Identification of the Quaternary low gas-saturation reservoirs in the Sanhu area of the Qaidam Basin

Pet. Sci.20118:4954DOI10.1007/s]2182-011-0llIdentification of the Quaternarylow gas-saturation reservoirs in theSanhu area of the Qaidam Basin, ChinaLi Xiongyan,, Li Hongqil-, 2, 3, Zhou Jinyu, He Xu Chen Yihan. 2 andYu Hongyan,2State Key Laboratory of Petroleum Resource and Prospecting, China University of Petroleum, Beijing 102249, China2 Key Laboratory of Earth Prospecting and Information Technology, China University of Petroleum, Beijing 102249, ChinaDepartment of Computer Science and Technology, China University of Petroleum, Beijing 102249, ChinaExploration and Development Institute, Changqing Petroleum Company, Shaanxi 710021, Chinac China University of Petroleum( Beijing) and Springer-Verlag Berlin Heidelberg 2011Abstract: Low gas-saturation reservoirs are gas bearing intervals whose gas saturation is less than 47%.They are common in the Quaternary of the Sanhu area in the Qaidam Basin. Due to the complex genesismechanisms and special geological characteristics, the logging curves of low gas-saturation reservoirsare characterized by ambiguity and diversity, namely without significant log response characteristicsTherefore, it is particularly difficult to identify the low gas-saturation reservoirs in the study area. Inddition, the traditional methods such as using the relations among lithology, electrical property, physicalproperty and gas bearing property, as well as their threshold values, can not effectively identify low gas-saturation reservoirs. To solve this problem, we adopt the decision tree, support vector machine andrough set methods to establish a predictive model of low gas-saturation reservoirs, which is capableof classifying a mass of multi-dimensional and fuzzy data. According to the transparency of learningprocesses and the understandability of learning results, the predictive model was also revised by absorbingthe actual reservoir characteristics. Practical applications indicate that the predictive model is effective inidentifying low gas-saturation reservoirs in the study areaKey words: Sanhu area, Qaidam Basin, low gas-saturation reservoir, decision tree, support vectormachine, rough set, predictive model, identification1 Introductionsaturation reservoirs turns out to be difficult. Decision treea number of typical shallow biogenic gas reservoirs existt (DT), support vector machine(SVM)and rough set(RS)techniques are the logical classification methods whichin the Quaternary of the Sanhu area in the Qaidam Basin. can be utilized to handle the logical structure. ComparedAfter several years'research, significant progress has been with artificial neural networks(ANN), Bayesian networksachieved in understanding the genesis and distribution of the (BN) and genetic algorithms (GA), the process is morebiogenic gas reservoirs, caprocks and reservoir characteristics clear, the rule is more comprehensible and the humanWei et al, 2005; Wang et al, 2007: Zhu and Kang, 2005, computer interaction function is stronger. However they areZhu et al, 2006: Guo et al, 2008; Cheng et al, 2008). In seldom used in reservoir evaluation and fluid identificationrecent years, exploration practices have shown that the (Sheremetov et al, 2007; Chen et al, 2008). Based on thelow gas-saturation reservoirs with gas saturation less than definition and characteristics of low gas-saturation reservoirs,47% exist widely in the region, accounting for about one we used the decision tree, support vector machine and roughthird of the total gas reserves. Because of their complicated set methods to construct a predictive model of low gasgenesis,low gas-saturation reservoirs do not have significant saturation reservoirs which can well and trulicro-features, and their log response characteristics showgas-saturation re中国煤化工 ation of decisionambiguity and diversity. Consequently identifying low gas- tree, support veCNMHGet techniques issignificant tonent of low gas-saturation reservoirs in the Quaternary of the Sanhu area inCorrespondingauthor.email:hq.li@cup.edu.cnthe qaidam BasinReceived october 26 200950Pet. Sc(20)849-542 Definition and characteristics of low gas- types: structure-deposition, diagenesis-compaction andsaturation reservoirsaccumulation. Different types of genesis lead to the diversityof low gas-saturation reservoirs, which results in comple2.1 Definitionmicro-features and unclear log response characteristicsThe lithology of low gas-saturation reservoirs is mainlyIn the study area, because of the presence of bound water argillaceous siltstone and argillaceous fine siltstone. Theand free water, the gas bearing interval with gas saturation porosity is generally from 25% to 40%, and the permeabilityless than 47% is named low