Three-dimensional reservoir attribute model resolution reconstruction method and device based on perception guidance generation adversarial network, electronic equipment and storage medium

Abstract

The present application relates to the technical field of reservoir geological modeling, and is a three-dimensional reservoir attribute model resolution reconstruction method and device based on a perception-guided generative adversarial network, an electronic device, and a storage medium. The method comprises the following steps: obtaining three groups of two-dimensional low-resolution multi-attribute slices; inputting the three groups of two-dimensional low-resolution multi-attribute slices into a resolution reconstruction model to obtain three groups of two-dimensional high-resolution multi-attribute slices, wherein the resolution reconstruction model is obtained by deep learning of a Geo-Real ESRGAN model through a plurality of samples; and decoding and fusing the three groups of two-dimensional high-resolution multi-attribute slices to obtain a high-resolution three-dimensional reservoir attribute model. The two-dimensional high-resolution multi-attribute slices predicted and output by the resolution reconstruction model are more superior in structure restoration and attribute consistency, so that the accuracy of the reconstructed high-resolution three-dimensional reservoir attribute model is higher, and reliable technical support can be provided for fine characterization and engineering design of unconventional reservoirs.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of reservoir geology modeling, and is a three-dimensional reservoir attribute model resolution reconstruction method and device based on a perception-guided generative adversarial network, an electronic device and a storage medium. BACKGROUND

[0002] A high-resolution three-dimensional reservoir attribute model is very important for the development of unconventional oil and gas reservoirs, and accurate characterization of reservoir attributes directly affects the design of drilling and completion strategies and significantly affects the effect of hydraulic fracturing.

[0003] The commonly used three-dimensional modeling method mainly relies on Kriging interpolation and geostatistical simulation methods, which have the advantages of mature theory and standardized process and are widely used in conventional reservoir modeling. However, they have obvious limitations in unconventional reservoirs. Specifically, on the one hand, the Kriging-based method relies on two-point statistical characteristics (such as the variogram function), which cannot fully capture the nonlinear and multiscale spatial structure of complex geological bodies. When applied to reservoir characteristics with dramatic spatial changes, it often produces overly smoothed attribute distributions, thereby weakening key structural and heterogeneous features. On the other hand, although random simulation methods such as sequential Gaussian simulation (SGS) and sequential indicator simulation (SIS) can quantify uncertainty, they have poor applicability for unconventional reservoir characteristics that exhibit non-Gaussian, multimodal, or asymmetric distributions. This mismatch often introduces simulation bias, which cannot meet the dual requirements of spatial precision and geological consistency in engineering practice. In addition, with the expansion of the modeling area and the refinement of the grid, the computational efficiency of traditional modeling methods also faces challenges. The computational complexity increases exponentially, and the iterative optimization process often faces challenges such as memory bottlenecks and excessive time consumption, making it difficult to meet the dual demands of rapid response and high-precision modeling in actual engineering.

[0004] In summary, the existing three-dimensional reservoir attribute model generally has low spatial resolution, which makes it difficult to support fine modeling of strong spatially heterogeneous reservoirs represented by shale and tight sandstone, thereby restricting the accuracy of reservoir evaluation and the effectiveness of engineering decision-making. Therefore, it is necessary to perform resolution reconstruction on the low-resolution three-dimensional reservoir attribute model to obtain a high-resolution three-dimensional reservoir attribute model.

[0005] In recent years, with the rapid development of artificial intelligence, breakthroughs in deep learning technology in the field of computer vision have provided new ideas for solving this problem. For example:

[0006] Song et al.developed a GANSim-3D framework based on a progressive GAN (PGGAN) that can automatically generate high-fidelity 3D facies structures using sparse well facies data and seismic probability maps (Song et al., 2022). This 3D GAN modeling method mainly focuses on the high-fidelity generation of facies structures, making it difficult to meet the actual scenarios based on the resolution improvement of the original model.

[0007] The existing patent document one with the publication number CN119832175A discloses a reservoir geological modeling method based on a progressive conditional generative adversarial network, relating to the field of reservoir geological modeling. The method includes: obtaining geological images and well logging data; constructing training sample pairs according to the geological images and well logging data; constructing a progressive conditional generative adversarial network based on well logging data; training the progressive conditional generative adversarial network through the training sample pairs combined with the progressive training method of resolution step-by-step improvement; obtaining the well logging data to be predicted; inputting the well logging data to be predicted into the progressive conditional generative adversarial network to generate a reservoir image that conforms to both the geological pattern and the well logging data constraints. The invention realizes high-precision reservoir geological modeling based on a generative adversarial network, but still has obvious deficiencies in three-dimensional spatial consistency, multi-attribute collaborative optimization, etc., making it difficult to be directly applied to multi-attribute, strongly heterogeneous unconventional reservoir modeling tasks.

[0008] The existing patent document two with the publication number CN108335263A discloses a microwave remote sensing image super-resolution reconstruction method based on VDSR, belonging to the field of microwave remote sensing and detection technology. The invention first simulates the microwave remote sensing process to generate a high-resolution microwave remote sensing image TB and a low-resolution microwave remote sensing image TA to form a data set; then pre-processes the data set to generate a VDSR training set; then constructs a deep convolutional neural network based on the VDSR training set; finally inputs the low-resolution microwave remote sensing image to be processed into the constructed deep convolutional neural network, and the output of the stacking layer of the deep convolutional neural network is the reconstructed high-resolution microwave remote sensing image. The invention realizes microwave remote sensing image super-resolution reconstruction, but is not used for three-dimensional reservoir attribute models, so there are still obvious deficiencies in three-dimensional spatial consistency, multi-attribute collaborative optimization, etc., making it difficult to be directly applied to multi-attribute, strongly heterogeneous unconventional reservoir modeling tasks. SUMMARY

[0009] The present application provides a three-dimensional reservoir attribute model resolution reconstruction method and device based on perception-guided generative adversarial networks, electronic equipment and storage medium, which overcomes the shortcomings of the prior art. It effectively solves the problem of not being able to consider three-dimensional spatial consistency, multi-attribute collaborative optimization, etc. when performing three-dimensional reservoir attribute model high-resolution recovery.

