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:

秦峰(1972—),女,山东滕州人,高级工程师,硕士,从事油气田开发工程工作。E-mail:QinFeng@cnooc.com.cn。

中图分类号:TE319+.2

文献标识码:A

文章编号:1009-9603(2022)01-0152-08

DOI:10.13673/j.cnki.cn37-1359/te.2022.01.019

参考文献 1
LÜ Zuobin,HUO Chunliang,GE Lizhen,et al.Application of an integrative new technique on modeling and numerical simulation for fractured reservoir based on unstructured grid:a case study of JZS buried hill reservoir[C].Beijing:International Petroleum Technology Conference,2019.
参考文献 2
SIVON A G,MORIDIS G J,BLASINGAME T A,et al.Developing guidelines for selection of appropriate fracture models in the nu⁃ merical simulation of well performance behavior for liquid rich ul⁃ tra-low permeability ULP reservoirs[C].Texas:SPE Hydraulic Fracture Technology Conference & Exhibition,2018.
参考文献 3
周贤文,汤达帧,张春书,等.精细油藏数值模拟研究现状及发展趋势[J].特种油气藏,2008,15(4):1-6,12. ZHOU Xianwen,TANG Dazhen,ZHANG Chunshu,et al.Present situation and growing tendency of detailed reservoir simulation technique[J].Special Oil & Gas Reservoirs,2008,15(4):1-6,12.
参考文献 4
曹国华,陈鹏,李彩霞,等.油藏数值模拟应用及未来发展趋势 [J].内蒙古师范大学学报:自然科学版,2019,48(5):465-470.CAO Guohua,CHEN Peng,LI Caixia,et al.Application and devel⁃ opment trend of reservoir numerical simulation[J].Journal of In⁃ ner Mongolia Normal University:Natural Science Edition,2019,48(5):465-470.
参考文献 5
王宝华,吴淑红,韩大匡,等.大规模油藏数值模拟的块压缩存储及求解[J].石油勘探与开发,2013,40(4):462-467. WANG Baohua,WU Shuhong,HAN Dakuang,et al.Block com⁃ pressed storage and computation in large-scale reservoir simula⁃ tion[J].Petroleum Exploration and Development,2013,40(4):462-467.
参考文献 6
AZLAN A A A,WAN Muda,MUBARAKI W M,et al.New ap⁃ proach of synergizing advanced well test deconvolution,rate tran⁃ sient analysis and dynamic modeling in evaluating reservoir com⁃ partmentalization uncertainty at k field in sarawak basin;A case study[C].Manama:SPE Middle East Oil and Gas Show and Con⁃ ference,2019.
参考文献 7
ANYANWU C,AFEKHUME D,ONOBRAKPEYA C,et al.Inte⁃ grated reservoir modeling and uncertainty assessment of a reser⁃ voir complex in the Niger Delta[C].Lagos:SPE Nigeria Annual In⁃ ternational Conference and Exhibition,2015.
参考文献 8
刘伟,赵辉,雷占祥,等.基于单井敏感性局域化EnKF的油藏辅助历史拟合方法[J].石油学报,2019,40(6):716-725. LIU Wei,ZHAO Hui,LEI Zhanxiang,et al.Reservoir assisted his⁃ tory matching method using a local ensemble Kalman filter based on single-well sensitivity region[J].Acta Petrolei Sinica,2019,40(6):716-725.
参考文献 9
张凯,马小鹏,王增飞,等.一种强非均质性油藏自动历史拟合混合求解方法[J].中国石油大学学报:自然科学版,2018,42(5):89-97. ZHANG Kai,MA Xiaopeng,WANG Zengfei,et al.A hybrid meth⁃ od for solving automated production history match of reservoirs with strong heterogeneity[J].Journal of China University of Petro⁃ leum:Edition of Natural Science,2018,42(5):89-97.
参考文献 10
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):1-21.
参考文献 11
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].Denver:SPE An⁃ nual Technical Conference and Exhibition,2020.
参考文献 12
BRUYELLE J,GUÉRILLOT D.Proxy model based on artificial in⁃ telligence technique for history matching-application to Brugge Field[C].Dubai:SPE Gas & Oil Technology Showcase and Con⁃ ference,2019.
参考文献 13
ZOTKIN O,OSOKINA A,SIMONOV M,et al.A novel approach to refinment reservoir proxy model using machine-learning tech⁃ niques[C].Baku:SPE Annual Caspian Technical Conference,2019.
参考文献 14
SHAHKARAMI Alireza,MOHAGHEGH Shahab.智能代理在油藏建模中的应用[J].石油勘探与开发,2020,47(2):372-382.SHAHKARAMI Alireza,MOHAGHEGH Shahab.Applications of smart proxies for subsurface modeling[J].Petroleum Exploration and Development,2020,47(2):372-382.
参考文献 15
王链,姚约东,王孔杰,等.基于代理辅助野草猴群算法的井位优快决策方法[J/OL].大庆石油地质与开发,2021:1-8[2021-11-10].https://kns.cnki.net/kcms/detai/23.1286.TE.20211014.1108.001.html.WANG Lian,YAO Yuedong,WANG Kongjie,et al.Optimal and fast decision making method for well location based on surrogateassisted weed monkey algorithm(WMA)[J/OL].Petroleum Geolo⁃ gy & Oilfield Development in Daqing,2021:1-8[2021-11-10].https://kns.cnki.net/kcms/detail/23.1286.TE.20211014.1108.001.html.
参考文献 16
李伟,唐放,侯博恒,等.基于神经网络的南海东部砂岩油藏采收率预测方法[J].油气藏评价与开发,2021,11(5):730-735,743.LI Wei,TANG Fang,HOU Boheng,et al.A method for oil recov⁃ ery prediction of sandstone reservoirs in the eastern South China Sea based on neural network[J].Reservoir Evaluation and Devel⁃ opment,2021,11(5):730-735,743.
参考文献 17
谷建伟,周梅,李志涛,等.基于数据挖掘的长短期记忆网络模型油井产量预测方法[J].特种油气藏,2019,26(2):77-81,131.GU Jianwei,ZHOU Mei,LI Zhitao,et al.Oil well production fore⁃ cast with long-short term memory network model based on data mining[J].Special Oil & Gas Reservoirs,2019,26(2):77-81,131.
参考文献 18
吴君达,李治平,孙妍,等.基于神经网络的剩余油分布预测及注采参数优化[J].油气地质与采收率,2020,27(4):85-93.WU Junda,LI Zhiping,SUN Yan,et al.Neural network-based pre⁃ diction of remaining oil distribution and optimization of injectionproduction parameters[J].Petroleum Geology and Recovery Effi⁃ ciency,2020,27(4):85-93.
参考文献 19
葛兰,蒲谢洋.页岩气田智能化生产辅助决策系统应用效果评价[J].油气藏评价与开发,2021,11(4):621-627.GE Lan,PU Xieyang.Evaluation of application effect of an intelli⁃ gent production assistant decision system for shale gas field[J]. Reservoir Evaluation and Development,2021,11(4):621-627.
参考文献 20
刘巍,刘威,谷建伟,等.利用卡尔曼滤波和人工神经网络相结合的油藏井间连通性研究[J].油气地质与采收率,2020,27(2):118-124.LIU Wei,LIU Wei,GU Jianwei,et al.Research on interwell con⁃ nectivity of oil reservoirs based on Kalman filter and artificial neu⁃ ral network[J].Petroleum Geology and Recovery Efficiency,2020,27(2):118-124.
参考文献 21
CANCHUMUNI S W,EMERICK A A,PACHECO M A C.History matching geological facies models based on ensemble smoother and deep generative models[J].Journal of Petroleum Science and Engineering,2019,177:941-958.
参考文献 22
KINGMA D P,WELLING M.Auto-encoding variational bayes[C].Banff:International Conference on Learning Representations,2013.
参考文献 23
WANG Z,XIAO D,FANG F,et al.Model identification of reduced order fluid dynamics systems using deep learning[J].Numerical Methods in Fluids,2018,86(4):255-268.
参考文献 24
GERS F A,SCHMIDHUBER J,CUMMINS F.Learning to forget:Continual prediction with LSTM[J].Neural Computation,2000,12(10):2 451-2 471.
参考文献 25
GONZALEZ F J,BALAJEWICZ M.Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems[D].Urbana:University of Illinois at Urbana-Cham⁃ paign,2018:1-28.
参考文献 26
TANG M,LIU Y,DURLOFSKY L J.A deep-learning-based sur⁃ rogate model for data assimilation in dynamic subsurface flow problems[J].Journal of Computational Physics,2020,413:109456.
参考文献 27
TEMIRCHEV P,SIMONOV M,KOSTOEV R,et al.Deep neural networks predicting oil movement in a development unit[J].Jour⁃ nal of Petroleum Science and Engineering,2020,184:106513.
参考文献 28
MO S,ZHU Y,ZABARAS N,et al.Deep convolutional encoderde⁃ coder networks for uncertainty quantification of dynamic multi⁃ phase flow in heterogeneous media[J].Water Resource,2019,55(1):703-728.
参考文献 29
RAISSI M,PERDIKARIS P,KARNIADAKIS G.Physics-in⁃ formed neural networks:a deep learning framework for solving for⁃ ward and inverse problems involving nonlinear partial differential equations[J].Journal of Computational Physics,2019,378:686-707.
参考文献 30
WONG Z Y,HORNE R,VOSKOV D.A geothermal reservoir simu⁃ lator in AD-GPRS[D].Stanford:Stanford University,2015:1-72.
参考文献 31
WATTER M,SPRINGENBERG J,BOEDECKER J,et al.Embed to control:A locally linear latent dynamics model for control from raw images[C].Montreal:International Conference on Neural In⁃ formation Processing Systems,2015.
参考文献 32
DONG Guozhu,LIU Huan.Feature engineering for machine learn⁃ ing and data analytics[M].New York:CRC Press,2018.
目录contents

