二维卷积神经网络驱动的砂地比地震预测方法  

2D convolutional neural network driven sandstone ratio prediction method with seismic data

作　　者：张宪国 储飞跃[1,2] 黄德榕 董春梅 刘晓宇 ZHANG Xianguo;CHU Feiyue;HUANG Derong;DONG Chunmei;LIU Xiaoyu(Shandong Provincial Key Laboratory of Deep Oil and Gas,Qingdao,Shandong 266580,China;School of Geosciences,China University of Petroleum(East China),Qingdao,Shandong 266580,China;Research Institute of Petroleum Exploration&Development,PetroChina,Beijing 100083,China;CNOOC Energy Development Corporation Ltd Engineering Branch,Tianjin 300450,China)

机构地区：[1]山东省深层油气重点实验室,山东青岛266580 [2]中国石油大学(华东)地球科学与技术学院,山东青岛266580 [3]中国石油勘探开发研究院,北京100083 [4]中海油能源发展有限公司工程技术分公司,天津300450

出　　处：《中国矿业大学学报》2022年第6期1128-1137,共10页Journal of China University of Mining & Technology

基　　金：国家自然科学基金项目(42172144);国家科技重大专项(2017ZX05009001)。

摘　　要：地震属性预测砂地比是油田勘探中沉积相研究的重要方法,但是缺乏有效的方法构建地震属性与砂地比之间的非线性关系,且少井区样本不足,制约了预测精度和可靠性.本文以辽河东部凹陷铁匠炉地区为例,探索少井区地震多属性预测扇三角洲砂地比分布的方法,提出用聚类分析优选地震属性,采用二维卷积神经网络方法可以有效地对少井区井震联合的复杂碎屑岩储层砂地比进行预测.研究结果表明:地震属性二维图像取值可以提高地震属性取值可靠性,通过图像旋转对称获取虚拟样本将样本量扩充为8倍,构建二维卷积神经网络模型,预测研究区砂地比平面分布,采用二维卷积神经网络方法与BP神经网络、支持向量机方法进行对照,预测精度有了显著提高,可达92.4%,有效解决了少井区样本量不足的问题.The prediction of sandstone ratio by seismic attributes is an important method for the study of sedimentary facies in oilfield exploration, but the lack of effective methods to construct the nonlinear relationship between seismic attributes and sandstone ratio and the lack of samples in sparse well areas restrict the prediction accuracy and reliability. Tiejianglu area in the Eastern Liaohe Sag is taking as an example to explore the method of seismic multi-attribute prediction of sandstone ratio distribution in fan delta in sparse well area. Cluster analysis is used to optimize seismic attributes;the method of two-dimensional image valuing is taken to improve the reliability of seismic attribute value;virtual samples obtention from image rotation symmetry is used to expand the sample set to 8 times. Based on the above process, 2 D convolution neural network model is built to predict the plane distribution of sandstone ratio in the study area. The result is compared with BP neural network and SVM. The results show that the 2 D convolution neural network method can significantly improve the prediction accuracy to 92.4%. This method provides an effective solution forthe problem of insufficient sample size in sparse well area.

关 键 词：二维卷积神经网络 扇三角洲 少井区 地震多属性 砂地比 

分 类 号：P631[天文地球—地质矿产勘探]