[0010] One of the technical solutions of this invention is achieved through the following measures: a method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network, comprising:

[0011] Three sets of two-dimensional low-resolution multi-attribute slices of a low-resolution three-dimensional reservoir property model were obtained. The three sets of two-dimensional low-resolution multi-attribute slices were obtained by processing the low-resolution three-dimensional reservoir property model through a three-way orthogonal slicing strategy and a multi-channel image encoding strategy.

[0012] Input three sets of two-dimensional low-resolution multi-attribute slices into the resolution reconstruction model to obtain three sets of two-dimensional high-resolution multi-attribute slices. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model through several samples. Each sample includes three sets of two-dimensional low-resolution multi-attribute slices of the three-dimensional reservoir attribute model and the label information of the corresponding three sets of two-dimensional high-resolution multi-attribute slices.

[0013] The three sets of two-dimensional high-resolution multi-attribute slices were decoded and fused to obtain a high-resolution three-dimensional reservoir attribute model.

[0014] The following are further optimizations and / or improvements to the above-mentioned technical solution:

[0015] The construction process of the above resolution reconstruction model includes:

[0016] Obtain at least one original three-dimensional reservoir attribute model and obtain the corresponding sample, wherein the sample includes three sets of two-dimensional low-resolution multi-attribute slices and the label information of the corresponding three sets of real two-dimensional high-resolution multi-attribute slices;

[0017] Data augmentation is performed on the samples to obtain multiple samples, which are then divided into training and test sets according to a certain ratio.

[0018] The Geo-RealESRGAN model is trained using a training set. A loss function is introduced during training, and training ends when the value of the loss function is stable, resulting in a resolution reconstruction model. The Geo-RealESRGAN model uses a generative adversarial network mechanism. The generator restores each set of input 2D low-resolution multi-attribute slices to predicted 2D high-resolution multi-attribute slices. The discriminator uses patch-wise discrimination to determine the detailed differences between the real 2D high-resolution multi-attribute slices and the predicted 2D high-resolution multi-attribute slices, and feeds back fine-grained spatial supervision signals to the generator.

[0019] The trained resolution reconstruction model is tested using a test set, the model parameters are optimized, and a resolution reconstruction model that meets the test evaluation requirements is output.

[0020] The above-mentioned acquisition of at least one original three-dimensional reservoir property model to obtain corresponding samples includes:

[0021] Obtain the original three-dimensional reservoir property model;

[0022] Using a three-dimensional orthogonal slicing strategy, two-dimensional slicing operations were performed on the original three-dimensional reservoir attribute model in three orthogonal directions (XY, YZ, and XZ) to obtain three sets of real two-dimensional high-resolution multi-attribute slices.

[0023] Three sets of real two-dimensional high-resolution multi-attribute slices were downsampled to obtain corresponding two-dimensional low-resolution multi-attribute slices. Samples were constructed using the LR-HR image pairing strategy. The samples included the label information of the three sets of two-dimensional low-resolution multi-attribute slices and the corresponding three sets of real two-dimensional high-resolution multi-attribute slices.

[0024] After linear normalization of the samples, a multi-channel image coding strategy is used to encode three sets of two-dimensional low-resolution multi-attribute slices in the samples.

[0025] The aforementioned enhancement processes include orientation transformation, cropping and scaling, blurring perturbation, and layer perturbation simulation.

[0026] The loss function is shown below:

[0027]

[0028] in, , , Weights for each loss; , , These are pixel reconstruction loss, perceptual loss, and adversarial loss, respectively.

[0029] The three sets of two-dimensional low-resolution multi-attribute slices used to obtain the low-resolution three-dimensional reservoir attribute model mentioned above include:

[0030] A low-resolution three-dimensional reservoir property model was obtained, and a three-dimensional orthogonal slicing strategy was used to perform two-dimensional slicing operations in three orthogonal directions (XY, YZ, and XZ) to obtain three sets of two-dimensional low-resolution multi-attribute slices.

[0031] After linear normalization of the three sets of low-resolution attribute slices, the corresponding three sets of two-dimensional low-resolution multi-attribute slices are obtained by using a multi-channel image coding strategy.

[0032] The above process decodes and fuses three sets of two-dimensional high-resolution multi-attribute slices to obtain a high-resolution three-dimensional reservoir attribute model, including:

[0033] Each set of two-dimensional high-resolution multi-attribute slices is subjected to multi-channel image decoding and inverse normalization.

[0034] Based on coordinate mapping lookup, each set of two-dimensional high-resolution multi-attribute slices is restored into a set of three-dimensional attribute point data, where each three-dimensional attribute point data includes three-dimensional physical coordinates (x, y, z) and its predicted attribute value;

[0035] A multi-source prediction fusion mechanism is introduced to weight and integrate multiple predicted attribute values ​​for each 3D coordinate point to obtain the final predicted attribute value.

[0036] A high-resolution three-dimensional reservoir attribute model is established based on the final predicted attribute values ​​of each three-dimensional coordinate point.

[0037] The second technical solution of the present invention is achieved through the following measures: a three-dimensional reservoir attribute model resolution reconstruction device based on a perception-guided generative adversarial network, comprising:

[0038] The low-resolution multi-attribute slice acquisition unit acquires three sets of two-dimensional low-resolution multi-attribute slices of the low-resolution three-dimensional reservoir attribute model. The three sets of two-dimensional low-resolution multi-attribute slices are obtained by processing the low-resolution three-dimensional reservoir attribute model through a three-way orthogonal slicing strategy and a multi-channel image encoding strategy.

