文章摘要
基于细菌觅食优化Stacking集成学习的钻孔效率预测模型研究
Research on a Drilling Efficiency Prediction Model Based on Bacterial Foraging Optimization Stacking Ensemble Learning
投稿时间:2025-07-17  修订日期:2025-12-25
DOI:
中文关键词: 钻孔效率  施工仿真  Stacking集成学习  XGBoost  LightGBM  MLP  支持向量机
英文关键词: drilling efficiency  construction simulation  Stacking ensemble learning  XGBoost  LightGBM  MLP  Support Vector Machine
基金项目:国家自然科学基金联合基金重点项目(U24B20111);国家自然科学基金项目(52279137);中国电力建设股份有限公司科技项目信息(DJ-ZDXM-2020-50)
作者单位邮编
关涛 天津大学 水利工程智能建设与运维全国重点实验室 300072
摘要点击次数: 49
全文下载次数: 0
中文摘要:
      钻孔效率是堆石坝料场开挖施工进度仿真的关键参数,其预测准确性直接关系到仿真模型的可靠性。针对现有数学方法预测效率较低、单一学习器难以满足仿真精度要求、集成学习研究中超参数调整方法局部搜索精细度不足的问题,本文提出了一种基于细菌觅食优化Stacking集成学习的钻孔效率预测模型。首先,以某堆石坝现场采集的钻孔效率数据为目标变量,以其影响因素(如钻孔深度、岩石性质、高程等)为特征变量构建数据集;其次,采用XGBoost、LightGBM和MLP三种异质基学习器并行训练,并引入细菌觅食优化算法模拟趋化和繁殖行为,通过R2曲线实时追踪,迭代优化各基学习器的超参数,确保输出稳定的“元特征”;最后,将各基学习器的预测结果输入支持向量回归(SVR)元学习器,通过整合多模型的互补信息,在抑制偏差与方差的同时获得集成预测结果。实验结果表明,经细菌觅食优化后,各基学习器的R2均可达到0.93以上,PCC值均超过0.97,集成模型在整个样本数据集上的学习曲线也平滑稳定,残差分析显示预测值与真实值的残差序列在零均值线附近均匀分布,最终结果的PCC值可接近0.98,可以满足施工过程仿真需求。
英文摘要:
      Drilling efficiency is a key parameter in progress simulation of rock-fill dam excavation, and its prediction accuracy directly affects the reliability of the simulation model. To address low prediction accuracy of existing mathematical methods, the inability of single learners to meet simulation precision requirements, and the insufficient local search granularity of hyperparameter tuning in ensemble learning research, this paper proposes a drilling efficiency prediction model based on Bacterial Foraging Optimization-optimized Stacking ensemble learning. First, a dataset was constructed using drilling efficiency data collected on a rock-fill dam site as the target variable and its influencing factors (e.g., drilling depth, rock properties, elevation) as feature variables. Second, three heterogeneous base learners (XGBoost, LightGBM, and MLP) were trained in parallel, and the Bacterial Foraging Optimization algorithm—simulating chemotaxis and reproduction—was introduced to iteratively optimize each base learner’s hyperparameters by tracking the R2 curve in real time, ensuring stable “meta-features” output. Finally, the base learners’ predictions were input to a Support Vector Regression (SVR) meta-learner; by integrating complementary information from multiple models, the ensemble prediction was obtained while suppressing bias and variance. Experimental results show that after Bacterial Foraging Optimization, each base learner’s R2 reaches above 0.93 and PCC exceeds 0.97; the ensemble model’s learning curve over the full sample set is smooth and stable, residual analysis indicates residuals are evenly distributed around the zero-mean line, and the final PCC approaches 0.98, meeting the requirements of construction process simulation.
  查看/发表评论  下载PDF阅读器
关闭