Shale oil lithofacies combination seismic phased inversion method based on deep learning
By employing a deep learning-based seismic facies inversion method for shale oil formations, the problem of identifying thin interlayers in shale oil formations in the Dongying Depression was solved, achieving high vertical and horizontal resolution shale oil reservoir prediction and improving the accuracy of reservoir description.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA PETROLEUM & CHEMICAL CORP
- Filing Date
- 2024-11-12
- Publication Date
- 2026-05-12
AI Technical Summary
Existing seismic inversion techniques are ineffective in identifying thin interlayers and highly heterogeneous shale oil reservoirs in the Dongying Depression shale oil formation. Conventional deterministic inversion methods have low vertical resolution, while stochastic inversion methods have low lateral resolution. Deep learning-based methods have high vertical resolution but insufficient lateral resolution.
A deep learning-based seismic facies inversion method for shale oil facies combinations was adopted. By preprocessing pre-stack angle gathers and well logging data, K-Means clustering algorithm was used to divide seismic facies, a three-dimensional convolutional deep learning model based on spatiotemporal attention mechanism was established, and the position coding theory in natural language processing was introduced. The seismic facies inversion deep learning model was trained by combining semi-supervised learning method to obtain the inversion results of P-wave and S-wave velocities and densities.
It improves the vertical and horizontal resolution of seismic inversion, and the thin layer thickness is less than one-quarter of the seismic wave wavelength, enabling more accurate prediction of the spatial distribution of shale oil reservoirs.
Smart Images
Figure CN122017958A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of petroleum geophysical exploration, and in particular to a deep learning-based seismic facies inversion method for shale oil facies combinations. Background Technology
[0002] With the continuous advancement of oil exploration and development technologies, the accuracy requirements for reservoir prediction are also gradually increasing. Seismic inversion technology is a key technology for describing underground reservoirs. By converting dimensionless seismic data into physically meaningful elastic parameter data through well logging data and other constraints, and even, under reasonable assumptions, ultimately into reservoir physical property parameter data, the Dongying Depression shale oil formation has the characteristics of thin interlayers and strong heterogeneity. Existing deterministic and stochastic seismic inversion technologies have difficulty identifying advantageous shale oil reservoirs.
[0003] Chinese patent application CN202110998575.7 discloses a deep learning-based seismic impedance inversion method driven by both data and intelligent optimization. It describes a seismic inversion method based on a deep learning network: a global optimization method is used to invert the impedance of a portion of post-stack seismic data, and the resulting data is used to pre-train a deep learning network to learn the mapping relationship from seismic data to impedance. The pre-trained network guides the global optimization method to invert the impedance data of another portion of the seismic data, accelerating its convergence to the optimal solution. The obtained optimal solution is then used to fine-tune the deep learning network. The fine-tuned deep learning network is then used to efficiently invert the impedance model of large-scale 3D seismic data. This invention significantly improves the inversion efficiency of impedance models, achieving a major breakthrough by enabling the application of global optimization methods in large-scale impedance inversion problems while maintaining acceptable computational costs. However, this patent only demonstrates the powerful nonlinear fitting ability of deep learning networks. A significant problem with using deep learning networks for seismic inversion is the limited number of well-seismic matching labels; therefore, research on the generalization ability of deep learning networks should be strengthened.
[0004] Chinese patent application CN202211206119.5 discloses a method and system for acoustic impedance inversion based on well-controlled semi-supervised deep learning. It describes a seismic inversion method based on a deep learning network: for seismic data with corresponding well logging curves, the predicted acoustic impedance is compared with the actual acoustic impedance using a mean square error calculation to obtain the well logging loss; for seismic data without corresponding well logging curves, the predicted acoustic impedance is used to obtain a synthetic seismic record through a seismic convolution model, and then compared with the actual seismic data using a mean square error calculation to obtain the seismic loss. The well logging loss and the seismic loss are weighted and summed to obtain the training loss, and the artificial neural network updates the model parameters by optimizing the training loss. This invention is a deep learning-based acoustic impedance inversion method. Compared with traditional methods, it requires less human-computer interaction, has a higher level of intelligence, and is also a semi-supervised learning method, eliminating the need for training the artificial neural network with large amounts of data. Furthermore, it incorporates a geophysical forward model, ensuring that the acoustic impedance inversion results conform to geophysical laws and are relatively reasonable. However, when applying deep learning networks to invert seismic impedance information, the generalization of deep learning networks only adopts a semi-supervised learning method, resulting in low lateral resolution of the inversion results of the deep learning network seismic inversion method.
[0005] Chinese patent application CN201310473076.1 discloses a phase-controlled seismic inversion method in geophysical exploration, mentioning the role of seismic phase control in seismic inversion: using seismic data and well logging impedance curves, under the control of seismic and geological facies results, reservoir impedance data is inverted. This method can improve vertical resolution while maintaining lateral resolution and can finely characterize the spatial distribution of reservoirs. However, this patented inversion method uses Gaussian co-simulation to retrieve seismic impedance data. This method is heavily constrained by prior model rules, and the inversion results have strong lateral randomness, leading to reduced reliability of the inversion results when the number of wells is small.