    摘要

    应用基于深度学习的代理模型进行油气藏模拟是油气藏仿真研究的一个新方向。针对高精度全阶油气藏模拟速度慢的问题,采用一种基于深度学习的嵌入式控制框架(E2C,Embed to Control)模型,通过“编码器+线性转化模型+解码器”的架构构建深度学习网络,将原始时刻的压力场、饱和度场数据与井控约束条件相结合来演化出新时刻的场数据。以南海东部番禺35-1气田为例,测试E2C模型与传统数值模拟器模拟结果的差别。测试结果显示 E2C模型误差较小,其中饱和度场的相对误差小于5%,压力场的平均相对误差为8%;在相同的CPU条件下,E2C模型运行100次算例时间为16 s,比传统数值模拟器(运行时间为6000 s)快375倍。实际应用结果表明E2C模型在保证模拟精度的条件下可以大幅度提升模拟速度。

    Abstract

    The application of agent models based on deep learning is a new direction of oil and gas reservoir simulation. Given the huge time cost of high-precision full-order oil and gas reservoir simulation,this paper adopts an Embed to Con- trol(E2C)model to construct a deep learning network through the architecture of“encoder+linear converter+decoder”. Data of the pressure field and saturation field at the original moment are integrated with well control constraints to generate the field data at a new moment. The PY35-1 Gas Field in the east of the South China Sea is discussed as an example to test the differences between the simulation results of the E2C model and those of the traditional numerical simulator. The test re- sults show that the E2C model has smaller errors,with a relative error in the saturation field of less than 5% and an average relative error in the pressure field of 8%. Under the same CPU,the E2C model takes 16 s to run 100 cases,which is 375 times faster than the traditional numerical simulator(the time cost is 6000 s). In conclusion,the E2C model can greatly re- duce the time cost under the condition of ensuring simulation precision.

