en
×

分享给微信好友或者朋友圈

使用微信“扫一扫”功能。
作者简介:

毕剑飞(1995—),男,安徽淮南人,在读博士研究生,从事非常规油气实验及模拟技术等方面的研究。E-mail:jianfei_bi@163.com。

通讯作者:

李靖(1990—),男,陕西延安人,副教授,博士。E-mail:lijingsuc@cup.edu.cn。

中图分类号:TE319

文献标识码:A

文章编号:1009-9603(2023)03-0104-11

DOI:10.13673/j.pgre.202205049

参考文献 1
史长林,魏莉,张剑,等.基于机器学习的储层预测方法[J].油气地质与采收率,2022,29(1):90-97.SHI Changlin,WEI Li,ZHANG Jian,et al.Reservoir prediction method based on machine learning[J].Petroleum Geology and Recovery Efficiency,2022,29(1):90-97.
参考文献 2
秦峰,闫正和,唐圣来,等.基于深度学习的代理模型在实际气藏三维模拟中的应用[J].油气地质与采收率,2022,29(1):152-159.QIN Feng,YAN Zhenghe,TANG Shenglai,et al.Application of agent models based on deep learning in actual three-dimension‐ al gas reservoir simulation[J].Petroleum Geology and Recovery Efficiency,2022,29(1):152-159.
参考文献 3
匡立春,刘合,任义丽,等.人工智能在石油勘探开发领域的应用现状与发展趋势[J].石油勘探与开发,2021,48(1):1-11.KUANG Lichun,LIU He,REN Yili,et al.Application and devel‐ opment trend of artificial intelligence in petroleum exploration and development[J].Petroleum Exploration and Development,2021,48(1):1-11.
参考文献 4
陈欢庆,唐海洋,吴桐,等.精细油藏描述中的大数据技术及其应用[J].油气地质与采收率,2022,29(1):11-20.CHEN Huanqing,TANG Haiyang,WU Tong,et al.Big data technology and its application in fine reservoir description[J].Pe‐ troleum Geology and Recovery Efficiency,2022,29(1):11-20.
参考文献 5
孙金声,刘凡,程荣超,等.机器学习在防漏堵漏中研究进展与展望[J].石油学报,2022,43(1):91-100.SUN Jinsheng,LIU Fan,CHENG Rongchao,et al.Research progress and prospects of machine learning in lost circulation control[J].Acta Petrolei Sinica,2022,43(1):91-100.
参考文献 6
王相,杨耀忠,何岩峰,等.基于深度学习的油井工况智能诊断技术研究及应用[J].油气地质与采收率,2022,29(1):181-189.WANG Xiang,YANG Yaozhong,HE Yanfeng,et al.Research and application of intelligent diagnosis technology of oil well working conditions based on deep learning[J].Petroleum Geolo‐ gy and Recovery Efficiency,2022,29(1):181-189.
参考文献 7
ZOTKIN O,OSOKINA A,SIMONOV M,et al.A novel ap‐ proach to refinment reservoir proxy model using machine-learn‐ ing techniques[C].Baku:SPE Annual Caspian Technical Confer‐ ence,2019.
参考文献 8
HAN D,KWON S,KIM J,et al.Comprehensive analysis for pro‐ duction prediction of hydraulic fractured shale reservoirs using proxy model based on deep neural network[C].Virtual:SPE An‐ nual Technical Conference and Exhibition,2020.
参考文献 9
KIM Y D,DURLOFSKY L J.A recurrent neural network-based proxy model for well-control optimization with nonlinear output constraints[J].SPE Journal,2021,26(4):1 837-1 857.
参考文献 10
李野,陈松灿.基于物理信息的神经网络:最新进展与展望[J].计算机科学,2022,49(4):254-262.LI Ye,CHEN Songcan.Physics-informed neural networks:re‐ cent advances and prospects[J].Computer Science,2022,49(4):254-262.
参考文献 11
CHEN Zhao,LIU Yang,SUN Hao.Physics-informed learning of governing equations from scarce data[J].Nature Communica‐ tions,2021,12(1):1-13.
参考文献 12
孙岿.基于改进KNN算法的潜山复杂岩性测井识别方法[J].特种油气藏,2022,29(3):18-27.SUN Kui.Logging identification method of complex lithology in buried hill based on the improved KNN algorithm[J].Special Oil & Gas Reservoirs,2022,29(3):18-27.
参考文献 13
KARNIADAKIS G E,KEVREKIDIS I G,LU Lu,et al.Physicsinformed machine learning[J].Nature Reviews Physics,2021,3(6):422-440.
参考文献 14
PSICHOGIOS D C,UNGAR L H.A hybrid neural network‐first principles approach to process modeling[J].AIChE Journal,1992,38(10):1 499-1 511.
参考文献 15
MEADE J A J,FERNANDEZ A A.The numerical solution of linear ordinary differential equations by feedforward neural net‐ works[J].Mathematical and Computer Modelling,1994,19(12):1-25.
参考文献 16
LAGARIS I E,LIKAS A,FOTIADIS D I.Artificial neural net‐ works for solving ordinary and partial differential equations[J].IEEE Transactions on Neural Networks,1998,9(5):987-1 000.
参考文献 17
KARPATNE A,ATLURI G,FAGHMOUS J H,et al.Theoryguided data science:a new paradigm for scientific discovery from data[J].IEEE Transactions on Knowledge and Data Engi‐ neering,2017,29(10):2 318-2 331.
参考文献 18
KARPATNE A,KANNAN R,KUMAR V.Knowledge guided machine learning:Accelerating discovery using scientific knowl‐ edge and data[M].Florida:CRC Press,2022.
参考文献 19
RAISSI M,PERDIKARIS P,KARNIADAKIS G E.Physics-in‐ formed neural networks:a deep learning framework for solving forward and inverse problems involving nonlinear partial differ‐ ential equations[J].Journal of Computational Physics,2019,378:686-707.
参考文献 20
WANG Nanzhe,ZHANG Dongxiao,CHANG Haibin,et al.Deep learning of subsurface flow via theory-guided neural network[J].Journal of Hydrology,2020,584:124700.
参考文献 21
XU Rui,ZHANG Dongxiao,RONG Miao,et al.Weak form theo‐ ry-guided neural network(TgNN-wf)for deep learning of sub‐ surface single-and two-phase flow[J].Journal of Computational Physics,2021,436:110318.
参考文献 22
陆至彬,瞿景辉,刘桦,等.基于物理信息神经网络的传热过程物理场代理模型的构建[J].化工学报,2021,72(3):1 496-1 503.LU Zhibin,QU Jinghui,LIU Hua,et al.Surrogate modeling for physical fields of heat transfer processes based on physics-in‐ formed neural network[J].CIESC Journal,2021,72(3):1 496-1 503.
参考文献 23
CAI Shengze,MAO Zhiping,WANG Zhicheng,et al.Physics-in‐ formed neural networks(PINNs)for fluid mechanics:a review[J].Acta Mechanica Sinica:English Series,2021,37(12):1 727-1 738.
参考文献 24
BAYDIN A G,PEARLMUTTER B A,RADUL A A,et al.Auto‐ matic differentiation in machine learning:a survey[J].Journal of Machine Learning Research,2018,18(153):1-43.
参考文献 25
RONG Miao,ZHANG Dongxiao,WANG Nanzhe.A lagrangian dual-based theory-guided deep neural network[J].Complex & In‐ telligent Systems,2022,8:4 849-4 862.
参考文献 26
ZHANG Dongxiao,LU Zhiming.An efficient,high-order pertur‐ bation approach for flow in random porous media via KarhunenLoève and polynomial expansions[J].Journal of Computational Physics,2004,194(2):773-794.
参考文献 27
GHANEM R G,SPANOS P D.Stochastic finite elements:A spectral approach[M].Massachusetts:Courier Corporation,2003.
参考文献 28
BAJPAI M K,SINGH K K,GIAKOS G.Machine vision and augmented intelligence:Heory and applications[M].Singapore:Springer,2021.
参考文献 29
ZHU Yinhao,ZABARAS N,KOUTSOURELAKIS P S,et al.Physics-constrained deep learning for high-dimensional surro‐ gate modeling and uncertainty quantification without labeled data [J].Journal of Computational Physics,2019,394:56-81.
参考文献 30
HE Qizhi,BARAJAS-SOLANO D,TARTAKOVSKY G,et al.Physics-informed neural networks for multiphysics data assimi‐ lation with application to subsurface transport[J].Advances in Water Resources,2020,141:103610.
参考文献 31
TAN Chuanqi,SUN Fuchun,KONG Tao,et al.A survey on deep transfer learning[C].Rhodes:International Conference on Artifi‐ cial Neural Networks,2018.
目录contents

