基于BO-CNN-LSTM的锡林郭勒草原干旱预测模型  

Drought prediction model for the Xilingol grassland based on BO-CNN-LSTM

作　　者：杜娟 董世杰 贺云 DU Juan;DONG Shu-jie;HE Yun(College of Statistics and Mathematics,Inner Mongolia University of Finance and Economics,Hohhot 010070,China;College of Water Resources and Civil Engineering,Inner Mongolia Agricultural University,Hohhot 010018,China)

机构地区：[1]内蒙古财经大学统计与数学学院,内蒙古呼和浩特010070 [2]内蒙古农业大学水利与土木建筑工程学院,内蒙古呼和浩特010018

出　　处：《草原与草坪》2024年第4期64-75,共12页Grassland and Turf

基　　金：内蒙古自治区高等学校科学研究项目(NJZY23053);内蒙古自然科学基金(2023QN01006,2023LHMS01012);内蒙古财经大学自治区直属高校基本科研业务费项目(NCYWT23027);一流学科科研专项(YLXKZXNCD-010)。

摘　　要：【目的】建立基于BO-CNN-LSTM耦合神经网络的干旱预测模型,探索干旱预测的适用性。【方法】首先,基于长短期记忆网络(LSTM)的记忆功能,将其嵌入卷积神经网络(CNN)全连接层。其次,为确定最优超参数,将贝叶斯优化算法(BO)的概率代理模型和采集函数引入至LSTM。最后,建立BO-CNN-LSTM耦合神经网络模型用以预测干旱状况。【结果】1) BO-CNN-LSTM预测精度随时间尺度的增大而提高,对SPEI-12模拟精度最高,且判定系数R2均在98%以上;2)与LSTM模型SPEI-12的模拟结果进行比较,BO-CNN-LSTM表现出更高拟合精度。其中R2相对提高值为[4.63%,8.67%],MSE的数量级由10-2降至10-3;3)通过BO-CNN-LSTM预测2023年锡林郭勒草原干旱空间分布,结果显示该区域整体呈干旱态势。其中东乌珠穆沁旗站点区域属于中旱,其它区域均属于重旱。【结论】BO-CNN-LSTM具有较高的计算精度,尤其适用于预测SPEI-12,故可将其应用于年时间尺度干旱预测。【Objective】Reliable and effective monitoring can mitigate the impact of drought disasters on socio-economic development and natural ecosystems.This study adopted the BO-CNN-LSTM coupled neural network as adrought prediction model.【Method】First,the memory function of long short-term memory(LSTM)was inte-grated into the fully connected layer of the convolutional neural network(CNN).Second,to determine the optimal hyperparameters for LSTM,the probability surrogate model and acquisition function from the Bayesian optimization(BO)algorithm wereintroduced.Finally,a BO-CNN-LSTM coupled neural network model was constructed to predict the drought conditions in the Xilingol grassland.【Result】(1)The prediction accuracy of the BO-CNN-LSTM model increased with the time scale,withthe highest prediction accuracy observed under the 12-month scale for the Standardized Precipitation-Evapotranspiration Index(SPEI).The determination coefficient R^(2) of SPEI-12 for each site exceeded 98%.(2)Compared to the simulation results of the LSTM model for SPEI-12,the proposed model exhibited higher fitting accuracy,showinga relative improvement in R^(2) of[4.63%,8.67%].The order of magnitude of mean squared error(MSE)at each site had decreased from 10-2 to 10-3.(3)Using the model to predict the spatial distribution of drought in the Xilingol grassland for 2023.indicated that the region as a whole was experi-encing drought.Especially,the Dongwuzhumuqin Banner area was classified as experiencing moderate drought,while other areas were classified as severe drought.【Conclusion】The results demonstrate that the BO-CNN-LSTM model has high computational accuracy,making it particularly suitable for predicting SPEI-12.Therefore,the meth-ods in this study can be effectively applied to drought prediction on an annual time scale.

关 键 词：干旱预测 贝叶斯优化算法 卷积神经网络 长短期记忆网络 锡林郭勒草原 

分 类 号：S812.5[农业科学—草业科学]