Research On Porosity And Permeability Model Of Tight Sandstone Reservoirs Based On Machine Learning

Porosity and permeability are important parameters for evaluating reservoir performance and seepage characteristics,and are also the key factors to excavate reservoir productivity and enhance oil recovery.Generally,porosity and permeability can be obtained through core experiments,but due to the high cost of coring and the limited number of corings,it is impossible to characterize the porosity and permeability of the entire reservoir.Compared with the"fine and incomplete"core experimental data,the logging data contains more comprehensive reservoir information,but it is not accurate enough.The upper Paleozoic He 8member reservoir in the middle part of the western margin of the Ordos Basin is a typical tight sandstone reservoir with complex pore-throat structure.In the previous evaluation of reservoir physical properties,the existing models and logging interpretation methods are often used to predict the results.The results are often very biased,and the parctical application effect is not good.How to connect the"complete but not precise"conventional logging data with the"precise but incomplete"core experimental data,and how to fully excavate the physical property information hidden in conventional logging data by using cutting-edge mathematical algorithms,and how to establish porosity and permeability prediction models based on laboratory testing of pore-throat structure and conventional logging data is a key problem for tight sandstone reservoir to excavate productivity and improve recovery efficiency.This paper takes the tight sandstone reservoir in the eight member of Shihezi Formation（HE 8）as a research example.With the help of experimental methods such as casting thin sections,scanning electron microscopy,high pressure mercury intrusion,nuclear magnetic resonance and CT scanning,the main controlling factors of porosity and permeability of tight sandstone reservoirs are systematically studied.On the basis of that,machine learning technology is used to fully mine the relationship between casting thin section data,high-pressure mercury intrusion data,nuclear magnetic data and conventional logging data,so as to extract more porosity and permeability information from conventional logging data.The influencing factors of diagenesis and microscopic pore-throat structure on tight sandstone reservoir physical property were introduced for the first time in the model construction,so as to build a pore-permeability prediction model based on laboratory testing of pore-throat structure characteristics and logging data.A set of effective technical methods for constructing pore and permeability prediction model of tight sandstone reservoir based on machine learning integrated algorithm are formed.The results of sedimentary microfacies analysis and data cluster analysis of the tight sandstone reservoir in the He 8 Member show that its depositional environment and diagenesis are relatively simple,which is suitable for unified modeling.In this paper,the influencing factors of porosity and permeability model of tight sandstone reservoir are studied,and the following five conclusions and understandings are obtained:1.The diagenesis of the tight sandstone reservoirs of the He 8 Member in the study area mainly includes compaction,cementation,metasomatism and dissolution.Among them,the secondary dissolved pores and kaolinite intercrystalline pores are the main storage space,while the primary intergranular pores are the secondary reservoir space.Compaction and cementation are the main reasons for the deterioration of reservoir physical properties,among which hydromica cementation and chlorite cementation are significantly negatively correlated with reservoir physical properties.Dissolution can significantly improve the physical properties of the reservoir.2.The plastic minerals that can indirectly respond to the compaction have a significant response on the spontaneous potential and density logging,and the plastic minerals are negatively correlated with the spontaneous potential and density.Kaolinite cementation has a significant response on potassium,thorium and sonic transit time logs,and the kaolinite cementation factor is negatively correlated with potassium,thorium and sonic transit time.The carbonate cementation factor is positively related to density.When modeling based on machine learning,spontaneous potential,density,acoustic time difference,potassium and thorium can be selected as characteristic variables,and the influence factors of diagenesis on the physical properties of tight sandstone reservoirs can be introduced into the model.3.The pore types of tight sandstone reservoirs of the He 8 Member in the study area are diverse,mainly nano-scale and sub-micron pore throats,and the pore throat structure is extremely complex.The size of pore-throat radius and pore-throat radius ratio can reflect the microscopic seepage capability of tight sandstone reservoirs.The larger the pore-throat radius,the higher the coordination number,the stronger the pore-throat connectivity,and the greater the core permeability.The larger the pore-throat radius ratio,the smaller the permeability and the stronger the rock seepage ability.Based on the median pressure and sorting coefficient of high-pressure mercury intrusion,NMR-based T₂geometric mean（centrifugation）,movable fluid saturation,T₂arithmetic mean（saturation）,and bound fluid saturation are closely related to the physical properties of the reservoir.4.A new porosity and permeability model of tight sandstone reservoir is constructed based on high-pressure mercury injection and nuclear magnetic resonance experiments.On this basis,the characterization factors R₁₀and FZI,which contain the microscopic pore-throat structure information of tight sandstone reservoirs,are introduced into the machine learning porosity and permeability model as input feature variables,which lays a foundation for improving the prediction accuracy of the model.5.In the process of building a porosity and permeability based on machine learning model for tight sandstone reservoir,acoustic transit time,neutron,gamma,resistivity,kaolinite cementation factor Eg,and compaction characterization factor Ey are selected as the characteristic variables of the porosity model.Porosity,acoustic transit time,neutron,spontaneous potential,pore throat structure characterization factor R₁₀and FZI are selected as permeability characteristic variables.The porosity model based on ensemble learning XGBoost has the highest prediction accuracy and the strongest generalization performance.The permeability model based on the ensemble learning Stacking CVRegressor algorithm has the highest prediction accuracy and the strongest generalization performance.The porosity prediction accuracy is improved from 53%of the conventional multiple regression model to91%of the machine learning XGBoost model;the permeability prediction accuracy is improved from 32%of the conventional multiple regression model to 86%of the machine learning Stackingcvregressor model.The results show that the machine learning XGBoost algorithm and Stacking CVRegressor algorithm can solve the problem of porosity–permeability nonlinearity and poor generalization performance of pore-permeability prediction under sparse samples,and form an effective technical method for the construction of porosity and permeability evaluation models of tight sandstone reservoir.