A rock physics and deep learning dual-driven reservoir physical property parameter intelligent prediction method

Abstract

The application discloses a rock physics and deep learning dual driving reservoir physical property parameter intelligent prediction method, and belongs to the technical field of unconventional reservoir seismic exploration, and comprises the following steps: constructing a rock physics model based on physical property parameters and elastic parameter observation data of calibrated logging; estimating the probability distribution of the physical property parameters and the elastic parameters respectively by using a Gaussian mixture model, generating physical property parameter unlabeled data and elastic parameter unlabeled data to form a labeled data set; building a residual convolutional neural network ResNet model, and training an initial ResNet model by using the labeled data; inputting the elastic parameter unlabeled data into the ResNet model to form a full pseudo-labeled data set; iteratively training the initial ResNet model to obtain a final ResNet model; and inputting observation data to obtain predicted reservoir physical property parameters. The application is driven by rock physics and deep learning, combines the advantages of geological priori and data driving, and improves the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of seismic exploration technology for unconventional reservoirs, specifically to an intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning. Background Technology

[0002] Accurate prediction of reservoir physical properties (such as porosity, water saturation, and clay content) is crucial for oil and gas exploration. These parameters not only help predict and identify potential reservoirs, but also play an important role in reservoir production prediction, fluid identification, and oil production analysis in the petroleum industry.

[0003] However, for unconventional reservoirs such as tight sandstone, their complex pore structures pose a significant challenge to the accurate prediction of reservoir physical parameters. Tight sandstone and other unconventional reservoirs typically exhibit complex pore structures, leading to a complex mapping relationship between reservoir physical properties and elastic characteristics. Conventional rock physics models struggle to accurately describe the nonlinear relationship between physical properties and elastic parameters, hindering accurate prediction of physical parameters using classical rock physics modeling and inversion methods. Furthermore, while deep learning methods have made significant progress in multiple fields, automatically extracting complex features from data and possessing strong nonlinear fitting capabilities, providing new methods for predicting reservoir physical parameters, conventional deep learning methods lack utilization of prior geological information when processing geological data. Additionally, the limited availability of well logging observation data for unconventional reservoirs makes it difficult to provide sufficient labeled samples of physical and elastic parameters. This limits the physical interpretability and generalization ability of neural network models, restricting the applicability of conventional deep learning methods for predicting unconventional reservoir physical parameters. Summary of the Invention

[0004] Technical problem solved: To address the challenges in predicting reservoir physical parameters, this invention provides an intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning. By combining the advantages of geological prior knowledge and data-driven approaches, the method improves prediction accuracy.

[0005] Technical solution: The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning, as described in this invention, includes the following steps:

[0006] Step 1: Based on the physical property parameters and elastic parameter observation data of the calibration well logging, construct a rock physics model, predict the elastic parameters, and match them with the well logging elastic parameter observation data;

[0007] Step 2: Based on the physical and elastic parameters observed from the calibration logging, the probability distributions of the physical and elastic parameters are estimated using a Gaussian mixture model.

[0008] Step 3: Based on the probability distribution of physical property parameters and elastic parameters, generate unlabeled data of physical property parameters and elastic parameters through Monte Carlo random simulation, and combine them with well logging observation data to form a labeled dataset;

[0009] Step 4: Build a ResNet model containing convolutional layers, pooling layers, fully connected layers, and residual modules, and determine its hyperparameters;

[0010] Step 5: Train the ResNet model using labeled data to build the initial ResNet model;

[0011] Step 6: Input the unlabeled elastic parameter data into the initial ResNet model to generate physical property parameter data, forming the first part of pseudo-label data;

[0012] Step 7: For the unlabeled physical property data, predict the elastic parameter data through the rock physics model to form the second part of pseudo-labeled data, and combine it with the first part of pseudo-labeled data to form a full pseudo-labeled dataset;

[0013] Step 8: Merge the labeled dataset and the fully pseudo-labeled dataset, and iteratively train the initial ResNet model to obtain the final ResNet model;

[0014] Step 9: Input the pre-stack seismic inversion or well logging elastic parameter observation data into the final ResNet model to predict reservoir physical parameters.

