Oil reservoir physical property intelligent prediction method based on sand body superposition mode

Abstract

The invention discloses an oil reservoir physical property intelligent prediction method based on a sand body superposition mode, and relates to the technical field of electric digital data processing. The method comprises the following steps that 101, an oil reservoir framework model is established, and existing data are preprocessed; step 102, generating a sand body superposition mode library; 103, well bypass statistical wavelets are extracted; 104, forward modeling of a sand body superposition mode sample is carried out; step 105, constructing a correlation graph and an index graph; 106, screening a high-quality training sample; step 107, training a neural network model; step 108, predicting oil reservoir physical property attributes; according to the method, disturbance functions under different geological constraint conditions are solved, then sand body superposition modes with different geological features are generated randomly, then accurate estimation of oil reservoir physical parameters based on geology, logging and earthquakes is achieved by combining a forward modeling technology and a deep learning technology, the new technology path better conforms to the actual situation, and the method is suitable for large-scale popularization and application. And the accuracy of oil reservoir physical property attribute estimation is improved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to an intelligent prediction method for reservoir physical properties based on sand body stacking patterns. Background Technology

[0002] In the field of reservoir geophysics, three-dimensional physical property prediction is a crucial task, typically employing geostatistical methods for parameter estimation. However, traditional geostatistical methods have limitations when simulating the characteristics of geological bodies with complex structures. For example, two-point geostatistics and multi-point geostatistics use variograms and combined models of multiple points, respectively, to describe the structural information of geological bodies. However, their performance is not ideal in specific scenarios.

[0003] Currently, deep learning methods have made groundbreaking progress in many fields, excelling at handling complex nonlinear problems and possessing strong generalization capabilities. Deep learning-based physical property prediction methods can integrate hard logging data and soft seismic data to obtain high-resolution three-dimensional physical property estimation models. However, deep learning methods require a large number of high-quality labeled samples during supervised learning, while the high cost of data acquisition in the geophysical exploration industry results in insufficient sample quantities, hindering the development of deep learning methods.

[0004] The application, published under CN116629023A, is titled "A Novel Three-Dimensional Reservoir Characterization Method Integrating Stratigraphic Forward Modeling and Geostatistics." Using a SFM mudstone model as a constraint, a lithological model calibrated under the SFM mudstone model constraint is established using geostatistical methods. Sedimentary microfacies are divided based on core samples and well logging curves, and planar distribution maps of sedimentary microfacies are obtained based on well-connected sedimentary facies analysis, which are then transformed into a three-dimensional trend volume of sedimentary facies. A sedimentary microfacies model is obtained based on the three-dimensional trend volume and used as a constraint in the geological modeling process.

[0005] The application, published under CN113589363A, is titled "A Novel Method for Oil and Gas Prediction Integrating Artificial Neural Networks and Geostatistics." This novel method involves averaging seismic and well logging data across different layers, followed by Kendall correlation analysis of the two data sets. A suitable curve is selected to train the artificial neural network at the well site. Then, combined seismic and well logging data with geostatistical analysis is performed to obtain the three-dimensional distribution of reservoir physical parameters. Based on this, a second artificial neural network is trained at the well site for three-dimensional oil and gas prediction.

[0006] Application publication number CN107316341A, entitled "A Multi-Point Geostatistical Sedimentary Facies Modeling Method," proposes a geometric factor concept to improve upon existing multi-point geostatistical sedimentary facies modeling methods. This method uses the geometric factor to define a spatially confined space and calculate the conditional probability reflecting the scale of sedimentary facies zones. It then applies a corrected proportional identity to modify the conditional probability distribution curves obtained from scanning training images, thus improving upon the original multi-point geostatistical method. Modeling implementation using the improved algorithm demonstrates that the multi-point geostatistical method incorporating the geometric factor inherits the advantages of traditional multi-point geostatistical methods in reproducing sedimentary facies geometry while addressing its shortcomings in handling facies zone scale and continuity. Since this method is based on multi-point geostatistics, the selection of training images significantly impacts the final results.

[0007] A conference paper titled "Application of a Novel Method for Sandbody Overlay Pattern Recognition in Seismic Driven Modeling" establishes five geological conceptual models of fluvial facies sandbody overlay patterns with different morphologies. Then, by performing forward modeling on these five models, seismic waveforms and attribute information are selected as template data. Based on this, an identification algorithm and templates are used to identify the overlay patterns of sand bodies in the target layer on the seismic profile, obtaining the pattern distribution of the target sand bodies. During the modeling process, based on the different sandbody patterns obtained from the previous pattern recognition, well logging data under the same pattern are selected to participate in the modeling process, ultimately yielding the reservoir physical property model. This paper builds upon the identification results of sandbody overlay patterns by selecting well logging data under the same pattern to participate in the physical property modeling process.

[0008] Existing reservoir physical property modeling suffers from a series of technical problems, which affect the accuracy of physical property parameter estimation and the reliability of the model. Specific technical problems are as follows:

[0009] First, there is the challenge of multi-data fusion. Reservoir property modeling requires the integration of various data, such as seismic and well logging data. Establishing conversion relationships between these data sources, while adhering to geological understanding, is crucial for achieving reliable estimation of physical property parameters. This is a pressing issue we currently face.

[0010] Second, there is a shortage of available data. The accuracy of reservoir property modeling depends on the quantity and quality of available data. However, in production practice, drilling and logging are time-consuming and expensive, and the number of well logs in highly developed oilfields is small. Therefore, how to fully explore the characteristics of current data and effectively expand the sample size is an urgent problem to be solved. Summary of the Invention

[0011] This invention provides an intelligent prediction method for reservoir physical properties based on sand body stacking patterns, which solves the technical problem of low accuracy in estimating reservoir physical properties.

[0012] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0013] A method for intelligent prediction of reservoir physical properties based on sand body stacking patterns includes the following steps:

[0014] Step 101: Establish and obtain the reservoir framework model. Based on the reservoir framework model, resample the logging curves, sand body labels, and seismic data respectively to obtain the logging curves, sand body labels, and reservoir seismic attributes at the model scale at the well points. According to the well point coordinates, extract the wellbore amplitude from the reservoir seismic attributes. The logging curves, sand body labels, and wellbore amplitude at the model scale at the well points form complete sample data.