gas-saturation reservoir. After is primarily from 1 mD to 10 mD, so a low gas-saturationmany years of study in this region, the figure 47% was reservoir is high porosity and middle-low permeabilityobtained and it can not be universally applied to other areas. reservoir. The pore type is simple, mostly primary pores andAs shown in Fig. 1 when the gas saturation increases to 47%, a few secondary pores. The primary pores are mainly primarythe gas relative permeability is high and the water relative intergranular pores and micropores, while the secondary porespermeability becomes very low. It indicates that the water is include dissolved pores and fractures. The reservoir thicknessentirely bound water at this moment. When the gas saturation is small, and the single layer thickness is basically from 0.5 mfurther increases, the water relative permeability is almost to 2 m, accounting for about 70%. The reservoirs whoseunchangedthickness is more than 5 m account for only about 15%. Thecontent of clay minerals, content of shale, and formationwater salinity are high. Carbonate and pyrite exist widelyr Gasin the study area, as shown in Table 1. the montmorilloniteand andreattite clay minerals have additional conductivityHigh shale content leads to the development of microporesIllite will expand when encountering water, and will blockmicropores. So it will adsorb more bound water, resulting ina higher bound water content in low gas-saturation reservoirsBecause of high formation water salinity, the concentrationof conductive ions is high. Formation water with highconductive network in the pore channels of the rock. WhenGas saturation.%6 80 100 concentration of conductive ions generates a we\\ _meRopethe distribution and content of pyrite meet specific conditions,it will severely influence the apparent resistivity of formationFig. 1 Relative permeability of gas and water in low gas-saturation reservoir All the factors mentioned above will significcantly decrease2.2 Characteristicsthe apparent resistivity of the formation. However, due tothe low conductivity of carbonate, the apparent resistivityIn the Qaidam Basin, the cold climate and high salinity of formation will increase. Additionally, because of theenvironment have slowed the degradation of organic matter. occurrence of thinly interbedded sands and shales and theBesides, the dark shale provides a sufficient gas source, and low resolution of the well logging, the measured resistivitythe combination of reservoir and cap rock forms a favorable can not adequately reflect the actual formation informationaccumulation condition. Therefore, the"source below and Therefore, the resistivity of low gas-saturation reservoirsreservoir above"as well as"self-source-reservoir"dynamic has a wide varying range, from 0. 27 22 m to 1.61 S2 m. Theaccumulation pattern has occurred in the Quaternary of the response characteristics of three porosity curves and theSanhu area (Dang et al, 2004; Zhao et al, 2008). The genesis apparent resistivity of formation can not be regarded as a trueof low gas-saturation reservoirs can be divided into three reflection of gas bearing reservoirsTable 1 Micro-features of low gas-saturation reservoirsClay mineral contentShale content, Formation water salinity, g/L Carbonate content, Pyrite content,%Kaolinite, Chlorite, Illite, Andreattite64.0046.005.0021.59111.6615.8222.67157.2914443 Method of identifying low gas-saturation which was or中国煤化工s Quinlan in 1986reservoirs(Quinlan, 19CNMHGimproved by laterresearchers. It adopts a local searching approach to learn3.1 Fundamental conceptsand analyze data sets, and uses the loss function with cross-a decision tree is a typical classification method, validation as the score function. At last, the simple binary treePes201184954structure is obtained, which will be applied in classification the upper approximation, the lower approximation and theand prediction. The decision tree method is composed of positive region. In principal, attributes are divided intotree construction and tree pruning a top-down recursive condition attributes and conclusion attributes in the databasedivide-and-conquer method is used to construct a tree. The Then according to the attribute value, the tuples will beinformation gain, gain ratio and Gini index are the evaluation divided into the corresponding subset. At last, the decisionparameters. The algorithm selects an appropriate splitting rules will be generated based on the relationships betweenattribute, and decides the next split node by analyzing other the upper approximation and the lower approximation in theattribute values. The training set is recursively partitioned subsets of condition attributes and conclusion attributes(Haninto smaller subsets as the tree is built. This process will be