[0015] Preferably, step 1 includes the following sub-steps:

[0016] Step 11: Calculate the bulk modulus K of the rock matrix using the VRH average model. m and shear modulus μ m The calculation formula is as follows:

[0017] ;

[0018] ;

[0019] Where: K mi μ mi v mi These represent the bulk modulus, shear modulus, and percentage content of the i-th mineral in the rock matrix, respectively, where n is the total number of mineral types.

[0020] Step 12: Calculate the bulk modulus K of the pore fluid using the Voigt model. fl The calculation formula is as follows:

[0021] ;

[0022] In the formula: Sw K represents water saturation. w K g These are the bulk moduli of liquid and gas in the pores, respectively.

[0023] Step 13: Calculate the bulk modulus K of the dry rock skeleton using the Kuster-Toksoz model. dry and shear modulus μ dry Its calculation formula is

[0024] ;

[0025] ;

[0026] Where: K * and μ * These represent the bulk modulus and shear modulus of the material, respectively; v is the pore volume fraction, α is the pore aspect ratio, and T... ijij and T iijj These are the pore structure factors;

[0027] Step 14: Calculate the bulk modulus K of fluid-saturated rock using the Biot-Gassmann equation. sat and shear modulus μ sat The calculation formula is as follows:

[0028] ;

[0029] ;

[0030] In the formula: ϕ Porosity;

[0031] Step 15: Based on K sat μ sat And the calculation of P-wave velocity V based on density ρ p Shear wave velocity V s The calculation formula is as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] In the formula: ρ fl and ρ m It is the density of the fluid and rock matrix.

[0036] Preferably, the probability distribution function of the Gaussian mixture model in step 2 is:

[0037] ;

[0038] In the formula: m is a physical property parameter or elastic parameter; N c ω represents the number of Gaussian components. k As weight; μ is the probability density function of the i-th Gaussian component. k and Let be the mean and covariance matrix.

[0039] Preferably, the unlabeled physical property data Y generated by the Monte Carlo random simulation in step 3 is... * And elasticity parameter unlabeled data X * It satisfies the probability distribution of a Gaussian mixture model.

[0040] Preferably, the ResNet model building process in step 4 is as follows:

[0041] ;

[0042] Here, Conv is a convolutional layer used to extract features; ReLU is an activation function used to insert non-linearity into the computation; Pool is a pooling layer used to reduce the spatial dimension of features and reduce computational cost; ResidualBlock is a combination of residual blocks used to solve the gradient vanishing problem in deep networks; FC is a fully connected layer used to perform classification or regression tasks.

[0043] Wherein: the structure of the residual block satisfies:

[0044] ;

[0045] In the formula: y represents the output of the convolutional layer and activation function operations in the residual block, where x is the input and y is the output.

[0046] Preferably, in step 5, an initial ResNet model is constructed, the expression of which is:

[0047] ;

[0048] In the formula: These are the parameters of the network structure; The training process for a residual convolutional neural network model. for Each piece of labeled data.

[0049] Preferably, in step 6, the first part of the pseudo-label data The expression is:

[0050] ;

[0051] In the formula: This is the initial ResNet model; For i elasticity parameters, there are no labels. This represents the prediction results of the initial ResNet model.

[0052] Preferably, in step 7, the second part of the pseudo-label data The expression is:

[0053] ;

[0054] In the formula: RPM is the rock physics model; For k unlabeled physical property parameters; These are parameters for the rock physics model; The results are predictions from a rock physics model;

[0055] Fully pseudo-labeled dataset The expression is:

[0056] ;

[0057] In the formula: This is the first part of the pseudo-label data; This is the second part of the pseudo-label data; the addition operation represents concatenation.

