLA格式引文
APA格式引文