    摘要

    渗流代理模型的构建是油气藏模拟技术研究的前沿方向,而目前广泛使用的纯数据驱动渗流代理模型无理论支撑,对数据数量和质量的要求较高,很大程度上限制了渗流代理模型的发展。为此提出了数据驱动与物理驱动相融合的双驱动渗流代理模型,其在纯数据驱动渗流代理模型的基础上,融合油气渗流理论,模拟预测油气渗流过程。结果表明:相较于纯数据驱动渗流代理模型,即使训练数据极度稀疏,双驱动渗流代理模型仍具有较高的预测精度;通过在训练数据中加入不同等级的干扰噪声,验证了双驱动渗流代理模型的鲁棒性优于纯数据驱动渗流代理模型;通过迁移学习,将训练好的双驱动渗流代理模型应用到新的渗流场,实现了快速收敛并节省了计算资源。

    Abstract

    The building of flow surrogate models is the frontier of simulation technology research for oil and gas reservoirs. However,the currently widely used pure data-driven flow surrogate models have no theoretical support and require a high data volume and data quality,which greatly limits the development of flow surrogate models. Therefore,this paper propos‐ es a flow surrogate model based on a data-driven and physics-driven method. On the basis of the pure data-driven flow sur‐ rogate model,it takes advantage of the flow theory to simulate and predict oil and gas flow processes. Firstly,the dual-driv‐ en flow surrogate model is compared with the pure data-driven model. The results show that the proposed model can still maintain high prediction accuracy even if the training data is extremely sparse. Secondly,the robustness of the dual-driven model is explored by adding different levels of noise interference to the training data,and it is verified that the proposed model outperforms the pure data-driven flow surrogate model. Finally,the trained dual-driven flow surrogate model is ap‐ plied to a new flow field through transfer learning. The model can achieve rapid convergence and save computing resources.