[0015] Step 102: Randomly select a disturbance function. Based on the sample data obtained in Step 101, randomly generate the starting depth and disturbance point of the sand-sludge boundary after disturbance. Solve the disturbance function and resample the disturbance function according to the set sampling interval. Input the resampled value into the mapping function of the logging curve to generate a sample of each sand body stacking pattern, forming a sand body stacking pattern library. The sand body stacking pattern samples do not contain amplitude data.

[0016] Step 103: Extract well logging and well access data from the complete sample data in Step 101, and optimize the initial Lake wavelet using the stochastic gradient descent method to obtain the well access statistical wavelet;

[0017] Step 104: Using the wellside statistical wavelet from Step 103, perform forward modeling on the sand body overlay pattern samples generated in Step 102 to obtain forward-modeled sand body overlay pattern samples, forming a forward-modeled sand body overlay pattern library.

[0018] Step 105: Based on the forward modeling sand body overlay model library obtained in Step 104, extract the amplitude data of each sample; based on the reservoir seismic attributes obtained in Step 101, extract the vertical amplitude data of each trace; perform correlation analysis on the two types of amplitude data obtained, record the correlation coefficient and index value of the amplitude data of samples with high correlation values, and after traversing all traces, form a correlation map and an index map.

[0019] Step 106: Obtain the correlation coefficient threshold. Select sand body stacking pattern samples that are above the correlation coefficient threshold from the correlation map obtained in step 105 to form high-quality training samples.

[0020] Step 107: Use the high-quality training samples obtained in step 106 as the training set to train the neural network until it converges, and obtain the trained neural network.

[0021] Step 108: Input the reservoir seismic attributes obtained in step 101 into the trained neural network obtained in step 107. The neural network outputs the predicted reservoir physical properties.

[0022] A further technical solution is as follows: In step 101, the logging curves are resampled based on the spatial coordinates of the reservoir grid model to obtain the logging curves at the model scale at the well point; the logging sand body interpretation data are labeled, and the sand body labels at the model scale are obtained based on the spatial coordinates of the well point at the reservoir grid model; the seismic data are resampled based on the spatial coordinates of the well point at the reservoir grid model to obtain the reservoir seismic attributes.

[0023] A further technical solution is as follows: In step 102, the perturbation function is a linear perturbation function or a nonlinear perturbation function, the depth threshold is 0.125 meters, and the mapping function is the logging curve extracted from the sample data obtained in step 101. The logging curve includes acoustic wave, density, porosity and corresponding spatial coordinate data. The piecewise function approximation method is used to establish the mapping function from spatial coordinates to acoustic wave, density and porosity, i.e., Equation (1.1).

[0024] Y = S(z) (1.1)

[0025] In equation (1.1), z is the depth coordinate corresponding to the geological body along the well trajectory, S(·) is the mapping function from depth coordinate data to acoustic, density and porosity data, and Y is the numerical value corresponding to acoustic, density and porosity.

[0026] The formula for the linear perturbation function is:

[0027] Z = f(Z′) = aZ′ + b (1.2)

[0028] The formula for the nonlinear perturbation function is:

[0029] Z = f(Z') = aZ' 2 +bZ'+c (1.3)

[0030] In the formula, Z represents the depth coordinates of the original geological body, and Z' represents the depth coordinates of the geological body after its position has changed.

[0031] A further technical solution is that, in step 102,

[0032] When replicating the sand body overlay pattern sample, a linear perturbation function, i.e., formula (1.2), is used, and the spatial coordinates of the geological body remain unchanged. While keeping the initial depth of the sand-mud boundary point unchanged, the linear perturbation function is solved to realize the process of replicating the sand body overlay pattern sample data.

[0033] When the thickness of the sand body in the stacked pattern is uniformly changed, a linear perturbation function, i.e., formula (1.2), is used. By changing the initial depth of the sand-mud boundary point, the perturbation function is solved to achieve the uniform change process of the sand body thickness in the stacked pattern.

[0034] When adding a sand body within an existing overlay pattern, a nonlinear perturbation function, i.e., formula (1.3), is used. By changing the initial depth of the sand-mud boundary point and randomly generating a perturbation point inside the original boundary point, the perturbation function is solved to realize the process of adding a sand body in the sand body overlay pattern sample.

[0035] When deleting a segment of the sand body in the overlay mode, a nonlinear perturbation function, i.e., formula (1.3), is used to randomly generate perturbation points inside the original boundary points. The depth of the perturbation points is taken to be outside the depth range of the overlay mode, thereby eroding away a segment of the sand body.

[0036] A further technical solution involves: in step 103, extracting wellbore amplitude, acoustic wave, and density data from the complete sample data; and obtaining reflection coefficient sequence data based on the acoustic wave and density data, using the following formula:

[0037]

[0038] In equation (1.4), ρ i This represents the density of the i-th layer of medium, expressed in g / cm³. 3 v i R represents the velocity of the i-th layer of medium, which is the reciprocal of the sound wave velocity, and its unit is m / s. i This represents the reflection coefficients of the i-th layer and the (i+1)-th layer;

[0039] Statistical wavelet extraction utilizes reflection coefficients and wellbore data. In the data-driven wavelet extraction process, the Ricker wavelet is used as the initial parameter of the convolution kernel, and a function from the reflection coefficient to the amplitude is constructed according to the convolution formula.

[0040]

[0041] In equation (1.5), s t It's amplitude data, r t It is the reflection coefficient obtained earlier, and N represents the wavelet ω. τ Length, ω τ This is the wavelet data to be extracted, with the initial value being the Ricker wavelet, and the formula is as follows:

[0042]

[0043] A further technical solution is as follows: In step 103, based on the data-driven wavelet extraction process, the actual coarsened well bypass is used as the result constraint, and the gradient descent method is adopted to optimize the initial result;

[0044] The gradient descent method is as follows:

[0045]

[0046] In equation (1.7), ω t Let η represent the value of the wavelet in the t-th iteration, and let η represent the learning rate. Describe the objective function The first derivative of the wavelet parameter ω is the gradient; by iteratively updating the wavelet parameter, an optimized wavelet is obtained.

[0047] A further technical solution is as follows: In step 104, the sand body stacking pattern sample generated in step 102 lacks amplitude data. Using the statistical wavelet obtained in step 103, the amplitude data is supplemented by forward modeling according to formulas (1.4) and (1.5). The sample data obtained after the supplementation is complete sample data. After this step, in the constructed sand body stacking pattern library, each sample contains logging curves, sand body labels and amplitude data.

[0048] A further technical solution is that, in step 105, the formula for calculating the correlation coefficient is as follows:

[0049]

[0050] In Equation (1.8), X and Y are the amplitude data of the overlaid sample and the longitudinal amplitude data of each trace of the reservoir seismic attribute, respectively.