[0006] Chinese patent application CN202210625625.1 discloses a pre-stack phasing inversion method and apparatus for reservoirs, mentioning the role of seismic phasing in seismic inversion: by using phasing interpolation technology, multiple attributes and sedimentary facies are incorporated into the establishment of the initial model for pre-stack inversion, resulting in an initial model body with elastic parameters conforming to sedimentary laws. This serves as the prior condition input for pre-stack inversion, overcoming the shortcomings of uniform interpolation in traditional modeling methods. By improving the accuracy of the initial input, more reasonable elastic parameter inversion results are obtained, thus improving computational accuracy. The use of a deterministic inversion method ensures stable computational results, compensating for the deficiencies of traditional inversion methods from multiple perspectives. Furthermore, by interpolating multiple parameters, the accuracy and reliability of reservoir prediction are improved, making the results more consistent with geological laws. However, the core algorithm of this patented phasing inversion method is deterministic, resulting in a relatively low vertical resolution of the inversion results.
[0007] Conventional deterministic inversion methods are limited by the resolution of seismic data, resulting in low vertical resolution. While stochastic inversion methods offer higher vertical resolution, their horizontal inversion results exhibit strong randomness. Deep learning-based seismic inversion methods possess strong nonlinear fitting capabilities and high vertical resolution, but their horizontal resolution remains low. To address these challenges, it is necessary to research a seismic inversion method that integrates the advantages of horizontal resolution in phase-controlled inversion with the advantages of vertical resolution in deep learning-based inversion, thus forming a deep learning-based seismic phase-controlled inversion technique for shale oil facies assemblages. Summary of the Invention
[0008] In view of the above problems, the present invention is proposed to provide a deep learning-based seismic phasing inversion method for shale oil facies combinations that overcomes or at least partially solves the above problems.
[0009] According to one aspect of the present invention, a deep learning-based seismic facies fusion inversion method for shale oil is provided, the inversion method comprising:
[0010] Preprocessing of pre-stack angle gathers and well logging data was performed, and the K-Means clustering algorithm was used to divide the seismic data into seismic phases.
[0011] A three-dimensional convolutional deep learning model based on spatiotemporal attention mechanism is established. The positional coding theory in natural language processing is introduced to transform the seismic facies classification results into temporal codes. The temporal codes are then introduced into the three-dimensional convolutional deep learning model to form a seismic facies-controlled deep learning model.
[0012] A semi-supervised learning method was used to train a seismic phased-array deep learning model to obtain the lateral resolution of the inverted P-wave and S-wave velocity and density results.
[0013] Optionally, the preprocessing of pre-stack angle gathers and logging data specifically includes:
[0014] Standardize and normalize the pre-stack angle gathers and logging data.
[0015] Optionally, the standardization and normalization of pre-stack angle gathers and logging data specifically includes:
[0016] Seismic data serves as sample data, while P-wave velocity, S-wave velocity, and density provided by well logging data serve as label data. The StandardScaler and MinMaxScaler functions are used to standardize and normalize the pre-stack angle gathers and well logging data to eliminate the differences between the dimensions of different well logging data and the differences between the amplitudes of different angles in the pre-stack angle gathers.
[0017] The functions StandardScaler and MinMaxScaler are:
[0018]
[0019] Among them, X i For well logging data or pre-stack angle gathers, X m and X σ These are the mean and variance of well logging data and pre-stack angle gathers, respectively; Y i These are the standardized data; Ymin and Ymax are the minimum and maximum values of the standardized data, respectively. i 'This is the result after normalization.'
[0020] Optionally, the step of using the K-Means clustering algorithm to divide seismic data into seismic facies specifically includes:
[0021] Seismic facies data are input into a deep learning model, and pre-stack angle gathers are fully stacked to obtain post-stack seismic data. The K-Means clustering algorithm is then used to divide the seismic data of the target layer into seismic facies.
[0022] Optionally, the conversion of seismic facies classification results into temporal codes using the location coding theory introduced from natural language processing specifically includes:
[0023] In natural language processing, positional encoding is a technique that encodes words or characters in a text sequence to represent their positional information in a deep learning model. Positional encoding helps deep learning models understand the relative positions and order of different elements in a sequence and capture the semantic information in the sequence.
[0024] Different seismic facies have different sedimentary stratigraphic structures. The location coding technique is used to convert the seismic facies classification results into spatial location codes, which are then used as a separate input to the neural network, thus serving as a seismic facies-controlled constrained deep learning model.
[0025] The positional coding theory formula is:
[0026]
[0027] Where pos represents the seismic phase category, i represents the dimension index of the location encoding vector, and d model Represents the embedding dimension of the model; PE m This represents the location code of the m-th seismic trace.