  • 随着高性能计算硬件的不断完善,以纯数据驱动为基础的人工智能技术在科学和商业领域取得了巨大的成功。近年来,纯数据驱动人工智能技术在油气行业获得广泛关注,中外石油公司纷纷朝着数字化、信息化和智能化方向转型,以期提高油气勘探开发的质量和效率[1-6]。目前,纯数据驱动渗流代理模型发展迅速,已经被用于解决油气渗流的模拟和预测问题,节省了大量计算资源[7-9]。然而,在某些复杂的工况条件下,数据的获取成本高、信噪比低,这种稀疏且伴有噪声的数据使得模型的训练易产生过拟合情况,模型预测精度面临较大挑战[10-12]。此外,纯数据驱动渗流代理模型的“黑盒” 属性及其对训练数据数量和质量的双重依赖性还难以解决。在纯数据驱动渗流代理模型技术快速发展之前,油气行业都是采用物理驱动模型模拟油气渗流过程,同时积累了丰富的油气渗流理论知识和渗流模拟技术。但物理驱动模型解决问题的过程难以深度融合实际的观测信息,而现行的纯数据驱动渗流代理模型又没有考虑油气渗流理论知识,所以2种方法都造成了信息资源的巨大浪费[13]

  • 在油气渗流过程中,弹性多孔介质中的单相可压缩流体不稳定渗流数学模型中的导压方程是一个二阶抛物线型偏微分方程。1992年,含有物理意义的微分方程已被嵌入到神经网络中,但由于当时计算能力的限制,并未对该方法进行进一步研究[14-16]。KARPATNE等提出了物理知识和数据驱动相融合的方法,但只是将简单的先验知识加入到模型的损失函数中,并未融合更具有物理意义的微分方程[17-18]。RAISSI 等在前人研究的基础上,提出了 PINN(Physics-informed Neural Networks)算法,用于求解偏微分方程的正问题及反问题[19],真正意义上将物理知识和深度学习模型融合,从此开启了人工智能融合专业知识的新范式[1120-23]。基于此,笔者以二维非均质渗流场为例,利用神经网络的自动微分技术将油气渗流理论嵌入到深度学习模型中,形成双驱动渗流代理模型。该模型融合了数据驱动与物理驱动,改变了纯数据驱动渗流代理模型的 “黑盒”属性,使其具有理论支撑,在降低模型数据依赖性的同时提高了其鲁棒性,为渗流代理模型的构建提供了新思路。

  • 1 双驱动渗流代理模型

  • 1.1 基本原理

  • 微分方程是随着微积分发展起来的,是连接数学方法与工程问题之间的桥梁。对于描述油气渗流过程的复杂偏微分方程,可采用模拟器求其数值解,如有限元法、有限体积法和有限差分法,以上方法的共同特点是对求解区域进行离散,离散程度越高则所得的数值解越逼近真实解。对于油藏等大尺度模型,为了减少计算复杂度,常常对离散网格进行粗化。即便如此,数值方法求解偏微分方程仍是一个费时费力的过程。而相比数值微分,自动微分是一种基于链式求导法则的微分技术,目前已成为深度学习的核心算法之一。如图1 所示,深度学习中的自动微分技术是通过保留计算图,在反向传播过程中调用中间每一步的微分结果并连乘,从而计算输出结果对输入时空变量的导数[24]

  • 图1 神经网络中输出压力对输入时空变量自动微分过程

  • Fig.1 Automatic differential calculation of output pressure value against input space-time variable in neural network

  • 与模拟器普遍采用差分代替微分的数值算法不同,深度学习通过自动微分技术计算导数,其无论是精度还是速度都优于数值方法。通过自动微分技术,油气渗流理论更容易融合到深度学习模型中,从而实现双驱动渗流代理模型。

  • 1.2 模型结构

  • 将描述油气渗流过程的导压方程嵌入到深度学习模型中。对于二维均质渗流场,单相渗流导压方程的表达式为:

  • Ctpt=Kμ2px2+2py2
    (1)
  • 微分方程的初始条件、定压边界条件和封闭边界条件的表达式分别为:

  • p(x,y,t)t=0=pic (x,y)
    (2)
  • p(x,y,t)(x,y)Γ=pbc(x,y)
    (3)
  • p(x,y,t)yy=±b=0
    (4)
  • 双驱动渗流代理模型是由纯数据驱动渗流代理模型演化而来。纯数据驱动渗流代理模型虽然可以在大量训练数据的驱动下通过反向传播、重复迭代和不断降低损失函数的方法学习输入与输出数据间的内在逻辑,但模型具有“黑盒”属性。而双驱动渗流代理模型的损失函数除了包括纯数据驱动渗流代理模型所含有的训练数据残差项外,还加入了能够体现物理驱动的正则项,包括初始条件残差项、边界条件残差项和控制方程残差项,其模型结构如图2所示。

  • 图2 双驱动渗流代理模型结构

  • Fig.2 Structure of dual-driven flow surrogate model

  • 为了有效训练模型使其满足物理规律,除了需要 Ntrain个训练数据点外,还需要采集 Nc个配点,Nbc 个边界条件点和Nic个初始条件点,其中配点是无标签的训练数据点。各残差项均采用均方误差 (MSE)计算,表达式分别为:

  • ldata =1Ntrain i=1Ntrain Nxi,yi,ti,θ-pxi,yi,ti2
    (5)
  • lic=1Nici=1Nic Nxi,yi,ti,θt=0-pic(x,y)2
    (6)
  • lbc=1Nbci=1Nbe Nxi,yi,ti,θ(x,y)Γ-pbc(x,y)2
    (7)
  • lpde=1Nci=1Nc Kμ2Nxi,yi,ti,θx2+2Nxi,yi,ti,θy2-CtNxi,yi,ti,θt2
    (8)
  • 损失函数定义为:

  • Loss =λ1ldata+λ2lic+λ3lbc+λ4lpde
    (9)
  • 其中,ldataliclbc是由标签数据点训练得到的残差项,具有相同的物理含义和相似的数量级,因此可令 λii = 1,2,3)= 1。lpde是由无标签的配点根据微分方程得到的残差项,其无论是物理含义还是数量级均与前 3 个残差项具有较大的差别,因此需要权重系数 λ4使 lpde和前 3 个残差项有效融合。为实现自动优化权重系数 λ4,避免根据经验手动调节超参数,本文将最小化损失函数问题转化成为拉格朗日对偶问题[25]