  • 数值模拟技术目前已是大型油气藏开发决策常用的辅助分析手段之一[1-5],然而由于控制方程的非线性性质和地质描述的多尺度特性,油气藏模型包含了复杂的渗流机理和巨量网格,当模型应用于优化、不确定性量化和数据同化时可能需要运行数千次模拟计算[6-9],这样模型的计算成本极高。应用基于深度学习的代理模型进行油气藏模拟是油气藏仿真研究的一个新方向[10-20],代理模型比高保真油气藏模型的模拟运算速度大幅提升,因此计算成本更加低廉,在实际应用中可替代高保真油气藏模型,实现大型复杂油气藏开发方案的快速迭代。

  • 目前国外对构建地下流体流动模拟的代理模型已进行了广泛的研究,其中基于图像处理网络架构的代理模型发展较快,这些构建代理模型的方法已被应用于地质参数化、不确定性量化和代理/降阶建模等方面。CANCHUMUNI 等将基于图像处理的网络架构应用于地质建模,使用变分自动编码器 (VAE,Variational Auto Encoder)将随机化的低维隐藏变量生成新的地质实现[21-22]。在流体流动模拟方面,WANG 等使用长短期记忆(LSTM,Long-Short Term Memory)循环神经网络模型来逼近由本征正交分解(POD,Proper-Orthogonal-Decomposition)方法所建立的流动动力学模型[23-24]。GONZALEZ 等用自编码器(AE,Auto-Encoder)架构代替了 POD步骤进行低维表示[25]。TANG等引入一种基于深度学习的代理模型 R-U-Net[26],该模型同时包含了卷积和递归神经网络,在井控条件不变的情况下用于预测不同地质模型的流量响应,在油气藏辅助历史拟合过程中应用较好。TEMIRCHEV等构造了一个与降阶模型相似的模型,用VAE来表示低维的储层状态[27]。模型结合线性回归、LSTM 或门控循环单元 (GRU)进行了实际油藏动态模拟测试,但结果显示在不考虑井控条件的情况下,模型结果的相对误差在某些验证方案中相对较大。上述模型只适用于油气藏内边界条件不变的情况,即没有考虑井控条件的变化,而实际油气藏开采过程中井控条件一般都会随时间发生变化,因此上述模型对油气藏实际流动模拟问题适应性较差。研究前期许多模型都是端到端的纯数据驱动方法,没有考虑流体流动的基本控制方程。为了提高模型性能,最新进展中已有许多方法将物理约束纳入到深度神经网络模型构建过程中。MO等将VAE模型扩展到处理多相流问题,并通过引入额外的物理约束进一步提高性能[28]。RAISSI 等提出了使用密集连接前馈神经网络的物理知识深度学习框架(被称为物理信息神经网络或PINN)[29]。在PINN中,将与控制偏微分方程相关联的残差函数引入神经网络的损失函数中,并利用控制方程的残差构建了用于不确定性量化的代理模型。

  • 笔者采用嵌入式控制框架(E2C,Embed to Con⁃ trol)模型构建流动控制方程的降阶模型,并对模型架构进行了调整优化,将编码的卷积输出结果加入到解码的卷积输入中,这样解码器可以获得更多的油气藏属性空间信息;同时将油气藏井控条件作为物理约束加入到模型中,因此该模型适用于油气藏边界条件可变的情况,更加符合油气藏实际流动模拟状态。

  • 1 油气藏流动控制方程

  • 达西定律描述了在孔隙介质中流体的宏观流动方式,因此,采用基于达西定律的常规油气水三相流动模型。从各组分的质量守恒出发,结合达西定律并忽略毛管压力的影响,导出非混相油气水流动的控制方程为:

  • ϕSjρj-λjρjKp+w ρjqjw=0
    (1)
  • 其中:

  • λj=Krjμj
    (2)
  • (1)式中,ϕK表示地质特征,Sjp为求解变量。一般采用有限体积方法对(1)式求解,并针对每个网格块计算它们的解。在本次研究中,使用特雷西能源科技有限公司的油气藏数值模拟器 Cloud[30] 来构建所有训练数据。

  • 2 E2C模型架构

  • WATTER 等提出了 E2C 模型,在机器人规划系统的背景下将直接感觉数据(图像)和时变控制作为输入来预测系统状态的演化[31]。E2C 模型将 VAE模型和生成模型相结合,VAE模型是将系统状态投影到低维子空间的推理模型,生成模型是对预测结果进行重构生成的模型和线性转化模型。后者基于时变控制输入近似低维状态的演化。借鉴 E2C模型的设计思路,开发一个包含物理约束(井控条件)的嵌入式控制框架的降阶模型,主要包括3个处理部分:将系统变量从高维空间投射到低维子空间(这里称为隐藏空间)的编码器、低维近似系统动力学的线性转化模型、将解映射回高维(全阶)空间的解码器或生成模型。该模型与 WATTER 所设计模型的区别在于对模型架构进行了调整优化,将编码的卷积输出结果加入到解码的卷积输入中,这样解码器可以获得更多的油气藏属性空间信息;同时将油气藏井控条件作为物理约束加入到模型中,更适用于油气藏实际流动模拟。

  • 2.1 编码器

  • 编码器提供了全序状态变量的低维表示。在训练时编码器可简单地表示为:

  • zt=Qϕencxt
    (3)
  • E2C模型中编码器结构如图1所示。这里使用的编码器网络由3个卷积块、3个残差块和1个全连接层堆积组成。在训练过程中,压力和饱和度被输入到编码器网络,隐藏状态的特征矩阵zt被生成。

  • 图1 编码器结构

  • Fig.1 Structure of encoder

  • 2.2 线性转化模型

  • 线性转化模型根据给定的控制将隐藏变量从一个时间步发展到下一个时间步,用于描述模型由于外界作用而发生变化的过程,文中主要用于描述模型由于井控条件的变化导致压力场、饱和度场的变化过程。线性转化模型(图2)由 3 个转化块和 2 个全连接层组成。线性转化模型遵循一个 Densebatchnorm-ReLU架构,这被认为是全连接网络的标准处理步骤。该模型的输入包括:隐藏状态的特征矩阵 zt、井控条件 ut + 1 和预测的时间间隔 Δt。模型输出为下一时刻隐藏状态的特征矩阵z^t+1

