基于机器学习算法和ARIMA模型的旱地春小麦产量预测  

Comparison of Spring Wheat Yield Prediction Models in Dryland Based on Machine Learning and ARIMA Model

作　　者：董莉霞[1] 张博 李广[2] 燕振刚[1] 逯玉兰 DONG Lixia;ZHANG Bo;LI Guang;YAN Zhengang;LU Yulan(College of Information Science and Technology,Gansu Agricultural University,Lanzhou,Gansu 730070,China;College of Forestry,Gansu Agricultural University,Lanzhou,Gansu 730070,China)

机构地区：[1]甘肃农业大学信息科学技术学院,甘肃兰州730070 [2]甘肃农业大学林学院,甘肃兰州730070

出　　处：《麦类作物学报》2024年第12期1551-1559,共9页Journal of Triticeae Crops

基　　金：甘肃省重点研究发展计划项目(22YF7FA116);甘肃省财政专项(GSCZZ 20160909);甘肃省高等学校产业支撑项目(2021CYZC-15);甘肃省高等学校产业支撑项目(2022CYZC-41)。

摘　　要：为探究机器学习算法与时间序列结合预测春小麦产量的可行性,使用HP滤波算法,将1971-2021年甘肃省定西市安定区和2014-2021年定西市渭源县的产量数据分离为气象产量和趋势产量,利用研究区的气象数据,分别基于随机森林(random forest,RF)、循环神经网络(recurrent neural network,RNN)、支持向量机(support vector machine,SVM)、反向传播神经网络(back propagation neural network,BPNN)和长短期记忆网络(long short-term memory,LSTM)5种机器学习算法实现春小麦气象产量的预测对比;构建了ARIMA时间序列模型,探究旱地春小麦趋势产量的最佳预测模型。结果表明,旱地春小麦产量中气象产量占比大,趋势产量占比小,且气象产量的变化趋势基本与总产量一致。LSTM模型对研究区内气象产量的模拟效果最佳。通过参数率定,获得最优的趋势产量预测模型为ARIMA(4,1,2)模型,其残差基本为白噪声,符合正态分布。利用LSTM和ARIMA(4,1,2)模型组合,对2014-2021年渭源县的春小麦产量进行预测,并与研究区的实际产量数据比较,预测结果与实际值接近,模型精度较高,其R^(2)为0.96,MAPE为1.22%,MAE为32.12 kg·hm^(-2),RMSE为35.32 kg·hm^(-2)。在本研究条件下,LSTM和ARIMA(4,1,2)组合模型具有良好的预测精度,可实现旱地春小麦产量的预测。In order to explore the feasibility of combining machine learning algorithm and time series to predict spring wheat yield,the yield data of Anding District,Dingxi City,Gansu Province from 1971 to 2021 and Weiyuan County,Dingxi City,Gansu Province from 2014 to 2021 were separated into meteorological yield and trend yield using HP filtering algorithm in this study.Using meteorological data from the study area,Random forest,RF,Recurrent Neural Network(RNN),Support Vector Machine(SVM),Back Propagation Neural Network(Back Propagation Neural Network),and Long Short-Term Memory(LSTM)the five machine learning algorithms(BPNN)were used to predict and compare spring wheat meteorological yield.The ARIMA time series model was constructed to explore the best prediction model for the trend yield of spring wheat in dryland.The results showed that the proportion of meteorological yield was large while the proportion of trend yield was small.The change trend of meteorological yield was basically consistent with the total yield.LSTM model had the best simulation effect on meteorological yield in the study area.Through parameter calibration,the optimal trend yield prediction model is ARIMA(4,1,2)model,where residual is basically white noise and accords with normal distribution.The combination of LSTM and ARIMA(4,1,2)models was used to forecast the spring wheat yield in Weiyuan County from 2014 to 2021,and compared with the actual yield data in the study area,the predicted result was close to the actual value,with high model accuracy(R^(2)=0.96,MAPE=1.22%,MAE=32.12 kg·hm^(-2),and RMSE=35.32 kg·hm^(-2)).Under the conditions of this study,the combined model of LSTM and ARIMA(4,1,2)has good prediction accuracy,which can realize the prediction of the yield of spring wheat in dryland.

关 键 词：产量分离 时间序列 机器学习 旱地 春小麦 产量预测 

分 类 号：S512.1[农业科学—作物学] S311