[0028] Optionally, the establishment of the three-dimensional convolutional deep learning model based on the spatiotemporal attention mechanism specifically includes: a three-dimensional convolutional deep learning model based on the spatiotemporal attention mechanism, which contains a total of 9 layers, with the input layer: well logging data, seismic data preprocessing, and seismic phase temporal coding dataset being input into the three-dimensional convolutional deep learning model;
[0029] The first part consists of three parallel convolutional layers, with kernel sizes and numbers of [specific values to be filled in]. and in This indicates a convolutional layer with n kernels, kernel size (x, y, z), where x, y, and z represent the main seismic data line, connecting seismic data line, and depth direction, respectively, and a sliding step size of 1.
[0030] Spatial features at different spatial scales are extracted from seismic data using convolution kernels of different scales. These extracted spatial features are then stitched together and input into the system. To further extract feature information across different scales of seismic data;
[0031] Spatial attention layer, used to improve the sensitivity of deep learning models to the spatial features of seismic data;
[0032] The second part consists of three parallel convolutional layers, with kernel sizes and numbers of [specific values to be filled in]. and Temporal features at different scales are extracted from seismic data using convolutional kernels of different scales. These extracted temporal features are then concatenated and input into a database. To further extract temporal characteristic information from earthquake data;
[0033] Temporal attention layer, used to improve the sensitivity of deep learning models to temporal features;
[0034] Dropout layers disable neurons with probability p during training;
[0035] The fully connected layer has 9 input channels and 3 output channels. It is used to improve the fitting ability of the nonlinear relationship between the input and output layers and to output the prediction results of the 3D convolutional deep learning fusion network.
[0036] Optionally, the seismic phase-controlled deep learning model specifically includes:
[0037] Establish a loss function for a deep learning model based on prior information constraints from well logging and seismic data.
[0038] Optionally, the loss function for establishing the deep learning model based on prior information constraints from well logging and seismic data specifically includes:
[0039] The training backpropagation of the deep learning model is constrained by three parts: well logging data loss, seismic data loss, and low-frequency model data loss.
[0040] The first part of the formula is the loss of elastic parameter data in well logging data. A deep learning model is trained using well logging and seismic data at well points. The input seismic data is used to predict elastic parameter data, and the loss error between the true value and the predicted value of the elastic parameter data at well points is obtained.
[0041] Under the constraint of weight λ1, the middle and high frequency components of well logging data are compensated to seismic data. The seismic inversion technology based on deep learning breaks through the limitation of the effective bandwidth of seismic data, and the inversion result has a high vertical resolution. Moreover, the larger λ1 is, the higher the vertical resolution of the prediction result is.
[0042] The second part of the formula is the loss of seismic data. The convolution model can establish the relationship between rock elastic parameter data and seismic data. For unlabeled seismic data, the convolution model is used to forward model the pre-stack seismic trace gather predicted by the deep learning model for elastic parameter data, and the loss error between the synthetic seismic data and the true seismic data is obtained. The weight λ2 controls the constraint degree of seismic data. The larger λ2 is, the more the lateral characteristics of the prediction result conform to the seismic data.
[0043] The third part of the formula is the loss of low frequency components, which is the loss error between the low frequency components of the prediction result obtained using a low-pass filter and the low frequency components obtained using joint interpolation of seismic and well logging data. It belongs to the constraint of prior information, and inverse distance weighted interpolation of seismic and well logging data is used to obtain the initial value. The weight λ3 controls the constraint degree of low frequency components. The lower the signal-to-noise ratio of seismic data is, the larger λ3 is.
[0044] The objective function of the network is expressed as:
[0045]
[0046] where M and N are the numbers of seismic data and well logging data respectively. During the training process, N << M to meet the situation of insufficient training samples in the actual work area.
[0047] During the inversion process of model data and actual data, the weight coefficients of the objective function are respectively set as
[0048] λ1 = 0.4, λ2 = 0.3, λ3 = 0.3.
[0049] Optionally, the specific steps of training the seismic facies controlled deep learning model using the semi-supervised learning method include:
[0050] Inputting the pre-stack seismic trace gather data, seismic facies data, and well logging data training set into a three-dimensional convolutional deep learning model;
[0051] Training the deep learning model using the loss function, adopting Adam with momentum as the optimization algorithm, and outputting the deep learning model after the loss function tends to be stable.
[0052] Optionally, the inversion method further includes: testing and evaluating the seismic phase-controlled deep learning model.
[0053] Optionally, the testing and evaluation of the seismic phase-controlled deep learning model specifically includes:
[0054] The test set of pre-stack seismic gather data, seismic facies data, and well logging data is input into the trained seismic facies-controlled deep learning model, and the coefficient of determination R is used. 2 Quantitatively evaluate deep learning models;
[0055] Coefficient of determination R 2 Represented as:
[0056]
[0057] Among them, y i Indicates the true label, The mean of the true values. This represents the network's predicted value, where n is the number of samples.
