基于深度学习加速的油藏数值模拟自动历史拟合方法  

Deep-learning-based acceleration method for automatic history matching of reservoir numerical simulation

作　　者：王森[1,2] 向杰[2,3] 冯其红 杨雨萱[2] 王振 王相 WANG Sen;XIANG Jie;FENG Qihong;YANG Yuxuan;WANG Zhen;WANG Xiang(State Key Laboratory of Deep Oil and Gas,China University of Petroleum(East China),Qingdao 266580,China;School of Petroleum Engineering in China University of Petroleum(East China),Qingdao 266580,China;Nanhai East Petroleum Research Institute,CNOOC China Ltd.Shenzhen,Shenzhen 518000,China;Shandong Institute of Petroleum and Chemical Technology,Dongying 257061,China;SINOPEC Shengli Oilfield Digital Intelligent Management Service Center,Dongying 257015,China;School of Petroleum and Natural Gas Engineering,Changzhou University,Changzhou 213164,China)

机构地区：[1]深层油气全国重点实验室(中国石油大学(华东)),山东青岛266580 [2]中国石油大学(华东)石油工程学院,山东青岛266580 [3]中海石油(中国)有限公司深圳分公司南海东部研究院,广东深圳518000 [4]山东石油化工学院,山东东营257061 [5]中国石化胜利油田数智化管理服务中心,山东东营257015 [6]常州大学石油与天然气工程学院,江苏常州213164

出　　处：《中国石油大学学报（自然科学版）》2024年第5期103-114,共12页Journal of China University of Petroleum(Edition of Natural Science)

基　　金：国家自然科学基金项目(52204027);山东省自然科学基金项目(ZR2022YQ50)。

摘　　要：历史拟合是降低油藏模型不确定性的重要方法,是对油藏进行生产动态预测和开发方案设计的基础。由于油藏模型往往包含数十万甚至数百万个不确定参数,重复调用油藏数值模拟器将对历史拟合的计算效率造成严重影响。针对该问题,提出一种基于多样视角深度全卷积编码-解码神经网络的油藏数值模拟代理模型构建方法。模型包含编码-解码单元和时间处理单元两部分,嵌入多样视角网络(VoVNet)的编码-解码单元实现输入参数的空间特征提取,而时间处理单元用来捕获时间的影响。经过训练的代理模型能够以图像-图像的形式实现从油藏渗透率场到压力场和饱和度场的预测,从而为自动历史拟合提供快速的生产动态响应。将所构建的代理模型与多重数据同化集合平滑方法(ES-MDA)结合,形成基于深度学习加速的油藏数值模拟自动历史拟合方法。结果表明:所提出的代理模型能够有效预测油藏压力场和饱和度场的动态变化;与传统油藏数值模拟相比,代理模型预测的生产动态与之相吻合,同时运算速度大大提升;基于代理模型的自动历史拟合方法能够实现油藏渗透率场的准确反演,且在计算效率上表现出较大优势。History matching is an important technique to reduce geological uncertainty in reservoir modeling,and it is the basis for oil field production prediction and development scheme design.Since a reservoir geo-model contains a lot of parameters with thousands or even millions of uncertain data,a repeated invocation of the reservoir numerical simulator tremendously impacts the computational efficiency during history matching.To solve this problem,a multi-view deep convolution coding-decoding neural network model was proposed for surrogate reservoir modeling.The model consists of a coding-decoding unit and a time-processing unit.The coding-decoding unit embedded in the variety of view networks(VoVNet)can extract the spatial features of the input parameters,while the time processing unit was used to capture the influence of time series.The trained surrogate model can predict pressure and saturation from permeability data in an image-to-image form,providing fast production performance prediction for automatic history matching.Moreover,the proposed surrogate model was incorporated into a multiple data assimilation ensemble smoother(ES-MDA)framework to create a fast deep-learning-based automatic history matching method.The results show that the proposed surrogate model can effectively predict the pressure and saturation distributions in the reservoir at a given time.The production performance predicted using the surrogate model is consistent with that calculated using the traditional reservoir simulation models,while the calculation efficiency is improved extensively.The surrogate-based automatic history matching method can provide accurate inversion of the permeability distribution and demonstrated superiorities in computational efficiency.

关 键 词：自动历史拟合 油藏数值模拟 代理模型 深度学习 

分 类 号：TE319[石油与天然气工程—油气田开发工程]