一种基于ConvLSTM神经网络的TOC含量地震预测方法  

Seismic prediction method of TOC content based on a ConvLSTM neural network

作　　者：李键 陈桂 David Cova 孙永壮 孙宇航 刘洋[2,3] LI Jian;CHEN Gui;David Cova;SUN YongZhuang;SUN YuHang;LIU Yang(CNOOC Shanghai Branch,Shanghai 200335,China;State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum(Beijing),Beijing 102249,China;Karamay Campus,China University of Petroleum(Beijing),Karamay 834000,China)

机构地区：[1]中海石油(中国)有限公司上海分公司,上海200335 [2]中国石油大学(北京)油气资源与探测国家重点实验室,北京102249 [3]中国石油大学(北京)克拉玛依校区,克拉玛依834000

出　　处：《地球物理学进展》2023年第3期1143-1151,共9页Progress in Geophysics

基　　金：中海油“七年行动计划”课题(CNOOC-KJ135ZDXM39S002)资助。

摘　　要：TOC(Total Organic Carbon,总有机碳)含量所表征的有机质丰度是评价烃源岩生烃潜力的重要指标之一.地震弹性参数与岩石的TOC含量之间存在一定的相关性,而且这种关系可能是非线性的.由于线性回归和XGBoost(Extreme Gradient Boosting,极端梯度提升)方法的复杂映射能力有限,以致TOC含量预测精度有限.本文提出了一种基于ConvLSTM(Convolutional Long Short-Term Memory,卷积长短时记忆)神经网络的TOC含量地震预测方法,主要包括网络搭建、数据预处理、网络训练和测试四个部分.通过训练具有强复杂映射能力的ConvLSTM神经网络来建立TOC含量与地震弹性参数(纵波阻抗、速度比以及泊松比)之间的非线性关系,以有效地预测TOC含量的区域性分布.L区的数据试算结果表明,相比于线性回归和XGBoost方法,本文方法具有更高的TOC含量预测精度.The abundance of organic matters represented by TOC(Total Organic Carbon)content is one of the important indicators to evaluate the hydrocarbon generation potential of source rocks.There is a certain correlation between seismic elastic parameters and TOC content of rocks,and this relationship may be nonlinear.Due to the limited complex mapping ability of linear regression and XGBoost(Extreme Gradient Boosting)methods,the accuracy of TOC content prediction is circumscribed.In this paper,we propose a TOC content seismic prediction method based on a ConvLSTM(Convolutional Long Short-Term Memory)neural network,which mainly includes the network construction,data preprocessing,network training and testing.The nonlinear relationship between TOC content and seismic elastic parameters(P-wave impedance,ratio of P-wave velocity to S-wave velocity,and Poisson's ratio)is established by training the ConvLSTM neural network with strong complex mapping ability to effectively predict the regional distribution of TOC content.Tests in area L show that the proposed method has higher accuracy of TOC content prediction than the linear regression and XGBoost methods.

关 键 词：TOC含量预测 ConvLSTM神经网络 地震弹性参数 

分 类 号：P631[天文地球—地质矿产勘探]