Scholarlink

<p>The content of total organic carbon (TOC) is a key indicator used to evaluate the hydrocarbon-generation capacity of shale formations and to identify potential sweet spots for exploration. To overcome the high cost and limited continuity of conventional core-based analyses, as well as the inadequate accuracy and interpretability of existing logging-based prediction models, this study investigates the Chang7 shale interval in the southwestern Yishan Slope of the Ordos Basin. Using pyrolysis data from 357 cores and five conventional logging suites from 12 wells, we developed an integrated machine learning framework that couples Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and XGBoost. The SHAP framework was incorporated to enhance model interpretability. Through stepwise synergistic optimization using GA and PSO, the method efficiently identified the optimal hyperparameter configurations for XGBoost, thereby markedly improving model convergence. The results show that the GA-PSO-XGBoost model achieves outstanding TOC prediction performance, yielding a cross-validation coefficient of determination (<i>R</i><sup>2</sup>) of 0.92 and a validation set <i>R</i><sup>2</sup> of 0.89. The average relative error and root mean square error are 17.83% and 3.26%, respectively, which significantly outperform traditional approaches, including multiple regression, support vector machines, and ΔLogR. SHAP analysis further indicates that the relative contributions of logging parameters to TOC prediction follow the order GR &gt; DEN &gt; CNL &gt; RT &gt; AC, with GR serving as the most influential predictor and a primary indicator of organic matter enrichment. The predictive accuracy of the multi-parameter model (<i>R</i><sup>2</sup> = 0.84) is substantially higher than that achieved using GR alone (<i>R</i><sup>2</sup> = 0.43). Validation using newly drilled wells shows that the model maintains a stable mean relative error of 17.44 ~ 18.12%, demonstrating robust generalization capability. The proposed method improves both TOC prediction accuracy and model interpretability, offering a reliable analytical approach for high-resolution reservoir-quality evaluation in terrestrial shale oil and gas exploration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluation method of total organic carbon content in shale oil reservoir sections based on combinatorial optimization and interpretable machine learning driven by logging data—a case study of Chang 7 reservoir in Ordos Basin

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Jianhong Guo,
  • Yanmei Wang,
  • Qing Zhao,
  • Changsheng Wang,
  • Hongyuan Wei,
  • Xin Nie

摘要

The content of total organic carbon (TOC) is a key indicator used to evaluate the hydrocarbon-generation capacity of shale formations and to identify potential sweet spots for exploration. To overcome the high cost and limited continuity of conventional core-based analyses, as well as the inadequate accuracy and interpretability of existing logging-based prediction models, this study investigates the Chang7 shale interval in the southwestern Yishan Slope of the Ordos Basin. Using pyrolysis data from 357 cores and five conventional logging suites from 12 wells, we developed an integrated machine learning framework that couples Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and XGBoost. The SHAP framework was incorporated to enhance model interpretability. Through stepwise synergistic optimization using GA and PSO, the method efficiently identified the optimal hyperparameter configurations for XGBoost, thereby markedly improving model convergence. The results show that the GA-PSO-XGBoost model achieves outstanding TOC prediction performance, yielding a cross-validation coefficient of determination (R2) of 0.92 and a validation set R2 of 0.89. The average relative error and root mean square error are 17.83% and 3.26%, respectively, which significantly outperform traditional approaches, including multiple regression, support vector machines, and ΔLogR. SHAP analysis further indicates that the relative contributions of logging parameters to TOC prediction follow the order GR > DEN > CNL > RT > AC, with GR serving as the most influential predictor and a primary indicator of organic matter enrichment. The predictive accuracy of the multi-parameter model (R2 = 0.84) is substantially higher than that achieved using GR alone (R2 = 0.43). Validation using newly drilled wells shows that the model maintains a stable mean relative error of 17.44 ~ 18.12%, demonstrating robust generalization capability. The proposed method improves both TOC prediction accuracy and model interpretability, offering a reliable analytical approach for high-resolution reservoir-quality evaluation in terrestrial shale oil and gas exploration.