  • 图2 线性转化模型结构

  • Fig.2 Structure of linear converter

  • 2.3 解码器

  • 解码器的结构类似于编码器,其结构如图3 所示,具体可以表示为:

  • x^t+1=pθdeczt+1
    (4)
  • 解码器结构如图3所示。将线性转化模型的输出zt + 1与编码器的卷积输出Et结合作为解码器的输入,预测高维状态的输出,即油气藏的压力场及饱和度场。解码器结构与编码器结构类似,只是顺序是颠倒的,由 1个全连接层、3个残差块、4个逆卷积块和1个卷积层堆叠组成。密集层将一个低维的潜在向量转换为特征映射的堆积,这些特征图先通过残差块和逆卷积块转化得到多个特征图,再通过卷积层将多个特征映射转换为压力和饱和度场作为最终输出。

  • 图3 解码器结构

  • Fig.3 Structure of decoder

  • 2.4 模型实现及训练细节

  • E2C模型的输入数据为数值模拟算例中包含的压力场、饱和度场及时序类型的井控条件数据。对输入数据的预处理主要是对数据进行归一化处理,可以提高学习过程和输出质量[32]。饱和度S ∈ [ 0,1 ] 不需要归一化,井控条件的压力数据需要归一化。标准压力 p0(包括网格压力和井底流压)的表达式为:

  • p0=p-pminpmax-pmin
    (5)
  • E2C模型的输出数据为压力场、饱和度场,为了便于实际应用,在计算出压力场、饱和度场结果的基础上,利用皮斯曼方程计算单井的产量。关于多少个训练样本可以使模型达到最佳训练效果,即使模型达到精准度与训练时间的平衡还没有进一步的深入研究,文中使用了数值模拟结果的 500 个算例作为训练数据,模型训练后达到了预期效果。

  • 3 基于E2C模型的气藏三维模拟

  • 3.1 区块概况及算例

  • 对实际气藏进行三维模拟以测试 E2C 模型效果。番禺35-1气田位于珠江口盆地流花区块,区域上位于白云凹陷北坡,北部紧靠番禺低隆起。气田所在海域水深约为 195~202 m,整体为受岩性控制的层状边水砂岩气藏。

  • 利用特雷西能源科技有限公司的油气藏数值模拟器 Cloud 建立气水模拟的算例,作为 E2C 模型的训练集。具体算例的网格模型(图4)包含 3口生产井,其中 2 口直井 A2 和 A3,一口水平井 A1H,实际生产为衰竭式开发。网格块的尺寸为 50 m×50 m×0.5 m,总网格数为 154×81×124 共 1 546 776。模型渗透率的算术平均值为 25.4 mD,孔隙度平均值为0.147。相对渗透率的表达式为:

  • 图4 番禺35-1气田算例的网格模型

  • Fig.4 Grid model of PY35-1 Gas Field case

  • KroSw=Kro01-1-Sw-Sor 1-Swr-Sora
    (6)
  • KrwSw=Krw01-Sw-Swr1-Swr-Sorb
    (7)
  • (6)式和(7)式中,Kro0为1.0,Krw0为0.7,S or为0.3, S wr为0.1,a为3.6,b为1.5。其他流体参数包括:ρg为 0.738 kg/m3ρw为1 000 kg/m3μg为0.02 mPa·s,μw为 0.91 mPa·s,毛管压力的影响可以忽略。

  • 模型采用了来自于研究区 1 400 d 的实际生产数据。模拟时间间隔对于训练和测试运行都是相同的。生产井采用变井底流压控制,井底流压数据按实际输入。

  • 3.2 应用效果

  • 通过对比 E2C模型与 Cloud模拟器模拟结果的差异及模拟用时,评价 E2C 模型的实际应用效果。在相同的生产时间节点,对比 E2C 模型与 Cloud 模拟器模拟的产量曲线、饱和度场、压力场,并定量分析其差异。

  • 产量曲线对比  分析实际工区3口生产井的井底流压曲线(图5),其井底流压为 0~32 MPa,Cloud 模拟器和 E2C 模型在相同的井底流压曲线即井控条件的约束下进行运算,E2C 模型在训练阶段的日产气量曲线整体与Cloud模拟器差别很小(图5),日产气量曲线的整体误差绝对值为 0~0.5 Sm3 /d,日产水量曲线的整体误差绝对值为 0~1.2 Sm3 /d,E2C 模型训练效果理想。预测阶段 Cloud模拟器和 E2C模型都在井底流压为25 MPa条件下定压生产,E2C模型在预测阶段误差稍微增加,日产气量曲线的整体误差绝对值为0~0.95 Sm3 /d,日产水量曲线的整体误差绝对值为 0~3.6 Sm3 /d。由此可见,E2C 模型模拟的生产曲线较 Cloud 模拟器差别不大,曲线形态与 Cloud模拟器匹配较好。

  • 饱和度场对比  以番禺 35-1 气田生产 1 400 d 的饱和度场(图6)为例说明。由于 Cloud 模拟器模拟结果(图6a)和 E2C 模型模拟结果(图6b)从视觉上看是相似的,因此对两者之间的差异进行了定量描述(图6c),可以看出偏差基本为-0.05~0.01,平均相对误差小于5%,饱和度场模拟效果较好。

  • 压力场对比  以番禺 35-1气田生产 1 400 d的压力场(图7)为例说明。由于 Cloud 模拟器模拟结果(图7a)和 E2C 模型模拟结果(图7b)从视觉上看是相似的,对两者之间的差异进行了定量描述(图7c),E2C 模型局部位置的压力场与 Cloud 模拟器模拟的压力场存在差异,且 Cloud 模拟器计算的压力场更加光滑连续,分析原因主要是基于物理渗流过程的 Cloud 模拟器可以再现渗流过程,基于数据驱动的 E2C 模型对于某些渗流过程的细节特征难以提取与保留,这也是未来 E2C模型的改进方向。与 Cloud模拟器模拟结果相比,E2C模型模拟的压力场平均相对误差为8%,效果较好。

