煤层含气量预测中的机器学习算法优选  

Selection of machine learning algorithms in coalbed methane content predictions

作　　者：郭彦省[1] Guo Yan-Sheng(School of Fundamental Education,Beijing Polytechnic College,Beijing 100042,China)

机构地区：[1]北京工业职业技术学院基础教育学院,北京100042

出　　处：《Applied Geophysics》2023年第4期518-533,672,共17页应用地球物理（英文版）

基　　金：sponsored by Beijing Educational Science Planning Project CDHB18383);Key Research Fund Projects(No.BGZYKY 201842Z);Top Talent Program((No.107512200)of Beijing Polytechnic College.

摘　　要：准确预测煤层气含量对煤层气开发具有重要指导意义。机器学习方法对提高煤层含气量预测精度有良好的效果。因此,筛选出一套可以准确且高效预测煤层含气量的机器学习算法就显得尤为重要。基于文献调研结果,选取了深度神经网络(DNN)、卷积神经网络(CNN)、深度信念网络(DBN)、深度交叉网络(DCN)、传统梯度提升树(GBT)、分类梯度提升树(CatBoost)和随机森林(RF)7种机器学习算法开展煤层含气量的预测;选取了测井响应特征、煤质参数和煤层储层特征作为机器学习模型的输入变量。采用DBSCAN聚类算法进行离群值的鉴别与剔除工作。基于统计学显著性检验结果,DCN模型是煤层气含量预测的最佳模型,其平均绝对百分比误差为3.7826%。本研究为煤层气勘探及储层评价提供了新的思路与方法,研究中涉及到的建模策略及思想可用于解决地球物理、石油工业中的其他问题。Accurate prediction of coalbed methane(CBM)content plays an essential role in CBM development.Several machine learning techniques have been widely used in petroleum industries(e.g.,CBM content predictions),yielding promising results.This study aims to screen a machine learning algorithm out of several widely applied algorithms to estimate CBM content accurately.Based on a comprehensive literature review,seven machine learning algorithms,i.e.,deep neural network,convolutional neural network,deep belief network,deep&cross network(DCN),traditional gradient boosting decision tree,categorical boosting,and random forest,are implemented and tuned in this study.Well-logging(i.e.,gamma ray,density,acoustic,and deep lateral resistivity)and coal-seam(i.e.,moisture,ash,volatile matter,fxed carbon,cover depth,porosity,and thickness)properties are selected as the input features of the above machine learning models.Density-based spatial clustering of applications with a noise algorithm is implemented before the training process to identify outliers.Prediction results reveal that DCN is the best model in CBM content predictions(among the ones examined in this study),with a mean absolute percentage error of 3.7826%.

关 键 词：CBM含量 机器学习 DBSCAN 深度交叉网络 集成学习算法 

分 类 号：P618.13[天文地球—矿床学]