基于双输入输出卷积神经网络代理模型的油藏自动历史拟合研究  

Automatic history matching of reservoirs based on dual input-output convolutional neural network agent model

作　　者：陈旭 张凯[1,2] 刘晨[3,4] 张金鼎 张黎明 姚军[1] CHEN Xu;ZHANG Kai;LIU Chen;ZHANG Jinding;ZHANG Liming;YAO Jun(School of Petroleum Engineering,China University of Petroleum(East China),Qingdao City,Shandong Province,266580,China;Civil Engineering School,Qingdao University of Technology,Qingdao City,Shandong Province,266520,China;CNOOC Research Institute Co.,Ltd.,Beijing City,100028,China;State Key Laboratory of Offshore Oil Exploitation,Beijing City,100028,China)

机构地区：[1]中国石油大学(华东)石油工程学院,山东青岛266580 [2]青岛理工大学土木工程学院,山东青岛266520 [3]中海油研究总院有限责任公司,北京100028 [4]海洋石油开发国家重点实验室,北京100028

出　　处：《油气地质与采收率》2024年第3期165-177,共13页Petroleum Geology and Recovery Efficiency

基　　金：国家自然科学基金面上项目“基于强化学习的离线-在线交互式油藏开发生产实时优化方法”(52274057);“基于迁移学习的油藏开发注采优化方法研究”(52074340);“基于电磁支撑剂的水力压裂裂缝监测理论与方法”(51874335)。

摘　　要：传统油藏自动历史拟合方法需进行多次计算耗时的油藏数值模拟,而深度学习代理模型可以实现高效且精度近似的油藏数值模拟替代计算。在基于深度学习代理模型的油藏自动历史拟合方法中,通常将采用油藏自动历史拟合方法进行调整的油藏不确定性参数作为深度学习代理模型的输入参数。现有的深度学习代理模型常为单一输入输出的神经网络模型架构,并未考虑油藏自动历史拟合方法需要对多个油藏不确定性参数进行调整,且需要训练多个深度学习代理模型以实现对油藏含水饱和度场分布及压力场分布的预测。为此,提出了一种基于双输入输出卷积神经网络代理模型的油藏自动历史拟合方法,将油藏渗透率场分布及相对渗透率参数作为输入,使用双输入输出卷积神经网络同时对油藏含水饱和度场分布及压力场分布进行预测,利用Peaceman方程计算产量,并耦合到多重数据同化集合平滑器(ES-MDA)方法中,对油藏渗透率场分布及相对渗透率参数进行反演更新,实现较为高效的油藏自动历史拟合求解。研究结果表明:双输入输出卷积神经网络代理模型在指定时间步的油藏含水饱和度场分布、压力场分布的预测精度均为93%以上。相较于传统油藏自动历史拟合方法,基于双输入输出卷积神经网络代理模型的油藏自动历史拟合方法避免了多次调用油藏数值模拟器的计算耗时问题,提高了拟合效率。The conventional reservoir automatic history matching method requires multiple computationally time-consuming reservoir numerical simulations.Deep learning agent models can perform alternative reservoir numerical simulation calculations with approximate accuracy and greater computational efficiency.In the reservoir automatic history matching method based on the deep learning agent model,the reservoir uncertainty parameters adjusted by the reservoir automatic history matching method are usually used as the input parameters of the deep learning agent model.Existing deep learning agent models are often single input-output neu‐ral network model architectures.They do not consider that reservoir automatic history matching methods require the adjustment of multiple reservoir uncertainty parameters.Multiple deep learning agent models need to be trained to predict water saturation field distribution and pressure field distribution in reservoirs.For solving this problem,a reservoir automatic history matching method based on a dual input-output convolutional neural network agent model was proposed to simultaneously predict the water saturation field distribution and pressure field distribution in a reservoir by using a dual input-output convolutional neural network,with the reservoir permeability field distribution and phase permeability parameters as input.The production was calculated with the help of the Peaceman equation.It was coupled to the ensemble smoother with multiple data assimilation(ES-MDA)methods to invert the reservoir permeability field distribution and phase permeability parameters to achieve a more efficient reservoir automatic history matching solution.The results of the study show that the prediction accuracy of the reservoir water saturation field distribution and pressure field distribution is above 93%at the specified time step based on the dual input-output convolutional neural network agent model.Compared with the traditional reservoir automatic history matching method,the proposed reservoir autom

关 键 词：油藏自动历史拟合 油藏数值模拟 深度学习 代理模型 双输入输出卷积神经网络 

分 类 号：TE319[石油与天然气工程—油气田开发工程]