A method and system for predicting depositional reservoirs based on three-dimensional seismic data

Abstract

The application discloses a sedimentary reservoir prediction method based on three-dimensional seismic data, which comprises the following steps: step one, feature extraction; step two, feature dynamic fusion; step three, model establishment and training; step four, optimization of storage prediction results; and step five, visualization interaction. The system comprises a data processing module, a feature fusion module, a model training module, a reservoir prediction module and a visualization interaction module. The dynamic weight distribution algorithm is used to fuse multi-dimensional features such as seismic waveforms, frequency spectra and geometric attributes, so that the vertical resolution of thin interbedded layers is significantly improved, the identification difficulty of hidden reservoirs is solved, and the loss risk is reduced. A double-channel neural network model is established to automatically extract seismic features and sedimentary facies knowledge, so that the deviation of artificial feature engineering is avoided, the rapid modeling demand of a complex sedimentary environment is met, and the prediction results are constrained and corrected, so that the continuity of a vertical sedimentary sequence is ensured.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a method and system for predicting sedimentary reservoirs based on three-dimensional seismic data. Background Technology

[0002] In the field of oil and gas exploration, 3D seismic data interpretation is the core technical means for reservoir prediction. The current mainstream prediction methods mainly rely on seismic attribute analysis, inversion technology and machine learning algorithms, aiming to infer the distribution of subsurface sedimentary facies and reservoir physical parameters through seismic response characteristics. However, as exploration targets shift to complex lithology, deep to ultra-deep and unconventional reservoirs, the existing technical system faces severe challenges in terms of prediction accuracy, geological fit and computational efficiency.

[0003] Traditional methods often use linear weighting to fuse seismic attributes (such as amplitude, frequency, and geometric attributes), lacking the ability to model nonlinear relationships between attributes, leading to identification errors. At the same time, existing intelligent prediction methods using CNN models and mainstream deep learning frameworks have obvious limitations. The former's input features are usually manually selected by experts, resulting in the model's generalization ability being limited by the geological features of the training area. The latter's framework does not embed sedimentary dynamics, and the prediction results may violate the principle of geological spatial continuity.

[0004] While existing systems support 3D modeling, their outdated geological knowledge bases make it difficult to automatically identify special sedimentary patterns. Therefore, this invention proposes a sedimentary reservoir prediction method and system based on 3D seismic data to address the problems in existing technologies. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose a sedimentary reservoir prediction method and system based on 3D seismic data. This method and system integrates multi-dimensional features such as seismic waveforms, spectra, and geometric attributes through a dynamic weight allocation algorithm, significantly improving the vertical resolution of thin interbedded layers, solving the problem of identifying concealed reservoirs, and reducing the risk of leakage. A dual-channel neural network model is established to automatically extract seismic features and sedimentary facies knowledge, avoiding the bias of manual feature engineering, adapting to the rapid modeling needs of complex sedimentary environments, and constraining and correcting the prediction results to ensure the continuity of the vertical sedimentary sequence.

[0006] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for predicting sedimentary reservoirs based on three-dimensional seismic data, comprising the following steps:

[0007] Step 1: Feature extraction. The acquired 3D seismic data is preprocessed, and then multi-dimensional features, including seismic waveforms, spectra, and geometric attributes, are extracted from the preprocessed 3D seismic data.

[0008] Step 2: Dynamic feature fusion. The extracted multi-dimensional features are fused using a dynamic weight allocation algorithm to generate optimized 3D seismic attribute data.

[0009] Step 3: Model building and training. Construct a dual-channel deep neural network model, and use known well control data and seismic data to build a training set. Use deep learning algorithms to train the model.

[0010] Step 4: Optimize the stored prediction results. Input the seismic data of the area to be predicted into the trained model, and based on the three-dimensional geological knowledge map containing sedimentary patterns and diagenetic rules, generate spatial constraints through graph neural networks to correct and optimize the prediction results output by the model, and obtain the predicted results of sedimentary reservoir distribution in the area.

[0011] Step 5: Visual Interaction. The revised reservoir prediction results are dynamically displayed through VR interactive modules, and uncertainty analysis, multi-scheme comparison, and real-time optimization selection of well location deployment are provided.

[0012] Further improvements are made in the following aspects: the data preprocessing in step one includes anisotropic noise suppression of the data using an improved Curvelet transform algorithm, filtering of the data using median filtering, and amplitude compensation processing.

