基于深度学习的钢套管长电极电法页岩压裂成像监测

DEEP LEARNING-BASED ELECTRICAL IMAGING MONITORING OF SHALE GAS FRACTURIN GUSING STEEL WELL CASINGS AS LONG ELECTRODES

李寅初

11849188

硕士

力学

杨迪琨

2020-05-30

2020-05-30

哈尔滨工业大学

深圳

页岩水力压裂过程中需要将大量的水压入地下。监测页岩开发中压裂液随时间变化的注入、返排和滞留吸收等信息对于页岩气可持续大规模安全开采具有重要意义。压裂液与周围地层相比通常具有较高的电导率，因此地球物理电法、电磁法可以发挥独特的作用。然而传统的地面电法、电磁法难以采集到深部页岩层内小尺度压裂液流动产生的信号，而井中方法价格昂贵并且会干扰正常的生产操作，因此亟需一种对深部储层具备高分辨率的地面电场监测方法。随着越来越多地球物理调查利用场区中的钢套管井作为长电极增大探测深度，以及近几年国际勘探地球物理领域对油井钢套管效应正演模拟的深入研究，长电极电磁法逐渐成为一种油藏开发监测的有力工具。长电极法与常规电磁法相比，钢套管作为人造极端高导体增加了在地层尺度下正演数值模拟的难度，并且压裂监测对快速成像的要求也进一步提高了反演成像的难度。目前对长电极电法、电磁法压裂液高时空分辨率监测的研究多集中于正演模拟研究阶段，而相应的反演成像问题尚未得到充分研究。本文以非常规油气藏水力压裂为背景，将长电极电法和深度学习技术引入到压裂液分布高分辨率反演成像监测和钢套管电导率反演成像问题中。本文建立典型页岩层地电模型并采用套管井口供电、井口周围观测的测量方法，利用三维等效 电阻网络法（RESnet）研究了长电极电法中地表电场时移差分数据对储层内压裂 液流动的可探测性，以及对反映钢套管完整性的异常电导率分布的可探测性。基于可探测性研究得到的定性认识以及目标异常可探测的条件，将深度学习以全卷积网络的形式引入到长电极电法压裂液分布和钢套管电导率分布的反演成像中，并 有针对性地设计深度神经网络算法。相较于常规的正演方法，RESnet 在保证精度 要求的同时避免了对小尺度物体的精细模拟，便于快速大量生成深度学习所需要的数据集。压裂液电导率分布的反演成像结果表明，本文设计的深度神经网络反演算法可以从井口周围地面电场数据中解析出水平井段压裂面上压裂液分布的有效信息，能够取得合理的反演成像结果。鲁棒性试验同样表明该深度学习的反演算法面对噪声等不确定性干扰也具有一定的稳定性。一旦训练完成，该基于深度学习的算法可以很好地应对水力压裂过程中压裂液实时成像的要求。在钢套管电导率的反演成像研究中，在关于套管完整性的几个地球物理电场模型上的测试表明经过训练后的网络模型能够很好地从地面数据反演钢套管上由破损或腐蚀导致的电导率异常。本论文第一次将深度学习应用到压裂液成像监测及钢套管完整性监测，结果表明这种通过大量数值模拟进行线下训练的方法为压裂液分布和钢套管电导率实时监测提供了一种新的思路与技术路线。

Hydraulic fracturing requires a large volume of water to be injected into the earth. Monitoring the injection, flowback, retention, and absorption of the fracturing fluid over time is of great significance for the safe and sustainable large-scale production and developmentofshalegas. Thefracturingfluidusuallyhasanelectricalconductivityhigher thanthesurroundingformation,whichallowstheutilizationofgeophysicalelectricaland electromagnetic methods. However, it is difficult for surface electrical and electromagneticmethodsconventionallycarriedoutabovetheanomalousobjectstodetectsignalsof small-scalefracturingfluidinthedeepshalelayer. Inaddition,thedownholemethodsare expensive and may interrupt the normal operation of the wells. Thus, new surface electrical field monitoring methods for deep reservoirs are urgently needed. In recent years, manygeophysicalsurveyshavebeenusingsteel-casedwellsaslongelectrodestoenhance deep signals. The international exploration geophysics community has also conducted many in-depth studies on the forward modeling of steel well casing’s effect in the oil fields. Thelongelectrodeelectromagneticmethodhasgraduallybecomeapowerfultool forreservoirdevelopmentmonitoring. Compared with the conventional electromagnetic methods,thepresenceofsteelcasingsasman-madeinfrastructureinextremelyhighconductivitymountsachallengetotheforwardnumericalsimulationatthescaleofgeological stratum. Andtherequirementoffastfracturingimagingmonitoringfurtherincreasesthe difficulty of inversion and imaging. At present, the research on long electrode electrical andelectromagneticmethodswithhighspatialandtemporalresolutionismostlyfocused ontheforwardmodeling,butonlyveryfewhavestudiedtheimagingandinverseproblem. Withinthecontextofthehydraulicfracturingofunconventionaloil&gasreservoirs scenario, this paper introduces the long electrode electrical method and deep learning techniqueinto(1)high-resolutionimagingmonitoringofdirectionalfracturingfluidflow and(2)wellcasing’sconductivityinversion. Firstly,Atypicalgeoelectricmodelconsideringshalelayerandasurveyconfigurationusingtop-casingsourceandreceiversaround thewellheadareestablishedinthispaper. The3Dequivalentresistancenetworkmethod (RESnet)isusedtostudythedetectabilityoftime-lapsedifferentialsurfaceelectricfield dataforthefracturingfluidflowinthereservoir,aswellasfortheanomalousconductivity distribution reflecting the casing integrity. From the detectability studies, some qualitative knowledge and the condition of target anomaly detection are obtained. This paper then introduces the deep learning technique in the form of a fully connected network to the long electrode electric imaging for fracturing distribution and casing’s conductivity distribution with a specifically designed deep neural network algorithm as the inversion method. Compared with the conventional forward modeling methods, RESnet avoids mesh refinement for small-scale targets without loss of necessary accuracy. This feature isusefulinthefastdatagenerationfordeeplearning. The results of deep learning-based inversion of fracturing fluid conductivity distributionshowthatthedeepneuralnetworkalgorithmdesignedinthispapercaneffectively decode the information reflecting the fracturing fluid distribution at the horizontal well fromsurfacedataandobtainreasonableimagingresults. Therobustnessexperimentsalso demonstrate the stability of this algorithm against uncertain disturbances such as noise. Once trained, this deep learning-based inversion algorithm can meet the requirement of real-timeimagingoffracturingfluidduringhydraulicfracturing. Insolvingtheinversion ofcasing’sconductivitydistribution,testsondifferentgeoelectricalmodelsaboutcasing integrity have shown that the trained network can recover the anomalous conductivity distributioncausedbydamageorcorrosiononthecasingfromsurfacedata. This paper is the first attempt to apply deep learning techniques to fracturing fluid imaging monitoring and casing integrity monitoring. The results show that the method ofusingmassivenumericalsimulationforofflinetrainingprovidesanewtechnicalroute forfracturingfluiddistributionandcasingconductivityinversion.

电法勘探 水力压裂监测 套管完整性 深度学习 全卷积网络

electricalexploration hydraulicfracturingmonitoring casingintegrity deep learning fullyconvolutionalnetwork

中文

联合培养

南方科技大学

李寅初. 基于深度学习的钢套管长电极电法页岩压裂成像监测[D]. 深圳. 哈尔滨工业大学,2020.