[0051] The overlay sample with high amplitude correlation is used to approximate the actual sand body overlay situation of the corresponding trace. The correlation coefficient of the overlay sample, its index in the sand body overlay model library, and the corresponding coordinates are recorded. This process is repeated until all traces are calculated to form a correlation map and an index map.

[0052] A further technical solution is as follows: In step 106, the correlation coefficient threshold is 0.9. Points with a correlation coefficient greater than 0.9 are selected from the correlation map, and the corresponding points are found in the sand body overlay pattern library, i.e., high correlation coefficient samples, in combination with the index map, to form high-quality training samples that reflect the structural information of actual complex geological bodies.

[0053] A further technical solution is as follows: In step 107, the neural network is a deep neural network. During the training process, amplitude data is extracted from the sand body stacked samples as the input of the neural network, and qualitative parameters are extracted from the sand body stacked samples as labels. The mean squared error loss function is selected.

[0054]

[0055] In equation (1.9), N is the output length of the neural network. Let y be the label sequence and y be the output sequence of the neural network; stochastic gradient descent is selected as the optimizer.

[0056] The beneficial effects of adopting the above technical solution are as follows:

[0057] A method for intelligent prediction of reservoir properties based on sand body stacking patterns includes the following steps: Step 101: Establish and obtain a reservoir framework model; based on the reservoir framework model, resample well logging curves, sand body labels, and seismic data respectively to obtain well logging curves, sand body labels, and reservoir seismic attributes at the model scale at the well point; extract wellbore amplitude from the reservoir seismic attributes according to the well point coordinates; the well logging curves, sand body labels, and wellbore amplitude at the model scale at the well point form complete sample data; Step 102: Randomly select a perturbation function, based on step 1... 01. The obtained sample data is used to randomly generate the initial depth and disturbance point of the sand-sludge boundary after disturbance. The disturbance function is solved and resampled at a set depth sampling interval. The mapping function from coordinates to logging curves is used to generate samples of each sand body stacking pattern, forming a sand body stacking pattern library. The sand body stacking pattern samples do not contain amplitude data. Step 103: The logging and wellbore access data are extracted from the complete sample data in Step 101. The initial Ricker wavelet is optimized using the stochastic gradient descent method to obtain the wellbore access statistical wavelet. Step 1 04: Using the wellside statistical wavelet from step 103, perform forward modeling on the sand body overlay model samples generated in step 102 to obtain forward-modeled sand body overlay model samples, forming a forward-modeled sand body overlay model library; Step 105: Based on the forward-modeled sand body overlay model library obtained in step 104, extract the amplitude data for each sample; based on the reservoir seismic attributes obtained in step 101, extract the vertical amplitude data for each trace; perform correlation analysis on the two types of amplitude data obtained, and record the correlation coefficient and index value of the amplitude data of samples with high correlation values. After traversing all channels, a correlation map and an index map are generated. In step 106, a correlation coefficient threshold is obtained, and sand body stacking pattern samples above the threshold are selected from the correlation map obtained in step 105 to form high-quality training samples. In step 107, the high-quality training samples obtained in step 106 are used as the training set to train the neural network until convergence, resulting in a trained neural network. In step 108, the reservoir seismic attributes obtained in step 101 are input into the trained neural network obtained in step 107, and the neural network outputs the predicted reservoir physical properties. The method optimizes the initial Ricker wavelet using stochastic gradient descent to obtain a well-side statistical wavelet, which better reflects the actual situation and improves the accuracy of reservoir physical property estimation.

[0058] See the detailed implementation section for further description. Attached Figure Description

[0059] Figure 1a This is a graph of the disturbance function when replicating sand bodies in the sand body stacking pattern sample generated in Embodiment 2 of the present invention;

[0060] Figure 1bThis is a graph of the disturbance function when the sand body thickness is changed in the sand body stacking mode sample generated in Embodiment 2 of the present invention;

[0061] Figure 1c This is a graph of the disturbance function when a new sand body is added to the sand body stacking pattern sample generated in Embodiment 2 of the present invention;

[0062] Figure 1d This is a graph of the disturbance function when deleting sand bodies in the sand body stacking pattern sample generated in Embodiment 2 of the present invention;

[0063] Figure 2 This is a comparison diagram of the unoptimized Ricker wavelet and the optimized statistical wavelet in Embodiment 2 of the present invention;

[0064] Figure 3a This is a correlation diagram showing the representativeness of the quantitative model library samples to actual geological conditions in Embodiment 2 of the present invention;

[0065] Figure 3b This is an index map showing the representativeness of the quantitative model library samples to actual geological conditions in Embodiment 2 of the present invention;

[0066] Figure 4 This is a structural diagram of the neural network in Embodiment 2 of the present invention;

[0067] Figure 5a This is a rendering of the reservoir porosity properties in Embodiment 2 of the present invention;

[0068] Figure 5b This is a comparison diagram of reservoir porosity properties on a well-connected profile in Embodiment 2 of the present invention;

[0069] Figure 6 This is a flowchart of Embodiment 1 of the present invention. Detailed Implementation

[0070] The significance of this technical solution lies in its novel approach, which fully leverages geological, seismic, and well logging information to achieve high-accuracy and high-compliance estimation of reservoir physical parameters. This innovative algorithm, based on a perturbation function, constructs a process for generating sand body overlay model samples from geological understanding, thereby reasonably expanding the sample size to form a sand body overlay model library with diverse and complex geological characteristics. Then, deep learning methods are employed to establish the relationship between seismic amplitude and reservoir physical parameters. In practical applications within the reservoir geophysics industry, this has effectively improved the accuracy and efficiency of 3D physical property modeling.

[0071] The application, published under CN116629023A, is titled "A Novel 3D Reservoir Characterization Method Integrating Stratigraphic Forward Modeling and Geostatistics." This method emphasizes the use of sedimentary microfacies models to constrain geostatistical methods, which is fundamentally different from the method used in this application, which generates a sand body overlay model library and employs deep learning for property prediction.

[0072] The application publication number is CN113589363A, entitled "A Novel Method for Oil and Gas Prediction Integrating Artificial Neural Networks and Geostatistics." The essential difference between this method and the present application is that the present application adopts the concept of sand body stacking models, generates a large number of sand body stacking model samples to form a model library, and uses convolutional networks to directly predict the three-dimensional physical parameters of reservoirs on the reservoir grid.