  • 基于纯数据驱动渗流代理模型构建的双驱动渗流代理模型耦合了油气渗流理论,使得深度学习模型由纯数据驱动的“黑盒”模型变成数据物理双驱动的综合智能模型。此外,根据K-L分解法生成符合某种协方差的随机渗透率场 Kxy[2026-27]。因此,考虑二维非均质渗流场的单相渗流导压方程的表达式为:

  • μCtp(x,y,t)t=xK(x,y)p(x,y,t)x+yK(x,y)p(x,y,t)y
    (10)
  • 2 模型验证

  • 根据 K-L 分解法生成 50 m×50 m 的随机渗透率场(图3a)。渗流区域的左右两边为线性的压力边界,压力分别为 5 和 10 MPa,上下两边定为封闭边界,中间渗流区域的初始压力为 5 MPa。流体黏度为 20 mPa·s,综合压缩系数为 33×10-4 MPa-1,孔隙度为 20%,模拟时长为 70 d 且步长为 1 d,并以模拟器计算得到的压力场结果(图3b)作为参考解。

  • 图3 随机渗透率场及模拟压力场

  • Fig.3 Random permeability field and simulated pressure field

  • 纯数据驱动和双驱动渗流代理模型均采用层全连接的神经网络,每层 64 个神经元,并以谷歌大脑提出的 Swish函数作为激活函数[28]。为了验证模型性能,分别将纯数据驱动和双驱动渗流代理模型的预测结果与模拟器的计算结果进行对比,并通过在渗流区域均匀取点的方式进行相关性分析。采用相对误差 L2 和决定系数 R2 作为评价指标[29-30],表达式分别为:

  • L2=Ω [p(x,y,t)-N(x,y,t,θ)]2dxdyΩ p2(x,y,t)dxdy
    (11)
  • R2=1-Ω [p(x,y,t)-N(x,y,t,θ)]2dxdyΩ [p(x,y,t)-p-(x,y,t)]2dxdy
    (12)
  • 对随机渗透率场分别采用纯数据驱动和双驱动渗流代理模型进行训练,2 个模型采用的训练数据量和训练轮次均为 10 000。需要注意的是,虽然网络权重是随机初始化的,但从图4可以看出,在训练10 000个轮次后2个模型均可以较好的收敛。由图5可以看出纯数据驱动和双驱动渗流代理模型在一定的训练条件下均接近模拟器计算结果。双驱动渗流代理模型具有很高的模拟预测精度,相对误差和决定系数分别为 1.74×10-5 和 0.999 39,而纯数据驱动渗流代理模型模拟预测精度略低,相对误差和决定系数分别为2.32×10-5 和0.998 545。

  • 3 结果与讨论

  • 3.1 稀疏数据训练的渗流代理模型

  • 现场观测数据的稀疏性是纯数据驱动渗流代理模型在石油工业应用的限制条件之一。为分析双驱动渗流代理模型对训练数据量的依赖性,纯数据驱动和双驱动渗流代理模型均设定4种不同数据量的训练情景,数据量分别为10,40,400和4 000。

  • 图4 不同初始化条件下2种渗流代理模型损失值随训练轮次的变化

  • Fig.4 Loss vs. training epochs for two surrogate models under different initial conditions

  • 由纯数据驱动渗流代理模型预测结果(图6)和双驱动渗流代理模型预测结果(图7)可以看出,在训练数据较多的情况下,纯数据驱动和双驱动渗流代理模型预测结果与模拟器计算结果的差异均较小。随着训练数据减少,纯数据驱动渗流代理模型预测效果变得越来越差,而双驱动渗流代理模型性能几乎没有变化,预测结果仍与模拟器计算结果保持较小的误差。

  • 图5 2种渗流代理模型预测结果

  • Fig.5 Prediction results of two flow surrogate models

  • 图6 不同训练数据量下纯数据驱动渗流代理模型预测结果

  • Fig.6 Prediction results of pure data-driven flow surrogate model under different training data volumes

  • 图7 不同训练数据量下双驱动渗流代理模型预测结果

  • Fig.7 Prediction results of dual-driven flow surrogate model under different training data volumes

  • 由图8可以看出,在训练数据充足的情况下,纯数据驱动和双驱动渗流代理模型预测精度均保持较高的水平。当训练数据量从 4 000 降至 10 时,双驱动渗流代理模型的决定系数仅下降0.7%,而纯数据驱动渗流代理模型的决定系数下降56.3%。由此可见,相较于纯数据驱动渗流代理模型的高训练数据量依赖性,双驱动渗流代理模型可以有效解决石油工业稀疏数据下的模型训练问题。

  • 图8 不同训练数据量下2种渗流代理模型准确性

  • Fig.8 Accuracy of two flow surrogate models under different training data volumes

  • 3.2 噪声数据训练的渗流代理模型

  • 纯数据驱动渗流代理模型除了在训练数据量上要求较高外,对训练数据的质量也有较高的要求。如果训练数据中干扰噪声较大并存在较多的异常值时,模型则难以收敛。然而,现场观测数据中都存在噪声,因此应用于现场的渗流代理模型需有较强的鲁棒性。为此,本文将高斯白噪声加入到 4 000 个训练数据中模拟现场观测数据。在训练数据中加入高斯白噪声的表达式为:

  • p^=p×(1+ gauss )
    (13)
  • 其中,gauss为满足均值为 0且方差为 σ的高斯白噪声。当 σ 越大时,高斯白噪声就越大。设定 4 种等级的高斯白噪声(图9),其中 σ 分别为 0.05, 0.10,0.15和0.20。

