基于BiLSTM-NFC的地下水埋深预测方法研究  

Research on Groundwater Depth Prediction Method Based on BiLSTM-NFC

作　　者：刘鑫 韩宇平 刘中培 黄会平 LIU Xin;HAN Yuping;LIU Zhongpei;HUANG Huiping(School of Water Conservancy,North China University of Water Resources and Electric Power,Zhengzhou 450046,China;College of Surveying and Geo-Informatics,North China University of Water Resources and Electric Power,Zhengzhou 450046,China)

机构地区：[1]华北水利水电大学水利学院,河南郑州450046 [2]华北水利水电大学测绘与地理信息学院,河南郑州450046

出　　处：《人民黄河》2021年第6期80-85,97,共7页Yellow River

基　　金：国家自然科学基金资助项目(51679089);水利部“948”项目(201328)。

摘　　要：为了提高地下水埋深预测的精度,提出了双向长短时记忆循环神经网络(BiLSTM)融合非全连接神经网络(NFC)的深度学习模型。使用自适应矩估计优化函数(Adam),耦合双曲正切(Tanh)、软最大逻辑回归(Softmax)和线性整流单元(ReLU)3个激活函数,且将学习率设置为动态的,以黄河下游人民胜利渠灌区1993—2018年的地下水埋深预测为例,将BiLSTM-NFC与BiLSTM、长短时记忆循环神经网络(LSTM)及LSTM-NFC的预测结果进行对比分析。结果表明:双向网络的性能优于单向网络,NFC可以防止过拟合,还能明显降低模型的均方误差(MSE);与BiLSTM、LSTM-NFC和LSTM相比,BiLSTM-NFC的学习能力、稳定性、可靠性及泛化能力最强;BiLSTM-NFC在测试集上的准确率(Acc)可以达到100%,最接近无偏估计,MSE比LSTM的减小96.60%,平均相对误差(MRE)减小85.63%,相关系数(r)增大34.81%;模型在图形处理单元(GPU)上比在中央处理单元(CPU)上训练时间明显缩短,合理设置多种激活函数可以解决单一激活函数的弊端;使用BiLSTM-NFC可以准确地预测地下水埋深的变化情况。In order to improve the accuracy of groundwater depth prediction,a deep learning model based on bi-directional long short term memory recurrent neural network(BiLSTM)with non-fully connected neural network(NFC)was proposed.The model used adaptive moment estimation optimization function(Adam),hyperbolic tangent(Tanh),softmax logistic regression(Softmax)and rectified linear unit(ReLU),and the learning rate was set to be dynamic.Taking the underground depth prediction of the People’s Victory Canal irrigation area in the lower Yellow River from 1993 to 2018 as an example,the prediction results of BiLSTM-NFC,BiLSTM,long short term memory recurrent neural network(LSTM)and LSTM-NFC models were compared.The results show that the performance of the bi-directional network is better than that of the unidirectional network,NFC can prevent over-fitting and significantly reduce the mean square error(MSE)of the models.Compared with BiLSTM,LSTM-NFC and LSTM,the BiLSTM-NFC model has the strongest learning ability,stability,reliability and generalization ability.The accuracy(Acc)of BiLSTM-NFC model on the test set can reach 100%,which is the closest to unbiased estimation.Compared with LSTM model,the MSE of BiLSTM-NFC model is reduced by 96.60%,the mean relative error(MRE)is reduced by 85.63%and the correlation coefficient(r)is increased by 34.81%.The training time of the model on the graphics processing unit(GPU)is significantly shorter than that on the central processing unit(CPU).The disadvantages of single activation function can be solved by setting multiple activation functions reasonably.Using the model structure and algorithm set up in this study,the change of groundwater depth can be predicted accurately.

关 键 词：地下水埋深预测 双向长短时记忆循环神经网络 非全连接神经网络 深度学习模型 自适应矩估计优化函数 耦合激活函数 动态学习率 

分 类 号：TV211.1[水利工程—水文学及水资源] TV551.412