  • 运行时间对比  在 Intel Xeon ES-2670 双 CPU (24 核)的计算节点上,Cloud 模拟 100 个测试算例,每个耗时 60 s,运行用时共计 6 000 s。在分配了 8 GB内存的Tesla V100 GPU节点上,使用E2C模型可以在大约 1.25 s内评估完 100 个测试用例。这种条件下进行比较可以得到6 000/1.25=4 800倍的加速,但是这个比较没有考虑 GPU 和 CPU 的差别。在相同的 CPU(Intel Xeon ES-2670)条件下,E2C 模型的 100 次测试算例运行时间大约需要 16 s,这相当于 6 000/16=375倍的加速,虽然比GPU的加速要小,但仍然非常可观。

  • 4 结论

  • 采用一种基于深度学习的嵌入式控制框架模型模拟油气藏地下三维属性场的变化,探索了深度学习模型在油气藏模拟方面的应用。利用 E2C 模型对实际气藏进行三维模拟,评价了深度学习模型在真实三维气藏条件下的性能和适用性。算例结果显示:在准确度上,E2C模型可以很好地模拟气藏在不同的井控条件下压力场、饱和度场的变化,与 Cloud 模拟器结果对比误差较小;在时间性能上, E2C模型模拟耗时是传统方法的1/375。综上所述, E2C模型在优化和历史拟合方面比传统数值模拟方法更具潜力。

  • 研究中实际应用的 E2C 模型预测耗时不包括前期数值模拟模型准备、批量计算、代理模型训练等耗时,如果综合考虑全部耗时,E2C模型的时间优势仍然比较明显:第一,前期模型的准备、批量计算、代理模型训练等工作可以在机器空闲时进行,即线下进行;训练好的模型由于具备极快的求解速度,可以在一些智能油藏的场景,比如生产方案实时优化、实时决策等,无延迟完成,即线上完成。第二,通过相对有限的训练算例得到的代理模型,具备在更广泛参数空间采样后快速求解的能力,因此在运用智能优化算法(如粒子群算法)时,能够避免成百上千优化算例的直接求解,这部分时间的节省也非常显著。模型泛化性方面,本次实际应用的 E2C 模型主要针对具体的一个气藏,目前的研究进展表明,代理模型的实际应用在现阶段还不具备泛化性,需要针对具体的地质模型和井位部署、在一定边界的生产制度下完成替代训练。目前正在探索具有泛化能力的网络模型路线,但需要建立极其海量的训练集(不同区块、不同地质模型的多个实现,以及不同生产制度下的多个算例),经过超大型的训练方能完成。

  • 图5 番禺35-1气田3口生产井的日产量曲线对比

  • Fig.5 Comparison of daily production curves of three producing wells in PY35-1 Gas Field

  • 图6 番禺35-1气田饱和度场对比及误差分析

  • Fig.6 Comparison of saturation fields of PY35-1 Gas Field and error analysis

  • 图7 番禺35-1气田压力场对比及误差分析

  • Fig.7 Comparison of pressure fields of PY35-1 Gas Field and error analysis

  • 符号解释

  • a ——油相相对渗透率指数;

  • b ——水相相对渗透率指数;

  • AtBt ——中间生成矩阵;

  • Et ——编码器的卷积输出;

  • 下标 j ——流体的相,j为 o,w,g 分别表示油相、水相、气相;

  • K ——岩石渗透率,D;

  • K rj ——j相的相对渗透率;

  • K ro——标准化油相相对渗透率;

  • K ro 0 ——标准化前油相相对渗透率;

  • K rw——标准化水相相对渗透率;

  • K rw 0 ——标准化前水相相对渗透率;

  • p ——网格块压力或者生产井井底流压,MPa;

  • p0 ——标准压力,无量纲;

  • p min——网格压力或者生产井井底流压的下边界值, MPa;

  • p max——可观测到的最大压力的 1.1 倍(1.1 倍用于确保所有数据在小于1的范围之内),MPa;

  • pdec θ ——解码器;

  • qw j ——井流量,m3

  • Qenc ϕ ——编码器;

  • R2nb ——状态矩阵(压力和饱和度)的维度,是时间步t上的全阶状态变量;

  • Rlz ——隐藏层维度;

  • Rnw——井的维度;

  • Sj——流体相饱和度;

  • S or——残余油饱和度;

  • S w——标准化含水饱和度;

  • S wr——束缚水饱和度;

  • t ——时间步;

  • Δt ——预测的时间间隔,d;

  • ut + 1——t +1时刻的井控条件,ut + 1Rnw

  • w ——井别,w为i和o分别表示注入井、采油井;

  • xt ——输入的特征矩阵,xtR2nb

  • x^t+1——t+1时刻的状态值;

  • zt ——提取的隐藏状态的特征矩阵,ztRlz

  • zt + 1——经过转化器输出的新特征;

  • z^t+1——下一时刻隐藏状态的特征矩阵,z^t+1Rlz

  • μg——干气黏度,mPa•s;

  • μj ——j相的黏度,mPa•s;

  • μw——水相黏度,mPa•s;

  • ρg——干气密度,kg/m3

  • ρj ——相密度,kg/m3

  • ρw——水相密度,kg/m3

  • ϕ——岩石孔隙度。

  • 参考文献

    • [1] LÜ Zuobin,HUO Chunliang,GE Lizhen,et al.Application of an integrative new technique on modeling and numerical simulation for fractured reservoir based on unstructured grid:a case study of JZS buried hill reservoir[C].Beijing:International Petroleum Technology Conference,2019.

    • [2] SIVON A G,MORIDIS G J,BLASINGAME T A,et al.Developing guidelines for selection of appropriate fracture models in the nu⁃ merical simulation of well performance behavior for liquid rich ul⁃ tra-low permeability ULP reservoirs[C].Texas:SPE Hydraulic Fracture Technology Conference & Exhibition,2018.

    • [3] 周贤文,汤达帧,张春书,等.精细油藏数值模拟研究现状及发展趋势[J].特种油气藏,2008,15(4):1-6,12. ZHOU Xianwen,TANG Dazhen,ZHANG Chunshu,et al.Present situation and growing tendency of detailed reservoir simulation technique[J].Special Oil & Gas Reservoirs,2008,15(4):1-6,12.

