基于深度学习集合优化模型的径流区间预测研究  

Research on streamflow interval prediction based on deep learning ensemble optimization model

作　　者：黄靖涵 王兆才[1] 吴俊豪[2] 姚之远 HUANG Jinghan;WANG Zhaocai;WU Junhao;YAO Zhiyuan(College of Information,Shanghai Ocean University,Shanghai 201306,China;State Key Laboratory of Estuarine Coastal Science,East China Normal University,Shanghai 200241,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]华东师范大学河口海岸学国家重点实验室,上海200241

出　　处：《水利学报》2025年第2期240-252,265,共14页Journal of Hydraulic Engineering

基　　金：国家自然科学基金项目(11701363,52279079);中国水利水电科学研究院流域水循环模拟与调控国家重点实验室开放研究基金项目(IWHR-SKL-201905);水利部泥沙科学与北方河流治理重点实验室开放基金项目(IWHR-SEDI-2023-10)。

摘　　要：由于极端天气事件趋多趋强和径流变化的复杂性,实现准确的径流预测具有挑战性,且以往研究多基于确定数值的点预测,难以考虑不确定性影响,导致预测结果缺乏实用性。本研究开发了基于气象和水文变量的径流区间预测深度学习集合模型。首先通过皮尔逊相关系数(PCC)筛选出影响径流的关键驱动变量;接着将原始数据通过变分模态分解(VMD)分解为多个模态分量(IMFs);然后利用互补集合经验模态分解法(CEEMD)对分量进行二次分解,捕捉更多的数据细节;径流的点预测结果由融合注意力机制的门控循环单元(AM-GRU)来取得,并使用改进的麻雀优化算法(ISSA)优化GRU的学习率、隐藏层维数等超参数以提升模型性能;最后,引入了非参数核密度估计(NKDE)进行径流区间预测。采用构建的组合模型VMD-CEEMD-ISSA-AM-GRU(VCIAG)对嘉陵江流域的9个水文站点进行多期预测。计算结果表明:本文模型在短期尺度表现优异,多个站点的纳什效率系数(NSE)接近1;在洪水预报方面,模型在东津沱站、武胜站、金溪站的NSE分别为0.73、0.92和0.92;此外,通过沙普利值法(Shapley)量化了输入变量对径流的影响。本研究提出的VCIAG模型不仅在径流点预测精度方面表现出色,而且在不确定性的区间预测方面也有显著优势,可为管理者提供更加准确、可靠的径流信息,从而在实践中更好地支持径流风险评估和科学决策方案的制定。Due to the increasing occurrence of extreme weather events and the complexity of streamflow variations,it is challenging to realize accurate streamflow prediction,and previous studies are mostly based on point prediction of determinate values,which is hard to take into account the effect of uncertainty and lead to the lack of practical applicability of the prediction results.In this study,a deep learning ensemble model for streamflow interval prediction based on meteorological and hydrological variables is developed.The model first filters out the key driving variables affecting streamflow through the Pearson correlation coefficient(PCC).Then the raw data are decomposed into intrinsic mode functions(IMFs)by variational modal decomposition(VMD).The components are then quadratically decomposed using complementary ensemble empirical modal decomposition(CEEMD)to capture more details of the data.The point prediction results of streamflow are obtained by a gated recurrent unit(GRU)incorporating an attention mechanism(AM),and an improved sparrow search algorithm(ISSA)is used to optimize hyperparameters such as the learning rate of the GRU and the number of hidden layer dimensions to enhance the model performance.Finally,nonparametric kernel density estimation(NKDE)is introduced for interval prediction.The combined model VMD-CEEMD-ISSA-AM-GRU(VCIAG)constructed in this study performs advance multi-period prediction for nine hydrological stations in the Jialing River basin.The results of the study showed that the model performs well in the short term,with Nash efficiency coefficients(NSE)close to 1 for several stations.For flood forecasting,the model’s NSE for Dongjintuo,Wusheng,and Jinxi stations are 0.73,0.92,and 0.92,respectively.In addition,the effects of the input variables on runoff are quantified by the Shapley value method(Shapley).The VCIAG model proposed in this study not only performs well in streamflow prediction accuracy,but also has significant advantages in interval prediction of uncertainty,which can prov

关 键 词：深度学习集合模型 径流区间预测 模态分解 改进的麻雀优化算法 注意力机制 

分 类 号：P338[天文地球—水文科学]