[0058] This invention provides a deep learning-based seismic facies retrieval method for shale oil facies combinations. The method includes: preprocessing pre-stack angle gathers and well logging data; using the K-Means clustering algorithm to classify seismic data into seismic facies; establishing a three-dimensional convolutional deep learning model based on a spatiotemporal attention mechanism; introducing positional coding theory from natural language processing to convert the seismic facies classification results into temporal codes; and incorporating these temporal codes into the three-dimensional convolutional deep learning model to form a seismic facies-controlled deep learning model; training the seismic facies-controlled deep learning model using a semi-supervised learning method to obtain the lateral resolution of the retrieved P-wave and S-wave velocity and density results. The lateral resolution of the P-wave and S-wave velocity and density results obtained by the facies-controlled deep learning inversion is superior to that of the inversion without facies control, and the retrieved thin layer thickness is much less than one-quarter of the seismic wave wavelength.
[0059] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description
[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0061] Figure 1 A flowchart illustrating a deep learning-based seismic facies inversion method for shale oil facies combinations, provided as an embodiment of the present invention;
[0062] Figure 2 A schematic diagram of a geological model of thin interbedded sandstone and mudstone provided in an embodiment of the present invention;
[0063] Figure 3 The filling speed and density of the thin interlayered geological model provided in the embodiments of the present invention;
[0064] Figure 4 Superimposed seismic profiles at 10°, 20° and 30° for thin interbedded sandstone and mudstone provided in embodiments of the present invention;
[0065] Figure 5 The k-means seismic facies clustering results for shale oil facies combinations provided in this embodiment of the invention;
[0066] Figure 6 This is a location coding sequence profile for seismic facies category conversion of shale oil facies assemblage provided in an embodiment of the present invention;
[0067] Figure 7 The inversion results of phase-controlled elastic parameters based on deep learning provided in the embodiments of the present invention;
[0068] Figure 8 The inversion results of phased elastic parameters based on deep learning provided in the embodiments of the present invention;
[0069] Figure 9 The structure of a three-dimensional convolutional deep learning model based on a spatiotemporal attention mechanism provided in this embodiment of the invention;
[0070] Figure 10 The deep learning-based inversion results of phase-controlled elastic parameters (longitudinal wave velocity) provided in this embodiment of the invention;
[0071] Figure 11 The inversion results (transverse wave velocity) of elastic parameters without phase control based on deep learning provided in the embodiments of the present invention;
[0072] Figure 12 The density is the result of inversion of phase-controlled elastic parameters based on deep learning provided in the embodiments of the present invention.
[0073] Figure 13 The inversion results (longitudinal wave velocity) of phased elastic parameters based on deep learning provided in the embodiments of the present invention;
[0074] Figure 14 The inversion results (transverse wave velocity) of phased elastic parameters based on deep learning provided in the embodiments of the present invention;
[0075] Figure 15 The inversion results (density) of phased elastic parameters based on deep learning are provided for embodiments of the present invention. Detailed Implementation
[0076] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0077] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.
[0078] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0079] Example 1
[0080] A deep learning-based seismic facies-based inversion method for shale oil facies combinations includes:
[0081] (1) Data preprocessing: Based on the deep learning-based seismic phased inversion method, seismic data is used as sample data, and P-wave velocity, S-wave velocity, and density provided by well logging data are used as label data. The StandardScaler and MinMaxScaler functions are used to standardize and normalize the pre-stack angle gathers and well logging data to eliminate the differences between different dimensions of well logging data and the differences between different angle amplitudes of pre-stack angle gathers.
[0082] The functions StandardScaler and MinMaxScaler are as follows:
[0083]
[0084] Where X i For well logging data or pre-stack angle gathers, X m and X σ These represent the mean and variance of well logging data and pre-stack angle gathers, respectively. i These are the standardized data, where Ymin and Ymax are the minimum and maximum values of the standardized data, respectively. i 'This is the result after normalization.'
[0085] (2) K-Means clustering algorithm for seismic facies division: Seismic data is an important basis for the division of sedimentary facies between wells. Geologists can directly convert seismic facies into three-dimensional spatial sediments through experience. This shows that strata with similar sedimentary characteristics have similar seismic response characteristics. Therefore, it is proposed to input seismic facies data into a deep learning model. In order to improve the lateral constraint capability of seismic prior information, the pre-stack angle gathers are first fully stacked to obtain post-stack seismic data. Then, the K-Means clustering algorithm is used to divide the seismic data of the target layer into seismic facies.
[0086] (3) Location coding technology converts seismic facies into temporal feature data: In natural language processing, location coding technology encodes words or characters in a text sequence to represent their positional information in a deep learning model. The role of location coding is to help the deep learning model understand the relative positions and order of different elements in the sequence, thereby better capturing the semantic information in the sequence. Different seismic facies have different sedimentary stratigraphic structures. To improve the fitting accuracy of the deep learning model, location coding technology is used to convert the seismic facies classification results into spatial location codes as a separate input to the neural network, which plays the role of seismic facies-controlled constraint of the deep learning model and improves the lateral extrapolation ability of the deep learning model.