  • 利用添加了不同等级高斯白噪声的训练数据对纯数据驱动和双驱动渗流代理模型进行训练。由图10可以看出,纯数据驱动渗流代理模型的预测结果对训练数据中的高斯白噪声较为敏感,随着高斯白噪声等级增大,纯数据驱动渗流代理模型很快失效。而图11 显示双驱动渗流代理模型受训练数据中高斯白噪声的干扰较弱,即便是在设定的最高等级高斯白噪声下,仍可与模拟器的计算结果保持较好的吻合。由图12也可以看出,在低等级高斯白噪声下,纯数据驱动和双驱动渗流代理模型性能差异不明显,但随着高斯白噪声等级的增大,纯数据驱动渗流代理模型的预测精度快速降低,而双驱动渗流代理模型的预测精度虽然略有下降,但仍保持在较高水平。其中当训练数据中的高斯白噪声从等级 1上升至等级4时,双驱动渗流代理模型的相对误差只提高了 0.000 15,而纯数据驱动渗流代理模型的相对误差提高了0.021。由此,在噪声数据的训练条件下,双驱动渗流代理模型由于存在理论知识的约束使其比纯数据渗流代理模型具有更强的鲁棒性,为低信噪比数据的利用提供了新的解决方案。

  • 图9 4种不同等级高斯白噪声

  • Fig.9 Four different levels of Gaussian white noise

  • 图10 不同等级高斯白噪声训练数据下纯数据驱动渗流代理模型预测结果

  • Fig.10 Prediction results of pure data-driven flow surrogate model under different levels of Gaussian white noise in training data

  • 图11 不同等级高斯白噪声训练数据下双驱动渗流代理模型预测结果

  • Fig.11 Prediction results of dual-driven flow surrogate model under different levels of Gaussian white noise in training data

  • 3.3 不同渗透率场下模型迁移能力

  • 油藏工程专家对新储层进行模拟时需要重新构建油藏模型,而不能重复使用以往的油藏模型。相比而言,深度学习模型具有一定的迁移能力[31]。通过迁移学习,深度学习模型可以将其在相关问题中学习到的知识应用到新问题中,实现知识的重复利用。

  • 构建一个新随机渗透率场(图13a),将基于图3a 随机渗透率场上训练好的双驱动渗流代理模型作为预训练模型,并通过迁移学习应用到新随机渗透率场中。在新随机渗透率场上模拟的初始条件、边界条件等均与图3a随机渗透率场的保持一致,同时将模拟器计算结果作为参考解(图13b)。通过实验发现固定神经网络的前两层,然后训练后五层可以有效实现模型迁移。为显示迁移学习的优越性,对比不同训练时间下双驱动渗流代理模型通过重头训练和迁移学习训练后的预测结果,并计算相应训练时间下模型的预测精度。

  • 由图14、图15和图16可以看出,双驱动渗流代理模型无论是采用重头训练方法还是迁移学习方法,其准确性都随着训练时间的增加而提高,且最终两者的预测精度都达到较高水平。双驱动渗流代理模型通过100 s的迁移学习训练后,决定系数可达 0.996,超过了重头训练 400 s 模型的决定系数 0.995。由此可见,采用迁移学习方法,模型只需较短的训练时间就可以达到长时间重头训练的效果,且实现了快速收敛并节省了计算资源。此外,这也说明渗透率场的相关知识已存储于神经网络的中深层结构中。

  • 图12 不同等级高斯白噪声训练数据下 2种渗流代理模型准确性

  • Fig.12 Accuracy of two flow surrogate models under different levels of Gaussian white noise in training data

  • 图13 新随机渗透率场及模拟压力场

  • Fig.13 New random permeability field and simulated pressure field

  • 图14 不同训练时间下双驱动渗流代理模型重头训练后的预测结果

  • Fig.14 Prediction results of dual-driven flow surrogate model after ab initio training at different training time

  • 图15 不同训练时间下双驱动渗流代理模型迁移学习训练后的预测结果

  • Fig.15 Prediction results of dual-driven flow surrogate model after transfer learning at different training time

  • 图16 不同训练时间下双驱动渗流代理模型准确性

  • Fig.16 Accuracy of dual-driven flow surrogate model at different training time

  • 4 结论

  • 将油气渗流理论与纯数据驱动渗流代理模型融合建立了双驱动渗流代理模型,其在相同的训练条件下具有更高的预测精度和更低的数据依赖性,有效地解决了深度学习模型在数据稀疏情况下应用的难题。与纯数据驱动渗流代理模型不同,双驱动渗流代理模型的强鲁棒性可以较好地克服训练数据中存在干扰噪声的情况,为低信噪比训练数据的利用提供了新思路。此外,双驱动渗流代理模型具有较好的迁移性,为解决相似的油气藏模拟问题提供了重复利用已有模型的可能性。

  • 符号解释

  • b——渗透率场边界坐标,m;

  • Ct ——综合压缩系数,MPa-1

  • gauss——高斯白噪声;

  • i——训练数据的编号,无量纲;

  • K——渗透率,mD;

  • lbc——边界条件残差项,MPa2

  • ldata——训练数据残差项,MPa2

  • lic——初始条件残差项,MPa2

  • lpde——方程残差项,无量纲;

  • Loss——训练损失值,MPa2

  • L2——相对误差,无量纲;

  • N(·)——神经网络模型;

  • Nbc——边界条件训练数据量,个;

  • Nc——配点数据量,个;

  • Nic——初始条件训练数据量,个;

  • Ntrain——训练数据量,个;

  • p——无高斯白噪声的压力,MPa;

  • p^——添加高斯白噪声后的压力,MPa;

  • p-——模拟压力均值,MPa;