[0013] The further improvement lies in the following: the amplitude compensation process is specifically as follows:

[0014] S1. First, the pre-stack earthquake single-shot record data is processed by seismic wave cut-off to obtain the earthquake first arrival wavelet data;

[0015] S2. Then, autocorrelation processing is performed on the first arrival wavelet data of the earthquake to obtain the earthquake record wavelet data;

[0016] S3. Subsequently, the amplitude value of each seismic trace in the seismic record wavelet data is decomposed to obtain the amplitude compensation factor.

[0017] S4. Finally, amplitude compensation is performed on the pre-stack seismic single-shot record data based on the amplitude compensation factor.

[0018] A further improvement is made in that the dynamic weight allocation algorithm in step two satisfies the following formula:

[0019]

[0020] Where S attrLet S be the seismic attribute matrix, W be the weight matrix, and S be the weight matrix. well Let F be the seismic data matrix near the well, where F represents the Frobenius norm, Λ is the regularization matrix of geometric constraints between response attributes, and α and β are adaptive adjustment coefficients, with α+β=1. α controls the matching accuracy between seismic attributes and well data, and β controls the strength of geological rule constraints.

[0021] Further improvements are made in the following: the dual-channel deep neural network model in step three includes a first channel layer, a second channel layer, and a feature fusion layer. The first channel layer uses a 3D convolutional neural network to extract spatial features of seismic data, the second channel layer uses a graph neural network to process sedimentary facies knowledge graphs, and the feature fusion layer introduces an attention mechanism to perform weighted fusion of the outputs of the first channel layer and the second channel layer.

[0022] A further improvement lies in the following: the spatial constraints in step four include vertical sedimentary sequence constraints and reservoir property parameter constraints. The vertical sedimentary sequence constraint requires that for any spatial location (i, j, k), the predicted result yk+1 must satisfy the vertical transfer rule of sedimentary facies.

[0023]

[0024] Where T(·) represents the preset set of sedimentary facies transfer relationships, and N t This represents the number of vertical sampling points;

[0025] Reservoir physical property parameters are constrained by porosity The wave impedance Z satisfies the following nonlinear mapping relationship:

[0026]

[0027] Parameters a and b are calibration parameters obtained through well logging data calibration, ∈ represents the random error term, N(0, σ 2 ) represents a normal distribution, σ 2 This represents the variance, specifically the range of fluctuations in the quantification error.

[0028] A sedimentary reservoir prediction system based on 3D seismic data includes a data processing module, a feature fusion module, a model training module, a reservoir prediction module, and a visualization and interaction module. The data processing module processes the acquired 3D seismic data and extracts multi-dimensional features. The feature fusion module performs feature fusion based on dynamic weight allocation. The model training module constructs and trains a dual-channel deep neural network model. The reservoir prediction module performs constraint optimization on the sedimentary reservoir distribution prediction results output by the model. The visualization and interaction module uses VR technology to dynamically and interactively display the sedimentary reservoir distribution prediction results and provides multiple optimization options.

[0029] Further improvements are made in that: the model training module includes a model building submodule, a training set construction submodule, and a training submodule. The model building submodule establishes a dual-channel deep neural network model based on 3D convolutional neural networks and graph neural networks and introduces an attention mechanism. The training set construction submodule constructs a training set based on known well control data and seismic data. The training submodule trains the dual-channel deep neural network model based on deep learning algorithms.

[0030] Further improvements are made in that: the storage prediction module includes a model application submodule and a constraint correction submodule. The model application submodule imports the established dual-channel deep neural network model into the prediction system, uses the seismic data of the area to be predicted as input for prediction application, and outputs the prediction results. The constraint correction submodule corrects and optimizes the prediction results based on the three-dimensional geological knowledge map and generates spatial constraints through graph neural networks.

[0031] Further improvements are made in that: the visualization interaction module includes a VR visualization submodule and an optimization selection submodule. The VR visualization submodule visualizes the predicted results of sedimentary reservoir distribution based on VR visualization technology, and the optimization selection submodule is used to provide uncertainty analysis, multi-scheme comparison and real-time optimization selection of well location deployment.

[0032] The beneficial effects of this invention are as follows: This invention integrates multi-dimensional features such as seismic waveforms, spectrum, and geometric properties through a dynamic weight allocation algorithm, which significantly improves the vertical resolution of thin interbedded layers, solves the problem of identifying concealed reservoirs, and reduces the risk of leakage.

