Modeling study of sandstone permeability under true triaxial stress based on backpropagation neural network, genetic programming, and multiple regression analysis
Introduction
As a new, clean, and efficient energy source, tight sandstone gas has become a new area of focus in the exploration and development of unconventional natural gas worldwide, and it plays an important role in oil and gas production in China as well (Zhang, 2008; Zhang et al., 2009, 2012; Dai et al., 2012; Li et al., 2012). In the exploitation of tight sandstone gas reservoirs, multiphysics coupling problems, such as fluid-solid coupling, are frequently encountered with the rock mass (Yang et al., 2012; Tian et al., 2018). The initial in situ stress and pore pressure significantly affect the seepage characteristics of fluid in the rock mass (Zou et al., 2012; Chen et al., 2019; Peng et al., 2020). Permeability is the key parameter determining the development of tight sandstone gas and reflects the difficulty of fluid migration in sandstone reservoirs. The evolution law of reservoir rock permeability has long been the focus of domestic and foreign scholars. Gray and Fatt (1963) studied the influence of pore pressure and confining pressure on the permeability of sandstone cores, and found that permeability increased with pore pressure and decreased with increase in confining pressure. Zoback and Byerlee (1975) further verified the relationship between the permeability, confining pressure, and pore pressure of Berea sandstone, and found that the effect of pore pressure on permeability was more significant than that of confining pressure. Liu et al. (2015) conducted laboratory experiments to study the seepage characteristics of sandstone along the complete stress-strain curve under different confining pressure and pore pressure combinations, and field tests to study the seepage law of sandstone in the vertical direction, and obtained the relationships between permeability and volumetric strain, pore pressure, and confining pressure. Zhang et al. (2017) studied the effect of stress concentration on the seepage characteristics of sandstone and established an exponential relationship between peak gas flow and volume strain of sandstone after failure. Yu et al. (2019) predicted sandstone permeability using scanning electron microscopy and studied its relationship with microscopic characteristics using fractal dimensions. The research results showed that the predicted permeability fit well with the measured data.
Currently, the most commonly used method for testing the permeability of rock mass is laboratory testing. However, precision instruments and equipment are frequently required to obtain more realistic permeability results; this often entails significant investment of time and energy to obtain the permeability of rock mass under the combined effect of principal stress and pore pressure via experiments (Gokceoglu, 2002; Taylor and Appleby, 2006). The use of artificial intelligence systems, such as neural networks, GP, and regression analysis, to solve such highly nonlinear, complex, and fuzzy geomechanical problems has become a global trend (Grima and Babuška, 1999; Ni and Wang, 2000; Özcan et al., 2009; Asadi et al., 2011; Cevik et al., 2011). Ali and Chawathé (2000) established an artificial neural network (ANN) model to predict the permeability of arkosic sandstone using its porosity and mineral composition, and the model exhibited good applicability in predicting the permeability of the South Lucky Lake field. Yin et al. (2013) constructed a BPNN model to predict the gas permeability of coal at various effective stresses and temperatures; they found that the predicted permeability was very close to the experimental permeability and that the maximum relative error did not exceed 5%. Rezaee et al. (2006) used regression analysis to obtain a permeability prediction formula for carbonate rock based on porosity and pore throat radius. The accuracy evaluation showed that the formula had good permeability prediction ability. Nazari and Riahi (2011) established ANN and GP models to predict the tensile strength and water permeability of high-strength concrete. The training results showed that both models had good prediction abilities, but GP had less dependence on users and more reasonable prediction results. These artificial intelligence technologies have strong robustness, generalization error ability, and nonlinear dynamic processing ability, which provide a new way to solve the seepage problem of reservoir rock with multi-factor, complexity, randomness, and nonlinearity.
At present, the experimental testing and prediction modeling for sandstone permeability are usually based on the conventional triaxial stress path or uniaxial stress condition, which belong to the simple stress state. The actual reservoir rock is in an unequal three-dimensional pressure state (σ1 > σ2 > σ3). The significant influence of the intermediate principal stress on the strength, deformation, seepage, and other characteristics of rock has been confirmed (Mogi, 1971; Liu et al., 2018, 2019). In the present study, cubic sandstones are taken as the research objects, and the large number of permeability test data obtained from seepage experiments under the true triaxial stress path are used to construct the sandstone permeability prediction models under the coupling effect of true triaxial stress field and pore pressure using BPNN, GP, and multiple regression analysis to study the seepage law of sandstone.