  • pbc——边界条件压力,MPa;

  • pic——初始条件压力,MPa;

  • p1p2,···,pn——神经网络中间节点值,无量纲;

  • Δp——压力绝对误差,MPa;

  • R2 ——决定系数,无量纲;

  • t——时间,d;

  • ti ——第i个训练数据点的时间,d;

  • xy——空间坐标,m;

  • xiyi ——第i个训练数据点的空间坐标,m;

  • Γ——定压边界;

  • ε——损失函数阈值,无量纲;

  • θ——神经网络参数,无量纲;

  • λii=1,2,3,4)——权重系数,无量纲;

  • σ——高斯白噪声方差,无量纲;

  • μ——流体黏度,mPa•s;

  • Ω——渗流区域。

  • 参考文献

    • [1] 史长林,魏莉,张剑,等.基于机器学习的储层预测方法[J].油气地质与采收率,2022,29(1):90-97.SHI Changlin,WEI Li,ZHANG Jian,et al.Reservoir prediction method based on machine learning[J].Petroleum Geology and Recovery Efficiency,2022,29(1):90-97.

    • [2] 秦峰,闫正和,唐圣来,等.基于深度学习的代理模型在实际气藏三维模拟中的应用[J].油气地质与采收率,2022,29(1):152-159.QIN Feng,YAN Zhenghe,TANG Shenglai,et al.Application of agent models based on deep learning in actual three-dimension‐ al gas reservoir simulation[J].Petroleum Geology and Recovery Efficiency,2022,29(1):152-159.

    • [3] 匡立春,刘合,任义丽,等.人工智能在石油勘探开发领域的应用现状与发展趋势[J].石油勘探与开发,2021,48(1):1-11.KUANG Lichun,LIU He,REN Yili,et al.Application and devel‐ opment trend of artificial intelligence in petroleum exploration and development[J].Petroleum Exploration and Development,2021,48(1):1-11.

    • [4] 陈欢庆,唐海洋,吴桐,等.精细油藏描述中的大数据技术及其应用[J].油气地质与采收率,2022,29(1):11-20.CHEN Huanqing,TANG Haiyang,WU Tong,et al.Big data technology and its application in fine reservoir description[J].Pe‐ troleum Geology and Recovery Efficiency,2022,29(1):11-20.

    • [5] 孙金声,刘凡,程荣超,等.机器学习在防漏堵漏中研究进展与展望[J].石油学报,2022,43(1):91-100.SUN Jinsheng,LIU Fan,CHENG Rongchao,et al.Research progress and prospects of machine learning in lost circulation control[J].Acta Petrolei Sinica,2022,43(1):91-100.

    • [6] 王相,杨耀忠,何岩峰,等.基于深度学习的油井工况智能诊断技术研究及应用[J].油气地质与采收率,2022,29(1):181-189.WANG Xiang,YANG Yaozhong,HE Yanfeng,et al.Research and application of intelligent diagnosis technology of oil well working conditions based on deep learning[J].Petroleum Geolo‐ gy and Recovery Efficiency,2022,29(1):181-189.

    • [7] ZOTKIN O,OSOKINA A,SIMONOV M,et al.A novel ap‐ proach to refinment reservoir proxy model using machine-learn‐ ing techniques[C].Baku:SPE Annual Caspian Technical Confer‐ ence,2019.

    • [8] HAN D,KWON S,KIM J,et al.Comprehensive analysis for pro‐ duction prediction of hydraulic fractured shale reservoirs using proxy model based on deep neural network[C].Virtual:SPE An‐ nual Technical Conference and Exhibition,2020.

    • [9] KIM Y D,DURLOFSKY L J.A recurrent neural network-based proxy model for well-control optimization with nonlinear output constraints[J].SPE Journal,2021,26(4):1 837-1 857.

    • [10] 李野,陈松灿.基于物理信息的神经网络:最新进展与展望[J].计算机科学,2022,49(4):254-262.LI Ye,CHEN Songcan.Physics-informed neural networks:re‐ cent advances and prospects[J].Computer Science,2022,49(4):254-262.

    • [11] CHEN Zhao,LIU Yang,SUN Hao.Physics-informed learning of governing equations from scarce data[J].Nature Communica‐ tions,2021,12(1):1-13.

    • [12] 孙岿.基于改进KNN算法的潜山复杂岩性测井识别方法[J].特种油气藏,2022,29(3):18-27.SUN Kui.Logging identification method of complex lithology in buried hill based on the improved KNN algorithm[J].Special Oil & Gas Reservoirs,2022,29(3):18-27.

    • [13] KARNIADAKIS G E,KEVREKIDIS I G,LU Lu,et al.Physicsinformed machine learning[J].Nature Reviews Physics,2021,3(6):422-440.

    • [14] PSICHOGIOS D C,UNGAR L H.A hybrid neural network‐first principles approach to process modeling[J].AIChE Journal,1992,38(10):1 499-1 511.

    • [15] MEADE J A J,FERNANDEZ A A.The numerical solution of linear ordinary differential equations by feedforward neural net‐ works[J].Mathematical and Computer Modelling,1994,19(12):1-25.

    • [16] LAGARIS I E,LIKAS A,FOTIADIS D I.Artificial neural net‐ works for solving ordinary and partial differential equations[J].IEEE Transactions on Neural Networks,1998,9(5):987-1 000.

    • [17] KARPATNE A,ATLURI G,FAGHMOUS J H,et al.Theoryguided data science:a new paradigm for scientific discovery from data[J].IEEE Transactions on Knowledge and Data Engi‐ neering,2017,29(10):2 318-2 331.