[0073] The application publication number is CN107316341A, entitled "A Multi-Point Geostatistical Sedimentary Facies Modeling Method". The essential difference between this method and this application is that this application adopts the concept of sand body stacking patterns, generates a large number of sand body stacking pattern samples to form a pattern library, and uses convolutional networks to directly predict the three-dimensional physical parameters of reservoirs on the reservoir grid.

[0074] The conference paper titled "Application of a Novel Method for Sand Body Overlay Pattern Recognition in Seismic Driven Modeling" differs from this application in that this application uses randomly generated effective sand body overlay pattern samples containing various complex conditions for deep learning modeling, rather than identifying overlay pattern regions and selecting well logging data to participate in modeling.

[0075] Based on the analysis of existing related technical solutions, the intelligent prediction method for reservoir physical properties based on sand body stacking patterns disclosed in this application is original. This intelligent prediction method for reservoir physical properties based on sand body stacking patterns can effectively expand the number of samples and features, and improve the accuracy of neural networks in predicting physical properties.

[0076] To address the aforementioned technical challenges, this application proposes an intelligent prediction method for reservoir physical properties based on sand body stacking models. The aim is to comprehensively utilize geological, seismic, and well logging information to achieve high accuracy and consistency in reservoir physical property parameter estimation. This method fully integrates forward modeling, numerical analysis, and deep learning techniques, thereby significantly improving the accuracy of reservoir physical property estimation. It provides a new technical approach for reservoir physical property estimation and offers reliable data support for reservoir development and management.

[0077] Key Inventions:

[0078] This application aims to address the problem of estimating physical property parameters in reservoir physical property modeling. By fusing geological, seismic, and well logging data, it improves the accuracy of reservoir physical property estimation. Its core aspects are as follows:

[0079] The first point is to adopt data-driven seismic wavelet extraction. In this process, the traditional process of obtaining the optimal wavelet parameters by the least squares method is transformed into a data-driven stochastic gradient descent optimization process, and the Ricker wavelet is used as the initial parameter to participate in the optimization. The final extracted statistical wavelet is more in line with the actual situation.

[0080] The second point is the adoption of a sand body superposition pattern library generation method. This method randomly generates sand body superposition patterns with different geological significance by solving the perturbation function under different constraints, thereby expanding the number of samples based on geological understanding and improving the sample characteristics.

[0081] The third point is to use correlation maps and index maps to measure the approximation of the overlay model library to underground geological bodies, thereby quantifying the application effect of the model library.

[0082] The fourth point is the entire intelligent prediction process for reservoir physical properties based on sand body stacking patterns. This technical process establishes a new methodological path for predicting reservoir physical property parameters, and improves the accuracy of reservoir physical property estimation under multiple data constraints.

[0083] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this application or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0084] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0085] Example 1:

[0086] like Figure 6 As shown, this invention discloses an intelligent prediction method for reservoir physical properties based on sand body stacking patterns, comprising the following steps:

[0087] Step 101: Establish a reservoir grid model and preprocess existing data. Resample the sequence data such as logging curves, wellbore amplitude, and sand body labels to form a complete sample data.

[0088] Step 102: Generate a sand body overlay pattern library. Randomly select a perturbation function, randomly generate the changed initial depth of the sand-mud boundary and the perturbation point, solve the perturbation function, and resample the perturbation function at 0.125 meters. Substitute the mapping function from coordinates to well logging curves to generate different sand body overlay pattern samples, forming a sand body overlay pattern library.

[0089] Step 103: Extract the wellbore access statistical wavelet. Well logging and wellbore access data are extracted from the complete sample data in Step 101. The initial Ricker wavelet is optimized using the stochastic gradient descent method to obtain the final statistical wavelet.

[0090] Step 104, Forward Modeling of Sand Body Overlay Pattern Samples. The statistical wavelet from Step 103 is used to perform forward modeling on the sand body overlay pattern samples generated in Step 102.

[0091] Step 105: Construct correlation and index maps. Select the sand body overlay pattern library obtained in Step 104 and extract the amplitude data for each sample; simultaneously, use the reservoir seismic attribute data obtained in Step 101 through preprocessing to extract the longitudinal amplitude data at a certain trace. Perform correlation analysis on the two types of amplitude data, record the correlation coefficient and index value of the amplitude data of samples with high correlation values, and after traversing all traces, form correlation and index maps.

[0092] Step 106: Select high-quality training samples. Determine the correlation coefficient threshold, and select sand body overlay pattern samples from Step 105 whose correlation graphs are above the threshold to form high-quality training samples.

[0093] Step 107, Train the neural network model. Use the high-quality training samples from Step 106 as the training set to train the neural network until it converges.

[0094] Step 108: Predict reservoir physical properties. The reservoir seismic attribute data obtained in step 101 is preprocessed and input into the convergent neural network obtained in step 107. The output of the network is the reservoir physical property model to be estimated.

[0095] Application scenario of Example 1:

[0096] Accurate physical property modeling is crucial in the early stages of reservoir development. The method presented in this application leverages multi-data fusion to improve the accuracy of reservoir physical property estimation, thereby providing reliable foundational data for reservoir development planning. By accurately predicting reservoir physical parameters such as porosity, permeability, and water saturation, it helps in formulating efficient development strategies and well placement plans.

[0097] For reservoirs with complex geological conditions, such as those with multi-layered superimposed sand bodies, the sand body superposition pattern library generation method and multi-data fusion technology proposed in this application can effectively handle the problem of physical property modeling under complex geological structures. By generating and applying sand body superposition patterns with different geological significance, the diversity and characteristics of the samples are improved, ensuring accurate modeling even under complex geological conditions.

[0098] In the oil and gas exploration phase, accurate physical property models are crucial for reserve calculation and economic evaluation. The method presented in this application improves the accuracy of physical property estimation, providing more precise data support for reserve calculation, helping to assess the economic feasibility and development potential of reservoirs, and thus providing a scientific basis for investment decisions.

[0099] In summary, this application's intelligent prediction of reservoir physical properties based on sand body stacking models is applicable to multiple fields, from reservoir development planning to reservoir management and production optimization. By improving the accuracy and reliability of physical property modeling, it provides a more comprehensive and precise reservoir analysis method, significantly enhancing the efficiency and effectiveness of reservoir development and management.