[0039] The resolution reconstruction unit takes three sets of two-dimensional low-resolution multi-attribute slices as input to the resolution reconstruction model and obtains three sets of two-dimensional high-resolution multi-attribute slices. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model through several samples. Each sample includes three sets of two-dimensional low-resolution multi-attribute slices of the three-dimensional reservoir attribute model and the label information of the corresponding three sets of two-dimensional high-resolution multi-attribute slices.

[0040] The high-resolution three-dimensional reservoir attribute model forming unit decodes and fuses three sets of two-dimensional high-resolution multi-attribute slices to obtain a high-resolution three-dimensional reservoir attribute model.

[0041] The third technical solution of the present invention is achieved through the following measures: an electronic device, including a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps in the method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network.

[0042] The fourth technical solution of the present invention is achieved through the following measures: a storage medium storing a computer program that can be read by a computer, the computer program being configured to execute the steps in the resolution reconstruction method of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network when it is run.

[0043] This invention obtains a resolution reconstruction model by deep learning the Geo-RealESRGAN model. The Geo-RealESRGAN model is an improved RealESRGAN model. Based on the Real-ESRGAN framework, this model combines the spatial distribution patterns and structural features of attribute images in reservoir modeling tasks, introducing multi-scale feature extraction, spatial context modeling, and local structure perception mechanisms to achieve high-precision reconstruction of sparse, low-resolution geological slices. Unlike traditional high-resolution image methods that pursue visual clarity, this model first maps each reservoir attribute to a multi-channel image. Then, while retaining the original perceptual loss and adversarial loss structure, it introduces pixel-wise L1 loss to improve attribute recovery accuracy. Furthermore, it intentionally eliminates visual optimization strategies such as color enhancement, ensuring that the network focuses more on the numerical reconstruction of the attributes themselves. Finally, the model performance is evaluated through attribute value error rather than image perceptual metrics, which better aligns with the actual needs of reservoir modeling. Therefore, the two-dimensional high-resolution multi-attribute slices predicted by the resolution reconstruction model obtained by deep learning on the Geo-RealESRGAN model are superior in terms of structure restoration and attribute consistency, resulting in higher accuracy of the reconstructed high-resolution three-dimensional reservoir attribute model, which can provide reliable technical support for the fine characterization and engineering design of unconventional reservoirs. Attached Figure Description

[0044] Appendix Figure 1 This is a schematic diagram of the implementation environment provided for an embodiment of the present invention.

[0045] Appendix Figure 2 This is a schematic diagram of the resolution reconstruction method for three-dimensional reservoir property model provided in Embodiment 1 of the present invention.

[0046] Appendix Figure 3 This is a flowchart of a method for obtaining three sets of two-dimensional low-resolution multi-attribute slice images in a low-resolution three-dimensional reservoir attribute model, as provided in Embodiment 2 of the present invention.

[0047] Appendix Figure 4 This is a flowchart of the resolution reconstruction model construction method provided in Embodiment 3 of the present invention.

[0048] Appendix Figure 5 This is a schematic diagram of the sample acquisition method provided in Embodiment 3 of the present invention.

[0049] Appendix Figure 6 This is a schematic diagram of the data augmentation process provided in Embodiment 3 of the present invention.

[0050] Appendix Figure 7 This is a schematic diagram of the Geo-RealESRGAN model structure provided in Embodiment 3 of the present invention.

[0051] Appendix Figure 8This is a schematic diagram of the three-dimensional reservoir property model resolution reconstruction device provided in Embodiment 5 of the present invention. Detailed Implementation

[0052] The present invention is not limited to the following embodiments, and the specific implementation can be determined according to the technical solution of the present invention and the actual situation.

[0053] Those skilled in the art will understand that, unless specifically stated otherwise, in the embodiments of the present invention, a "module" or "unit" refers to a computer program or part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0054] In addition, in the embodiments of the present invention, "multiple" refers to two or more, and "first" and "second" are used to distinguish descriptions and should not be construed as implying relative importance.

[0055] This invention provides a method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network. The method obtains a resolution reconstruction model by performing deep learning on a Geo-RealESRGAN model using several samples. Three sets of two-dimensional low-resolution multi-attribute slices of the low-resolution three-dimensional reservoir attribute model are obtained and input into the resolution reconstruction model to obtain three sets of two-dimensional high-resolution multi-attribute slices. The three sets of two-dimensional high-resolution multi-attribute slices are then decoded and fused to obtain a high-resolution three-dimensional reservoir attribute model.

[0056] The method provided in this embodiment of the invention may involve artificial intelligence (AI) technology and may be implemented based on artificial intelligence technology, such as using deep learning to train a corresponding model using samples.

[0057] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0058] Deep learning (DL) specifically refers to machine learning based on deep neural network models and methods. It has developed from statistical machine learning, artificial neural network algorithms, and other algorithms, combined with the advancements in big data and computing power. The most important technical feature of deep learning is its ability to automatically extract features.

[0059] The aforementioned machine learning and deep learning typically include techniques such as neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0060] In deep learning, the loss function is used to predict the target value by comparing the predicted value with the target value. This is done by updating the weight vector of each layer of the neural network based on the difference between the two values ​​(usually with an initialization process before the first update, where parameters are pre-configured for each layer) until the network can predict the target value or a value very close to it. Therefore, deep learning requires pre-defining "how to compare the difference between the predicted value and the target value," which is the loss function.

[0061] As attached Figure 1 The diagram illustrates an implementation environment provided by an embodiment of the present invention. This implementation environment may include: training equipment and usage equipment.

[0062] Both the training equipment and the equipment used are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, tablet computer, PC (Personal Computer) or other electronic devices; or, the computer device is a server, which can be a single server, a server cluster composed of multiple servers, or a cloud computing service center. This embodiment of the invention does not limit this.