[0058] Preferably, the expression for merging the datasets in step 8 is:

[0059] ;

[0060] In the formula: θ * These are the parameters for the final network structure; ResNet train This describes the training process of a residual convolutional neural network model, where (X,Y) is a labeled dataset.

[0061] Preferably, the expression for the reservoir physical property parameters predicted in step 9 is:

[0062] ;

[0063] In the formula: This represents the final residual convolutional neural network model; x represents the observed elastic parameters including P-wave velocity, S-wave velocity, and density. These are the predicted reservoir physical properties, including porosity, water saturation, and clay content.

[0064] Compared with the prior art, the present invention has at least the following beneficial effects:

[0065] 1. This invention combines rock physics models with deep learning technology to form a dual-driven method of rock physics and deep learning. It makes full use of prior geological knowledge to improve physical interpretability, supplements labeled data to improve generalization ability, and also leverages the nonlinear fitting advantages of deep learning, thereby achieving more accurate and reliable prediction of reservoir physical parameters and providing technical support for oil and gas exploration and development.

[0066] 2. This invention establishes a forward model that links reservoir physical parameters with seismic response by combining rock physics models and seismic reflectivity equations. It directly estimates the physical parameters of tight sandstone reservoirs from seismic data, and more accurately describes the nonlinear relationship between physical parameters and elastic parameters.

[0067] 3. In processing geological data, this invention uses a Gaussian mixture model as a prior model to handle the complex prior distribution of reservoir physical parameters; at the same time, it adds a pseudo-label dataset to train the model, which makes up for the limited label data of physical property-elastic parameters and improves the physical interpretability and generalization ability of the deep learning model. Attached Figure Description

[0068] Figure 1 This is a flowchart of the intelligent prediction process for reservoir physical parameters according to the present invention;

[0069] Figure 2 The data curves from well logging observations in this embodiment of the invention are shown in the graphs ((a) porosity; (b) water saturation; (c) clay content; (d) P-wave velocity; (e) S-wave velocity; (f) density).

[0070] Figure 3 The following diagram shows the predicted physical property parameters of the initial residual convolutional neural network model in the embodiments of this application ((a) porosity; (b) water saturation; (c) clay content).

[0071] Figure 4 The following diagram shows the predicted physical property parameters of the final residual convolutional neural network model in the embodiments of this application ((a) porosity; (b) water saturation; (c) clay content). Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will be described in conjunction with the accompanying drawings. Figures 1-4 The technical solutions of the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0073] like Figure 1As shown, this invention discloses an intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning, comprising the following steps:

[0074] I. Based on the physical and elastic parameters observed from the calibration logging, a rock physics model is constructed to predict elastic parameters and match them with the observed elastic parameters from the logging.

[0075] (1) Set the basic rock physical parameters, including the bulk modulus, shear modulus, and density of minerals, and the bulk modulus and density of fluids; use the VRH averaging model to calculate the bulk modulus K of the rock matrix. m and shear modulus μ m The calculation formula is as follows:

[0076] ;

[0077] ;

[0078] Where: K mi μ mi v mi denoted as the bulk modulus, shear modulus, and percentage content of the i-th mineral in the rock matrix, respectively, where n is the total number of mineral types.

[0079] (2) The bulk modulus K of the pore fluid was calculated using the Voigt model. fl The calculation formula is as follows:

[0080] ;

[0081] In the formula: S w K represents water saturation. w K g These are the bulk moduli of liquid and gas in the pores, respectively.

[0082] (3) The bulk modulus K of the dry rock skeleton was calculated using the Kuster-Toksoz model. dry and shear modulus μ dry Its calculation formula is

[0083] ;

[0084] ;

[0085] Where: K * and μ * These represent the bulk modulus and shear modulus of the material, respectively; v is the pore volume fraction, α is the pore aspect ratio, and T... ijij and T iijj These are the pore structure factors.

[0086] (4) The bulk modulus K of fluid-saturated rock was calculated using the Biot-Gassmann equation. sat and shear modulus μ sat The calculation formula is as follows:

[0087] ;

[0088] ;

[0089] In the formula: Porosity.