    • [4] 曹国华,陈鹏,李彩霞,等.油藏数值模拟应用及未来发展趋势 [J].内蒙古师范大学学报:自然科学版,2019,48(5):465-470.CAO Guohua,CHEN Peng,LI Caixia,et al.Application and devel⁃ opment trend of reservoir numerical simulation[J].Journal of In⁃ ner Mongolia Normal University:Natural Science Edition,2019,48(5):465-470.

    • [5] 王宝华,吴淑红,韩大匡,等.大规模油藏数值模拟的块压缩存储及求解[J].石油勘探与开发,2013,40(4):462-467. WANG Baohua,WU Shuhong,HAN Dakuang,et al.Block com⁃ pressed storage and computation in large-scale reservoir simula⁃ tion[J].Petroleum Exploration and Development,2013,40(4):462-467.

    • [6] AZLAN A A A,WAN Muda,MUBARAKI W M,et al.New ap⁃ proach of synergizing advanced well test deconvolution,rate tran⁃ sient analysis and dynamic modeling in evaluating reservoir com⁃ partmentalization uncertainty at k field in sarawak basin;A case study[C].Manama:SPE Middle East Oil and Gas Show and Con⁃ ference,2019.

    • [7] ANYANWU C,AFEKHUME D,ONOBRAKPEYA C,et al.Inte⁃ grated reservoir modeling and uncertainty assessment of a reser⁃ voir complex in the Niger Delta[C].Lagos:SPE Nigeria Annual In⁃ ternational Conference and Exhibition,2015.

    • [8] 刘伟,赵辉,雷占祥,等.基于单井敏感性局域化EnKF的油藏辅助历史拟合方法[J].石油学报,2019,40(6):716-725. LIU Wei,ZHAO Hui,LEI Zhanxiang,et al.Reservoir assisted his⁃ tory matching method using a local ensemble Kalman filter based on single-well sensitivity region[J].Acta Petrolei Sinica,2019,40(6):716-725.

    • [9] 张凯,马小鹏,王增飞,等.一种强非均质性油藏自动历史拟合混合求解方法[J].中国石油大学学报:自然科学版,2018,42(5):89-97. ZHANG Kai,MA Xiaopeng,WANG Zengfei,et al.A hybrid meth⁃ od for solving automated production history match of reservoirs with strong heterogeneity[J].Journal of China University of Petro⁃ leum:Edition of Natural Science,2018,42(5):89-97.

    • [10] 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):1-21.

    • [11] 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].Denver:SPE An⁃ nual Technical Conference and Exhibition,2020.

    • [12] BRUYELLE J,GUÉRILLOT D.Proxy model based on artificial in⁃ telligence technique for history matching-application to Brugge Field[C].Dubai:SPE Gas & Oil Technology Showcase and Con⁃ ference,2019.

    • [13] ZOTKIN O,OSOKINA A,SIMONOV M,et al.A novel approach to refinment reservoir proxy model using machine-learning tech⁃ niques[C].Baku:SPE Annual Caspian Technical Conference,2019.

    • [14] SHAHKARAMI Alireza,MOHAGHEGH Shahab.智能代理在油藏建模中的应用[J].石油勘探与开发,2020,47(2):372-382.SHAHKARAMI Alireza,MOHAGHEGH Shahab.Applications of smart proxies for subsurface modeling[J].Petroleum Exploration and Development,2020,47(2):372-382.

    • [15] 王链,姚约东,王孔杰,等.基于代理辅助野草猴群算法的井位优快决策方法[J/OL].大庆石油地质与开发,2021:1-8[2021-11-10].https://kns.cnki.net/kcms/detai/23.1286.TE.20211014.1108.001.html.WANG Lian,YAO Yuedong,WANG Kongjie,et al.Optimal and fast decision making method for well location based on surrogateassisted weed monkey algorithm(WMA)[J/OL].Petroleum Geolo⁃ gy & Oilfield Development in Daqing,2021:1-8[2021-11-10].https://kns.cnki.net/kcms/detail/23.1286.TE.20211014.1108.001.html.

    • [16] 李伟,唐放,侯博恒,等.基于神经网络的南海东部砂岩油藏采收率预测方法[J].油气藏评价与开发,2021,11(5):730-735,743.LI Wei,TANG Fang,HOU Boheng,et al.A method for oil recov⁃ ery prediction of sandstone reservoirs in the eastern South China Sea based on neural network[J].Reservoir Evaluation and Devel⁃ opment,2021,11(5):730-735,743.

    • [17] 谷建伟,周梅,李志涛,等.基于数据挖掘的长短期记忆网络模型油井产量预测方法[J].特种油气藏,2019,26(2):77-81,131.GU Jianwei,ZHOU Mei,LI Zhitao,et al.Oil well production fore⁃ cast with long-short term memory network model based on data mining[J].Special Oil & Gas Reservoirs,2019,26(2):77-81,131.

    • [18] 吴君达,李治平,孙妍,等.基于神经网络的剩余油分布预测及注采参数优化[J].油气地质与采收率,2020,27(4):85-93.WU Junda,LI Zhiping,SUN Yan,et al.Neural network-based pre⁃ diction of remaining oil distribution and optimization of injectionproduction parameters[J].Petroleum Geology and Recovery Effi⁃ ciency,2020,27(4):85-93.

    • [19] 葛兰,蒲谢洋.页岩气田智能化生产辅助决策系统应用效果评价[J].油气藏评价与开发,2021,11(4):621-627.GE Lan,PU Xieyang.Evaluation of application effect of an intelli⁃ gent production assistant decision system for shale gas field[J]. Reservoir Evaluation and Development,2021,11(4):621-627.

    • [20] 刘巍,刘威,谷建伟,等.利用卡尔曼滤波和人工神经网络相结合的油藏井间连通性研究[J].油气地质与采收率,2020,27(2):118-124.LIU Wei,LIU Wei,GU Jianwei,et al.Research on interwell con⁃ nectivity of oil reservoirs based on Kalman filter and artificial neu⁃ ral network[J].Petroleum Geology and Recovery Efficiency,2020,27(2):118-124.