Section snippets
Sample data from true triaxial seepage experiments
The sandstone for the specimens used in this study was obtained from the outcrop of Chayuan, Chongqing. The Young's modulus, E, of the sandstone is 10.6 GPa and its Poisson's ratio, v, is 0.31. The nuclear magnetic resonance (NMR) test of the sandstone used showed that the porosity is 4.6% and the permeability is 0.3638 mD, the sandstone is mainly composed of quartz and feldspar and possesses relatively stronger homogeneity, with no obvious joints and fissures on the surface. After they were
BPNN
Artificial neural network is a type of nonlinear data modeling technology based on neuron design. It originated from the study of the thought process of the human brain (Sietsma and Dow, 1991; Basheer and Hajmeer, 2000). It contains three main parameters: weight, bias, and activation function. The most widely used neural network model is the BPNN (Rezaee et al., 2006; Özcan et al., 2009; Ding et al., 2011; Yin et al., 2013). The gradient descent algorithm is used by reducing the slope of the
Sensitivity analysis
From the comparison between the experimental and predicted permeability results of the three models, it can be seen that σ1, σ2, σ3, and p affect the variation in permeability. When each principal stress increases, the pores and fractures in the sandstone are compressed and the seepage channel narrows, causing permeability to decline.
Gas injection causes the effective stress in the pores and fractures to decrease; hence, the pore space increases and leads to the expansion of sandstone. This
Conclusion
In this research, BPNN, GP, and multiple regression analysis were used to construct the prediction models for sandstone permeability under true triaxial stress conditions. The conclusions are as follows:
- (1)
The permeability predicted by the BPNN, GP, and multiple regression analysis models were in good agreement with the experimental data, and showed the variation of sandstone permeability with principal stress and pore pressure.
- (2)
Based on the cosine amplitude method, we investigated the effect of
Credit author statement
Beichen Yu: Conceptualization, Methodology, Writing - Original Draft, Writing - Review & Editing; Honggang Zhao: Investigation, Methodology; Jiabao Tian: Investigation, Software; Chao Liu: Methodology, Resources, Project administration, Writing - Review & Editing; Zhenlong Song: Software; Yubing Liu: Formal analysis; Minghui Li: Supervision.
Declaration of competing interest
We declare we have no competing interests.
Acknowledgments
This study was financially supported by the National Natural Science Foundation of China (51874053), the Graduate Research and Innovation Foundation of Chongqing, China (CYS19013, CYB19046, CYB19045).
References (55)
- et al.
Using artificial intelligence to predict permeability from petrographic data
Comput. Geosci.
(2000) - et al.
Evaluating the strength of intact rocks through genetic programming
Appl. Soft Comput.
(2011) - et al.
Artificial neural networks: fundamentals, computing, design, and application
J. Microbiol. Methods
(2000) - et al.
Prediction of compressive and tensile strength of limestone via genetic programming
Expert Syst. Appl.
(2008) - et al.
Application of genetic programming to predict the uniaxial compressive strength and elastic modulus of carbonate rocks
Int. J. Rock Mech. Min. Sci.
(2013) - et al.
Modeling of the uniaxial compressive strength of some clay-bearing rocks using neural network
Appl. Soft Comput.
(2011) - et al.
Experimental study of permeability change of organic-rich gas shales under high effective stress
J. Nat. Gas Sci. Eng.
(2019) - et al.
Estimating DEM microparameters for uniaxial compression simulation with genetic programming
Int. J. Rock Mech. Min. Sci.
(2019) A fuzzy triangular chart to predict the uniaxial compressive strength of the Ankara agglomerates from their petrographic composition
Eng. Geol.
(2002)- et al.
Influence of geological conditions on the powder factor for tunnel blasting
Int. J. Rock Mech. Min. Sci.