[0033] A dual-channel neural network model was established to automatically extract seismic features and sedimentary facies knowledge, avoiding the bias of manual feature engineering, adapting to the rapid modeling needs of complex sedimentary environments, and constraining and correcting the prediction results based on a three-dimensional geological knowledge map to ensure the continuity of vertical sedimentary sequences and avoid economic losses caused by geological conflicts in well location deployment.

[0034] Finally, VR interactive technology is used to achieve collaborative decision-making among geologists, engineers, and algorithms, thereby improving the success rate of exploration in complex blocks. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0036] Figure 2 This is a system architecture diagram of Embodiment 2 of the present invention. Detailed Implementation

[0037] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0038] Example 1

[0039] according to Figure 1 As shown, this embodiment provides a method for predicting sedimentary reservoirs based on three-dimensional seismic data, including the following steps:

[0040] Step 1: Preprocess the acquired 3D seismic data, and then extract multi-dimensional features, including seismic waveforms, spectra, and geometric attributes, from the preprocessed 3D seismic data.

[0041] The data preprocessing includes using an improved Curvelet transform algorithm to suppress anisotropic noise, using median filtering to filter the data, and performing amplitude compensation processing.

[0042] The amplitude compensation process is as follows:

[0043] S1. First, the pre-stack earthquake single-shot record data is processed by seismic wave cut-off to obtain the earthquake first arrival wavelet data;

[0044] S2. Then, autocorrelation processing is performed on the first arrival wavelet data of the earthquake to obtain the earthquake record wavelet data;

[0045] S3. Subsequently, the amplitude value of each seismic trace in the seismic record wavelet data is decomposed to obtain the amplitude compensation factor.

[0046] S4. Finally, amplitude compensation is performed on the pre-stack seismic single-shot record data based on the amplitude compensation factor.

[0047] Step 2: The extracted multi-dimensional features are fused using a dynamic weighting algorithm to generate optimized 3D seismic attribute data;

[0048] The dynamic weight allocation algorithm satisfies the following formula:

[0049]

[0050] Where S attr This is the seismic attribute matrix, with dimensions M×K (M represents spatial location, such as each seismic trace or voxel; K represents the seismic attribute category, such as instantaneous amplitude, curvature, spectral slope, etc.).

[0051] W is the weight matrix with dimensions K×P (K corresponds to the number of seismic attribute categories; P is the dimension of the target attribute after fusion, which is usually defined by geological requirements, such as reservoir thickness, porosity, etc.).

[0052] S well This is a matrix of seismic data near wells, with dimensions equal to S. attr same;

[0053] F represents the Frobenius norm;

[0054] Λ is the regularization matrix for the geometric constraints between reaction attributes, with diagonal elements Λ ii Represents the geological contribution weight of the i-th attribute. Off-diagonal elements can encode the spatial correlation between attributes (such as the correlation between curvature attributes and fault distance).

[0055] α and β are adaptive adjustment coefficients. α controls the matching accuracy between seismic attributes and well data, and β controls the constraint strength of geological rules (the larger the value, the more the attribute fusion conforms to the prior geological laws). α+β=1 normalization ensures optimization stability.

[0056] Step 3: Construct a dual-channel deep neural network model, and build a training set using known well control data and seismic data, and train the model using deep learning algorithms;

[0057] The dual-channel deep neural network model consists of a first channel layer, a second channel layer, and a feature fusion layer. The first channel layer uses a 3D convolutional neural network to extract spatial features of seismic data, the second channel layer uses a graph neural network to process sedimentary facies knowledge graphs, and the feature fusion layer introduces an attention mechanism to perform weighted fusion of the outputs of the first and second channel layers.

[0058] Step 4: Input the seismic data of the area to be predicted into the trained model, and based on the three-dimensional geological knowledge map containing sedimentary patterns and diagenetic rules, generate spatial constraints through graph neural networks to correct and optimize the prediction results output by the model, and obtain the predicted results of sedimentary reservoir distribution in the area.

[0059] Spatial constraints include vertical sedimentary sequence constraints and reservoir property parameter constraints. The vertical sedimentary sequence constraint requires that for any spatial location (i, j, k), the predicted result yk+1 must satisfy the vertical transfer rule of sedimentary facies.

