Scholarlink

<p>Coal body structure (CBS) is a critical geological factor governing coalbed methane (CBM) enrichment and reservoir stimulation potential. Accurate CBS identification is essential for optimizing CBM exploration and development strategies. While geophysical logging serves as a practical tool for CBS characterization, its effectiveness is constrained by geological complexity and multifactorial logging influences. Traditional logging interpretation methods face limitations such as intricate feature engineering, poor adaptability, and interpretive ambiguities. To overcome these limitations, this study developed an asymmetric convolutional neural network (CNN)-based framework that autonomously extracts spatial correlations and key information from logging data and establishes robust nonlinear relationships between CBS classifications and logging responses. The methodology employs five diagnostic logging curves (density, acoustic, caliper, resistivity &lt; natural gamma ray) to construct 5 × 5 feature matrices preserving ± 0.2&#xa0;m vertical structural context. The experimental results demonstrated exceptional model performance, achieving 97.94% validation accuracy—surpassing the basic multilayer perceptron (MLP) model's performance (90.41%). In field testing, the CNN-based model achieved a prediction accuracy of 81.61% in test wells, outperforming the conventional approaches including MLP (71.35%) and K-means clustering (57.86%). Field implementation in the Panjiwaiwei CBM exploration area, Huainan Coalfield, China, revealed a distinct tectonic coal distribution pattern: intense development in the north, moderate in the south, and centralized accumulation in the central zone. The proposed CNN-driven workflow establishes a new technical pathway for high-resolution CBS identification, enabling quantitative characterization of reservoir heterogeneity while providing critical geological constraints for CBM sweet spot prediction and hydraulic fracturing optimization, demonstrating significant theoretical and practical value.</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

An Improved Convolutional Neural Network-Based Coal Structure Identification Method in Complex Tectonic Areas

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Xuehua Chen,
  • Huijing Fang,
  • Xiaozhi Zhou,
  • Shuaiqi Wang,
  • Hai Ding,
  • Shiliang Zhu,
  • Yu Xiong

摘要

Coal body structure (CBS) is a critical geological factor governing coalbed methane (CBM) enrichment and reservoir stimulation potential. Accurate CBS identification is essential for optimizing CBM exploration and development strategies. While geophysical logging serves as a practical tool for CBS characterization, its effectiveness is constrained by geological complexity and multifactorial logging influences. Traditional logging interpretation methods face limitations such as intricate feature engineering, poor adaptability, and interpretive ambiguities. To overcome these limitations, this study developed an asymmetric convolutional neural network (CNN)-based framework that autonomously extracts spatial correlations and key information from logging data and establishes robust nonlinear relationships between CBS classifications and logging responses. The methodology employs five diagnostic logging curves (density, acoustic, caliper, resistivity < natural gamma ray) to construct 5 × 5 feature matrices preserving ± 0.2 m vertical structural context. The experimental results demonstrated exceptional model performance, achieving 97.94% validation accuracy—surpassing the basic multilayer perceptron (MLP) model's performance (90.41%). In field testing, the CNN-based model achieved a prediction accuracy of 81.61% in test wells, outperforming the conventional approaches including MLP (71.35%) and K-means clustering (57.86%). Field implementation in the Panjiwaiwei CBM exploration area, Huainan Coalfield, China, revealed a distinct tectonic coal distribution pattern: intense development in the north, moderate in the south, and centralized accumulation in the central zone. The proposed CNN-driven workflow establishes a new technical pathway for high-resolution CBS identification, enabling quantitative characterization of reservoir heterogeneity while providing critical geological constraints for CBM sweet spot prediction and hydraulic fracturing optimization, demonstrating significant theoretical and practical value.