(2004)
Deformation and permeability evolution of coals considering the effect of beddings
Int. J. Rock Mech. Min. Sci.
Permeability evolution of anthracite coal considering true triaxial stress conditions and structural anisotropy
J. Nat. Gas Sci. Eng.
Evolving neural network using a genetic algorithm for predicting the deformation modulus of rock masses
Int. J. Rock Mech. Min. Sci.
Prediction split tensile strength and water permeability of high strength concrete containing TiO2 nanoparticles by artificial neural network and genetic programming
Compos. B Eng.
Prediction of tunneling-induced ground movement with the multi-layer perceptron
Tunn. Undergr. Space Technol.
Prediction of compressive strength of concrete by neural networks
Cement Concr. Res.
Comparison of artificial neural network and fuzzy logic models for prediction of long-term compressive strength of silica fume concrete
Adv. Eng. Software
An ANN application for water quality forecasting
Mar. Pollut. Bull.
Computed tomography analysis on cyclic fatigue and famage properties of rock salt under gas pressure
Int. J. Fatig.
Creating artificial neural networks that generalize
Neural Network.
Estimation of rock modulus: for intact rocks with an artificial neural network and for rock masses with a new empirical equation
Int. J. Rock Mech. Min. Sci.
The threshold pressure gradient effect in the tight sandstone gas reservoirs with high water saturation
Fuel
Predicting intact rock strength for mechanical excavation using multivariate statistics, artificial neural networks, and regression trees
Eng. Geol.
Tight gas sandstone reservoirs in China: characteristics and recognition criteria
J. Petrol. Sci. Eng.
Genetic Programming: an Introduction: on the Automatic Evolution of Computer Programs and its Applications
Stream temperature modelling using artificial neural networks: application on Catamaran Brook, New Brunswick, Canada
Hydrol. Process.
Tight gas in China and its significance in exploration and exploitation
Petrol. Explor. Dev.
Cited by (22)
Genetic programming-based predictive model for the Cr removal effect of in-situ electrokinetic remediation in contaminated soil
2023, Journal of Hazardous MaterialsFormation damage and improved recovery in kaolinitic high enthalpy gas fields with fabric geological settings
2023, Gas Science and EngineeringExperimental study on stress and permeability response with gas depletion in coal seams
2022, Journal of Natural Gas Science and EngineeringCitation Excerpt :The pores and fissures of the two coal specimens are well developed and have a typical double pore structure. The multifunctional true triaxial geophysical (TTG) apparatus, developed by Chongqing University, was utilized in this true triaxial stress and permeability response test under gas extraction (Li et al., 2016b; Yu et al., 2020b), as shown in Fig. 2. The system can provide a load of 6000 kN in two directions, 4000 kN in one direction, and a maximum fluid pressure of 60 MPa.
Modeling of true triaxial strength of rocks based on optimized genetic programming
2022, Applied Soft ComputingCitation Excerpt :The above findings showed that the optimized GP prediction model has better performance than the established multiple regression model. In our previous study, the multiple regression model of permeability prediction was more accurate than GP [12], but after improving the GP function, the optimized GP model was more accurate. This confirmed the effectiveness of our method: dynamically limiting the size of individuals, locally searching the space near the optimal individual, and using multithreaded evaluation effectively improved the algorithm running speed and increased the accuracy of the results.
Data-driven estimation for permeability of simplex pore-throat reservoirs via an improved light gradient boosting machine: A demonstration of sand-mud profile, Ordos Basin, northern China
2022, Journal of Petroleum Science and EngineeringCitation Excerpt :ML (machine learning) under modern conditions has displayed its outstanding capability in solving fitting issues. As the mapping relationship between independent and dependent variables can be analyzed in an implicit pattern rather than a certain expression by ML, this technique is welcomed and preferred to be employed for the nonlinear problems (Zhang et al., 2018; Yu et al., 2020). The function ML presents meets the need of permeability prediction, hence arousing a rapid development of corresponding researches (Akande et al., 2016; Gu et al., 2021).
Experimental study on deformation and fracture characteristics of coal under different true triaxial hydraulic fracture schemes
2022, Journal of Petroleum Science and Engineering