[0063] Training equipment refers to computer equipment capable of training and learning Geo-RealESRGAN models. Optionally, the training equipment has the ability to acquire Geo-RealESRGAN models and train and learn them according to application requirements. For example, the training equipment acquires a Geo-RealESRGAN model from other devices via a network and then trains it using training samples according to application requirements, enabling the Geo-RealESRGAN model to generate three sets of two-dimensional high-resolution multi-attribute slices. Optionally, the training equipment has the ability to build Geo-RealESRGAN models. It can build a Geo-RealESRGAN model itself according to application requirements and then train and learn it. For example, to generate three sets of two-dimensional high-resolution multi-attribute slices from three sets of two-dimensional low-resolution multi-attribute slices, the training equipment builds a Geo-RealESRGAN model and then trains and learns it using samples according to application requirements.

[0064] The device used refers to a computer device that has the capability to use the Geo-RealESRGAN model. Optionally, the device uses the Geo-RealESRGAN model from other devices via the network according to application requirements. For example, if the device has the capability to obtain three sets of two-dimensional high-resolution multi-attribute slices, it can obtain the Geo-RealESRGAN model that has been trained and learned to obtain three sets of two-dimensional high-resolution multi-attribute slices from other devices via the network, and use the Geo-RealESRGAN model to obtain the three sets of two-dimensional high-resolution multi-attribute slices.

[0065] Based on this, the technical solution of the present invention will be described and explained below with reference to several examples.

[0066] Example 1: As shown in the attached document Figure 2 As shown in the figure, this invention discloses a method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network, comprising:

[0067] Step S110: Obtain three sets of two-dimensional low-resolution multi-attribute slices of the low-resolution three-dimensional reservoir attribute model. The three sets of two-dimensional low-resolution multi-attribute slices are obtained by processing the low-resolution three-dimensional reservoir attribute model through a three-way orthogonal slicing strategy and a multi-channel image encoding strategy.

[0068] Step S120: Input three sets of two-dimensional low-resolution multi-attribute slices into the resolution reconstruction model to obtain three sets of two-dimensional high-resolution multi-attribute slices. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model through several samples. Each sample includes three sets of two-dimensional low-resolution multi-attribute slices of the three-dimensional reservoir attribute model and the label information of the corresponding three sets of two-dimensional high-resolution multi-attribute slices.

[0069] Step S130: Decode and fuse the three sets of two-dimensional high-resolution multi-attribute slices to obtain a high-resolution three-dimensional reservoir attribute model.

[0070] This invention discloses a resolution reconstruction method for three-dimensional reservoir attribute models based on a perception-guided generative adversarial network. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model, an improved version of the RealESRGAN model. The RealESRGAN model incorporates a generative adversarial network, which not only enables detail restoration and high-frequency enhancement of attribute images but also improves the model's representational accuracy and image fidelity. Furthermore, it exhibits stronger generalization ability in multi-attribute collaborative modeling. Therefore, the two-dimensional high-resolution multi-attribute slices predicted using the resolution reconstruction model are superior in terms of structural restoration and attribute consistency, resulting in higher accuracy of the reconstructed high-resolution three-dimensional reservoir attribute model. This provides reliable technical support for the fine characterization and engineering design of unconventional reservoirs.

[0071] Example 2: As shown in the attached document Figure 3 As shown, this embodiment of the invention is a further optimization of the above embodiment, wherein obtaining three sets of two-dimensional low-resolution multi-attribute slices of a low-resolution three-dimensional reservoir attribute model includes:

[0072] Step S210: Obtain a low-resolution three-dimensional reservoir attribute model, and use a three-dimensional orthogonal slicing strategy to perform two-dimensional slicing operations in three orthogonal directions (XY, YZ, XZ) to obtain three sets of two-dimensional low-resolution multi-attribute slices.

[0073] Specifically, the low-resolution three-dimensional reservoir property model is structured point cloud data. Each data point includes spatial coordinates (x, y, z) and its corresponding reservoir physical property parameters, including porosity, permeability, oil saturation, triaxial geostress, lithofacies, and oil-bearing facies.

[0074] Specifically, in order to preserve the three-dimensional spatial structure features of the low-resolution three-dimensional reservoir attribute model as much as possible, this embodiment adopts a three-dimensional orthogonal slicing strategy from the low-resolution three-dimensional reservoir attribute model. Two-dimensional slicing was performed along three orthogonal directions (XY, YZ, and XZ) to obtain three sets of two-dimensional low-resolution multi-attribute slices, i.e., three sets of two-dimensional low-resolution attribute matrices. This strategy aims to capture the spatial variation characteristics of reservoir attributes from different perspectives, providing structurally diverse and directionally complementary input data for subsequent high-resolution reconstruction.

[0075] Step S220: After linear normalization of the three sets of low-resolution attribute slices, the corresponding three sets of two-dimensional low-resolution multi-attribute slices are obtained by using a multi-channel image coding strategy.

[0076] Specifically, to express multiple reservoir attributes in a format suitable for image super-resolution modeling, this embodiment constructs an image mapping strategy based on two-dimensional low-resolution multi-attribute slices. All attribute slices are normalized before image encoding and uniformly standardized to 256x256 resolution. To handle the encoding requirements of continuous attributes and data types, a multi-channel image encoding and grayscale image marking strategy is adopted, as follows:

[0077] (1) Reservoir property map (RGB color map): R channel is porosity, G channel is oil saturation, and B channel is permeability. The reservoir property parameters are used to express the reservoir's storage capacity, pore structure and fluid saturation distribution characteristics, and are the basis for the analysis of the transformation potential.

[0078] (2) Geostress state diagram (RGB color diagram): The R channel is the minimum horizontal principal stress, the G channel is the maximum horizontal principal stress, and the B channel is the vertical stress. The geostress state diagram reflects the tensor characteristics of the reservoir geostress field and is an important basis for fracturing design.