and Micheline, 2007; lan and Eibe, 2006)repeated until the division stops. In the tree pruning includlng 3.2 Modeling ideaspre-pruning and post-pruning, a statistical measure is appliedto cut off the least reliable branches, which results in faster With different types of reservoirs as the class labels, weclassification, as well as improvement of the classification and can use abundant information to classify various reservoirsprediction capabilities of decision tree(Han and Micheline, That is the essence of using logging, core and well testing data2007; lan and Eibe, 2006)to identify low gas-saturation reservoirs. When the algorithmThe support vector machine method was introduced obtains the optimal learning model, the fluid in unknownin 1963 by Vladimir Vapnik(Cortes and Vapnik, 1995). It reservoirs can be predicted. This is a nonlinear, complex andis a very promising classification method and can replace high-dimensional classification problem. The decision tree,artificial neural networks. Its algorithm can construct a support vector machine and rough set methods can all gain themaximum marginal hyperplane in the training sample classification rules by analyzing a group of non-order, non-with limited data. The maximum marginal hyperplane can rule and complicated instances, as shown in Fig. 2. Comparedseparate two types of data sample as much as possible, while with the black box model of artificial neural networks andmaximizing the difference between them. The support vector Bayesian networks, they have relatively transparent processesmachine finds the maximum marginal hyperplane through and easily understandable rules. In addition, they can alsonuclear techniques in high-dimensional space so the data clearly show the sensitivity of each attributeample must be projected onto the high-dimensional space. Therefore, the logging, core and well testing data canConsequently, it has a strong capacity for distinguishing two be considered as the attributes, which are analyzed andtypes of attributes and a good ability to generalize the high- compared by different methods to decide their sensitivdimensional sample properties(Han and Micheline, 2007; Due to the ambiguity of logging information in low gasIan and Eibe 2006; Shi, 2008; Cortes and Vapnik, 1995; saturation reservoirs, the traditional three porosity curvesScholkopf et al, 2000; Platt, 1999and resistivity property are not necessarily the most sensitiveThe rough set method was first described by the Polish attributes. We use decision tree, support vector machinescholar Zdzislaw Pawlak in 1982. It is a useful data analysis and rough set methods to process respectively a variety ofmethod to deal with imprecise or fuzzy data. By analyzing attribute combinations based on the most sensitive attributesinaccurate, inconsistent and incomplete data, the rough set The rationality and accuracy are used to evaluate the analysiscan discover hidden knowledge and reveal underlying rules. results of different parameter combinations. Besides, theThe rough set method is based on the knowledge reduction integration of effective rules and reduction of invalid rulesaims to classify and maintains the strong classification are indispensable processes. As a result, we will obtain thecapability. The basic concepts of rough set theory include optimal attribute combination and corresponding predictivethe basic knowledge block, the indistinguishable relation, model of low gas-saturation reservoirsDistinguishing low gas-saturationThis is a nonlinearIdentifying lowservoirs, gas bearing reservoirscomplex andlogging, core and well testing dataclassification problemClassification problemThey can all gain the classification rules byDecisionof non order, non-rule and中国煤化工complicated instances. They can be used forCNMHGFig. 2 Feasibility analysis of identifying low gas-saturation reservoirs using DT, SVM and Rs3.3 Modeling processphysical parameters including POR, PERM, RQI and SHIn the study area, we select 86 key wells and 336 well have no effect on the fluid recognition. It indicates that theretesting intervals including 146 low gas-saturation reservoirs are no significant differences between the micro-structure ofLGS), 50 gas bearing reservoirs(GB), 118 water layers low gas-saturation reservoirs and that of other reservoirs. The(WL)and 22 dry layers(DL). There are a great number of small displacement pressure of low gas-saturation reservoirslow gas-saturation reservoirs, which have great potential for results in an undersize hole. Consequently, CAL becomes oneexploitation. The information of logging and core data in of main parameters in identifying different fluids. with thewell testing intervals is expressed by 16 parameters, namely seven core parameters as well as others, we comprehensivelyspontaneous