[0060]

[0061] Where T(·) represents the preset set of sedimentary facies transfer relationships, and N t The number of vertical sampling points determines the number of iterations for the vertical sequence constraint, for example, N. t If the value is 500, then sedimentary facies transfer verification needs to be performed on 499 adjacent layers.

[0062] Reservoir physical property parameters are constrained by porosity The wave impedance Z satisfies the following nonlinear mapping relationship:

[0063]

[0064] Parameters a and b are calibration parameters, obtained through logging data. Parameter a controls the sensitivity of wave impedance to porosity (e.g., a>0 indicates that the impedance increases while the porosity decreases), and parameter b is a correction term to eliminate the influence of impedance baseline offset.

[0065] ∈ represents the random error term, N(0, σ) 2 Let σ represent a normal distribution, where 0 is the mean and σ is the expected value of the error. 2 Variance represents the range of error fluctuations, and σ represents the standard deviation, which directly affects the width of the confidence interval for porosity prediction (e.g., σ = 0.5% means a 95% confidence interval of ±1%).

[0066] Step 5: The revised reservoir prediction results are dynamically displayed through the VR interactive module, and uncertainty analysis, multi-scheme comparison and real-time optimization selection of well location deployment are provided;

[0067] The 3D dynamic rendering for VR interaction is achieved through GPU acceleration to create volumetric rendering, providing reservoir 3D visualization, well location design, and real-time optimization functions.

[0068] Example 2

[0069] according to Figure 2 As shown, this embodiment provides a sedimentary reservoir prediction system based on 3D seismic data, including a data processing module, a feature fusion module, a model training module, a reservoir prediction module, and a visualization interaction module. The data processing module processes the acquired 3D seismic data and extracts multi-dimensional features. The feature fusion module performs feature fusion processing based on dynamic weight allocation. The model training module constructs and trains a dual-channel deep neural network model. The reservoir prediction module performs constraint optimization on the sedimentary reservoir distribution prediction results output by the model. The visualization interaction module uses VR technology to dynamically and interactively display the sedimentary reservoir distribution prediction results and provides multiple optimization options.

[0070] The model training module includes a model building submodule, a training set construction submodule, and a training submodule. The model building submodule is based on 3D convolutional neural networks and graph neural networks and introduces an attention mechanism to build a dual-channel deep neural network model. The training set construction submodule is based on known well control data and seismic data to build a training set. The training submodule is based on deep learning algorithms to train the dual-channel deep neural network model.

[0071] The prediction storage module includes a model application submodule and a constraint correction submodule. The model application submodule imports the established dual-channel deep neural network model into the prediction system, uses the seismic data of the area to be predicted as input for prediction application, and outputs the prediction results. The constraint correction submodule uses a three-dimensional geological knowledge map and generates spatial constraints through a graph neural network to correct and optimize the prediction results.

[0072] The visualization and interaction module includes a VR visualization submodule and an optimization selection submodule. The VR visualization submodule uses VR visualization technology to visualize the predicted results of sedimentary reservoir distribution and uses GPU acceleration to achieve volume rendering. The optimization selection submodule is used to provide uncertainty analysis, multi-scheme comparison, and real-time optimization selection of well location deployment.

[0073] The systematic design makes the entire forecasting process more automated and improves work efficiency.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method of sedimentary reservoir prediction based on three-dimensional seismic data, characterized in that, Includes the following steps: Step 1: Feature extraction. The acquired 3D seismic data is preprocessed, and then multi-dimensional features, including seismic waveforms, spectra, and geometric attributes, are extracted from the preprocessed 3D seismic data. Step 2: Dynamic feature fusion. The extracted multi-dimensional features are fused using a dynamic weight allocation algorithm to generate optimized 3D seismic attribute data. Step 3: Model building and training. Construct a dual-channel deep neural network model, and use known well control data and seismic data to build a training set. Use deep learning algorithms to train the model. The dual-channel deep neural network model includes a first channel layer, a second channel layer, and a feature fusion layer. The first channel layer uses a 3D convolutional neural network to extract spatial features of seismic data, the second channel layer uses a graph neural network to process sedimentary facies knowledge graphs, and the feature fusion layer introduces an attention mechanism to weightedly fuse the outputs of the first and second channel layers. Step 4: Optimize the stored prediction results. Input the seismic data of the area to be predicted into the trained model, and based on the three-dimensional geological knowledge map containing sedimentary patterns and diagenetic rules, generate spatial constraints through graph neural networks to correct and optimize the prediction results output by the model, and obtain the predicted results of sedimentary reservoir distribution in the area. The spatial constraints include vertical sedimentary sequence constraints and reservoir physical property parameter constraints. Step 5: Visual Interaction. The revised reservoir prediction results are dynamically displayed through VR interactive modules, and uncertainty analysis, multi-scheme comparison, and real-time optimization selection of well location deployment are provided.