[0079] (3) Lithology field map (grayscale map): The reservoir lithology is mapped to grayscale values ​​according to the coding form of 0-mudstone, 1-sandstone to form a single-channel image. This geostress state map is used to express the lithological boundary and tectonic transition.

[0080] (4) Oil-bearing facies map (grayscale map): The reservoir is mapped to grayscale values ​​according to the coding form of 0-mudstone, 1-oil layer, 2-poor oil layer, 3-dry layer to form a single-channel image. Here, the oil-bearing facies map is used to express the distribution of oil-bearing properties of the reservoir and can be used to help identify sweet spots.

[0081] All attribute values Before image encoding, all images are linearly normalized and converted to pixel values. ;

[0082]

[0083] in, For attribute values, , These are the minimum and maximum values ​​of the attribute in the current slice, respectively.

[0084] The normalization process remains monotonically reversible, and the predicted image output can be deduced back to the physical properties using the following formula;

[0085] .

[0086] Example 3: As shown in the attached document Figure 4 As shown, the embodiments of the present invention are further optimizations of the above embodiments, wherein the process of constructing the resolution reconstruction model includes:

[0087] Step S310: Obtain at least one original three-dimensional reservoir attribute model and obtain the corresponding sample, wherein the sample includes three sets of two-dimensional low-resolution multi-attribute slices and the label information of the corresponding three sets of real two-dimensional high-resolution multi-attribute slices.

[0088] As attached Figure 5 As shown, step S310 above includes:

[0089] Step S311: Obtain the original three-dimensional reservoir attribute model;

[0090] Specifically, the original 3D reservoir attribute model is a high-resolution 3D reservoir attribute model, specifically structured point cloud data. Each data point contains spatial coordinates (x, y, z) and its corresponding reservoir physical parameters, including porosity, permeability, oil saturation, triaxial geostress, lithofacies, and oil-bearing facies. The overall size of the original 3D reservoir attribute model can be, but is not limited to, 1300m × 2000m × 200m, with a sampling resolution of 10m in each direction. The corresponding grid dimensions are 131 × 201 × 21, as shown in Table 1.

[0091] Table 1 Raw Data

[0092] .

[0093] Step S312: Using the three-way orthogonal slicing strategy, perform two-dimensional slicing operations on the original three-dimensional reservoir attribute model in three orthogonal directions (XY, YZ, XZ) to obtain three sets of real two-dimensional high-resolution multi-attribute slices.

[0094] Specifically, to construct samples suitable for resolution reconstruction modeling and to preserve as much of the three-dimensional spatial structural features of the reservoir attribute model as possible, this embodiment adopts a three-dimensional orthogonal slicing strategy from the original three-dimensional attribute model. Two-dimensional slicing was performed along three orthogonal directions (XY, YZ, and XZ) to obtain three sets of real two-dimensional high-resolution multi-attribute slices. This strategy aims to capture the spatial variation characteristics of geological attributes from different perspectives, providing structurally diverse and directionally complementary training samples for subsequent modeling, as detailed below (using an original resolution of 10m as an example):

[0095] XY plane slice (along the Z-axis): Fixed For each Extracting the matrix , representing the reservoir property distribution on the k-th horizontal plane;

[0096] YZ plane slice (along the X-axis): Fixed For each Extracting the matrix , represents the vertical profile attribute distribution at the i-th east-west position;

[0097] XZ plane slice (along the Y-axis): Fixed For each Extracting the matrix , represents the vertical profile attribute distribution at the j-th north-south position;

[0098] The slices mentioned above correspond to the three-dimensional distribution features of the model, systematically capturing reservoir structure differences from the horizontal, vertical and interlayer directions. Each slice in each direction constitutes an independent two-dimensional slice matrix, which serves as the basis for subsequent training sample construction and image encoding.

[0099] Step S313: Downsample the three sets of real two-dimensional high-resolution multi-attribute slices respectively to obtain the corresponding two-dimensional low-resolution multi-attribute slices, and construct samples using the LR-HR image pairing strategy. The samples include the label information of the three sets of two-dimensional low-resolution multi-attribute slices and the corresponding three sets of real two-dimensional high-resolution multi-attribute slices.

[0100] The specific implementation process includes:

[0101] (1) The three sets of real two-dimensional high-resolution multi-attribute slices are regarded as "high-resolution ground truth images" (HR), representing the ideal accuracy of the model, and serving as the label information of the samples.

[0102] (2) Generate two-dimensional low-resolution multi-attribute slices;

[0103] We perform equidistant downsampling in a two-dimensional plane on a real two-dimensional high-resolution multi-attribute slice. Specifically, using a sampling strategy with a step size of 2, we select a value from the original slice every two grid points, forming a low-resolution (LR) matrix with a spatial resolution of 20m. For example, an original matrix of size 131×201 becomes 66×101 after downsampling. Mathematically, this is expressed as:

[0104]

[0105] in, and These are the original high-resolution attribute matrix (i.e., the true two-dimensional high-resolution multi-attribute slice) and the downsampled low-resolution matrix (i.e., the two-dimensional low-resolution multi-attribute slice), respectively. , These are the original height and width of the slice, respectively.

[0106] (3) Spatial coordinate mapping record

[0107] To preserve the spatial correspondence between each matrix point and the original 3D reservoir attribute model, a coordinate mapping table is generated synchronously during downsampling for physical reconstruction of subsequent training outputs.

[0108] Based on this, a set of spatially consistent LR-HR two-dimensional attribute matrix samples was obtained, laying the foundation for subsequent image encoding and neural network input construction.

[0109] Step S314: After linear normalization of the samples, the three sets of two-dimensional low-resolution multi-attribute slices in the samples are encoded using a multi-channel image coding strategy.

