基于随机森林特征选择与POA-LSTM组合的参考作物腾发量预测方法  

Random Forest Feature Selection and POA-LSTM Combination Based Prediction Method for Reference Evapotranspiration

作　　者：李越 岳春芳[1] 陈大春[1] LI Yue;YUE Chun-fang;CHEN Da-chun(College of Hydraulic and Civil Engineerin g,Xinjiang Agricultural University,Urumqi 830052,Xinjiang,China)

机构地区：[1]新疆农业大学水利与土木工程学院,新疆乌鲁木齐830052

出　　处：《节水灌溉》2025年第1期120-128,共9页Water Saving Irrigation

基　　金：新疆维吾尔自治区重大科技专项项目(2022A02003-5)。

摘　　要：为了更好地捕捉参考作物腾发量(ET_(0))数据的非线性特点及有效影响因素,实现对气象资料缺乏时的ET_(0)精准预测,基于融合建模思想提出了一种随机森林特征选择与鹈鹕优化算法(POA)优化长短期记忆神经网络(LSTM)组合的ET_(0)预测方法。首先,采用随机森林特征选择方法筛选出有效气象因子作为模型输入;随后,通过POA搜索最优超参数组合用于优化LSTM模型;最后,基于最优超参数下的LSTM模型进行ET_(0)预测。结果表明,POA-LSTM模型整体优于其余模型,其中POA-LSTM1(u_(2)、N、R_(H)、T_(mean))预测精度最高,测试集R^(2)、RMSE和MAE分别为0.927、0.778和0.400 mm/d;POA-LSTM4(u_(2)、N)也能较好地适应少量气象参数估算ET_(0),测试集R^(2)、RMSE和MAE分别为0.881、0.995和0.510 mm/d,相较于其他方法,具有更高的预测精度和稳定性。In order to better capture the nonlinear characteristics and effective influencing factors of reference crop evapotranspiration(ET_(0))data,and to achieve accurate ET_(0)prediction when meteorological information is lacking,an ET_(0)prediction method based on the fusion modeling idea of a combination of Random Forest Feature Selection and Pelican Optimization Algorithm(POA)optimized Long Short Term Memory Neural Network(LSTM)is proposed.First,Random forest feature selection was used to evaluate the importance of the features and filter the effective weather factors as model inputs;subsequently,the optimal hyperparameter combinations are searched by POA for optimizing the LSTM model;finally,the ET_(0)prediction was performed based on the LSTM model under the optimal hyperparameters.The results show that the POA-LSTM model outperformed the other models,among which POA-LSTM1(u_(2)、N、R_(H)、T_(mean))has the highest prediction accuracy,with the test set R^(2),RMSE and MAE of 0.927,0.778,and 0.400 mm/d,respectively;POA-LSTM4(u_(2)、N)also demonstrated good performance in estimating ET_(0)with fewer meteorological inputs,with the test set R^(2),RMSE and MAE of 0.881、0.995 and 0.510 mm/d,with higher prediction accuracy and stability compared to other methods.

关 键 词：参考作物腾发量 长短期记忆神经网络 随机森林 特征选择 鹈鹕优化算法 

分 类 号：S161.4[农业科学—农业气象学]