[0087] The positional coding theory formula is:
[0088]
[0089] Where pos represents the seismic phase category, i represents the dimension index of the location encoding vector, and d model Represents the embedding dimension of the model; PE m This represents the location code of the m-th seismic trace.
[0090] (4) Establishment of a three-dimensional convolutional deep learning model based on spatiotemporal attention mechanism: Prestack seismic data has spatiotemporal sequence data features. The time sequence features are the features of each seismic data changing over time, representing the geological features of the underground sedimentary strata changing with depth. The spatial sequence change features are the AVO change features of the prestack seismic data, which are the changes in the propagation features of seismic waves in the strata. Different strata have different AVO prestack seismic response features. By inferring the strata features through the AVO prestack seismic response features, in order to improve the fitting ability of the complex nonlinear relationship between input and output data, a three-dimensional convolutional deep learning model based on spatiotemporal attention mechanism is proposed. The model contains a total of 9 layers. Input layer: Well logging data, seismic data preprocessing and seismic phase time-series encoded dataset are input into the STA3DCNN model.
[0091] (2) The first part consists of three parallel convolutional layers, with kernel sizes and numbers of kernels as follows: and ( This indicates a convolutional layer with n kernels and kernel size (x, y, z), where x, y, and z represent the main seismic line, connecting seismic line, and depth direction, respectively, and a sliding step of 1. Spatial features at different spatial scales of the seismic data are extracted using convolutional kernels of different sizes. The extracted spatial features are then stitched together and input into the system. To further extract feature information across different scales of seismic data;
[0092] (3) Spatial attention layer, used to improve the sensitivity of deep learning models to the spatial features of seismic data.
[0093] (4) The second part consists of three parallel convolutional layers, with kernel sizes and numbers of kernels as follows: and Temporal features at different scales are extracted from seismic data using convolutional kernels of different scales. These extracted temporal features are then concatenated and input into a database. To further extract temporal characteristic information from earthquake data;
[0094] (6) Temporal attention layer, used to improve the sensitivity of deep learning models to temporal features; (7) Dropout layer, which disables neurons with a certain probability p during training; (8) Fully connected layer, with 9 input channels and 3 output channels, used to improve the fitting ability of the nonlinear relationship between the input and output layers and output the prediction results (elasticity parameter data) of the STA3DCNN fusion network.
[0095] (5) Establishment of loss function for deep learning model based on prior information constraints of well logging and seismic data: In order to improve the stability and accuracy of prediction of lateral heterogeneity of elastic parameters of shale oil reservoir, three parts are proposed to constrain the training backpropagation of deep learning model: well logging data loss, seismic data loss and low frequency model data loss. The three loss functions fully consider the longitudinal and lateral constraints of prior information of well logging and seismic data, and control the deep learning model to have high longitudinal and lateral resolution.
[0096] The first part is the loss of elastic parameter data in the well logging data (first term of the formula). A deep learning model is trained using well logging and seismic data at the well point. The input seismic data is used to predict the elastic parameter data, resulting in the loss error between the true value (Yreal) and the predicted value (Ypredict) of the elastic parameter data at the well point. Under the constraint of weight λ1, the mid-to-high frequency components of the well logging data are compensated for as seismic data. Therefore, the seismic inversion technique based on deep learning can overcome the limitation of the effective bandwidth of seismic data, and the inversion results have high vertical resolution. Furthermore, the larger λ1 is, the higher the vertical resolution of the prediction results.
[0097] The second part is the loss of seismic data (the second term of the formula). Seismic data is essentially only related to the elastic parameter data of rocks, and the convolution model can establish the relationship between the elastic parameter data of rocks and seismic data. To solve the problem of insufficient samples in the phased seismic inversion technology based on deep learning and enhance the lateral resolution of the prediction results, for the unlabeled seismic data, the convolution model is used to forward model the elastic parameter data predicted by the deep learning model to obtain the pre-stack seismic trace gather (Ssyn), thereby obtaining the loss error between the synthetic seismic data (Ssyn) and the real seismic data (Sreal). The weight λ2 controls the degree of constraint of the seismic data. The larger λ2 is, the more the lateral characteristics of the prediction results conform to the seismic data.
[0098] The third part is the loss of the low-frequency component (the third term of the formula), that is, the loss error between the low-frequency component (Ypredict_filter) of the prediction result obtained using the low-pass filter and the low-frequency component (Yreal_initial) obtained using the joint interpolation of seismic and logging data. This part belongs to the constraint of prior information. Assuming that the sedimentation law of large formations is known, this assumption conforms to the geophysical basic theory assumption. The inverse distance weighted interpolation of seismic and logging data is used to obtain the initial value to improve the generalization ability of the proposed network. The weight λ3 controls the degree of constraint of the low-frequency component. The lower the signal-to-noise ratio of the seismic data, the larger λ3 is.