[0110] Specifically, to express multiple reservoir attributes in a format suitable for image super-resolution modeling, this embodiment constructs an image mapping strategy based on two-dimensional attribute slices. All attribute slices are normalized before image encoding and uniformly standardized to a 256x256 resolution. To handle the encoding requirements of continuous attributes and data types, a multi-channel image encoding and grayscale image marking strategy is adopted, as follows:

[0111] (1) Reservoir property map (RGB color map): R channel is porosity, G channel is oil saturation, and B channel is permeability. The reservoir property parameters are used to express the reservoir's storage capacity, pore structure and fluid saturation distribution characteristics, and are the basis for the analysis of the transformation potential.

[0112] (2) Geostress state diagram (RGB color diagram): The R channel is the minimum horizontal principal stress, the G channel is the maximum horizontal principal stress, and the B channel is the vertical stress. The geostress state diagram reflects the tensor characteristics of the reservoir geostress field and is an important basis for fracturing design.

[0113] (3) Lithology field map (grayscale map): The reservoir lithology is mapped to grayscale values ​​according to the coding form of 0-mudstone, 1-sandstone to form a single-channel image. This geostress state map is used to express the lithological boundary and tectonic transition.

[0114] (4) Oil-bearing facies map (grayscale map): The reservoir is mapped to grayscale values ​​according to the coding form of 0-mudstone, 1-oil layer, 2-poor oil layer, 3-dry layer to form a single-channel image. Here, the oil-bearing facies map is used to express the distribution of oil-bearing properties of the reservoir and can be used to help identify sweet spots.

[0115] All attribute values Before image encoding, all images are linearly normalized and converted to pixel values. ;

[0116]

[0117] in, For attribute values, , These are the minimum and maximum values ​​of the attribute in the current slice, respectively.

[0118] The normalization process remains monotonically reversible, and the predicted image output can be deduced back to the physical properties using the following formula;

[0119] .

[0120] All normalized parameters and coordinate mapping tables are saved synchronously during the data processing stage to ensure that the model training output data has reproducibility and spatial positioning capability.

[0121] Step S320: Perform data augmentation on the samples to obtain multiple samples, and divide them into training set and test set according to the ratio.

[0122] As attached Figure 6 As shown, to mitigate the overfitting risk caused by small sample size and improve model robustness and spatial structure recognition capabilities, this embodiment designs a physically interpretable data augmentation strategy for multi-channel reservoir attribute images. Unlike general image tasks, three-dimensional reservoir attribute slices are generated by encoding spatial attribute fields, exhibiting the following characteristics: First, the sample size is limited, and the number of images is constrained by the slice direction and spacing; second, the pixel values ​​in the images carry physical properties, stress, or lithological information, possessing clear physical meaning; third, image transformations must maintain the consistency between pixel space and physical space coordinate mapping. Therefore, the augmentation strategy must not only increase the data scale and image diversity but also strictly adhere to the spatial constraints of the reservoir images to avoid disrupting the true attribute distribution. This embodiment uses the following image spatial augmentation method to generate expanded samples based on the original slice images:

[0123] (1) Orientation transformation: including clockwise rotation (90°, 180°, 270°) and horizontal / vertical mirror reversal. This type of transformation does not change the topological relationship between pixels and can simulate changes in the orientation of strata or the symmetrical structure of images, thereby improving the model's ability to model the rotation invariance of structures;

[0124] (2) Cropping and scaling: A fixed-scale random cropping window is used to extract local regions of the image and restore them to the same size as the original slice, thereby enhancing the model's ability to recognize local heterogeneous structures.

[0125] (3) Blur perturbation: Apply low-order texture transformations such as Gaussian blur and contrast perturbation to the image to simulate different logging response conditions or imaging quality differences;

[0126] (4) Layer perturbation simulation: When slicing, the Z-axis layer is offset by ±1 or 2 layers to construct perturbation images of adjacent layers, so as to improve the model's recognition accuracy of subtle changes between layers.

[0127] All the above enhancement operations are applied simultaneously to three image types (physical property maps, stress maps, and grayscale images), ensuring that each enhanced sample maintains channel alignment and spatial consistency between images. Using this enhancement strategy, each original slice can be expanded into 16 variant images, effectively increasing the training sample size. Furthermore, while maintaining the original physical property structure, the enhancement strategy significantly improves the data diversity during training, providing a rich and stable training foundation for subsequent high-resolution reconstruction models.

[0128] Step S330: Train the Geo-RealESRGAN model using the training set. During training, a loss function is introduced. When the value of the loss function is stable, the training ends, and the resolution reconstruction model is obtained. The Geo-RealESRGAN model adopts the Generative Adversarial Network (GAN) mechanism. The generator restores each set of input two-dimensional low-resolution multi-attribute slices to predicted two-dimensional high-resolution multi-attribute slices. The discriminator uses patch-wise discrimination to determine the detailed differences between the real two-dimensional high-resolution multi-attribute slices and the predicted two-dimensional high-resolution multi-attribute slices, and feeds back fine-grained spatial supervision signals to the generator.

[0129] Specifically, the Geo-RealESRGAN model (Geological Real-World Super-ResolutionGenerative Adversarial Network) is a neural network model that combines perceptual-guided adversarial learning. Based on the Real-ESRGAN framework, this model incorporates the spatial distribution patterns and structural features of attribute images in reservoir modeling tasks, introducing multi-scale feature extraction, spatial context modeling, and local structure perception mechanisms to achieve high-precision reconstruction of sparse, low-resolution geological slices. Unlike traditional high-resolution image methods that prioritize visual clarity, this model first maps each reservoir attribute to a multi-channel image. Then, while preserving the original perceptual and adversarial loss structures, it introduces pixel-wise L1 loss to improve attribute recovery accuracy. Furthermore, it intentionally eliminates visual optimization strategies such as color enhancement, ensuring that the network focuses more on the numerical reconstruction of the attributes themselves. Finally, the model performance is evaluated through attribute value error rather than image perceptual metrics, making it more aligned with the actual needs of reservoir modeling.

