摘要：

鄂尔多斯盆地东缘深部煤层气资源丰富，开发潜力巨大。以光亮煤岩为主体的优质储层产气贡献显著，对煤层气高效开采具有关键作用。为实现对此类储层的精准、高效判识，本文提出了一种基于元素特征的深部煤层气水平井储层智能评价方法。以鄂尔多斯盆地大牛地气田8^#煤层为例，通过扫描电镜与能谱分析实验，明确了不同煤岩类型的孔隙结构与元素分布等储层特征；综合皮尔逊相关系数和互信息法，优选Al、S、Ti等10项元素含量作为特征参数，采用人工智能－数据驱动方法，构建了K-近邻、随机森林、梯度提升等8种机器学习分类器模型。研究表明，与非光亮型煤相比，光亮型煤整体呈现Al、Si、Ti、V、Zr等元素含量低，P、S、Ca等元素含量高的特征。根据混淆矩阵分类评估指标，综合得出支持向量机－深部煤岩类型判识模型性能最优，其F₁ Score为85.9%。将得到的最优模型应用于区块内3口水平井的煤岩类型预测分析，与实钻轨迹对比，其准确率超过80%，表明该方法能够为压裂施工的甜点优选与高效选段定簇提供参考。

关键词: 鄂尔多斯盆地, 深部煤层气, 宏观煤岩类型, 孔隙结构, 元素特征, 机器学习

Abstract:

The eastern margin of the Ordos Basin is rich in deep coalbed methane (CBM) resources and possesses enormous development potential. Field practice has shown that high-quality reservoirs dominated by bright coal make a significant contribution to gas production and play a key role in efficient CBM extraction. To achieve accurate and efficient identification of such reservoirs, this paper proposes an intelligent evaluation method for deep CBM horizontal well reservoirs based on elemental characteristics. Taking the No.8 coal seam in the Daniudi Gas Field of the Ordos Basin as an example, the pore structure and elemental distribution of different coal lithotypes were characterized through scanning electron microscopy (SEM) and energy-dispersive spectroscopy (EDS) experiments. By integrating the Pearson correlation coefficient and mutual information method, ten elemental contents including Al, S, and Ti were selected as feature parameters. Using an artificial intelligence-data-driven approach, eight machine-learning classifier models such as K-Nearest Neighbors, Random Forest, and Gradient Boosting were constructed. Research indicates that bright coal generally exhibits lower contents of elements such as Al, Si, Ti, V, and Zr, and higher contents of P, S, and Ca compared with non-bright coal. Based on confusion matrix classification evaluation metrics, the support vector machine-based deep coal rock type identification model demonstrates the optimal performance, with an F₁ Score of 85.9%. Applying the optimized model to predict coal lithotypes in three horizontal wells within the block yields an accuracy exceeding 80% compared with actual drilling trajectories, which can provide a reference for fracture-stage optimization and efficient cluster placement during fracturing operations.

Key words: Ordos Basin, deep coalbed methane, macroscopic coal lithotype, pore structure, elemental characteristics, machine learning

中图分类号: 

• P618.13

• TE122