[0099] The objective function of the proposed network is expressed as:
[0100]
[0101] Where M and N are the numbers of seismic data and logging data respectively. During the training process, N << M to meet the situation of insufficient training samples in the actual work area. During the inversion process of the model data and the actual data, the weight coefficients of the objective function are set to λ1 = 0.4, λ2 = 0.3, and λ3 = 0.3 respectively.
[0102] (6) Deep learning model training: Input the training sets such as pre-stack seismic trace gather data, seismic facies data, and logging data into the deep learning model in step (4), and use the loss function in step (5) to train the deep learning model. Use Adam with momentum as the optimization algorithm and output the deep learning model after the loss function tends to be stable.
[0103] (7) Deep learning model testing and evaluation: Input the test sets such as pre-stack seismic trace gather data, seismic facies data, and logging data into the well-trained deep learning model in step (6), and quantitatively evaluate the deep learning model using the coefficient of determination (R2).
[0104] The coefficient of determination can be expressed as:
[0105]
[0106] Among them, y i Indicates the true label, The mean of the true values. This represents the network's predicted value, where n is the number of samples.
[0107] Example 2
[0108] Thin interbedded geological models with different sedimentary characteristics (such as...) were established. Figure 2 As shown in the figure, the sandstone (H1, H3, H4, H5, H6, H7, H8) and the dark colors represent mudstone (H0, H2, H9).
[0109] The geological model includes 150 CDPs (Common Depth Point Gathers), with 200 sampling points and a sampling interval of 0.5 ms. The P-wave velocity, S-wave velocity, and density of each layer in the thin interbedded geological model are as follows: Figure 3 As shown, the reflection coefficient is calculated using the Aki approximation formula and synthesized with 30Hz Ricker wavelet convolution to form pre-stack seismic data (e.g., Figure 4 (As shown). Three wells were established on the geological model, representing constrained wells in different facies zones, and named w25, w75, and w125 respectively, in conjunction with specific example data.
[0110] like Figure 1 As shown, a deep learning-based seismic facies inversion method for shale oil facies combinations includes:
[0111] (1) Data preprocessing: standardize and normalize seismic and well logging data to eliminate the gap between different data units.
[0112] (2) Seismic facies division: The K-Means clustering algorithm is used to analyze the seismic data of the target layer (e.g., Figure 4 (As shown) Seismic facies classification of shale oil rock facies assemblage. Figure 5 The results show the seismic facies classification of shale oil rock facies assemblage.
[0113] (3) Location coding is introduced to encode the seismic facies types of shale oil facies combinations. This is used to map discrete words or tags to continuous vector representations, thus solving the coding problem of sequential data. Drawing on absolute location coding in the Transformer model, spatial location coding is introduced into the geophysical field. Sine and cosine functions are used to generate spatial location codes for the divided shale oil facies combination seismic facies categories, which serve as separate inputs to the neural network, thereby playing a role in facies control constraints. Figure 6 This is a location-coded sequence profile for seismic facies category conversion of shale oil facies assemblage.
[0114] The positional coding theory formula is expressed as:
[0115]
[0116] Among them, pos represents the category of seismic facies, i represents the dimension index of the position encoding vector, and d model represents the embedding dimension of the model; PE m represents the position encoding of the m-th seismic trace.
[0117] (4) Establishment of a deep learning network: A three-dimensional convolutional deep learning model based on a spatio-temporal attention mechanism is established. This network consists of an input layer, a three-dimensional convolutional layer, a spatio-temporal attention mechanism layer, a fully connected layer, and an output layer. This deep learning model takes into account the spatio-temporal characteristics of seismic data and can establish complex non-linear relationships between input and output data.
[0118] (5) Establishment of the loss function of the deep learning model constrained by prior information of well logging and seismic data: To improve the prediction stability and accuracy of the lateral heterogeneity of elastic parameters in shale oil reservoirs, three parts of well logging data loss, seismic data loss, and low-frequency model data loss are proposed to constrain the training backpropagation of the deep learning model. The three loss functions fully consider the vertical and horizontal constraints of the prior information of well logging and seismic data, and control the deep learning model to have high vertical and horizontal resolutions.
[0119] The objective function of the network is expressed as:
[0120]
[0121] Among them, M and N are the numbers of seismic data and well logging data respectively. During the training process, N << M to meet the situation of insufficient training samples in the actual work area. In the inversion process of model data and actual data, the weight coefficients of the objective function are set to λ1 = 0.4, λ2 = 0.3, and λ3 = 0.3 respectively.