[0130] As attached Figure 7 As shown, it specifically includes:

[0131] (a) Generator

[0132] The Geo-RealESRGAN model generator is based on the SRVGGNet and RRDB (Residual in Residual DenseBlock) modules, aiming to efficiently extract multi-scale features from low-resolution attribute images and achieve high-precision attribute reconstruction. The specific structure is as follows:

[0133] (1) Downsampling module (Pixel Unshuffle): The Pixel Unshuffle operation is used at the front end of the network to divide the spatial information of the input high-resolution image into blocks and remap it to the channel dimension, so as to efficiently compress the image into a low-resolution multi-channel feature map and provide optimized input for subsequent deep network modeling.

[0134] (2) Initial convolutional layer: The input image is initially encoded using a 3×3 convolutional kernel to extract shallow texture and attribute change features, providing basic feature representation for subsequent deep networks.

[0135] (3) Residual Dense Modules (RRDB Stack): The main network consists of multiple RRDBs, and each RRDB module contains multiple Dense Blocks. Through dense feature fusion and residual connections, the network's ability to capture deep features is enhanced, thereby improving its expressive power for complex geological structural changes. RRDBs introduce a residual scaling factor β (usually 0.2) to suppress gradient explosion and enhance training stability.

[0136] (4) Spatial Pyramid Pooling Module (SPP): The SPP module is added after the RRDB module. It captures contextual information at different spatial scales through multi-scale (1×1, 2×2, 4×4) pooling operations and fuses it with the original feature map through interpolation upsampling. This enhances the model's ability to express complex structures such as sand body boundaries, fault fissures, and fine-grained interlayers, and improves the quality of structure reconstruction.

[0137] (5) Upsampling module (PixelShuffle Upsampling): The PixelShuffle method is used to perform upsampling step by step. Before each upsampling step, a 3×3 convolutional layer is added to expand the feature channels to ensure the effectiveness of pixel rearrangement and gradually restore the feature map to the target high resolution.

[0138] (6) Output convolutional layer: A 3×3 convolutional kernel is used to map the upsampled features to the final output channel to generate a high-resolution attribute image. Color attribute images output three channels (such as porosity, saturation, and permeability), while grayscale images output a single channel.

[0139] (7) No Normalization and No Tanh: All BatchNormalization (BN) layers are removed from the generator to enhance generalization performance and avoid artifacts caused by statistical offset in the data domain. At the same time, the Tanh activation function at the end is removed to avoid pixel intensity oversaturation and prevent excessive compression of attribute values, thereby improving the accuracy of continuous physical quantity recovery.

[0140] This generator structure has efficient feature extraction capabilities, a robust residual learning path, and spatial augmentation effects suitable for sparse geological data. It can effectively capture microscopic changes and structural boundary features between different layers and achieve accurate distribution restoration at the attribute level.

[0141] (ii) Discriminator

[0142] (1) Discriminator Network: To enhance the model's ability to recognize the realism of local structures and spatial textures in generated images, a U-Net discriminator structure based on the PatchGAN concept was constructed. The discriminator provides fine-grained spatial supervision signals to the generator by determining the realism of each local region in the input image, thereby improving the perceptual quality and structural consistency of the predicted image. The overall structure includes the following key parts:

[0143] (2) Adversarial Discriminant Framework (PatchGAN): The discriminator does not output a true / false classification result at the whole-image level. Instead, it divides the input image into several local regions (patch), performs true / false classification on each patch, and finally outputs a patch-wise probability map. This design improves the model's sensitivity to problems such as local artifacts and texture deviations, and helps guide the generator to generate more natural and continuous structures in space.

[0144] (3) Encoder module (Encoder with Downsampling): Feature extraction is performed by combining multi-layer 3×3 convolutions with the Leaky ReLU activation function. Spatial downsampling is achieved between every two convolution modules through a convolution operation with a stride of 2. At the same time, the features output by each encoder stage are retained through skip connections for fusion in the decoder stage.

[0145] (4) Decoder module (Decoder with Upsampling): The feature map is upsampled step by step by interpolation and fused with the encoder feature through skip connections. Each decoding module contains two layers of 3×3 convolution and Leaky ReLU activation to restore spatial details and enhance feature expression.

[0146] (5) Patch-wise Output: Finally, a single-channel Patch-wise real and fake rating map is generated through a 1×1 convolutional layer. Each pixel corresponds to the realism evaluation of the local area of ​​the original image. The real image is close to 1, and the generated image is close to 0, providing fine feedback for adversarial loss.

[0147] (6) No BatchNorm: To enhance the network’s adaptability to different data distributions, the discriminator does not use BN operations, thereby avoiding the risk of artifacts caused by statistical shifts and improving training stability.

[0148] The adversarial training mechanism and loss function are detailed below:

[0149] The Geo-RealESRGAN model is trained using a Generative Adversarial Network (GAN) mechanism, where the generator and discriminator compete against each other to optimize the structural consistency and visual realism of the generated images. The generator takes a low-resolution input and produces a high-resolution attribute image, while the discriminator uses patch-wise discrimination to provide feedback on the subtle differences between the real and generated images, helping the generator achieve more realistic structural reconstruction. Unlike the binary classification mechanism of traditional GANs that judges the authenticity of images, this embodiment employs the Relativistic Average GAN (RaGAN) strategy. Its core idea is to enable the discriminator to learn whether an image is "more realistic than the average generated image." This mechanism, by introducing modeling of relative realism, can alleviate problems such as training instability and discriminator overfitting, while improving the model's ability to express boundary structures and detailed textures, making it particularly suitable for reservoir image reconstruction tasks with transition zones and intra-layer heterogeneity.