2. The method of claim 1, wherein: The data preprocessing in step one includes using an improved Curvelet transform algorithm to suppress anisotropic noise, using median filtering to filter the data, and performing amplitude compensation processing.

3. The method for predicting sedimentary reservoirs based on three-dimensional seismic data according to claim 2, characterized in that: The amplitude compensation process is specifically as follows: S1. First, the pre-stack earthquake single-shot record data is processed by seismic wave cut-off to obtain the earthquake first arrival wavelet data; S2. Then, autocorrelation processing is performed on the first arrival wavelet data of the earthquake to obtain the earthquake record wavelet data; S3. Subsequently, the amplitude value of each seismic trace in the seismic record wavelet data is decomposed to obtain the amplitude compensation factor. S4. Finally, amplitude compensation is performed on the pre-stack seismic single-shot record data based on the amplitude compensation factor.

4. The method for predicting sedimentary reservoirs based on three-dimensional seismic data according to claim 1, characterized in that: The dynamic weight allocation algorithm in step two satisfies the following formula: in Let W be the earthquake attribute matrix and W be the weight matrix. This is the well-side seismic data matrix, where F represents the Frobenius norm. This is the regularization matrix for the geometric constraints between reaction attributes. It is an adaptive adjustment coefficient, and + =1, Controlling the matching accuracy between seismic attributes and well data. Control the intensity of geological rule constraints.

5. The method for predicting sedimentary reservoirs based on three-dimensional seismic data according to claim 1, characterized in that: In step four, the vertical deposition sequence constraint is applied to any spatial location. The predicted result yk+1 must satisfy the vertical transfer rule of sedimentary facies: in This represents a predefined set of sedimentary facies transfer relationships. This represents the number of vertical sampling points; Reservoir physical property parameters are constrained by porosity The wave impedance Z satisfies the following nonlinear mapping relationship: Parameters a and b are calibration parameters, obtained through well logging data calibration. Represents the random error term. Represented by a normal distribution, This represents the variance, specifically the range of fluctuations in the quantification error.

6. A prediction system for a sedimentary reservoir prediction method based on three-dimensional seismic data according to any one of claims 1-5, characterized in that: The system includes a data processing module, a feature fusion module, a model training module, a reservoir prediction module, and a visualization and interaction module. The data processing module processes the acquired 3D seismic data and extracts multi-dimensional features. The feature fusion module performs feature fusion based on dynamic weight allocation. The model training module constructs and trains a dual-channel deep neural network model. The reservoir prediction module performs constraint optimization on the sedimentary reservoir distribution prediction results output by the model. The visualization and interaction module uses VR technology to dynamically and interactively display the sedimentary reservoir distribution prediction results and provides multiple optimization options.

7. A sedimentary reservoir prediction system based on three-dimensional seismic data according to claim 6, characterized in that: The model training module includes a model building submodule, a training set construction submodule, and a training submodule. The model building submodule establishes a dual-channel deep neural network model based on 3D convolutional neural networks and graph neural networks, and introduces an attention mechanism. The training set construction submodule constructs a training set based on known well control data and seismic data. The training submodule trains the dual-channel deep neural network model based on deep learning algorithms.

8. A sedimentary reservoir prediction system based on three-dimensional seismic data according to claim 6, characterized in that: The reservoir prediction module includes a model application submodule and a constraint correction submodule. The model application submodule imports the established dual-channel deep neural network model into the prediction system, uses the seismic data of the area to be predicted as input for prediction application, and outputs the prediction results. The constraint correction submodule corrects and optimizes the prediction results based on a three-dimensional geological knowledge map and generates spatial constraints through a graph neural network.

9. A sedimentary reservoir prediction system based on three-dimensional seismic data according to claim 6, characterized in that: The visualization and interaction module includes a VR visualization submodule and an optimization selection submodule. The VR visualization submodule uses VR visualization technology to visualize the predicted results of sedimentary reservoir distribution. The optimization selection submodule is used to provide uncertainty analysis, multi-scheme comparison, and real-time optimization selection of well location deployment.

Images