测井技术2025，Vol.49Issue(6)：890-906,17.DOI:10.16489/j.issn.1004-1338.2025.06.007
基于斯通利波与监督学习的砂砾岩储层渗透率预测方法
Permeability Prediction Method for Glutenite Reservoirs Using Stoneley Waves and Supervised Learning
摘要
Abstract
Focus to the strong heterogeneity and complex porosity and permeability relationship in glutenite reservoirs,and poor stoneley wave applicability for permeability inversion,the shortcomings of the theoretical inversion model are analyzed.Furthermore,based on core experiments and logging data,properties sensitive to permeability characterization such as Stoneley wave slowness,amplitude,center frequency,and attenuation are obtained.Then taking core permeability as the label,three supervised machine learning permeability prediction models are constructed with support vector machine regression(SVR),extreme gradient boosting decision tree(XGBoost),and one-dimensional convolutional neural network(1DCNN),which using techniques such as random cross validation,adaptive moment estimation(Adam),and transfer learning to optimize the model hyperparameters.Also,the contribution of input variables to different models and the rationality of the models are analyzed using shapley additive explanations(SHAP)analysis,finally the models application effectiveness is comparatively analyzed.The research results show that:① The time difference,amplitude,center frequency,and attenuation of Stoneley waves are sensitive characteristics of permeability in glutenite reservoirs,and their correlation with permeability conform to the principles of slow compress wave theory;② The three machine learning models can significantly improve the accuracy of permeability estimation in glutenite reservoirs,compared with artificially optimized inversion results,the logarithmic mean square error of XGBoost prediction values is reduced by more than 30%,and the prediction error of 1DCNN model using transfer learning is reduced by more than 50%,The accuracy of SVR model is relatively low compared to the them;③ SHAP analysis shows that the 1DCNN and XGBoost models can correctly learn knowledge,and the Stoneley wave slowness,amplitude,and center frequency are the main features that affect the model output.However,the SVR model learns knowledge poorly,there is a big difference between features contribution to model output and sensitive attributes correlation analysis;④ After discarding the Stoneley wave attenuation as the input feature,there is no significant change in the accuracy of the XGBoost model,while the errors of the 1DCNN and SVR models increase significantly,which indicating that Stoneley wave attenuation also is one effective feature for characterizing permeability.The conclusion is that using Stoneley wave sensitivity attributes combined with supervised learning algorithms,especially using XGBoost and 1DCNN transfer learning,can effectively overcome the application difficulties of traditional inversion algorithms in glutenite reservoirs,improving the accuracy of permeability prediction significantly,and provides an effective new method for permeability evaluation in heterogeneous reservoir using Stoneley wave attributes.关键词
砂砾岩储层/渗透率预测/斯通利波/监督学习/极限梯度提升决策树(XGBoost)/一维卷积神经网络(1DCNN)/支持向量机回归(SVR)/SHAP分析分类
天文与地球科学引用本文复制引用
SHANG Suogui,SHAO Cairui,CHENG Jiajun,LI Ruijuan,GAO Qiangyong,WANG Ruihong,ZHANG Fuming,WANG Miao..基于斯通利波与监督学习的砂砾岩储层渗透率预测方法[J].测井技术,2025,49(6):890-906,17.基金项目
国家科技重大专项课题"海上油气富集规律与新领域勘探开发关键技术"(2025ZD1402800) （2025ZD1402800）
中国海洋石油有限公司课题"渤海大中型天然气形成条件、富集规律与勘探方向研究"(KJZH-2024-2107) （KJZH-2024-2107）
