坝基灌浆量预测ISSA-Stacking集成学习代理模型研究  

Research on an ISSA-Stacking Ensemble Learning Surrogate Model of Dam Foundation Grouting Volume Prediction

作　　者：祝玉珊 王晓玲[1] 崔博[1] 陈文龙[1] 轩昕祺 余红玲 Zhu Yushan;Wang Xiaoling;Cui Bo;Chen Wenlong;Xuan Xinqi;Yu Hongling(State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation,Tianjin University,Tianjin 300350,China)

机构地区：[1]天津大学水利工程智能建设与运维全国重点实验室,天津300350

出　　处：《天津大学学报（自然科学与工程技术版）》2024年第2期174-185,共12页Journal of Tianjin University：Science and Technology

基　　金：国家自然科学基金资助项目(51839007,51779169).

摘　　要：灌浆量预测对坝基灌浆施工具有重要意义.由于灌浆工程隐蔽且复杂,传统方法难以实现准确高效的灌浆量预测.代理模型是一种能够建立影响因素与响应值之间近似关系的快速求解方法,然而单一代理模型的预测稳定性和准确性较低,组合代理模型仅将单一模型结果进行加权平均,预测精度仍有待提高.为解决上述问题,本文提出一种ISSA-Stacking集成学习代理模型新方法用于灌浆量预测研究.首先,针对灌浆量预测具有数据量小、影响因素与灌浆量之间非线性关系复杂且预测不确定性较大等特性,基于Stacking集成学习策略,选取在小样本预测中表现优越的支持向量回归(SVR)、具有良好非线性拟合能力的BP神经网络(BPNN)和预测泛化性能及稳定性高的随机森林(RF)等算法作为基学习器,采用自适应学习和不确定性处理能力强的自适应神经模糊推理系统(ANFIS)作为元学习器以集成上述机器学习算法的优势,构建具有更优预测性能和泛化能力的Stacking集成学习方法作为代理模型;其次,为进一步提高模型预测精度,采用混沌理论和Lévy飞行策略改进的麻雀搜索算法(ISSA)对集成学习代理模型进行参数同步优化;最后,将所提ISSA-Stacking集成学习代理模型应用于某实际灌浆工程的灌浆量预测并与其他方法进行对比分析.结果表明,所提方法具有较高的预测精度,绝对平均误差仅为0.21 m^(3);与组合代理模型及单一代理模型(SVR、BPNN和RF)相比,平均精度分别提高24.34%、30.84%、32.68%和26.56%,为灌浆量预测提供了一种新思路.The prediction of grouting volume is crucial to the construction of dam foundation grouting.Due to the concealment and complexity of grouting engineering,conventional methods have difficulty predicting the accurate and efficient grouting volume.A surrogate model is a type of fast solution method that can establish the approximate relationship between influencing factors and response values.However,single surrogate models have low prediction stability and accuracy,and combined surrogate models can only perform a weighted average of the results of single models,whose prediction accuracy still requires improvement.To address these problems,this paper proposes a new method of the improved sparrow search algorithm(ISSA)-Stacking ensemble learning surrogate model for grouting volume prediction.Grouting volume prediction is characterized by limited data amount,complex nonlinear relationship between influencing factors and grouting amount,and large prediction uncertainty.Thus,based on the Stacking ensemble learning strategy,the support vector regression(SVR)with excellent performance in small sample prediction,BP neural network(BPNN)with good nonlinear fitting ability,and random forest(RF)with high prediction generalization performance and stability are selected as base learners.An adaptive neuro-fuzzy inference system with adaptive learning and uncertainty processing ability is selected as the meta learner.The aim is to integrate the advantages of these machine learning algorithms and build a Stacking ensemble learning method with better prediction performance and generalization ability as a surrogate model.Second,to further improve the prediction accuracy of the model,based on chaos theory and Lévy flight strategy,the improved sparrow search algorithm(ISSA)is developed and used to synchronously optimize the parameters of the stacking ensemble learning surrogate model.Finally,the proposed ISSA-Stacking ensemble learning surrogate model is applied to grouting volume prediction in practical grouting engineering and compa

关 键 词：灌浆量预测 Stacking集成学习方法 代理模型 麻雀搜索算法 

分 类 号：TV52[水利工程—水利水电工程]