    • [18] KARPATNE A,KANNAN R,KUMAR V.Knowledge guided machine learning:Accelerating discovery using scientific knowl‐ edge and data[M].Florida:CRC Press,2022.

    • [19] RAISSI M,PERDIKARIS P,KARNIADAKIS G E.Physics-in‐ formed neural networks:a deep learning framework for solving forward and inverse problems involving nonlinear partial differ‐ ential equations[J].Journal of Computational Physics,2019,378:686-707.

    • [20] WANG Nanzhe,ZHANG Dongxiao,CHANG Haibin,et al.Deep learning of subsurface flow via theory-guided neural network[J].Journal of Hydrology,2020,584:124700.

    • [21] XU Rui,ZHANG Dongxiao,RONG Miao,et al.Weak form theo‐ ry-guided neural network(TgNN-wf)for deep learning of sub‐ surface single-and two-phase flow[J].Journal of Computational Physics,2021,436:110318.

    • [22] 陆至彬,瞿景辉,刘桦,等.基于物理信息神经网络的传热过程物理场代理模型的构建[J].化工学报,2021,72(3):1 496-1 503.LU Zhibin,QU Jinghui,LIU Hua,et al.Surrogate modeling for physical fields of heat transfer processes based on physics-in‐ formed neural network[J].CIESC Journal,2021,72(3):1 496-1 503.

    • [23] CAI Shengze,MAO Zhiping,WANG Zhicheng,et al.Physics-in‐ formed neural networks(PINNs)for fluid mechanics:a review[J].Acta Mechanica Sinica:English Series,2021,37(12):1 727-1 738.

    • [24] BAYDIN A G,PEARLMUTTER B A,RADUL A A,et al.Auto‐ matic differentiation in machine learning:a survey[J].Journal of Machine Learning Research,2018,18(153):1-43.

    • [25] RONG Miao,ZHANG Dongxiao,WANG Nanzhe.A lagrangian dual-based theory-guided deep neural network[J].Complex & In‐ telligent Systems,2022,8:4 849-4 862.

    • [26] ZHANG Dongxiao,LU Zhiming.An efficient,high-order pertur‐ bation approach for flow in random porous media via KarhunenLoève and polynomial expansions[J].Journal of Computational Physics,2004,194(2):773-794.

    • [27] GHANEM R G,SPANOS P D.Stochastic finite elements:A spectral approach[M].Massachusetts:Courier Corporation,2003.

    • [28] BAJPAI M K,SINGH K K,GIAKOS G.Machine vision and augmented intelligence:Heory and applications[M].Singapore:Springer,2021.

    • [29] ZHU Yinhao,ZABARAS N,KOUTSOURELAKIS P S,et al.Physics-constrained deep learning for high-dimensional surro‐ gate modeling and uncertainty quantification without labeled data [J].Journal of Computational Physics,2019,394:56-81.

    • [30] HE Qizhi,BARAJAS-SOLANO D,TARTAKOVSKY G,et al.Physics-informed neural networks for multiphysics data assimi‐ lation with application to subsurface transport[J].Advances in Water Resources,2020,141:103610.

    • [31] TAN Chuanqi,SUN Fuchun,KONG Tao,et al.A survey on deep transfer learning[C].Rhodes:International Conference on Artifi‐ cial Neural Networks,2018.

  • 参考文献

    • [1] 史长林,魏莉,张剑,等.基于机器学习的储层预测方法[J].油气地质与采收率,2022,29(1):90-97.SHI Changlin,WEI Li,ZHANG Jian,et al.Reservoir prediction method based on machine learning[J].Petroleum Geology and Recovery Efficiency,2022,29(1):90-97.

    • [2] 秦峰,闫正和,唐圣来,等.基于深度学习的代理模型在实际气藏三维模拟中的应用[J].油气地质与采收率,2022,29(1):152-159.QIN Feng,YAN Zhenghe,TANG Shenglai,et al.Application of agent models based on deep learning in actual three-dimension‐ al gas reservoir simulation[J].Petroleum Geology and Recovery Efficiency,2022,29(1):152-159.

    • [3] 匡立春,刘合,任义丽,等.人工智能在石油勘探开发领域的应用现状与发展趋势[J].石油勘探与开发,2021,48(1):1-11.KUANG Lichun,LIU He,REN Yili,et al.Application and devel‐ opment trend of artificial intelligence in petroleum exploration and development[J].Petroleum Exploration and Development,2021,48(1):1-11.

    • [4] 陈欢庆,唐海洋,吴桐,等.精细油藏描述中的大数据技术及其应用[J].油气地质与采收率,2022,29(1):11-20.CHEN Huanqing,TANG Haiyang,WU Tong,et al.Big data technology and its application in fine reservoir description[J].Pe‐ troleum Geology and Recovery Efficiency,2022,29(1):11-20.

    • [5] 孙金声,刘凡,程荣超,等.机器学习在防漏堵漏中研究进展与展望[J].石油学报,2022,43(1):91-100.SUN Jinsheng,LIU Fan,CHENG Rongchao,et al.Research progress and prospects of machine learning in lost circulation control[J].Acta Petrolei Sinica,2022,43(1):91-100.

    • [6] 王相,杨耀忠,何岩峰,等.基于深度学习的油井工况智能诊断技术研究及应用[J].油气地质与采收率,2022,29(1):181-189.WANG Xiang,YANG Yaozhong,HE Yanfeng,et al.Research and application of intelligent diagnosis technology of oil well working conditions based on deep learning[J].Petroleum Geolo‐ gy and Recovery Efficiency,2022,29(1):181-189.

    • [7] ZOTKIN O,OSOKINA A,SIMONOV M,et al.A novel ap‐ proach to refinment reservoir proxy model using machine-learn‐ ing techniques[C].Baku:SPE Annual Caspian Technical Confer‐ ence,2019.

