摘要
为了科学准确预测参考作物蒸腾量(ET0),提高预测精度,减少输入变量的数量,从而降低智能节水灌溉系统的建设成本,采用深度学习和人工神经网络方法分别建立ET0智能预测模型,采用局部敏感性分析、模糊曲线和模糊曲面等方法研究ET0预测中各输入变量对预测结果的影响,以影响因子大小为依据,构建8种不同气象因子输入组合,利用日照气象站的逐日气象资料,对采用不同方法和不同输入变量组合的预测模型进行训练和测试,并以彭曼公式的计算结果作为参考,对预测模型的性能进行评估。结果表明,在以完整变量作为预测输入时,深度学习预测模型R2为0.980,高于人工神经网络模型(0.963),获得了更高的预测精度;而在缺省输入变量的ET0预测中,深度学习预测模型的性能均优于人工神经网络,以平均温度和日照时数作为输入变量的深度学习预测模型R2仍达到0.935,表明在仅有少量气象参数的情况下,深度学习预测模型仍能获得较好的预测结果。综合分析R2、RMSE、RMSRE、MRE、MAPE等结果,若研究区域气候数据有限,采用输入组合分别为(n、T、RH、Tmin、Ws)和(n、T、RH、Ws)的深度学习模型预测,其结果与彭曼公式的计算结果相比,MAPE分别为8.753和8.404,R2均大于0.98,可以作为标准预测模型。
In order to predict crop reference transpiration(ET0)scientifically and accurately,improve the prediction accuracy,and reduce the number of input variables,so as to reduce the construction cost of intelligent water-saving irrigation system,in this paper,deep learning and artificial neural network methods were used to establish ET0 intelligent prediction models,and local sensitivity analysis,fuzzy curve,and fuzzy surface methods were used to study the influence of each input variable in ET0 prediction on the prediction results.Based on the influence factor size,eight input combinations of different meteorological factors were constructed.The daily meteorological data of Rizhao weather station were used to train and test the prediction models with different methods and input variable combinations.The results of Penman formula were used as reference to evaluate the performance of the prediction models.The results showed that the R2 of the deep learning prediction model was 0.980,which was higher than that of the artificial neural network model which was 0.963 when the complete variables were used as the prediction input,and higher prediction accuracy was obtained.For ET0 prediction of default input variables,the performance of the deep learning prediction model was superior to that of the artificial neural network,and the R2 of the deep learning prediction model with the input parameters of average temperature and Sunshine duration still reached 0.935,indicating that the deep learning prediction model could still obtain good prediction results when there were only a few meteorological parameters.Through comprehensive analysis of R2,RMSE,RMSRE,MRE,MAPE,and other results,it could be seen that if the regional climate data of the study was limited,the deep learning model with input combinations of(n,T,RH,Tmin,Ws)and(n,T,RH,Ws)could be used to predict,compared with the results calculated by Penman formula,the MAPE values were 8.753 and 8.404,respectively,and the R2 values were both greater than 0.98,which could be used as a standard prediction model.
作者
潘振华
刘子菡
沈欣
张钟莉莉
史凯丽
张石锐
PAN Zhenhua;LIU Zihan;SHEN Xin;ZHANG Zhongii;SHI Kaii;ZHANG Shirui(Intelligent Equipment Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;Key Laboratory for Quality Testing of Software and Hardware Products on Agricultural Information,Ministry of Agriculture and Rural Affairs,Beijing 100097,China;Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;National Agro-Technical Extension and Service Center,Beijing 100125,China)
出处
《山西农业科学》
2023年第8期942-952,共11页
Journal of Shanxi Agricultural Sciences
基金
财政部和农业农村部:国家现代农业产业技术体系建设项目(CARS-02-87)
八师石河子市财政科技计划项目(2021ZD01)
北京市农林科学院创新能力建设专项储备性研究项目(KJCX20210411)。
关键词
参考作物蒸散量
深度学习
模糊曲线
模糊曲面
预测模型
决定系数
reference crop evapotranspiration
deep learning
fuzzy curves
fuzzy surfaces
prediction models
coefficient of determination