[0090] (5) Based on K sat μ sat And the calculation of P-wave velocity V based on density ρ p Shear wave velocity V s The calculation formula is as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] In the formula: ρ fl and ρ m It is the density of the fluid and rock matrix.

[0095] In this embodiment of the invention, the rock physical parameters used are as follows: , , , , , , , , , .

[0096] 2. Based on the physical and elastic parameters observed from the calibration logging, the probability distributions of the physical and elastic parameters are estimated using a Gaussian mixture model.

[0097] The probability distribution function of the Gaussian mixture model is:

[0098] ;

[0099] In the formula: m is a physical property parameter or elastic parameter; N c ω represents the number of Gaussian components. k As weight; μ is the probability density function of the i-th Gaussian component. k and Let be the mean and covariance matrix.

[0100] III. Based on the probability distributions of physical property parameters and elasticity parameters, unlabeled data for physical property parameters and elasticity parameters are generated through Monte Carlo random simulation. The unlabeled data Y for physical property parameters generated by Monte Carlo random simulation is shown below. * And elasticity parameter unlabeled data X * The probability distribution satisfies a Gaussian mixture model. A labeled dataset (X,Y) is formed based on the well logging physical property parameter observation data Y and the elastic parameter observation data X.

[0101] In this embodiment of the invention, the number N of Gaussian components c =3, the total number of Monte Carlo samples is 5000.

[0102] IV. Construct a ResNet model containing convolutional layers, pooling layers, fully connected layers, and residual modules, and determine its hyperparameters.

[0103] (1) The process of building a ResNet model is as follows:

[0104] ;

[0105] Here, Conv is a convolutional layer used to extract features; ReLU is an activation function used to insert non-linearity into the computation; Pool is a pooling layer used to reduce the spatial dimension of features and reduce computational cost; ResidualBlock is a combination of residual blocks used to solve the gradient vanishing problem in deep networks; and FC is a fully connected layer used to perform classification or regression tasks.

[0106] (2) The structure of the residual block satisfies:

[0107] ;

[0108] In the formula: y represents the output of the convolutional layer and activation function operations in the residual block, representing the residual learned by the network. x is the input and y is the output.

[0109] 5. Train the ResNet model using labeled data to construct the initial ResNet model, whose expression is:

[0110] ;

[0111] In the formula: These are the parameters of the network structure; The training process for a residual convolutional neural network model. There are i labeled data points.

[0112] In this embodiment of the invention, the parameters for constructing the residual convolutional neural network model are as follows: the convolutional kernel size is 3, the number of input and output channels is the same, the stride is 1, the padding is 1, and a ReLU activation function is applied afterward; the output of the residual block is directly added to the output of the convolutional layer to form a residual learning mechanism; and a fully connected layer is connected at the end of the network to compress the number of channels to the target dimension.

[0113] In this embodiment of the invention, the following methods are adopted: Figure 2 The well logging data shown is used as labeled data to train the residual convolutional neural network model, thus constructing the initial ResNet model.

[0114] 6. Input the unlabeled elastic parameter data into the initial ResNet model to generate physical property parameter data, forming the first part of pseudo-labeled data; the first part of pseudo-labeled data The expression is:

[0115] ;

[0116] In the formula: This is the initial ResNet model; For i elasticity parameters, there are no labels. This represents the prediction results of the initial ResNet model.

[0117] 7. For unlabeled physical property data, elastic parameter data is predicted using a rock physics model to form a second part of pseudo-labeled data, which is then combined with the first part of pseudo-labeled data to form a full pseudo-labeled dataset.

[0118] (1) The second part of pseudo-label data The expression is:

[0119] ;

[0120] In the formula: RPM is the rock physics model; For k unlabeled physical property parameters; These are parameters for the rock physics model; These are predictions from a rock physics model.