[0122] (6) Training and evaluation of the deep learning network: According to the synthetic pre-stack seismic trace set data, position encoding data, P-wave and S-wave velocities, and density elastic parameters, Adam with momentum is used as the optimization algorithm. The deep learning network is trained using the loss function in step (5). After the loss function tends to be stable, the deep learning network model is output, and the inversion result of the deep learning network without phase control is obtained ( Figure 7 ). Although it can invert the elastic parameters of the second facies belt well, the difference between the inversion result of the first facies belt and the elastic parameter information in the well is large. The reason is that the first facies belt and the third facies belt have similar seismic reflection characteristics, but the corresponding label data (elastic parameter information in the well) are quite different. Therefore, the deep learning network cannot establish complex non-linear relationships between the same input and different outputs. The prediction result of the phase-controlled deep learning network is basically consistent with the information in the well ( Figure 8This indicates that adding seismic facies combination seismic phase control to deep learning networks can improve the vertical and horizontal resolution of seismic data inversion results.
[0123] To better explain the implementation process of this patent, actual well logging and seismic data from the Dongying Depression shale oil formation were used to interpret the patent's prediction of favorable shale oil reservoirs. Figure 9 The deep learning model structure is based on practical commonality. The network consists of an input layer, a three-dimensional convolutional layer, a spatiotemporal attention mechanism layer, a fully connected layer, and an output layer. This deep learning model considers the spatiotemporal characteristics of seismic data and can establish complex nonlinear relationships between input and output data. Based on pre-stack seismic gather data, location-coded data, P-wave and S-wave velocities, and density elastic parameters, the momentum-based Adam algorithm is used as the optimization algorithm. The deep learning network is trained using the loss function in step (5). After the loss function tends to stabilize, the deep learning network model is output, and the inversion result of the phase-controlled deep learning network is obtained. Figures 10-12 ), and the prediction results of phased deep learning networks ( Figures 13-15 Compared to other methods, the P-wave velocity, S-wave velocity, and density predicted by the phase-controlled deep learning model are more consistent with the characteristics of underground strata in the horizontal direction, and also more consistent with the characteristics of seismic data. This indicates that adding shale oil facies combination seismic phase control to the deep learning network can improve the vertical and horizontal resolution of seismic data inversion results.
[0124] Beneficial effects: Test results of thin interlayer models and actual data show that blind wells have verified that the lateral resolution of the P-wave and S-wave velocity and density results obtained by phase-controlled deep learning inversion is better than that of the inversion results without phase control. Moreover, the thickness of the inverted thin layer is much less than one-quarter of the seismic wave wavelength. This proves that the seismic phase-controlled inversion technology based on deep learning for shale oil facies combination is a seismic inversion technology with high vertical and lateral resolution. It can predict the spatial distribution of dominant shale oil reservoirs and provide technical support for the three-dimensional exploration and development of shale oil reservoirs.
[0125] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A deep learning-based seismic facies inversion method for shale oil facies combinations, characterized in that, The inversion method includes: Preprocessing of pre-stack angle gathers and well logging data was performed, and the K-Means clustering algorithm was used to divide the seismic data into seismic phases. A three-dimensional convolutional deep learning model based on spatiotemporal attention mechanism is established. The positional coding theory in natural language processing is introduced to transform the seismic facies classification results into temporal codes. The temporal codes are then introduced into the three-dimensional convolutional deep learning model to form a seismic facies-controlled deep learning model. A semi-supervised learning method was used to train a seismic phased-array deep learning model to obtain the lateral resolution of the inverted P-wave and S-wave velocity and density results.
2. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 1, characterized in that, The preprocessing of pre-stack angle gathers and logging data specifically includes: Standardize and normalize the pre-stack angle gathers and logging data.
3. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 2, characterized in that, The standardization and normalization of pre-stack angle gathers and logging data specifically includes: Seismic data serves as sample data, while P-wave velocity, S-wave velocity, and density provided by well logging data serve as label data. The StandardScaler and MinMaxScaler functions are used to standardize and normalize the pre-stack angle gathers and well logging data to eliminate the differences between the dimensions of different well logging data and the differences between the amplitudes of different angles in the pre-stack angle gathers. The functions StandardScaler and MinMaxScaler are: Among them, X i For well logging data or pre-stack angle gathers, X m and X σ These are the mean and variance of well logging data and pre-stack angle gathers, respectively; Y i These are the standardized data; Ymin and Ymax are the minimum and maximum values of the standardized data, respectively. i 'This is the result after normalization.' 4. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 1, characterized in that, The specific steps of using the K-Means clustering algorithm to divide seismic data into seismic facies include: Seismic facies data are input into a deep learning model, and pre-stack angle gathers are fully stacked to obtain post-stack seismic data. The K-Means clustering algorithm is then used to divide the seismic data of the target layer into seismic facies.
5. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 1, characterized in that, The introduction of location coding theory from natural language processing to convert seismic facies classification results into temporal coding specifically includes: In natural language processing, positional encoding is a technique that encodes words or characters in a text sequence to represent their positional information in a deep learning model. Positional encoding helps deep learning models understand the relative positions and order of different elements in a sequence and capture the semantic information in the sequence. Different seismic facies have different sedimentary stratigraphic structures. The location coding technique is used to convert the seismic facies classification results into spatial location codes, which are then used as a separate input to the neural network, thus serving as a seismic facies-controlled constrained deep learning model. The positional coding theory formula is: Where pos represents the seismic phase category, i represents the dimension index of the location encoding vector, and d model Represents the embedding dimension of the model; PE m This represents the location code of the m-th seismic trace.
6. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 1, characterized in that, The establishment of the three-dimensional convolutional deep learning model based on the spatiotemporal attention mechanism specifically includes: a three-dimensional convolutional deep learning model based on the spatiotemporal attention mechanism, which contains a total of 9 layers. The input layer is: well logging data, seismic data preprocessing, and seismic phase time-series coding dataset are input into the three-dimensional convolutional deep learning model. The first part consists of three parallel convolutional layers, with kernel sizes and numbers of [specific values to be filled in]. and in This indicates a convolutional layer with n kernels, kernel size (x, y, z), where x, y, and z represent the main seismic data line, connecting seismic data line, and depth direction, respectively, and a sliding step size of 1. Spatial features at different spatial scales are extracted from seismic data using convolution kernels of different scales. These extracted spatial features are then stitched together and input into the system. To further extract feature information across different scales of seismic data; Spatial attention layer, used to improve the sensitivity of deep learning models to the spatial features of seismic data; The second part consists of three parallel convolutional layers, with kernel sizes and numbers of [specific values to be filled in]. and Temporal features at different scales are extracted from seismic data using convolutional kernels of different scales. These extracted temporal features are then concatenated and input into a database. To further extract temporal characteristic information from earthquake data; Temporal attention layer, used to improve the sensitivity of deep learning models to temporal features; Dropout layers disable neurons with probability p during training; The fully connected layer has 9 input channels and 3 output channels. It is used to improve the fitting ability of the nonlinear relationship between the input and output layers and to output the prediction results of the 3D convolutional deep learning fusion network.
7. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 1, characterized in that, The earthquake phase-controlled deep learning model specifically includes: Establish a loss function for a deep learning model constrained by prior information of logging and seismic data.
8. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 7, characterized in that, The establishment of the loss function for the deep learning model constrained by prior information of logging and seismic data specifically includes: Three parts: logging data loss, seismic data loss, and low-frequency model data loss, to constrain the training backpropagation of the deep learning model; The first part of the formula is the loss of elastic parameter data in logging data. The deep learning model is trained using logging and seismic data at well points. The input seismic data is used to predict elastic parameter data, and the loss error between the true value and the predicted value of elastic parameter data at well points is obtained; Under the constraint of the weight λ1, the mid-high frequency components of logging data are compensated to seismic data. The seismic inversion technology based on deep learning breaks through the limitation of the effective bandwidth of seismic data, and the inversion result has a higher vertical resolution. The larger λ1 is, the higher the vertical resolution of the prediction result; The second part of the formula is the loss of seismic data. The convolution model can establish the relationship between rock elastic parameter data and seismic data. For unlabeled seismic data, the convolution model is used to forward model the elastic parameter data predicted by the deep learning model to obtain the pre-stack seismic trace gather, and the loss error between the synthetic seismic data and the true seismic data is obtained. The weight λ2 controls the constraint degree of seismic data. The larger λ2 is, the more the lateral characteristics of the prediction result conform to seismic data; The third part of the formula is the loss of low-frequency components, which is the loss error between the low-frequency components of the prediction result obtained by using a low-pass filter and the low-frequency components obtained by using the joint interpolation of seismic and logging data; it belongs to the constraint of prior information, and the inverse distance weighted interpolation of seismic and logging data is used to obtain the initial value; the weight λ3 controls the constraint degree of low-frequency components. The lower the signal-to-noise ratio of seismic data, the larger λ3 is; The objective function of the network is expressed as: where M and N are the numbers of seismic data and logging data respectively. During the training process, N << M to meet the situation of insufficient training samples in the actual work area; During the inversion process of model data and actual data, the weight coefficients of the objective function are set as λ1 = 0.4, λ2 = 0.3, λ3 = 0.
3.
9. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 8, characterized in that, The training of the seismic facies-controlled deep learning model using the semi-supervised learning method specifically includes: Inputting the pre-stack seismic trace gather data, seismic facies data, and logging data training set into a three-dimensional convolutional deep learning model; Training the deep learning model using the loss function, and using Adam with momentum as the optimization algorithm, and outputting the deep learning model after the loss function tends to be stable.
10. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 1, characterized in that, The inversion method further includes: testing and evaluating the seismic facies-controlled deep learning model.
11. The deep learning-based seismic facies inversion method for shale oil facies combination according to claim 10, characterized in that, The testing and evaluation of the seismic facies-controlled deep learning model specifically includes: The test set of pre-stack seismic gather data, seismic facies data, and well logging data is input into the trained seismic facies-controlled deep learning model, and the coefficient of determination R is used. 2 Quantitatively evaluate deep learning models; Coefficient of determination R 2 Represented as: Among them, y i Indicates the true label, The mean of the true values. This represents the network's predicted value, where n is the number of samples.