基于TSA-DRNN模型的年径流预测研究  

Research on Annual Runoff Prediction Based on TSA-DRNN Model

作　　者：崔东文 杨琼波 CUI Dongwen;YANG Qiongbo(Water Affairs Bureau of Wenshan Prefecture,Yunnan Province,Wenshan 663000,China;Honghe Branch,Yunnan Provincial Hydrology and Water Resources Bureau,Honghe 661100,China)

机构地区：[1]云南省文山州水务局,云南文山663000 [2]云南省水文水资源局红河分局,云南红河661100

出　　处：《华北水利水电大学学报（自然科学版）》2021年第6期35-41,共7页Journal of North China University of Water Resources and Electric Power：Natural Science Edition

摘　　要：为了解决深度递归神经网络(DRNN)权值和阈值难以选取的问题,有效提高DRNN径流预测精度,提出了被囊群算法(TSA)与DRNN相融合的预测方法。选取4个标准测试函数对TSA进行仿真验证,并与粒子群优化(PSO)算法的仿真结果进行比较;通过主成分分析(PCA)对数据样本进行降维并构建DRNN_(2)(双隐层DRNN)、DRNN_(3)(三隐层DRNN)、DRNN_(4)(四隐层DRNN)模型,利用TSA优化DRNN权值和阈值,建立了TSA-DRNN_(2)、TSA-DRNN_(3)、TSA-DRNN_(4)径流预测模型,并构建TSA-Elman、Elman、DRNN_(2)、DRNN_(3)、DRNN_(4)、TSA-SVM模型作对比;利用云南省姑老河站年径流预测实例对TSA-DRNN_(2)、TSA-DRNN_(3)、TSADRNN_(4)、TSA-Elman、Elman、DRNN_(2)、DRNN_(3)、DRNN_(4)、TSA-SVM模型进行检验。结果表明:在不同维度条件下,TSA仿真效果优于PSO算法,具有较好的寻优精度和全局搜索能力;TSA-DRNN_(2)、TSA-DRNN_(3)、TSADRNN_(4)模型对实例年径流预测的平均相对误差分别为3.63%、2.81%、2.50%,预测精度优于TSA-Elman等其他6种模型,且随着隐含层数的增加,预测精度呈提高趋势。TSA-DRNN模型用于径流预测是可行的,模型及DRNN权、阈值优化方法可为相关预测研究提供参考。In order to solve the problem of difficult selection of the weights and thresholds of the deep recurrent neural network(DRNN),and effectively improve the accuracy of DRNN runoff prediction,a prediction method combining tunicate swarm algorithm(TSA)and DRNN is proposed.Four standard test functions are selected for simulation verification of TSA and compared with the simulation results of the particle swarm optimization(PSO)algorithm.Principal component analysis(PCA)is applied to reduce the dimensionality of data samples and build DRNN_(2)(dual hidden layer DRNN),DRNN_(3)(three hidden layer DRNN),DRNN_(4)(four hidden layer DRNN)models.TSA is used to optimize DRNN weights and thresholds,and TSA-DRNN_(2),TSA-DRNN_(3) and TSA-DRNN_(4) runoff prediction models are established,and TSA-Elman,Elman,DRNN_(2),DRNN_(3),DRNN_(4) and TSA-SVM models are constructed for comparison.The TSA-DRNN_(2),TSA-DRNN_(3),TSA-DRNN_(4),TSA-Elman,Elman,DRNN_(2),DRNN_(3),DRNN_(4) and TSA-SVM models are tested by the example of annual runoff prediction of Gulao River Station in Yunnan Province.The results are as follows.Under different dimensional conditions,the TSA simulation effect is better than the PSO algorithm,and it has better optimization accuracy and global search capability.The average relative errors of TSA-DRNN_(2),TSA-DRNN_(3) and TSA-DRNN_(4) models for the annual runoff prediction of the example are 3.63%,2.81%and 2.50%,respectively.The prediction accuracy is better than other six models such as TSA-Elman.As the number of hidden layers increases,the prediction accuracy tends to improve.The TSA-DRNN model is feasible for runoff prediction,and the model and DRNN weight threshold optimization method can provide reference for related prediction research.

关 键 词：径流预测 深度递归神经网络(DRNN) 被囊群算法(TSA) 仿真验证 数据降维 权、阈值优化 

分 类 号：P338.2[天文地球—水文科学]