基于U-Net和CNN深度学习求取地质属性建模变差函数参数  

Obtaining variation function parameters in modeling geological attribute based on U⁃Net and CNN deep learning

作　　者：冯国庆[1] 莫海帅 吴宝峰[2] FENG Gouqing;MO Haishuai;WU Baofeng(Petroleum Engineering School,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Research Institute of Exploration and Development,PetroChina Daqing Oilfield Co.,Ltd.,Daqing,Heilongjiang 163712,China)

机构地区：[1]西南石油大学石油与天然气工程学院,四川成都610500 [2]中国石油大庆油田有限责任公司勘探开发研究院,黑龙江大庆163712

出　　处：《石油地球物理勘探》2024年第4期692-701,共10页Oil Geophysical Prospecting

摘　　要：在油气藏地质属性建模中,变差函数的求取尤为关键,一般是通过拟合实验变差函数求取变程、方位角、基台值等参数的方式获得,但当研究区内样本点数量过少时,实验变差函数拟合效果较差,从而影响属性建模质量。为了克服传统方法的不足,最大限度地利用空间数据,文中提出一种基于U-Net和CNN网络求取变差函数参数的新方法:以孔隙度属性建模为例,首先选用球状模型并利用序贯高斯模拟(SGS)算法模拟生成多组孔隙度模型,以所得孔隙度平面模型抽取的数据点为基准构成样本集;然后采用U-Net进行孔隙度模型重构,保证孔隙度分布的空间相关性;最后利用CNN对样本集进行深度学习,从而建立求取变差函数的模型。实际应用表明,利用所提方法取得的主变程方位角与通过实验变差函数拟合求取的方位角仅相差1.52°,与沉积微相展布方向一致,得到的主次变程与实验变差函数非常契合,证明求取的变差函数结果可靠。同时,该方法简化了地质建模工作流程,减少了求取实验变差函数主观性,降低了研究区内样本点数量的局限性,为变差函数的预测提供了一种新的思路。In the modeling of geological attribute of oil and gas reservoirs,obtaining the variation function is es⁃pecially critical,which is generally obtained by fitting the experimental variation function to acquire parameters such as varve,azimuth,and abutment value.However,when the number of sample points in the study area is insufficient,it will lead to a poor fitting effect,thereby affecting the quality of attribute modeling.To overcome the shortcomings of traditional experimental variation function modeling and make the most use of spatial data,this paper proposes a new method based on U⁃Net and CNN networks to predict the parameters of the variation function.The data points extracted from the porosity plane model obtained by sequential Gaussian simulation are taken as the benchmark.Using the U⁃Net network structure,the porosity distribution is reconstructed to maintain spatial correlation.Subsequently,a CNN network structure is applied to the sample set for deep lear⁃ning,thereby developing a model to predict the variation function.The practical application shows that the prin⁃cipal range direction obtained by the proposed method in this paper deviates by only 1.52°from that obtained using the experimental range function,closely matching the distribution direction of sedimentary microfacies.Meanwhile,the obtained principal and secondary ranges closely align with the experimental variation function,confirming the reliability of the model’s variation function results.At the same time,the method also simplifies the geological modeling workflow,reduces the subjectivity of finding the experimental variation function,and reduces the limitations posed by a small number of data points in the study area.It offers a novel approach for the predictive research of the variation function.

关 键 词：属性建模 深度学习 模型重构 SGS算法 变差函数 

分 类 号：P631[天文地球—地质矿产勘探]