potential relative reading(ASP), natural gamma utilized all the logging information to identify low gas-ray(GR, unit: API), caliper(CAL, unit: cm), acoustic velocity saturation reservoirs, gas bearing reservoirs, water layers and(AC, unit: us/m), compensated neutron porosity(CNL, dry layersunit:%), densilog(DEN, unit: g/cm), deep investigationThe results of processing various parameter combinationslaterolog relative reading(ARD), shallow investigation with the decision tree, support vector machine and roughlaterolog relative reading(ARS), microlaterolog relative set methods are shown in Table 2. Decision tree and roughreading(ARMLL), deep investigation induction log(ILD, set methods are better than the support vector machine inunit: S2 m), medium investigation induction log(ILM, unit: the accuracy of modeling. With regard to the parameter22'm), porosity(POR, unit: %) permeability(PERM, unit: combination, none of the parameter combinations has amD), reservoir quality index(RQl, unit: w/m), displacement distinct advantage. Additionally, in Table 2, the accuracy ispressure(P, unit: MPa), shale content(SH, unit: %) As the the overall accuracy to identify all fluids, while the accuracymud resistivity has serious impact on the laterolog, we used of identifying low gas-saturation reservoirs requires furthertheir relative values. The sensitivity of main parameters to analysis. Because of some false IF-THEN rules producedfluid properties is P>ILD>AC>DEN>CAL>ILM>GR. The by the decision tree and rough set methods, the assistance ofTable 2 Processing results of different parameter combinationsParameter combinationsSVM1 DEN. AC P ILD2 DEN. AC P ILD. ILM8290%62.18%63.89%3 CAL, DEN, AC, P, ILD, ILM86.01%82.05%74.50%4 CNL, DEN, POR, PERM, RQL, P, ILD, ILM8497%82.05%7751%5 AC, CNL, DEN, POR, PERM, RQL, P, ILD, ILM86.53%71.50%9005%6 GR, AC, CNL, DEN, POR, PERM, RQL P, ILD, ILM86.53%71.50%8805%7 ASP, GR, AC, CNL, DEN, POR, PERM, RQL, P, ILD, ILM 88.08% 71.50% 78.06%ASP, GR, CAL, AC, CNL, DEN, POR, PERM, RQL, P, ILD,88.60%71.50%75.50%the support vector machine method and domain know ledgeTable 3 Identification rules of low gas-saturation reservoirsis essential. We further analyze the predictive model of allNumber of eligiblereservoirs based on parameter combinations from 3 to 8. Eightrules of identifying low gas-saturation reservoirs are shown inIdentification rulesLGS GB WL DLTable 32.28389.52,Because of the complex geological environment where low100%0.042.31, P>0.22 120.31%verifies the fact that only relying on the traditional threshold 4 CAL21 40 AC>381.08,involving multiple parameters of logging information. It also29000100%values to identify low gas-saturation reservoirs is not enoughThe eight rules are used to deal with 86 wells in the study area 5 AC<400.95,2.288033,P0.22and the processing results are compared with the well testing 6 AC>502YH中国煤化工208636%conclusions In the total 146 low gas-saturation reservoirs, 144can be identified with the accuracy of 98.63%. As each rule is 7CNMHGa necessary, but not sufficient, condition of the corresponding0.54101.93,1301092.86%Pet. Sci.(2011)8:49-54Due to the complex genesis of low gas-saturation situations. In the study area, on account of various genesis ofreservoirs, the same parameters have entirely different low gas-saturation reservoirs, the traditional threshold valuesthreshold values in various low gas-saturation reservoirs. do not existThe most sensitive parameter to the fluid propertiesdisplacement pressure( P), appearing respectively in 1, 2, 3, 4 Applications4, and 5 rules. It has significant difference in different rulesThe integrated identification rules were applied to processas shown in Fig 3. The displacement pressure of low gas- the actual logging data in the study area on the basis of thesaturation reservoirs is distributed from 0.01 MPa-1. 68 MPa low gas-saturation reservoir characteristics. Fig. 4 is the(two main ranges are 0.04 MPa-0. 22 MPa and 0. 22 MPa1.18 MPa and their accuracy are 67.21% and 53. 16%) In processing result of well A. LSG is the parameter to identifygas bearing reservoirs and water layers, the displacementlow gas-saturation reservoirs, and WL is the parameter topressure values have no significant difference. It shows that recognize water layers. When the value of LSG and WL is 1the displacement pressure value has no threshold value in and water layer, respectively. In Fig. 4, LSG shows the firstlow gas-saturation reservoirs. Other parameters have similarand second layers are low gas-saturation reservoirs and WLindicates the third layer is the water layer. The gas production0.010.22rate is 30, 095 m per day1.1The processing results of 86 key wells and 6 new wells0.221.68indicate that the overall accuracy is 88.60% and the accuracy0040.22of identifying low gas-saturation reservoirs is 91.70%.The04reason is that the predictive model of low gas-saturationreservoirs using the combination of decision tree, support1.8 vector machine and rough set methods is better than thatof water layer and gas layer' Compared with the traditionalP MPainterpretation method, the predictive model adds 276 low gassaturation reservoirs, whose effective thickness is 303. 