[0100] Technical effects of Example 1:

[0101] This application presents an intelligent prediction method for reservoir physical properties based on sand body stacking models, aiming to solve the problem of physical property parameter estimation in reservoir physical property modeling. Through multi-data fusion of geological, seismic, and well logging data, this application significantly improves the accuracy of reservoir physical property estimation. Specific technical effects are as follows:

[0102] First, this application improves the accuracy of wavelet extraction. It employs a data-driven seismic wavelet extraction method, overcoming the limitations of traditional least-squares methods for obtaining optimal wavelet parameters. By introducing a stochastic gradient descent optimization process and using the Ricker wavelet as the initial parameter, the extracted statistical wavelet more closely reflects reality. This improvement significantly enhances the accuracy of wavelet extraction, leading to more accurate subsequent seismic inversion.

[0103] Second, this application expands the sample features and increases the sample size. By solving the perturbation function under different geological constraints, this application achieves the generation of a sand body overlay pattern library based on geological understanding. The randomly generated sand body overlay patterns with different geological significance can effectively expand the sample features and increase the sample size. This random expansion method based on geological understanding provides rich data support for reservoir property modeling and further improves the reliability of the model.

[0104] Third, the application effect of the model library is quantified. This application introduces correlation diagrams and index diagrams to measure the approximation of the overlay model library to subsurface geological bodies. These diagrams quantify the application effect of the model library, providing an intuitive and quantitative way to evaluate the accuracy and applicability of the overlay models, ensuring that the generated models accurately reflect the characteristics of subsurface geological bodies.

[0105] Fourth, innovative methodology. The entire intelligent prediction process for reservoir physical properties based on sand body stacking models establishes a new path for predicting reservoir physical property parameters. Through multi-data constraints, including the comprehensive utilization of geological, seismic, and well logging information, this technical process significantly improves the accuracy of reservoir physical property estimation. Practical application results show that this method can more accurately predict reservoir physical parameters, providing reliable data support for reservoir development and management.

[0106] Example 2:

[0107] Example 2 is a further refinement of Example 1.

[0108] like Figures 1a to 5b As shown, this invention discloses an intelligent prediction method for reservoir physical properties based on sand body stacking patterns, comprising the following steps:

[0109] Step 101': Preprocess the existing data.

[0110] A reservoir framework model was established using the reservoir modeling software iloop+. Based on the spatial coordinates of the reservoir framework model, the actual drilling logging curves were resampled to obtain model-scale logging curves at the well points. The logging sand body interpretation data was labeled, and sand body label data at the model scale was obtained based on the spatial coordinates of the well points in the reservoir framework model. Seismic data was also resampled based on the spatial coordinates of the well points in the reservoir framework model to obtain reservoir seismic attributes. Simultaneously, the reservoir seismic attributes were extracted based on the well point coordinates to obtain wellbore data at the model scale. After these operations, all data were unified to the reservoir model scale. At this point, the set of data containing logging curves, sand body labels, and wellbore amplitude data at the well points was defined as a complete sample data set. Through step 101', reservoir seismic attribute data and multiple complete sample data sets were obtained.

[0111] Step 102': Generate a sand body overlay pattern library.

[0112] Sand and mud stack together, forming different sand body stacking patterns. The data from the sample in step 101' also contains information about sand and mud stacking; therefore, each sample can be considered a sand body stacking pattern sample. From the sample data in step 101', well logging curves, such as sonic logging, density, porosity, and corresponding spatial coordinate data, are extracted. Using a piecewise function approximation method, a mapping function from spatial coordinates to sonic logging, density, and porosity is established.

[0113] Y = S(z) (1.1)

[0114] In Equation (1.1), z is the depth coordinate corresponding to the geological body along the well trajectory, S(·) is the mapping function from depth coordinate data to acoustic, density and porosity data, and Y is the numerical value corresponding to acoustic, density and porosity.

[0115] A perturbation function for coordinates is established, and by solving the perturbation function under different constraints, random overlay samples are generated. The perturbation function is specifically divided into two categories: linear perturbation functions and nonlinear perturbation functions. The formula for the linear perturbation function is:

[0116] Z = f(Z′) = aZ′ + b (1.2)

[0117] The formula for the nonlinear perturbation function is:

[0118] Z = f(Z') = aZ' 2 +bZ'+c (1.3)

[0119] In the formula, Z represents the depth coordinates of the original geological body, and Z' represents the depth coordinates of the geological body after its position has changed.

[0120] By solving different perturbation functions under different constraints, it is possible to generate sand body stacking pattern samples.

[0121] like Figure 1a As shown, if you want to replicate a sand body overlay pattern sample, you need to select a linear perturbation function, i.e., formula (1.2), and keep the spatial coordinates of the geological body unchanged. While keeping the initial depth of the sand-mud boundary point unchanged, you can replicate the sand body overlay pattern sample data by solving the linear perturbation function.

[0122] like Figure 1b As shown, if you want to uniformly change the thickness of the sand body in the stacked pattern, you can use the linear perturbation function, i.e., formula (1.2). By changing the initial depth of the sand-mud boundary point and then solving the perturbation function, you can achieve a uniform change in the thickness of the sand body in the stacked pattern.

[0123] like Figure 1c As shown, if you want to add a sand body to an existing overlay pattern, you can use a nonlinear perturbation function, i.e., formula (1.3). By changing the initial depth of the sand-mud boundary point and randomly generating a perturbation point inside the original boundary point, you can solve the perturbation function and realize the process of adding a sand body to the sand body overlay pattern sample.

[0124] like Figure 1d As shown, similarly, if you want to delete a segment of the sand body in the overlay mode, you can select the nonlinear perturbation function, i.e., formula (1.3), and randomly generate perturbation points inside the original boundary points. In this process, the depth of the perturbation points is taken to be outside the depth range of the overlay mode, so as to achieve the effect of eroding a segment of the sand body.

[0125] The overall process of generating the sand body overlay model library first requires randomly selecting a perturbation function, then randomly generating the initial depth of the modified sand-mud boundary, followed by randomly generating perturbation points, and finally solving for the perturbation function. According to this process, the sample data in step 101' is used as input, and the output is a large number of randomly generated sand body overlay model sample data. It is worth noting that the sand body overlay model samples generated in this step do not include amplitude data; therefore, the generated sample data is incomplete.

[0126] Step 103', Data-driven seismic wavelet extraction.

