基于深度学习贝叶斯模型平均代理的油藏自动历史拟合研究  

Study on Automatic History Matching of Reservoirs Based on Deep Learning Bayesian Model Averaging Surrogates

作　　者：张凯[1,2] 陈旭[1] 刘丕养 张金鼎 张黎明 姚军[1] ZHANG Kai;CHEN Xu;LIU Piyang;ZHANG Jinding;ZHANG Liming;YAO Jun(School of Petroleum Engineering in China University of Petroleum(East China),Qingdao,Shandong 266580,China;Civil Engineering School,Qingdao University of Technology,Qingdao,Shandong 266520,China)

机构地区：[1]中国石油大学(华东)石油工程学院 [2]青岛理工大学土木工程学院

出　　处：《钻采工艺》2025年第1期147-156,共10页Drilling & Production Technology

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”(编号:52274057);“基于迁移学习的油藏开发注采优化方法研究”(编号:52074340)。

摘　　要：油藏自动历史拟合过程中,需要频繁调用数值模拟器进行正向计算,导致计算时间长、资源消耗大。基于深度学习的油藏数值模拟代理模型提供了一种快速计算油水井生产动态的替代方案。然而,单一神经网络产量预测代理模型在特征提取和学习能力方面存在局限性。基于空间特征构建的代理模型侧重于学习油藏渗流的空间特性,但忽视了时间维度;基于时空特征构建的模型虽然擅长捕捉时间序列特征,却在空间特征学习方面不足。为此,文章提出了一种基于深度学习的贝叶斯模型平均代理方法,利用贝叶斯模型平均方法对两种深度学习代理模型进行集成,结合二者优势,增强代理模型对油藏特征的多维度学习能力,从而提高预测精度。该方法进一步结合多重数据同化集合平滑器,应用于实际油藏历史拟合中。实验结果表明,基于深度学习贝叶斯模型平均代理的历史拟合方法能够在保证高效计算的同时,准确拟合油藏实际生产动态,为快速、精确的历史拟合提供了一种创新解决方案。During the process of automatic history matching for reservoirs,frequent invocation of numerical simulators for forward calculations results in extended computation times and significant resource consumption.A deep learning-based numerical simulation surrogate model for reservoirs offers a rapid computational alternative for predicting the production dynamics of oil and water wells.However,a single neural network production forecast surrogate model has limitations in feature extraction and learning capabilities.Spatial feature surrogate models focus on learning the spatial characteristics of reservoir flow but neglect the temporal dimension.While spatiotemporal feature models excel at capturing time series characteristics,they fall short in learning spatial features.To address this problem,this paper proposes a Bayesian model averaging surrogate method based on deep learning,which utilizes Bayesian model averaging algorithms to integrate spatial feature models with spatiotemporal feature models.By combining the strengths of both models,this approach enhances the surrogate model's multidimensional learning capability of reservoir characteristics,thereby improving prediction accuracy.This method is further combined with the Ensemble Smoother with Multiple Data Assimilation and applied to the history matching of actual reservoirs.Experimental results indicate that the history matching method based on the deep learning Bayesian model averaging surrogate can accurately match the actual production dynamics of reservoirs while ensuring computational efficiency.This provides an innovative solution for rapid and precise history matching.

关 键 词：深度学习 历史拟合 产量预测 贝叶斯模型平均方法 集成代理模型 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TE32[自动化与计算机技术—控制科学与工程]