[0121] (2) Fully pseudo-labeled dataset The expression is:

[0122] ;

[0123] In the formula: This is the first part of the pseudo-label data; This is the second part of the pseudo-label data; the addition operation represents concatenation.

[0124] 8. By fusing the labeled dataset and the fully pseudo-labeled dataset, the initial ResNet model is iteratively trained to obtain the final ResNet model, whose expression is:

[0125] ;

[0126] In the formula: θ * These are the parameters for the final network structure; ResNet train This describes the training process of a residual convolutional neural network model, where (X,Y) is a labeled dataset.

[0127] 9. Input the pre-stack seismic inversion or well logging elastic parameter observation data into the final ResNet model to predict reservoir physical parameters, the expression of which is:

[0128] ;

[0129] In the formula: y represents the final residual convolutional neural network model; x represents the observed data of elastic parameters including P-wave velocity, S-wave velocity, and density; y represents the predicted reservoir physical properties including porosity, water saturation, and clay content.

[0130] In this embodiment of the invention, Figure 2 The data curves from well logging observations are shown in the following figures: (a) porosity; (b) water saturation; (c) clay content; (d) P-wave velocity; (e) S-wave velocity; (f) density. Figure 3 The prediction results of the physical property parameters of the initial residual convolutional neural network model are compared with the well logging observation data ((a) porosity; (b) water saturation; (c) clay content). Although the trend of the predicted physical property parameters is basically consistent with the observation data, there is still a certain deviation in the values ​​(e.g., water saturation). Figure 4A comparison of the final residual convolutional neural network model's predicted physical property parameters with well logging observation data ((a) porosity; (b) water saturation; (c) clay content) shows that the predicted physical property parameters (red dashed line) of the final residual convolutional neural network model have good consistency with the observed data (blue solid line). Therefore, this invention, by combining a rock physics model and a seismic reflectivity equation, establishes a forward model that links reservoir physical property parameters with seismic response, directly estimating the physical property parameters of tight sandstone reservoirs from seismic data, and more accurately describing the nonlinear relationship between physical property parameters and elastic parameters. When processing geological data, a Gaussian mixture model is used as a prior model to handle the complex prior distribution of reservoir physical property parameters. Simultaneously, a pseudo-labeled dataset is added for model training to compensate for the limited label data of physical property-elastic parameters, improving the physical interpretability and generalization ability of the deep learning model. This can effectively improve the accuracy of reservoir physical property parameter prediction and is expected to play an important role in predicting complex reservoir physical property parameters, providing technical support for oil and gas exploration and development.

[0131] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of reservoir physical parameters driven by both rock physics and deep learning, characterized in that, Includes the following steps: Step 1: Based on the physical and elastic parameter observation data from the calibration logging, construct a rock physics model to predict elastic parameters and match them with the well logging elastic parameter observation data; the rock physics model specifically includes the following sub-steps: Step 11: Calculate the bulk modulus K of the rock matrix using the VRH average model. m and shear modulus μ m The calculation formula is as follows: ； ； Where: K mi μ mi v mi These represent the bulk modulus, shear modulus, and percentage content of the i-th mineral in the rock matrix, respectively, where n is the total number of mineral types. Step 12: Calculate the bulk modulus K of the pore fluid using the Voigt model. fl The calculation formula is as follows: ； In the formula: S w K represents water saturation. w K g These are the bulk moduli of liquid and gas in the pores, respectively. Step 13: Calculate the bulk modulus K of the dry rock skeleton using the Kuster-Toksoz model. dry and shear modulus μ dry Its calculation formula is ； ； Where: K * and μ * These represent the bulk modulus and shear modulus of the material, respectively; v is the pore volume fraction, α is the pore aspect ratio, and T... ijij and T iijj These are the pore structure factors; Step 14: Calculate the bulk modulus K of fluid-saturated rock using the Biot-Gassmann equation. sat and shear modulus μ sat The calculation formula is as follows: ； ； In the formula: ϕ is porosity; Step 15: Based on K sat μ sat And the calculation of P-wave velocity V based on density ρ p Shear wave velocity V s The calculation formula is as follows: ； ； ； In the formula: ρ fl and ρ m It is the density of the fluid and the rock matrix; Step 2: Based on the physical and elastic parameters observed from the calibration logging, the probability distributions of the physical and elastic parameters are estimated using a Gaussian mixture model. Step 3: Based on the probability distribution of physical property parameters and elastic parameters, generate unlabeled data of physical property parameters and elastic parameters through Monte Carlo random simulation, and combine them with well logging observation data to form a labeled dataset; Step 4: Build a ResNet model containing convolutional layers, pooling layers, fully connected layers, and residual modules, and determine its hyperparameters; Step 5: Train the ResNet model using labeled data to build the initial ResNet model; Step 6: Input the unlabeled elastic parameter data into the initial ResNet model to generate physical property parameter data, forming the first part of pseudo-label data; Step 7: For the unlabeled physical property data, predict the elastic parameter data through the rock physics model to form the second part of pseudo-labeled data, and combine it with the first part of pseudo-labeled data to form a full pseudo-labeled dataset; Step 8: Merge the labeled dataset and the fully pseudo-labeled dataset, and iteratively train the initial ResNet model to obtain the final ResNet model; Step 9: Input the pre-stack seismic inversion or well logging elastic parameter observation data into the final ResNet model to predict reservoir physical parameters.

2. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 1, characterized in that, The probability distribution function of the Gaussian mixture model in step 2 is: ； In the formula: m is a physical property parameter or elastic parameter; N c ω represents the number of Gaussian components. k As weight; μ is the probability density function of the i-th Gaussian component. k Harmony Mean and covariance matrices.

3. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 2, characterized in that, In step 3, the unlabeled data Y of the physical property parameters generated by the Monte Carlo random simulation * And elasticity parameter unlabeled data X * It satisfies the probability distribution of a Gaussian mixture model.

4. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 3, characterized in that, The process of building the ResNet model in step 4 is as follows: ； Here, Conv is a convolutional layer used to extract features; ReLU is an activation function used to insert non-linearity into the computation; Pool is a pooling layer used to reduce the spatial dimension of features and reduce computational cost; ResidualBlock is a combination of residual blocks used to solve the gradient vanishing problem in deep networks; FC is a fully connected layer used to perform classification or regression tasks. Wherein: the structure of the residual block satisfies: ； In the formula: W is the output of the convolutional layer and activation function operations in the residual block. i X represents the residual of network learning, where X is the input and Y is the output.

5. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 4, characterized in that, Step 5 constructs the initial ResNet model, whose expression is: ； In the formula: These are the parameters of the network structure; The training process for a residual convolutional neural network model. This is labeled data.

6. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 5, characterized in that, The first part of the pseudo-label data in step 6 The expression is: ； In the formula: This is the initial ResNet model; For an elastic parameter without label data, This represents the prediction results of the initial ResNet model.

7. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 6, characterized in that, The second part of the pseudo-label data in step 7 The expression is: ； In the formula: RPM is the rock physics model; For k unlabeled physical property parameters; These are parameters for the rock physics model; The results are predictions from a rock physics model; Fully pseudo-labeled dataset The expression is: ； In the formula: This is the first part of the pseudo-label data; This is the second part of the pseudo-label data; the addition operation represents concatenation.

8. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 7, characterized in that, The expression for merging the datasets in step 8 is: ； In the formula: θ * These are the parameters for the final network structure; ResNet train This describes the training process of a residual convolutional neural network model, where (X,Y) is a labeled dataset.

9. The intelligent prediction method for reservoir physical parameters driven by both rock physics and deep learning according to claim 8, characterized in that, The expression for the predicted reservoir physical property parameters in step 9 is: ； In the formula: y represents the final residual convolutional neural network model; x represents the observed data of elastic parameters including P-wave velocity, S-wave velocity, and density; y represents the predicted reservoir physical properties including porosity, water saturation, and clay content.

Images