Speaker
摘要
深部煤层兼具 CO2地质封存潜力与 CO2强化煤层气开采价值，其 CO2-水-煤三相体系润湿性直接影响 CO2吸附、甲烷置换、毛管行为及封存安全性。然而，受煤岩非均质性、实验条件差异及模型泛化能力不足等因素制约，煤润湿性的快速准确预测仍具有挑战。本文构建了包含 531 组 CO2-水-煤接触角数据的跨区域、多来源数据库，选取灰分含量、温度、压力和最大镜质体反射率作为输入变量，以接触角作为预测目标，建立并比较了随机森林、XGBoost、LightGBM、CatBoost 和GBDT等五类集成机器学习模型。各模型采用贝叶斯优化进行超参数寻优，并利用独立测试集评价预测性能。结果表明，CatBoost 模型表现最优，测试集决定系数 R2 达 0.881，RMSE为 9.089°，MAE为 6.641°。进一步采用 SHAP 方法揭示模型决策机制，发现压力是影响接触角的主控因素，其次为最大镜质体反射率、灰分含量和温度。当压力超过 7.38 MPa 后，CO2吸附状态和密度变化促使接触角显著增大；煤阶升高会增强煤表面疏水性，而灰分中的亲水性矿物组分则总体降低接触角。最后，利用准噶尔煤田 0-1100 m 深度梯度原位岩芯样品开展外部验证，模型取得 R2 = 0.7206、MAE = 6.904° 的预测效果。研究表明，融合集成学习与可解释性分析的方法可有效连接实验室数据与实际储层润湿性评价，为 CO2地质封存储层筛选、参数优化和工程设计提供数据驱动支撑。
Abstract
Deep coal seams are promising targets for geological CO2 storage and CO2-enhanced coalbed methane recovery, where the wettability of the CO2-water-coal system plays a key role in CO2 adsorption, methane displacement, capillary behavior, and storage security. Nevertheless, reliable wettability prediction remains difficult because coal properties are highly heterogeneous and existing models often show limited transferability. In this study, a cross-regional and multi-source database containing 531 CO2-water-coal contact-angle records was compiled. Four accessible variables, namely ash content, temperature, pressure, and maximum vitrinite reflectance, were used as model inputs, while contact angle was taken as the prediction target. Five ensemble learning algorithms, including random forest, XGBoost, LightGBM, CatBoost, and gradient boosting decision tree, were developed and optimized through Bayesian hyperparameter tuning. Their predictive capability was then evaluated on an independent test set. Among the tested models, CatBoost achieved the best overall performance, with an R2 of 0.881, RMSE of 9.089°, and MAE of 6.641° on the test dataset. SHAP analysis was further introduced to interpret the model behavior and quantify the contribution of each variable. The results indicate that pressure is the dominant factor controlling contact angle, followed by maximum vitrinite reflectance, ash content, and temperature. A pressure threshold near 7.38 MPa marks a transition in CO2 adsorption behavior, above which the contact angle increases sharply. Higher coal rank enhances hydrophobicity, whereas ash-related hydrophilic minerals tend to reduce the contact angle. External validation using in-situ core samples from the Junggar Coalfield along a 0-1100 m depth profile yielded an R2 of 0.7206 and an MAE of 6.904°, demonstrating encouraging transferability under field-relevant conditions. This work provides an interpretable data-driven framework for assessing coal wettability and supports reservoir screening and engineering design for CO2 geological sequestration.
| 关键词 | CO2地质封存；煤润湿性；接触角预测；集成机器学习；SHAP 可解释性；原位岩芯验证 |
|---|---|
| Keywords | CO2 geological storage; coal wettability; contact angle estimation; ensemble machine learning; SHAP interpretation; in-situ core validation |
