基于CNN-BiLSTM-AM的储层岩石脆性指数预测  

Prediction of Reservoir Rock Brittleness Index Based on CNN-BiLSTM-AM

作　　者：杜睿山[1,2] 李宏杰 孟令东 DU Rui-shan;LI Hong-jie;MENG Ling-dong(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China;Key Laboratory of Oil and Gas Reservoir and Underground Gas Storage Integrity Evaluation,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318 [2]油气藏及地下储库完整性评价黑龙江省重点实验室,黑龙江大庆163318

出　　处：《计算机技术与发展》2023年第10期28-34,共7页Computer Technology and Development

基　　金：国家自然科学基金青年科学基金(41702156);东北石油大学引导性创新基金(2020YDL-04)。

摘　　要：脆性指数是储层岩石的重要地质力学性质之一,但对于脆性指数至今为止都没有一个明确的定义,许多学者提出了不同的方法来测量该参数,一些方法如矿物分析等成本较高,然而机器学习和深度学习可以有效融合多元数据,充分利用数据去挖掘自变量与因变量之间的关系,且成本较低。因此,该文使用深度学习方法来构建测井曲线数据与储层岩石脆性之间的关系模型。因测井曲线是特殊的时序曲线,该文采用可以综合考虑过去和未来信息的BiLSTM(双向长短期记忆)模型,同时为了进一步的优化,在模型中添加1DCNN(一维卷积)用来提取特征,且引入了AM(注意力机制)。同时利用Pearson系数和XGBoost(极限梯度提升树)进行分析,研究了各测井曲线对脆性的敏感性等级以及重要性程度,最终选取的输入参数有AC(声波时差)、DEN(密度)、CAL(井径)、GR(伽马射线)和SP(自然电位)。与其它机器学习方法相比,该方法的MSE和RMSE最小,分别为0.0035和0.05916,表明CNN-BiLSTM-AM是一种预测精度更高、效果更好的方法。Brittleness index is one of the important geomechanical properties of reservoir rocks,but there is no clear definition for brittleness index so far.Many scholars had proposed different methods to measure this parameter.Some methods,such as mineral analysis,had higher costs.However,machine learning and depth learning can effectively integrate multivariate data,and make full use of the data to mine the relationship between independent variables and dependent variables with lower costs.Therefore,we used depth learning method to build the relationship model between logging curve data and reservoir rock brittleness.Because the logging curve was a special time series curve,we adopted the BiLSTM(Bi-directional Long Short-Term Memory)model that can comprehensively consider past and future information.At the same time,for further optimization,1DCNN(One-Dimensional Convolution)was added to the model to extract features,and AM(Attention Mechanism)was introduced.At the same time,Pearson coefficient and XGBoost(eXtreme Gradient Boosting)were used for analysis,and the sensitivity level and importance of each logging curve to brittleness are studied.The final selected input parameters were AC(Acoustic moveout),DEN(Density),CAL(Caliper),GR(Gamma Ray)and SP(Spontaneous Potential).Compared with other machine learning methods,the proposed method has the smallest MSE and RMSE,0.0035 and 0.05916 respectively.It is showed that CNN-BiLSTM-AM is a method with higher prediction accuracy and better effect.

关 键 词：测井曲线 脆性指数 深度学习 Pearson系数 BiLSTM 一维卷积 注意力机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]