    • [21] CANCHUMUNI S W,EMERICK A A,PACHECO M A C.History matching geological facies models based on ensemble smoother and deep generative models[J].Journal of Petroleum Science and Engineering,2019,177:941-958.

    • [22] KINGMA D P,WELLING M.Auto-encoding variational bayes[C].Banff:International Conference on Learning Representations,2013.

    • [23] WANG Z,XIAO D,FANG F,et al.Model identification of reduced order fluid dynamics systems using deep learning[J].Numerical Methods in Fluids,2018,86(4):255-268.

    • [24] GERS F A,SCHMIDHUBER J,CUMMINS F.Learning to forget:Continual prediction with LSTM[J].Neural Computation,2000,12(10):2 451-2 471.

    • [25] GONZALEZ F J,BALAJEWICZ M.Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems[D].Urbana:University of Illinois at Urbana-Cham⁃ paign,2018:1-28.

    • [26] TANG M,LIU Y,DURLOFSKY L J.A deep-learning-based sur⁃ rogate model for data assimilation in dynamic subsurface flow problems[J].Journal of Computational Physics,2020,413:109456.

    • [27] TEMIRCHEV P,SIMONOV M,KOSTOEV R,et al.Deep neural networks predicting oil movement in a development unit[J].Jour⁃ nal of Petroleum Science and Engineering,2020,184:106513.

    • [28] MO S,ZHU Y,ZABARAS N,et al.Deep convolutional encoderde⁃ coder networks for uncertainty quantification of dynamic multi⁃ phase flow in heterogeneous media[J].Water Resource,2019,55(1):703-728.

    • [29] RAISSI M,PERDIKARIS P,KARNIADAKIS G.Physics-in⁃ formed neural networks:a deep learning framework for solving for⁃ ward and inverse problems involving nonlinear partial differential equations[J].Journal of Computational Physics,2019,378:686-707.

    • [30] WONG Z Y,HORNE R,VOSKOV D.A geothermal reservoir simu⁃ lator in AD-GPRS[D].Stanford:Stanford University,2015:1-72.

    • [31] WATTER M,SPRINGENBERG J,BOEDECKER J,et al.Embed to control:A locally linear latent dynamics model for control from raw images[C].Montreal:International Conference on Neural In⁃ formation Processing Systems,2015.

    • [32] DONG Guozhu,LIU Huan.Feature engineering for machine learn⁃ ing and data analytics[M].New York:CRC Press,2018.

  • 参考文献

    • [1] LÜ Zuobin,HUO Chunliang,GE Lizhen,et al.Application of an integrative new technique on modeling and numerical simulation for fractured reservoir based on unstructured grid:a case study of JZS buried hill reservoir[C].Beijing:International Petroleum Technology Conference,2019.

    • [2] SIVON A G,MORIDIS G J,BLASINGAME T A,et al.Developing guidelines for selection of appropriate fracture models in the nu⁃ merical simulation of well performance behavior for liquid rich ul⁃ tra-low permeability ULP reservoirs[C].Texas:SPE Hydraulic Fracture Technology Conference & Exhibition,2018.

    • [3] 周贤文,汤达帧,张春书,等.精细油藏数值模拟研究现状及发展趋势[J].特种油气藏,2008,15(4):1-6,12. ZHOU Xianwen,TANG Dazhen,ZHANG Chunshu,et al.Present situation and growing tendency of detailed reservoir simulation technique[J].Special Oil & Gas Reservoirs,2008,15(4):1-6,12.

    • [4] 曹国华,陈鹏,李彩霞,等.油藏数值模拟应用及未来发展趋势 [J].内蒙古师范大学学报:自然科学版,2019,48(5):465-470.CAO Guohua,CHEN Peng,LI Caixia,et al.Application and devel⁃ opment trend of reservoir numerical simulation[J].Journal of In⁃ ner Mongolia Normal University:Natural Science Edition,2019,48(5):465-470.

    • [5] 王宝华,吴淑红,韩大匡,等.大规模油藏数值模拟的块压缩存储及求解[J].石油勘探与开发,2013,40(4):462-467. WANG Baohua,WU Shuhong,HAN Dakuang,et al.Block com⁃ pressed storage and computation in large-scale reservoir simula⁃ tion[J].Petroleum Exploration and Development,2013,40(4):462-467.

    • [6] AZLAN A A A,WAN Muda,MUBARAKI W M,et al.New ap⁃ proach of synergizing advanced well test deconvolution,rate tran⁃ sient analysis and dynamic modeling in evaluating reservoir com⁃ partmentalization uncertainty at k field in sarawak basin;A case study[C].Manama:SPE Middle East Oil and Gas Show and Con⁃ ference,2019.

    • [7] ANYANWU C,AFEKHUME D,ONOBRAKPEYA C,et al.Inte⁃ grated reservoir modeling and uncertainty assessment of a reser⁃ voir complex in the Niger Delta[C].Lagos:SPE Nigeria Annual In⁃ ternational Conference and Exhibition,2015.

    • [8] 刘伟,赵辉,雷占祥,等.基于单井敏感性局域化EnKF的油藏辅助历史拟合方法[J].石油学报,2019,40(6):716-725. LIU Wei,ZHAO Hui,LEI Zhanxiang,et al.Reservoir assisted his⁃ tory matching method using a local ensemble Kalman filter based on single-well sensitivity region[J].Acta Petrolei Sinica,2019,40(6):716-725.

    • [9] 张凯,马小鹏,王增飞,等.一种强非均质性油藏自动历史拟合混合求解方法[J].中国石油大学学报:自然科学版,2018,42(5):89-97. ZHANG Kai,MA Xiaopeng,WANG Zengfei,et al.A hybrid meth⁃ od for solving automated production history match of reservoirs with strong heterogeneity[J].Journal of China University of Petro⁃ leum:Edition of Natural Science,2018,42(5):89-97.

    • [10] 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):1-21.

