基于机器学习的煤层含气量测井评价方法——以沁水盆地柿庄南区块为例

Artificial-intelligence and machine-learning models of coalbed methane content based on geophysical logging data:A case study in Shizhuang South Block of Qinshui Basin,China

1.中海油研究总院有限责任公司勘探开发研究院,北京100028;
2.中海石油(中国)有限公司勘探部,北京100016

1. Exploration and Development Research Institute of CNOOC Research Institute Co.,Ltd.,Beijing 100028,China;
2. Exploration Department of CNOOC China Limited,Beijing 100016,China

由于煤层含气量受控因素多且成因机理复杂,其评价的准确性和泛化性问题一直是研究的热点与难点。为了提高煤层含气量评价的准确性和泛化性,基于沁水盆地柿庄南区块的测井资料、煤心分析资料、排采生产数据等,将地球物理测井数据作为输入,利用BP神经网络、支持向量机及随机森林3种机器学习方法,对训练集数据采用交叉验证与网格寻优方法确定各机器学习模型超参数,得到煤层含气量评价模型。引入盲井验证模型,对比了3种机器学习方法的优缺点和适用条件。结果表明,应用此3种机器学习方法能有效评价煤层含气量,随机森林模型在此区块应用效果最好,该模型能有效评价煤层含气量,为今后此方法的应用提供了选择依据,同时也提高了模型的泛化能力;进一步对沁水盆地柿庄南区块3号煤层开发井进行含气量评价预测,并将预测结果与实际排采生产数据进行对比,发现二者误差较小。研究结果对煤层气勘探开发、“甜点”找寻具有指导意义,实际应用价值突出。

An effective and accurate quantification of coalbed methane (CBM) content plays an important role in the exploration and development of CBM.Due to the effects of many factors on the CBM content and the complicated mechanism of its formation,the quality and generalizability of models of the CBM content have been a research focus and challenge.Derived from geophysical logs,data for core analysis and production were used as inputs in back propagation neural network (BNN),support vector machine (SVM),and random forest (RF) in order to predict the CBM content.Hyper-parameters of each of the artificial-intelligence and machine-learning algorithms were determined using cross-validation and grid optimization.To validate the models,the entire data were randomly partitioned into training and testing data.Although all three models predicted the CBM content with a high accuracy,RF outperformed the others.The best-fit model of the CBM content was validated against the coal seam production data of Shizhuang South Block wells of Qinshui Basin and led to a high predictive power.Given the above results,the model provides actionable insights into the exploration and development of CBM,in particular,in deserts.

10.3969/j.issn.1000-1441.2023.01.005