[0127] Extract wellbore amplitude, acoustic wave, and density data from the complete sample data in step 101'. Based on the acoustic wave and density data, the reflection coefficient sequence data can be obtained, using the following formula:

[0128]

[0129] In equation (1.4), ρ i This represents the density of the i-th layer of medium, expressed in g / cm³. 3 v i R represents the velocity of the i-th layer of medium, which is the reciprocal of the sound wave velocity, and its unit is m / s. i This represents the reflection coefficients of the i-th layer and the (i+1)-th layer.

[0130] Statistical wavelet extraction requires the use of reflection coefficients and wellbore data. In the data-driven wavelet extraction process, the Ricker wavelet is first used as the initial parameter of the convolution kernel, and a function from the reflection coefficient to the amplitude is constructed according to the convolution formula:

[0131]

[0132] In equation (1.5), s t It's amplitude data, r t It is the reflection coefficient obtained earlier, and N represents the wavelet ω. τ Length, ω τ The wavelet data to be extracted is initially set to the Ricker wavelet, and its formula is as follows:

[0133]

[0134] The data-driven wavelet extraction process uses the actual coarsened wellbore as a result constraint, and then employs gradient descent to optimize the initial results. The gradient descent method is as follows:

[0135]

[0136] In equation (1.7), ω tLet η represent the value of the wavelet in the t-th iteration, and let η represent the learning rate. Describe the objective function The first derivative with respect to the wavelet parameter ω is the gradient. By iteratively updating the wavelet parameter, the optimized wavelet is finally obtained.

[0137] like Figure 2 As shown, it illustrates the difference between the initial wavelet and the optimized wavelet.

[0138] Step 104', Forward modeling of sand body overlay mode samples.

[0139] The numerous sand body overlay model samples generated in step 102' lack amplitude data. Therefore, the statistical wavelet obtained in step 103' is used to supplement the amplitude data through forward modeling according to formulas (1.4) and (1.5). The supplemented sample data is then complete. After this step, each sample in the constructed sand body overlay model library contains logging curves, sand body labels, and amplitude data.

[0140] Step 105': Construct the related graph and index graph.

[0141] The sand body overlay model library after step 104' was selected, and amplitude data was extracted from each sample in the model library. Correlation analysis was performed with the longitudinal amplitude data of each trace of reservoir seismic attributes obtained in step 101'. The formula for calculating the correlation coefficient is as follows:

[0142]

[0143] In Equation (1.8), X and Y are the amplitude data of the overlaid sample and the longitudinal amplitude data of each trace of the reservoir seismic attribute, respectively.

[0144] The overlay sample with high amplitude correlation is used to approximate the actual sand body overlay situation of the trace. The correlation coefficient of the overlay sample, its index in the sand body overlay model library, and the corresponding coordinates are recorded. This process is repeated until all traces are calculated, forming a correlation plot and an index plot.

[0145] like Figure 3a As shown, the correlation diagram can measure the representativeness of the model library to the complex geological conditions of the work area.

[0146] like Figure 3b The image shown is an index diagram.

[0147] Step 106': Select high-quality training samples.

[0148] The correlation map obtained in step 105' records the representativeness of the model library for the work area and can provide structural information of complex geological bodies. Therefore, points with correlation coefficients greater than 0.9 are selected from the correlation map, and then the sample data of the point in the model library are found in combination with the index map. The resulting high correlation coefficient samples constitute high-quality training samples that reflect the structural information of actual complex geological bodies.

[0149] Step 107': Train the neural network model.

[0150] The high-quality training samples obtained in step 106' are used for training the deep neural network model.

[0151] like Figure 4 The diagram shows the structure of the neural network, which employs a convolutional network. The short arrow pointing to the right represents an equivalent convolutional layer with 64 kernels; the short arrow pointing downwards represents a max-pooling layer; the short arrow pointing upwards represents an upsampling layer; and the long arrow pointing to the right represents copying the data. During training, amplitude data is extracted from the sand body stacked samples as input to the neural network, and qualitative parameters are extracted from the sand body stacked samples as labels. The mean squared error loss function is selected.

[0152]

[0153] In equation (1.9), N is the output length of the neural network. y is the label sequence, and y is the output sequence of the neural network.

[0154] Stochastic gradient descent was chosen as the optimizer, and the optimization process of the gradient descent method for the parameters is shown in Equation (1.7). After training, a neural network model with converged parameters was obtained.

[0155] Step 108': Predict reservoir physical properties.

[0156] The converged network model in step 107' is selected, and the reservoir seismic attribute data in step 101' is input into the network model. The output obtained is the reservoir physical property obtained by this scheme.

[0157] like Figure 5a As shown, the results of the three-dimensional reservoir porosity properties predicted using the scheme of this application are presented.

[0158] like Figure 5b As shown, the prediction results match the actual well porosity on the well profile. It can be seen that the prediction results match the actual results very well, with a high accuracy.

[0159] Example 3:

[0160] Example 3 is a further refinement of Example 2. This intelligent prediction method for reservoir properties based on sand body stacking patterns is applicable to reservoir property prediction based on deep learning. It can effectively expand the features of training samples, reduce the high requirements of deep learning methods on datasets, and improve the model's ability to predict complex geological conditions.

[0161] This invention discloses an intelligent prediction method for reservoir physical properties based on sand body stacking patterns, comprising the following steps:

[0162] Step 101”: Establish a reservoir grid model based on the target layer and faults in the work area, and then obtain resampling data from 482 wells. The resampling data includes various well curves, well-side waveform curves, layer labels and sand body labels, etc., forming complete sample data.

[0163] Step 102”: Randomly select the perturbation function, change the initial depth of the sand-sludge boundary and the perturbation point, and solve for the perturbation function. Randomly perturb the dataset of 482 coarsened wells using the perturbation function to generate 20,000 different sand body stacking pattern samples, forming a pattern library.

[0164] Step 103” extracts the density and acoustic wave curve sequences from the complete sample data, calculates the wave impedance, and then obtains the reflection coefficient sequence. Based on the wellbore waveform and the corresponding reflection coefficient sequence, the seismic wavelet is extracted using the above method.

[0165] Step 104” uses the seismic wavelet extracted in Step 103” to perform convolution forward modeling on the sand body stacking model sample to supplement the amplitude data of the sand body stacking model sample.

[0166] Step 105” extracts the amplitude data of the sand body superposition pattern library sample and performs correlation analysis with the amplitude data of each trace of the reservoir seismic attribute data to form the correlation map and index map of the work area.

[0167] Step 106”: Select the correlation coefficient threshold, and filter out sand body superposition pattern samples with complex geological structure information based on the correlation map and index map to form training samples.

