土石方开挖钻孔效率预测可解释超级学习器集成学习模型  

Interpretable Super Learner Ensemble Learning Model for Earth-Rock Excavation Drilling Efficiency Prediction

作　　者：王晓玲[1] 胡亦宁 张君[1] 衣传宝 张捷[3] 李希稷 Wang Xiaoling;Hu Yining;Zhang Jun;Yi Chuanbao;Zhang Jie;Li Xiji(State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation,Tianjin University,Tianjin 300350,China;Shandong Weifang Pumped Storage Co.,Ltd.,Weifang 261000,China;Beijing Engineering Co.,Ltd.,Power Construction Corporation of China,Beijing 100024,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 [2]山东潍坊抽水蓄能有限公司,潍坊261000 [3]中国电建集团北京勘测设计研究院有限公司,北京100024

出　　处：《天津大学学报(自然科学与工程技术版)》2025年第4期331-342,共12页Journal of Tianjin University:Science and Technology

基　　金：国家自然科学基金青年科学基金资助项目(52309165);国网新源集团有限公司及国网新源山东潍坊抽水蓄能有限公司科技资助项目(SGXYKJ-2020-076).

摘　　要：动态预测钻孔效率并探究不同因素对钻孔效率的影响程度,对土石方开挖进度分析和风险管理具有重要意义.然而,现有土石方钻孔效率分析大都依赖人工经验,少数机器学习模型无法解释不同因素对钻孔效率的影响程度.针对上述问题,本研究提出土石方开挖钻孔效率预测可解释超级学习器(SL)集成学习模型.通过强化学习中的Q学习改进猎人猎物优化算法局部搜索过程与全局信息进行交互的能力,提出Q学习改进的猎人猎物优化(QIHPO)算法对SL的n_estimators、learning_rate、max_depth等超参数进行优化,进而利用SL能够通过具有互补特征的异构基学习器捕捉样本特征差异性的优势,建立基于QIHPO优化的超级学习器土石方开挖钻孔效率预测QIHPO-SL模型,以揭示地质、作业、环境和机械特性等众多因素与钻孔效率的复杂非线性映射关系.进一步将QIHPO-SL集成学习算法与可解释机器学习框架下的沙普利加性解释(SHAP)理论相结合,挖掘影响钻孔效率的关键特征,并解释不同因素对钻孔效率的影响程度.案例分析表明:QIHPO-SL具有较高的预测精度,相较于QIHPO-XGB、QIHPO-RF和SL等基准模型,本文所提方法的预测精度分别提高了12.94%、12.02%和1.58%,且SHAP理论提高了模型的可解释性和预测结果的可信度,为钻孔效率预测及致因分析提供了新思路和新途径.The dynamic prediction of the drilling efficiency and the exploration of the degree of influence of different factors on the drilling efficiency are of significance for the earth-rock excavation schedule analysis and risk management.However,most of the existing earth-rock drilling efficiency analysis methods rely on manual experience,and a few machine learning models cannot explain the degree of influence of different factors on the drilling efficiency.To address these problems,an interpretable super learner(SL)ensemble learning model for earth-rock excavation drilling efficiency prediction is proposed in this paper.The capability of the local search process of a hunter-prey optimization algorithm interacting with the global information is improved by Q-learning in reinforcement learning,and a Qlearning improved hunter-prey optimization(QIHPO)algorithm is put forward to optimize the hyper-parameters of the SL,such as n_estimators,learning_rate and max_depth.Accordingly,the QIHPO algorithm can take advantage of the SL which is capable of capturing the variability of sample features through the heterogeneous base-learner with complementary characteristics,and a QIHPO-SL model based on QIHPO optimization and SL for predicting the earth-rock excavation drilling efficiency is established,thereby revealing the complex nonlinear mapping relationship between the drilling efficiency and various factors including the geological,operation,environmental and mechanical characteristics.Furthermore,the QIHPO-SL ensemble learning algorithm is combined with the theory of Shapley additive explanation(SHAP)under the interpretable machine learning framework to excavate the key features affecting the drilling efficiency and explain the degree of influence of different factors on the drilling efficiency.The analysis of a case study shows that QIHPO-SL has a high prediction accuracy.Compared with those of the benchmark models such as QIHPO-XGB,QIHPO-RF and SL,the prediction accuracy of the proposed method is improved by 12.94%,12.02%an

关 键 词：钻孔效率 沙普利加性解释 超级学习器 强化学习 

分 类 号：TV53[水利工程—水利水电工程]