一种基于机器学习的测井与地震属性联合反演方法  

An Inversion Approach Combining Logging and Seismic Attributes Based on Machine Learning

作　　者：安振芳 AN Zhen-fang(Oil Reservoir Dynamic Monitoring Center,Shengli Oifield,SINOPEC,Dongying 257100,China)

机构地区：[1]中国石油化工股份有限公司胜利油田分公司油藏动态监测中心,山东东营257100

出　　处：《河北地质大学学报》2024年第3期51-58,共8页Journal of Hebei Geo University

基　　金：中国石化胜利油田博士后科研项目(YKB2115)。

摘　　要：整装油气藏随着多年的开发在不断减少,薄层油气藏逐渐成了精细开发的对象。深度采样间隔可以小到0.1 m的测井技术在纵向上有着较高的分辨率,但井间的地下地质情况由于没有可供测井仪器下入的井孔通道而不易洞明,作为寻找油气资源主要手段且有着较好横向连续性的地震勘探,因分辨不出小于地震波长度1/4的薄层而受到制约,测井与地震联合反演兼具测井的纵向分辨率高和地震的横向连续性好的优点。基于模型驱动的井震联合反演方法要求具有先验的确定性映射算子,而基于数据驱动的井震联合反演方法则不用精确的褶积模型表达式。从地震记录到测井曲线的映射反演,无需建立初始模型,但输入数据中只有地震记录一种信息致使泛化能力不高,将测井属性与地震属性构成的二维矩阵作为卷积神经网络的输入,以提高泛化能力。从测井曲线到地震记录的映射反演,不受褶积模型的限制,却依赖于初始模型且尚存多解性,将测井与地震属性联合反演结果作为初始模型,并在损失函数中添加一项正则化约束条件,以降低多解性。Marmousi-Ⅱ模型测试和Blackfoot地区的应用实例结果表明,所提方法反演速度更加接近实测速度,从预测的速度剖面上能够更好地辨认出薄层。Integrated reservoirs have been decreasing with many years of exploration and development.Thin reservoirs have gradually become the object of fine development.Logging technology has a high longitudinal resolution,because its depth sampling interval can be as small as O.I meter.But it can't ascertain the geological condition between the wells,because there is no way for the instrument to go down the well.Seismic exploration,the main means to search for oil and gas resources and has a good transverse continuity,is limited by the inability to distinguish thin layers less than a quarter of the length of seismic waves.The inversion combining logging and seismic data has the advantages of high longitudinal resolution of logging and good transverse continuity of seismic.The model-driven inversion method combining logging and seismic data requires a prior deterministic mapping operator,while the data-driven inversion method combining logging and seismic data does not requires an exact convolution model expression.The inversion mapping from seismic record to well logs does not need to establish an initial model,but only one information is seismicrecord in the input data which resulting in poor generalization ability.The two-dimensional matrix composed of logging attribute and seismic attribute is used as the input of the convolutional neural network to improve generalization ability.The inversion mapping from well logs to seismic record is not limited by convolution model,but depends on the initial model with multiple solutions.The result of inversion combining logging and seismic attributes is taken as the initial model,and a regularization constraint condition is added to the loss function to reduce the multiple solutions.The results of Marmousi-model testing and the application of field data from Blackfoot area show that the inversion velocity of the proposed method is closer to the measured velocity,and the thin layer can be identified better from the predicted velocity profile.

关 键 词：速度建模 井震联合反演 属性 机器学习 

分 类 号：P631[天文地球—地质矿产勘探]