48 m.Fig. 3 Threshold values of P in low gas-saturation reservoirsLithologyPorosityLaterolognduction IoLow saturation Water layerDepth, m500(us/m)20d0.1(Qm)CNLRS1(0m)1(01m1CALDENRMLLILM(cm)36195(gcm)2940.1(0m)101(9m)KHa中国煤化工CNMHGEFig. 4 Processing result of well APet. Sci(2018:49-45 ConclusionsDang Y Q, Zhang D W, Xu ZY, et al. Sedimentary facies and biogenicgas pool of the Quaternary of the Sanhu area in Qaidam Basin1)Although the decision tree, support vector machine Joumal of Palaeogeography. 2004. 6(1): 110-118(in Chinese)and rough set methods can quickly build a predictive Guo z Q, Li b l, zhang L, et al. Discussion on minimum closure formodel, a variety of parameter combinations can change thelow-amplitude structural nature gas pool: A case study from theidentification rules. There are also some false rules ThereforeSanhu area in the Qaidam Basin. Chinese Jourmal of Geology. 2008exhaustive parameter combinations based on the sensitive43(1): 34-49(in Chinese)parameters are essential. Other methods in combination withHan J W and Micheline K. Data Mining Concepts and Techniquesdomain knowledge can be used to remove the false rules.(Second Edition). Beijing: China Machine Press. 2007. 3(in2)If the optimal linear combinationof multiple attributes lan h and Eibe E. Data Mining Practical Machine Learning Tools andcould work as the derived parameter, it can not only greatlyTechniques(Second Edition). Beijing: China Machine Press. 2006. 2improve the performance of the predictive model, but (in Chinese)also integrate the linear regression of single parameter Platt J C. Fast training of support vector machines using sequentialand multiple parameters in the well logging evaluationminimal optimization. Advances in Kernel Methods: Support VectorConsequently, choosing a splitting parameter is one of newLeaming. Cambridge, MA, USA: MIT Press. 1999. 185-208development directions in the process of building predictive Quinlan JR Induction of decision trees. Machine Learning. 1986. I(D)models8l-106reservoirs in the study area. As a result the traditional algorithms. Neural Computation. 2000. 12(5): 1207-1245 ort vector3) There is no threshold value of low gas-saturationScholkopf B, Smola A J, williamson R C, et al. New supporeservoir evaluation methods could not be applied to the low Shererssefov l, Cosuitchi A, Batyrshin I, et al. New insights andapplications of soft computing on analysis of water production fromgas-saturation reservoirs with complex genesis mechanismsoil reservoirs. SPE108702 2007and special geological characteristics. New methods ShiGR. Application of support vector machine to multi-geological-and techniques are effective to evaluate and identify thefactor analysis. Acta Petrolei Sinica. 2008. 29(2): 195-198(incomplicated reservoirsChinese)Wang JP, Peng S M, Guan Z Q, et al. Mudstone caprock's sealingAcknowledgementsmechanism of biogenetic gas reservoir of Quatenary in QaidamBasin. Journal of Southwest Petroleum University. 2007. 29(6): 63This work is partially supported by the National High67(in ChineseTechnology Research and Development Program(863 Wei G Q, Liu D L, Zhang Y, et al. Formation mechanism, distributionProgram 2009AA062802). The authors would like to feature and exploration prospect of the Quatermary biogenic gas inappreciate Penny Peng for her kind assistance and supportQaidam Basin, Nw China Petroleum Exploration and Development.2005.32(4:84-89( Gn Chinese)ReferencesZhao WW. Zha M and Wu K Y Relationship between unconformityand hydrocarbon accumulation in the eastern Qaidam Basin. ActaChen D, Hamid S, Dix M, et al. Cooperative optimization-basedGeologica Sinica. 2008. 82(2): 247-263(in Chinese)dimensionality reduction for advanced data mining and visualization. Zhu HY, Ma Ln, Chen jJ, et al. Research on characteristics of reservoirSPEl17765.2008bed in Quaternary of Sebei-1 Gas Field. Natural Gas Industry. 2006Cheng F Q, Jin Q, Lin H x, et al. Controlling effect of reservoir-cap26(4): 29-31(in Chinese)combination quality on richment grade of biogas in the Sanhu area Zhu X M and Kang A Characteristics and evaluation of the Quaternaryof the Qaidam Basin. Chinese Journal of Geology. 2008. 43(2): 333-reservoirs in Qaidam Basin. Natural Gas Industry. 2005. 25(3): 29-31(in Chinese)Cortes C and Vapnik V. Support vector networks. Machine Learning.1995.20(3):273-297(Edited by Hao Jie)中国煤化工CNMHG