基于卷积神经网络的核磁共振测井渗透率预测

摘要

渗透率是地下流体流动分析、油气藏管理、烃类采收与二氧化碳封存中的关键参数。传统的渗透率测定方法依赖成本高、耗时长的室内实验或与井相关的资料。机器学习——特别是卷积神经网络（CNN）——因其能够挖掘输入变量与输出变量之间的内在联系，被认为是一种低成本、快速的渗透率预测手段。本文用卷积神经网络建立了渗透率的经验关联式：先对 40 条核磁共振（NMR）T2 谱与 89 个 NMR T2 对数平均分布（T2lm）作预处理与筛选，再用主成分分析（PCA）识别出关键谱段；随后采用自行设计的卷积神经网络结构，充分利用核磁共振数据中蕴含的空间模式与复杂关系来构建关联式；并用 k 折交叉验证方案对模型作了严格的训练与验证，以保证其稳健性与泛化能力。评价模型精度所用的指标包括决定系数（R2）、均方根误差（RMSE）、绝对误差（MAE）、标准差（SD）、绝对偏差（AD）、平均绝对偏差（AAD）、平均绝对百分比相对误差（AAPRE）与最大误差（Emax）。在所考察的各折当中，第 1 折表现最佳，R2 达到 0.9544。该基于卷积神经网络的关联式在 R2、Emax、AD、AAD、AAPRE 等指标上均优于常规模型与其他人工智能模型。总体而言，本研究表明卷积神经网络能够有效预测渗透率，可替代成本高昂且适用范围有限的传统方法，其中第 1 折的结果最为理想。

Abstract

Permeability is a critical parameter in subsurface fluid flow analysis, reservoir management, hydrocarbon recovery, and carbon dioxide sequestration. Traditional permeability measurement methods involve costly and time-consuming laboratory tests or well-related data. Machine learning (ML), specifically convolutional neural networks (CNN), is proposed as a cost-effective and rapid permeability prediction solution, harnessing interrelationships of input-output variables. In this study, empirical permeability correlation was developed using CNN. Forty nuclear magnetic resonance (NMR) T2 spectrums and 89 logarithmic mean NMR T2 distributions (T2lm) were preprocessed, screened and key spectra were identified using the principal component analysis (PCA). To develop the correlations, a custom-designed CNN architecture was employed to leverage the spatial patterns and intricate relationships embedded in the NMR data. The model was trained and validated rigorously using k-fold cross validation scheme to ensure robustness and generalization. Performance metrics like R-squared (R2), root mean squared error (RMSE), mean absolute error (MAE), standard deviation (SD), absolute deviation (AD), average absolute deviation (AAD), average absolute percentage relative error (AAPRE), and maximum error (Emax) were deployed to evaluate the model's accuracy and ability to predict permeability values accurately. Among the folds considered, the fold 1 emerged as the best-performing model with the highest R2 value of 0.9544. This CNN-based correlation outperformed conventional and other AI-based models in terms of R2, Emax, AD, AAD, AAPRE, among other metrics. Overall, the study demonstrates the effectiveness of CNN in predicting permeability, offering a superior alternative to costly and limited traditional methods, with fold 1 showing the most promising results.