    • [8] HAN D,KWON S,KIM J,et al.Comprehensive analysis for pro‐ duction prediction of hydraulic fractured shale reservoirs using proxy model based on deep neural network[C].Virtual:SPE An‐ nual Technical Conference and Exhibition,2020.

    • [9] KIM Y D,DURLOFSKY L J.A recurrent neural network-based proxy model for well-control optimization with nonlinear output constraints[J].SPE Journal,2021,26(4):1 837-1 857.

    • [10] 李野,陈松灿.基于物理信息的神经网络:最新进展与展望[J].计算机科学,2022,49(4):254-262.LI Ye,CHEN Songcan.Physics-informed neural networks:re‐ cent advances and prospects[J].Computer Science,2022,49(4):254-262.

    • [11] CHEN Zhao,LIU Yang,SUN Hao.Physics-informed learning of governing equations from scarce data[J].Nature Communica‐ tions,2021,12(1):1-13.

    • [12] 孙岿.基于改进KNN算法的潜山复杂岩性测井识别方法[J].特种油气藏,2022,29(3):18-27.SUN Kui.Logging identification method of complex lithology in buried hill based on the improved KNN algorithm[J].Special Oil & Gas Reservoirs,2022,29(3):18-27.

    • [13] KARNIADAKIS G E,KEVREKIDIS I G,LU Lu,et al.Physicsinformed machine learning[J].Nature Reviews Physics,2021,3(6):422-440.

    • [14] PSICHOGIOS D C,UNGAR L H.A hybrid neural network‐first principles approach to process modeling[J].AIChE Journal,1992,38(10):1 499-1 511.

    • [15] MEADE J A J,FERNANDEZ A A.The numerical solution of linear ordinary differential equations by feedforward neural net‐ works[J].Mathematical and Computer Modelling,1994,19(12):1-25.

    • [16] LAGARIS I E,LIKAS A,FOTIADIS D I.Artificial neural net‐ works for solving ordinary and partial differential equations[J].IEEE Transactions on Neural Networks,1998,9(5):987-1 000.

    • [17] KARPATNE A,ATLURI G,FAGHMOUS J H,et al.Theoryguided data science:a new paradigm for scientific discovery from data[J].IEEE Transactions on Knowledge and Data Engi‐ neering,2017,29(10):2 318-2 331.

    • [18] KARPATNE A,KANNAN R,KUMAR V.Knowledge guided machine learning:Accelerating discovery using scientific knowl‐ edge and data[M].Florida:CRC Press,2022.

    • [19] RAISSI M,PERDIKARIS P,KARNIADAKIS G E.Physics-in‐ formed neural networks:a deep learning framework for solving forward and inverse problems involving nonlinear partial differ‐ ential equations[J].Journal of Computational Physics,2019,378:686-707.

    • [20] WANG Nanzhe,ZHANG Dongxiao,CHANG Haibin,et al.Deep learning of subsurface flow via theory-guided neural network[J].Journal of Hydrology,2020,584:124700.

    • [21] XU Rui,ZHANG Dongxiao,RONG Miao,et al.Weak form theo‐ ry-guided neural network(TgNN-wf)for deep learning of sub‐ surface single-and two-phase flow[J].Journal of Computational Physics,2021,436:110318.

    • [22] 陆至彬,瞿景辉,刘桦,等.基于物理信息神经网络的传热过程物理场代理模型的构建[J].化工学报,2021,72(3):1 496-1 503.LU Zhibin,QU Jinghui,LIU Hua,et al.Surrogate modeling for physical fields of heat transfer processes based on physics-in‐ formed neural network[J].CIESC Journal,2021,72(3):1 496-1 503.

    • [23] CAI Shengze,MAO Zhiping,WANG Zhicheng,et al.Physics-in‐ formed neural networks(PINNs)for fluid mechanics:a review[J].Acta Mechanica Sinica:English Series,2021,37(12):1 727-1 738.

    • [24] BAYDIN A G,PEARLMUTTER B A,RADUL A A,et al.Auto‐ matic differentiation in machine learning:a survey[J].Journal of Machine Learning Research,2018,18(153):1-43.

    • [25] RONG Miao,ZHANG Dongxiao,WANG Nanzhe.A lagrangian dual-based theory-guided deep neural network[J].Complex & In‐ telligent Systems,2022,8:4 849-4 862.

    • [26] ZHANG Dongxiao,LU Zhiming.An efficient,high-order pertur‐ bation approach for flow in random porous media via KarhunenLoève and polynomial expansions[J].Journal of Computational Physics,2004,194(2):773-794.

    • [27] GHANEM R G,SPANOS P D.Stochastic finite elements:A spectral approach[M].Massachusetts:Courier Corporation,2003.

    • [28] BAJPAI M K,SINGH K K,GIAKOS G.Machine vision and augmented intelligence:Heory and applications[M].Singapore:Springer,2021.

    • [29] ZHU Yinhao,ZABARAS N,KOUTSOURELAKIS P S,et al.Physics-constrained deep learning for high-dimensional surro‐ gate modeling and uncertainty quantification without labeled data [J].Journal of Computational Physics,2019,394:56-81.

    • [30] HE Qizhi,BARAJAS-SOLANO D,TARTAKOVSKY G,et al.Physics-informed neural networks for multiphysics data assimi‐ lation with application to subsurface transport[J].Advances in Water Resources,2020,141:103610.

    • [31] TAN Chuanqi,SUN Fuchun,KONG Tao,et al.A survey on deep transfer learning[C].Rhodes:International Conference on Artifi‐ cial Neural Networks,2018.