摘要
机理模型预测评估土石坝渗流安全性态,物理意义明确、可解释性好,但是预测精度波动性较大。通过麻雀搜索算法(SSA)与径向基函数(RBF)对渗透系数进行反演并构建SSA-RBF渗压预测代理模型,得到模型预测值与残差序列;通过变分模态分解(VMD)将残差序列进行分解,并通过长短时记忆网络(LSTM)进行训练得到残差序列修正模型;将机理模型与数据驱动模型叠加构建得到SSA-RBF-VMD-LSTM融合模型,并实现对渗压水位准确预测。工程实例表明:本文提出的模型具有较高预测精度,相比于统计模型、LSTM模型和SSA-RBF-LSTM模型,其预测精度提高了89.64%、69.59%、60.45%,且在过程线出现较大幅度变动时,该模型仍能够及时给出准确的预测值,模型稳定性与外推能力较好,具有推广使用价值。
The mechanistic models can predict and evaluate the seepage safety state of earth-rock dams,which offer clear physical meaning and good interpretations,but their prediction accuracy fluctuates greatly.To enhance this accuracy,a fusion model that incorporates a data-driven deep learning approach was introduce in this study,and the Sparrow Search Algorithm(SSA)and Radial Basis Function(RBF)were employed to invert the permeability coefficient.This process constructs an SSA-RBF surrogate model for predicting seepage pressure,yielding both the model’s predictive values and a residual sequence.Then,the residual sequence was decomposed by using Variational Mode Decomposition(VMD),training a Long Short-Term Memory(LSTM)neural network to obtain a model for correcting the residual sequence.By overlaying the mechanistic model with the data-driven model,an SSA-RBF-VMD-LSTM fusion model was constructed,which enables accurate predictions of seepage water levels.The engineering case demonstrates that the model proposed in this paper possesses high predictive accuracy,with improvements of 89.64%,69.59%,and 60.45%in prediction accuracy compared to statistical models,LSTM models,and SSA-RBF-LSTM models,respectively.Notably,even when the seepage process line undergoes significant fluctuations,the model is still capable of providing timely and accurate predictions,showcasing good stability and extrapolation capabilities.These attributes make the model worthy of practical application and dissemination.
作者
黄昊冉
谷艳昌
陈斯煜
王士军
黄海兵
HUANG Haoran;GU Yanchang;CHEN Siyu;WANG Shijun;HUANG Haibing(Dam Safety and Management Department,Nanjing Hydraulic Research Institute,Nanjing 210029,China;Dam Safety Management Center of the Ministry of Water Resources,Nanjing 210029,China;Key Laboratory of Flood&Drought Disaster Defense of the Ministry of Water Resources,Nanjing 210029,China)
出处
《水利学报》
北大核心
2025年第3期398-410,共13页
Journal of Hydraulic Engineering
基金
国家自然科学基金项目(51979175,52309157)
南京水利科学研究院研究生学位论文基金项目(Yy724005)
南京水利科学研究院中央级公益性科研院所基本科研业务费(Y723008,Y722003,Y723002)。
关键词
土石坝
代理模型
麻雀搜索算法
变分模态分解
LSTM神经网络
机理-数据驱动融合
earth-rock dam
surrogate models
sparrow search algorithm
Variational Modal Decomposition
LSTM neural networks
mechanism-data-driven fusion
