基于常规测井与可解释机器学习的深部煤系地层煤层识别与含煤性评价

Coal Seam Identification and Coal-Bearing Evaluation in Deep Coal Measures Based on Conventional Logging and Interpretable Machine Learning

摘要: 针对深部煤系地层中薄煤层、夹矸发育及纵向变化快造成的煤层边界识别和层段含煤性评价问题，基于某盆地东部某区块多井常规测井资料与已有解释结果，建立煤层概率识别和含煤性评价流程。模型输入包括AC、CAL、DEN、CNL、GR、RD、RS、SP和PE共9条常规曲线，以已有解释结论构建煤/非煤标签，采用类别加权L2正则化逻辑回归和留一井验证评价井间适用性；在单点特征基础上，进一步采用11个采样点的滑动窗口提取均值和标准差，以表征纵向上下文。结果表明，单点模型总体AUC为0.9867、召回率为0.9601；窗口模型总体AUC和召回率分别为0.9931和0.9668，但精确率有所下降，说明邻域统计有助于减少漏判，同时会扩大待复核范围。传统阈值法对典型煤层响应较为保守，概率模型则更适合筛选薄层、边界和响应不典型层段。通过概率分级、连续厚度约束、短间隔合并及井径质量复核，可将深度点输出转化为层段含煤性评价依据。该流程用于常规解释基础上的概率复核和层段量化，不能替代人工综合解释。

Abstract: Thin coal seams, abundant partings and rapid vertical variations complicate coal-boundary identification and interval-scale coal-bearing evaluation in deep coal measures. Multi-well conventional logs and existing interpretations from an eastern block of a sedimentary basin were used to establish a probabilistic workflow. Nine logs (AC, CAL, DEN, CNL, GR, RD, RS, SP and PE) were used as predictors. Coal and non-coal labels were derived from existing interpretations, and class-weighted L2-regularized logistic regression was assessed by leave-one-well-out validation. Means and standard deviations within an 11-sample sliding window were further introduced to represent vertical context. The point model achieved an overall AUC of 0.9867 and recall of 0.9601. The context model increased AUC and recall to 0.9931 and 0.9668, respectively, while precision decreased, indicating fewer missed coal responses but a broader review range. The conventional threshold method remained conservative for typical coal responses, whereas probability output was more suitable for screening thin seams, boundaries and atypical intervals. Probability grading, continuity constraints, short-gap merging and borehole-quality checks were combined to support interval-scale coal-bearing evaluation. The workflow is intended for probabilistic review rather than as a replacement for integrated manual interpretation.