[0168] Step 107” uses the selected high-quality training samples for training the deep neural network model. The input is the sample amplitude data, and the label is the sample physical property parameter.

[0169] Step 108”: After the network converges, the reservoir seismic attribute data is used as the input to the neural network model to obtain the reservoir physical property data.

[0170] Practical applications show that the property model obtained by intelligent prediction of reservoir properties based on sand body stacking pattern is significantly better than the property model obtained by training the network with only 482 original wells.

[0171] Example 4:

[0172] Example 4 is a further refinement of Example 2. This intelligent prediction of reservoir properties based on sand body stacking patterns is applicable to predicting the three-dimensional porosity of this work area. This work area has only 6 actual drilled wells. Using this method, 10,000 samples were effectively expanded and used for deep learning training, thereby predicting the porosity property model of this work area.

[0173] This invention discloses an intelligent prediction method for reservoir physical properties based on sand body stacking patterns, comprising the following steps:

[0174] Step 101”': Based on the 10 seismic horizons and 11 fault models provided for this work area, a reservoir framework model is established. The size of the framework model is 298x196x908. Based on the spatial coordinates of the reservoir framework model, well logging curves, seismic amplitude data, layer labels, and sand body labels can be resampled to form complete sample data according to the well trajectory;

[0175] Step 102”': Randomly select the perturbation function, change the initial depth of the sand-mud boundary and the perturbation point, and solve for the perturbation function. Randomly perturb the complete sample data in Step 101”' using the perturbation function to generate 10,000 different sand body stacking pattern samples, forming a sand body stacking pattern library.

[0176] Step 103””: Extract the density and acoustic wave curve sequences from the complete sample data, calculate the wave impedance, and then obtain the reflection coefficient sequence. Based on the wellbore waveform and the corresponding reflection coefficient sequence, extract the seismic wavelet using the above method.

[0177] Step 104”': Using the seismic wavelet extracted in step 103”', perform forward modeling on each sample in the model library to supplement the amplitude data of the sand body overlay model samples.

[0178] Step 105””: Extract the amplitude data of the sand body overlay model library sample and perform correlation analysis with the amplitude data of each trace of the reservoir seismic attribute data to form the correlation map and index map of the work area.

[0179] Step 106”': Select the correlation coefficient threshold, and filter out sand body superposition pattern samples with complex geological structure information based on the correlation map and index map to form training samples.

[0180] Step 107”': The selected high-quality training samples are used for training the deep neural network model. The input is the sample amplitude data and the label is the sample physical property parameter.

[0181] Step 108”' After the network converges, the reservoir seismic attribute data is used as the input to the neural network model to obtain the reservoir physical property data.

[0182] The porosity model obtained based on this method matches the actual seismic waveform trend, and its effect is significantly better than the result based on the two-point geostatistical method.

[0183] Example 5:

[0184] Example 5 is a further refinement of Example 2. The intelligent prediction method for reservoir properties based on sand body stacking pattern is applicable to the prediction of the three-dimensional porosity model of this work area. There are only 8 actual drilled wells in this work area, of which 7 are usable. According to this method, 8000 samples are effectively expanded for deep learning, thereby predicting the porosity property model of this work area.

[0185] This invention discloses an intelligent prediction method for reservoir physical properties based on sand body stacking patterns, comprising the following steps:

[0186] Step 101””: Based on the two seismic horizons provided by the work area, the interlayer data is interpolated into an equal-length sequence to simulate the reservoir grid. Then, based on the well-seismic calibration results, the corresponding range of logging data is selected for resampling, and the corresponding logging curves, well-side waveform curves, layer labels and sand body labels are extracted to form complete sample data.

[0187] Step 102””: Randomly select a perturbation function, change the initial depth of the sand-mud boundary and the perturbation point, and solve for the perturbation function. Generate 8000 different sand body stacking pattern samples through the perturbation function to form a pattern library.

[0188] Step 103””, select the corresponding parameters of the Lake wavelet based on the well vibration calibration results.

[0189] Step 104”” uses the Reck wavelet from Step 103”” to perform convolution forward modeling on each dataset in the model library to supplement the amplitude data of the sand body superimposed model samples.

[0190] Step 105””: Extract the amplitude data of the sand body superposition pattern library sample, and perform correlation analysis with the inter-layer seismic amplitude data to form the correlation map and index map of the work area.

[0191] Step 106””, select the correlation coefficient threshold, and filter out sand body superposition pattern samples with complex geological structure information based on the correlation map and index map to form high-quality training samples.

[0192] Step 107”” uses the selected high-quality training samples for training the deep neural network model, with amplitude data as input and porosity data as label.

[0193] Step 108”” After the network converges, the actual inter-layer seismic data is resampled to the same length and used as the output of the neural network. The output of the neural network is then interpolated back to the original length of the inter-layer seismic data to finally obtain the porosity model.

Claims

1. A method for intelligent prediction of reservoir physical properties based on sand body stacking patterns, characterized in that: Includes the following steps, Step 101: Establish and obtain the reservoir framework model. Based on the reservoir framework model, resample the logging curves, sand body labels, and seismic data respectively to obtain the logging curves, sand body labels, and reservoir seismic attributes at the model scale at the well points. According to the well point coordinates, extract the wellbore amplitude from the reservoir seismic attributes. The logging curves, sand body labels, and wellbore amplitude at the model scale at the well points form complete sample data. Step 102: Randomly select a disturbance function. Based on the sample data obtained in Step 101, randomly generate the starting depth and disturbance point of the sand-sludge boundary after disturbance. Solve the disturbance function and resample the disturbance function according to the set sampling interval. Substitute the resampled value into the mapping function of the logging curve to generate a sample of each sand body stacking pattern, forming a sand body stacking pattern library. The sand body stacking pattern samples do not contain amplitude data. Step 103: Extract well logging and well access data from the complete sample data in Step 101, and optimize the initial Lake wavelet using the stochastic gradient descent method to obtain the well access statistical wavelet; Step 104: Using the wellside statistical wavelet from Step 103, perform forward modeling on the sand body overlay pattern samples generated in Step 102 to obtain forward-modeled sand body overlay pattern samples, forming a forward-modeled sand body overlay pattern library. Step 105: Based on the forward modeling sand body overlay model library obtained in Step 104, extract the amplitude data of each sample; based on the reservoir seismic attributes obtained in Step 101, extract the vertical amplitude data of each trace; perform correlation analysis on the two types of amplitude data obtained, record the correlation coefficient and index value of the amplitude data of samples with high correlation values, and after traversing all traces, form a correlation map and an index map. Step 106: Obtain the correlation coefficient threshold. Select sand body stacking pattern samples that are above the correlation coefficient threshold from the correlation map obtained in step 105 to form high-quality training samples. Step 107: Use the high-quality training samples obtained in step 106 as the training set to train the neural network until it converges, and obtain the trained neural network. Step 108: Input the reservoir seismic attributes obtained in step 101 into the trained neural network obtained in step 107. The neural network outputs the predicted reservoir physical properties.

2. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 1, characterized in that: In step 101, the logging curves are resampled based on the spatial coordinates of the reservoir grid model to obtain the logging curves at the model scale at the well points; the logging sand body interpretation data are labeled, and sand body labels at the model scale are obtained based on the spatial coordinates of the well points at the reservoir grid model; the seismic data are resampled based on the spatial coordinates of the well points at the reservoir grid model to obtain the reservoir seismic attributes.

3. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 1, characterized in that: In step 102, the perturbation function is a linear perturbation function or a nonlinear perturbation function, the depth sampling interval is 0.125 meters, and the mapping function is the logging curve extracted from the sample data obtained in step 101. The logging curve includes acoustic wave, density, porosity and corresponding spatial coordinate data. Using the piecewise function approximation method, the mapping function from spatial coordinates to acoustic wave, density and porosity is established, i.e., Equation (1.1). Y = S(z)(1.1) In equation (1.1), z is the depth coordinate corresponding to the geological body along the well trajectory, S(·) is the mapping function from depth coordinate data to acoustic, density and porosity data, and Y is the numerical value corresponding to acoustic, density and porosity. The formula for the linear perturbation function is: Z = f(Z) = aZ′ + b (1.2) The formula for the nonlinear perturbation function is: Z=f(Z')=aZ' 2 +bZ'+c(1.3) In the formula, Z represents the depth coordinates of the original geological body, and Z' represents the depth coordinates of the geological body after its position has changed.

4. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 3, characterized in that: In step 102, When replicating the sand body overlay pattern sample, a linear perturbation function, i.e., formula (1.2), is used, and the spatial coordinates of the geological body remain unchanged. While keeping the initial depth of the sand-mud boundary point unchanged, the linear perturbation function is solved to realize the process of replicating the sand body overlay pattern sample data. When the thickness of the sand body in the stacked pattern is uniformly changed, a linear perturbation function, i.e., formula (1.2), is used. By changing the initial depth of the sand-mud boundary point, the perturbation function is solved to achieve the uniform change process of the sand body thickness in the stacked pattern. When adding a sand body within an existing overlay pattern, a nonlinear perturbation function, i.e., formula (1.3), is used. By changing the initial depth of the sand-mud boundary point and randomly generating a perturbation point inside the original boundary point, the perturbation function is solved to realize the process of adding a sand body in the sand body overlay pattern sample. When deleting a segment of the sand body in the overlay mode, a nonlinear perturbation function, i.e., formula (1.3), is used to randomly generate perturbation points inside the original boundary points. The depth of the perturbation points is taken to be outside the depth range of the overlay mode, thereby eroding away a segment of the sand body.

5. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 1, characterized in that: In step 103, wellbore amplitude, acoustic wave, and density data are extracted from the complete sample data; reflection coefficient sequence data are obtained based on the acoustic wave and density data, using the following formula: In equation (1.4), ρ i This represents the density of the i-th layer of medium, expressed in g / cm³. 3 v i R represents the velocity of the i-th layer of medium, which is the reciprocal of the sound wave velocity, and its unit is m / s. i This represents the reflection coefficients of the i-th layer and the (i+1)-th layer; Statistical wavelet extraction utilizes reflection coefficients and wellbore data. In the data-driven wavelet extraction process, the Ricker wavelet is used as the initial parameter of the convolution kernel, and a function from the reflection coefficient to the amplitude is constructed according to the convolution formula. In equation (1.5), s t It's amplitude data, r t It is the reflection coefficient obtained earlier, and N represents the wavelet ω. τ Length, ω τ This is the wavelet data to be extracted, with the initial value being the Ricker wavelet, and the formula is as follows:

6. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 5, characterized in that: In step 103, based on the data-driven wavelet extraction process, the actual coarsened well bypass is used as the result constraint, and the gradient descent method is employed to optimize the initial result. The gradient descent method is as follows: In equation (1.7), ω t Let η represent the value of the wavelet in the t-th iteration, and let η represent the learning rate. Describe the objective function The first derivative of the wavelet parameter ω is the gradient; by iteratively updating the wavelet parameter, an optimized wavelet is obtained.

7. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 5, characterized in that: In step 104, the sand body stacking pattern sample generated in step 102 lacks amplitude data. Using the statistical wavelet obtained in step 103, the amplitude data is supplemented by forward modeling according to formulas (1.4) and (1.5). The sample data obtained after the supplementation is complete sample data. After this step, each sample in the constructed sand body stacking pattern library contains logging curves, sand body labels and amplitude data.

8. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 1, characterized in that: In step 105, the formula for calculating the correlation coefficient is as follows: In Equation (1.8), X and Y are the amplitude data of the overlaid sample and the longitudinal amplitude data of each trace of the reservoir seismic attribute, respectively. The overlay sample with high amplitude correlation is used to approximate the actual sand body overlay situation of the corresponding trace. The correlation coefficient of the overlay sample, its index in the sand body overlay model library, and the corresponding coordinates are recorded. This process is repeated until all traces are calculated to form a correlation map and an index map.

9. The intelligent prediction method for reservoir properties based on sand body stacking pattern according to claim 1, characterized in that: In step 106, the correlation coefficient threshold is 0.

9. Points with a correlation coefficient greater than 0.9 are selected from the correlation map. The corresponding points are then used to find sample data in the sand body overlay pattern library. This sample data is the high correlation coefficient sample, which constitutes a high-quality training sample that reflects the structural information of the actual complex geological body.

10. The intelligent prediction method for reservoir physical properties based on sand body stacking pattern according to claim 1, characterized in that: In step 107, the neural network is a deep neural network. During training, amplitude data is extracted from the sand body stacked samples as input to the neural network, and qualitative parameters are extracted from the sand body stacked samples as labels. The mean squared error loss function is selected. In equation (1.9), N is the output length of the neural network. Let y be the label sequence and y be the output sequence of the neural network; Stochastic gradient descent was chosen as the optimizer.

Images