    • [11] 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].Denver:SPE An⁃ nual Technical Conference and Exhibition,2020.

    • [12] BRUYELLE J,GUÉRILLOT D.Proxy model based on artificial in⁃ telligence technique for history matching-application to Brugge Field[C].Dubai:SPE Gas & Oil Technology Showcase and Con⁃ ference,2019.

    • [13] ZOTKIN O,OSOKINA A,SIMONOV M,et al.A novel approach to refinment reservoir proxy model using machine-learning tech⁃ niques[C].Baku:SPE Annual Caspian Technical Conference,2019.

    • [14] SHAHKARAMI Alireza,MOHAGHEGH Shahab.智能代理在油藏建模中的应用[J].石油勘探与开发,2020,47(2):372-382.SHAHKARAMI Alireza,MOHAGHEGH Shahab.Applications of smart proxies for subsurface modeling[J].Petroleum Exploration and Development,2020,47(2):372-382.

    • [15] 王链,姚约东,王孔杰,等.基于代理辅助野草猴群算法的井位优快决策方法[J/OL].大庆石油地质与开发,2021:1-8[2021-11-10].https://kns.cnki.net/kcms/detai/23.1286.TE.20211014.1108.001.html.WANG Lian,YAO Yuedong,WANG Kongjie,et al.Optimal and fast decision making method for well location based on surrogateassisted weed monkey algorithm(WMA)[J/OL].Petroleum Geolo⁃ gy & Oilfield Development in Daqing,2021:1-8[2021-11-10].https://kns.cnki.net/kcms/detail/23.1286.TE.20211014.1108.001.html.

    • [16] 李伟,唐放,侯博恒,等.基于神经网络的南海东部砂岩油藏采收率预测方法[J].油气藏评价与开发,2021,11(5):730-735,743.LI Wei,TANG Fang,HOU Boheng,et al.A method for oil recov⁃ ery prediction of sandstone reservoirs in the eastern South China Sea based on neural network[J].Reservoir Evaluation and Devel⁃ opment,2021,11(5):730-735,743.

    • [17] 谷建伟,周梅,李志涛,等.基于数据挖掘的长短期记忆网络模型油井产量预测方法[J].特种油气藏,2019,26(2):77-81,131.GU Jianwei,ZHOU Mei,LI Zhitao,et al.Oil well production fore⁃ cast with long-short term memory network model based on data mining[J].Special Oil & Gas Reservoirs,2019,26(2):77-81,131.

    • [18] 吴君达,李治平,孙妍,等.基于神经网络的剩余油分布预测及注采参数优化[J].油气地质与采收率,2020,27(4):85-93.WU Junda,LI Zhiping,SUN Yan,et al.Neural network-based pre⁃ diction of remaining oil distribution and optimization of injectionproduction parameters[J].Petroleum Geology and Recovery Effi⁃ ciency,2020,27(4):85-93.

    • [19] 葛兰,蒲谢洋.页岩气田智能化生产辅助决策系统应用效果评价[J].油气藏评价与开发,2021,11(4):621-627.GE Lan,PU Xieyang.Evaluation of application effect of an intelli⁃ gent production assistant decision system for shale gas field[J]. Reservoir Evaluation and Development,2021,11(4):621-627.

    • [20] 刘巍,刘威,谷建伟,等.利用卡尔曼滤波和人工神经网络相结合的油藏井间连通性研究[J].油气地质与采收率,2020,27(2):118-124.LIU Wei,LIU Wei,GU Jianwei,et al.Research on interwell con⁃ nectivity of oil reservoirs based on Kalman filter and artificial neu⁃ ral network[J].Petroleum Geology and Recovery Efficiency,2020,27(2):118-124.

    • [21] CANCHUMUNI S W,EMERICK A A,PACHECO M A C.History matching geological facies models based on ensemble smoother and deep generative models[J].Journal of Petroleum Science and Engineering,2019,177:941-958.

    • [22] KINGMA D P,WELLING M.Auto-encoding variational bayes[C].Banff:International Conference on Learning Representations,2013.

    • [23] WANG Z,XIAO D,FANG F,et al.Model identification of reduced order fluid dynamics systems using deep learning[J].Numerical Methods in Fluids,2018,86(4):255-268.

    • [24] GERS F A,SCHMIDHUBER J,CUMMINS F.Learning to forget:Continual prediction with LSTM[J].Neural Computation,2000,12(10):2 451-2 471.

    • [25] GONZALEZ F J,BALAJEWICZ M.Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems[D].Urbana:University of Illinois at Urbana-Cham⁃ paign,2018:1-28.

    • [26] TANG M,LIU Y,DURLOFSKY L J.A deep-learning-based sur⁃ rogate model for data assimilation in dynamic subsurface flow problems[J].Journal of Computational Physics,2020,413:109456.

    • [27] TEMIRCHEV P,SIMONOV M,KOSTOEV R,et al.Deep neural networks predicting oil movement in a development unit[J].Jour⁃ nal of Petroleum Science and Engineering,2020,184:106513.

    • [28] MO S,ZHU Y,ZABARAS N,et al.Deep convolutional encoderde⁃ coder networks for uncertainty quantification of dynamic multi⁃ phase flow in heterogeneous media[J].Water Resource,2019,55(1):703-728.

    • [29] RAISSI M,PERDIKARIS P,KARNIADAKIS G.Physics-in⁃ formed neural networks:a deep learning framework for solving for⁃ ward and inverse problems involving nonlinear partial differential equations[J].Journal of Computational Physics,2019,378:686-707.

    • [30] WONG Z Y,HORNE R,VOSKOV D.A geothermal reservoir simu⁃ lator in AD-GPRS[D].Stanford:Stanford University,2015:1-72.

    • [31] WATTER M,SPRINGENBERG J,BOEDECKER J,et al.Embed to control:A locally linear latent dynamics model for control from raw images[C].Montreal:International Conference on Neural In⁃ formation Processing Systems,2015.

    • [32] DONG Guozhu,LIU Huan.Feature engineering for machine learn⁃ ing and data analytics[M].New York:CRC Press,2018.