中国农业大学学报2025，Vol.30Issue(4)：38-50,13.DOI:10.11841/j.issn.1007-4333.2025.04.04
融合多头注意力机制的LSTM冬小麦需水量预测
LSTM winter wheat water demand prediction based on multi-head attention mechanism
摘要
Abstract
To realize the accurate prediction of winter wheat field water demand,based on the measured meteorological data of the experimental field in Yuanyang City,Henan Province,the multi-head attention mechanism was adopted to effectively obtain the relational features between water demand time series according to the size of the correlation coefficients between the meteorological factors and the crop water demand.The 11-,7-and 4-factor fused multi-head attention(MHA)convolutional-long-short-term memory network(CNN-LSTM)model for predicting winter wheat water demand and comparing with the actual values.The results showed that:1)The coefficient of determination(R2)between the predicted and measured values of crop water requirement was 0.914,the root mean square error(RMSE)was 0.627 mm,and the relative analysis error(RPD)was 4.243.The R2 was higher and RMSE was lower than that of the 7-factor and 4-factor prediction models.The accuracy of this prediction model increased with more inputs of meteorological factors related to crop water demand.The accuracy of the 4-factor prediction model was the lowest,and the R2 between the predicted value and the measured value of crop water requirement was 0.825,RMSE was 0.946 mm,and RPD was 3.124.In the absence of meteorological data,only the minimum temperature,maximum temperature,average temperature and atmospheric pressure with the highest correlation with the reference crop water requirement could be used to predict the crop water requirement.The number of sensors was reduced and the universality was improved compared with that of 11 factors.2)Compared with the classical machine learning model BP,the recurrent neural network model LSTM and the LSTM-Attention with attention mechanism,the CNN-LSTM-MHA model with multi-head attention mechanism has better parameters such as the R2,RMSE and RPD,and the prediction effect is closer to the actual situation of field production.In summary,the CNN-LSTM-MHA crop water requirement prediction model based on the fusion of 11 meteorological factors related to crop water requirement is established,which can improve the problem of insufficient internal feature extraction between data in the convolutional layer.The model can effectively improve the prediction accuracy of winter wheat water requirement and can be used for irrigation decision-making of winter wheat in the field.关键词
小麦/作物需水量/神经网络/多头注意力机制/预测模型Key words
wheat/crop water requirement/neural network/multi-head attention mechanism/prediction model分类
农业科技引用本文复制引用
汪强,李清阳,席磊,樊泽华,马新明,时雷,李美琳,卢建龙,熊淑萍..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):38-50,13.基金项目
"十四五"国家重点研发计划(2023YFD2301503) （2023YFD2301503）
河南省重大科技专项(221100110800) （221100110800）
河南省重点研发计划(241111111500) （241111111500）
