摘要
相比地震反演方法和技术,基于多属性回归方法的储层预测技术能够缓解分辨率有限、过于模型化等问题,但模型泛化能力不足常造成井间薄储层预测结果不合理。为此,提出基于多层感知机深度学习网络的多属性回归薄储层预测方法,即以地震数据(提供背景信息)、90°相移数据(提供储层结构近似信息)、储层不连续界限属性(提供储层空间分布信息)为输入,以井点高频自然伽马为期望输出,利用多层感知机深度学习网络训练模型,预测井间自然伽马值,利用自然伽马值与砂—泥岩性的高度相关特性刻画薄储层。A油田实际资料测试表明,自然伽马预测值与真实值平均相关系数达到86.4%(训练集,10口井)和85.5%(验证集,两口井),明显优于传统多属性回归方法。应用该方法解释重点层段6套小层,薄储层预测结果与156口井实钻砂岩厚度平均相关系数较相移数据提升约38%,证实该方法应用效果良好。
Characterization of thin reservoirs is significantly important in seismic exploration.Compared with the method based on seismic inversion,reservoir prediction based on seismic multiple-attribute regression(MAR)can enhance resolution and alleviate the over modeling issue,but the poor ability of gene-ralizing the trained model in MAR frequently causes the instability of the estimated result between wells.A multi-layer perceptron(MLP)-based MAR method is proposed for characterizing thin sandstone-shale reservoirs.This method takes seismic data(background information),90°-phase data(the framework of reservoir structure)and discontinuious reservor boundary(which is a selfdeve-loped seismic attribute measuring the discontinuity of reservoir)as inputs,and the gamma-ray(GA)logs of wells as expected outputs,uses the MLP deep neural network to train the model for estimating the GA data,and finally characterizes thin re-servoirs based on the close lithologic relationship between sandstone and shale.Applications to field data from an offshore oilfield A show that the correlated coefficient between the estimated and the true GA of wells has reached 0.855 in a training set with 10 wells,and 0.864 in a prediction set with 2 wells.They are significantly better than the results from the traditional MAR method.Based on the GA data,we interpreted 6 top surfaces for finely describing the reservoir in a target area,and extracted the reservoir-sensitive seismic attribute(Sum of Negative Amplitude,SNA)along the horizon to assess the association between the SNA and the sum thickness(ST)of the reservoir drilled in 156 wells.The SNA based on GA data shows relatively high association with the ST.The correlation coefficient between the SNA based on GA estimation and the ST is about 38%higher than that between the SNA based on the 90°-phase data and the ST,which further confirmed the feasibility of the proposed MLP-based MAR method.
作者
杜昕
范廷恩
董建华
聂妍
范洪军
郭泊洋
DU Xin;FAN Ting’en;DONG Jianhua;NIE Yan;FAN Hongjun;GUO Boyang(State Key Laboratory of Offshore Oil Exploitation,Beijing 100020,China;CNOOC Research Institute Ltd.,Beijing 100020,China;School of Energy Resources,China University of Geosciences(Beijing),Beijing 100083,China)
出处
《石油地球物理勘探》
EI
CSCD
北大核心
2020年第6期1178-1187,1159,共11页
Oil Geophysical Prospecting
关键词
薄层预测
多层感知机
深度学习
多属性回归
thin sand-shale reservoir
multi-layer perceptron
deep learning
seismic multiple-attribute regression
作者简介
杜昕,北京市朝阳区太阳宫南街6号院中海油研究总院有限责任公司,100020。Email:duxin7@cnooc.com.cn,助理工程师,1991年生,2015、2018年分别获中国石油大学(北京)勘查技术与工程专业学士学位和地质资源与地质工程专业硕士学位,现就职于中海油研究总院有限责任公司,主要从事储层地球物理与深度学习等领域科研及生产。