[0150] In the design of the loss function, the generator optimization objective consists of three parts: pixel reconstruction loss, perceptual loss, and adversarial loss. Specifically:

[0151] (1) Pixel-level L1 loss (pixel reconstruction loss): ensures that the generated image is close to the target image at the pixel level, avoiding excessive blurring.

[0152]

[0153] in, For pixel reconstruction loss, It is a high-resolution image output by the generator. These are true high-definition images.

[0154] (2) Perceptual Loss: Utilizes high-level features of pre-trained networks (such as VGG) to maintain the semantic consistency of generated images.

[0155]

[0156] in, In order to perceive loss, For the pre-trained network Layer features, These are the weighting coefficients.

[0157] (3) Adversarial Loss: Through the Generative Adversarial Network (GAN) framework, the generator is encouraged to output realistic images to deceive the discriminator. Therefore, it includes generator adversarial loss and discriminator adversarial loss.

[0158]

[0159] in, To help the generator combat loss, To help the discriminator combat loss, For real images, To generate an image, D is the discriminator output. To calculate the average of the joint distribution of real samples and generated samples, To calculate the average of the distribution of the real sample, To calculate the average of the distribution of the generated samples.

[0160] In summary, the total loss function is:

[0161]

[0162] in, , , For each loss weight, To combat losses, i.e. and sum.

[0163] In this embodiment, the loss weights can be set as follows: =0.1、 =1、 =0.1. This configuration references the empirical proportions of image super-resolution tasks and is adjusted in conjunction with the reconstruction requirements of geological attribute images. Perceptual loss is used as the dominant term, emphasizing the fidelity of structural information; pixel loss and adversarial loss are used as auxiliary terms, respectively controlling numerical restoration and local realism. The three together drive the network to converge to the optimal solution that balances structural quality and physical accuracy.

[0164] Step S340: Test the trained resolution reconstruction model using the test set, optimize the model parameters of the resolution reconstruction model, and output a resolution reconstruction model that meets the test evaluation requirements.

[0165] Example 4: This embodiment of the invention is a further optimization of the above embodiments. This embodiment sets up a back-mapping mechanism from image space to three-dimensional attribute points and integrates multiple prediction results from slices in different directions to improve the spatial continuity and local consistency of attribute recovery. Since the correspondence between each image pixel and the original three-dimensional grid coordinate points (Pixel–Voxel Mapping Table) has been recorded during the slice construction process in the model construction stage, the stage of inferring the high-resolution three-dimensional reservoir attribute model can be directly restored to the corresponding attribute point data through inverse normalization and mapping table, specifically including:

[0166] (1) Multi-channel image decoding and inverse normalization: Two-dimensional high-resolution multi-attribute slices are inversely transformed into the original attribute dimensions through the stored normalization parameters;

[0167] (2) Coordinate mapping lookup: Each pixel finds its corresponding three-dimensional physical coordinates (x, y, z) in the mapping table;

[0168] (3) Constructing the predicted attribute point set: Finally, a set of three-dimensional attribute point data is obtained from the two-dimensional high-resolution multi-attribute slices. Each point includes three-dimensional physical coordinates (x, y, z) and its predicted attribute value;

[0169] (4) Since attribute points at the same spatial location may be repeatedly predicted in multiple slice directions (XY, YZ, XZ), this embodiment further introduces a multi-source prediction fusion mechanism to weightedly integrate multiple predicted attribute values ​​for each 3D coordinate point. That is, firstly, for all spatial coordinate points that appear multiple times, the predicted attribute values ​​in all directions are recorded, and then the average value is used as the fusion strategy, i.e., the arithmetic mean of multiple prediction results for the same point is taken as the final predicted attribute value:

[0170]

[0171] in, For the final predicted attribute value, For the first i Predicted attribute values ​​in each direction n This represents the number of times the word repeats.

[0172] (5) Based on the final predicted attribute values ​​of each three-dimensional coordinate point, a high-resolution three-dimensional reservoir attribute model is established.

[0173] This multi-slice fusion strategy fully utilizes the complementarity of structural distribution in images from different directions, effectively improving the spatial consistency of prediction results, and showing stronger recovery ability, especially for complex faults, thin interlayers and other structures.

[0174] Example 5: As shown in the attached documentFigure 8 As shown, this embodiment of the invention discloses a three-dimensional reservoir attribute model resolution reconstruction device based on a perception-guided generative adversarial network, comprising:

[0175] The low-resolution multi-attribute slice acquisition unit acquires three sets of two-dimensional low-resolution multi-attribute slices of the low-resolution three-dimensional reservoir attribute model. The three sets of two-dimensional low-resolution multi-attribute slices are obtained by processing the low-resolution three-dimensional reservoir attribute model through a three-way orthogonal slicing strategy and a multi-channel image encoding strategy.

[0176] The resolution reconstruction unit takes three sets of two-dimensional low-resolution multi-attribute slices as input to the resolution reconstruction model and obtains three sets of two-dimensional high-resolution multi-attribute slices. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model through several samples. Each sample includes three sets of two-dimensional low-resolution multi-attribute slices of the three-dimensional reservoir attribute model and the label information of the corresponding three sets of two-dimensional high-resolution multi-attribute slices.

[0177] The high-resolution three-dimensional reservoir attribute model forming unit decodes and fuses three sets of two-dimensional high-resolution multi-attribute slices to obtain a high-resolution three-dimensional reservoir attribute model.

[0178] The specific implementation steps of each of the above units are the same as those in the above embodiments, and will not be repeated here.

[0179] Example 6: This embodiment of the invention discloses a storage medium storing a computer program that can be read by a computer. The computer program is configured to execute a method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network when it is run.

[0180] The aforementioned storage media may include, but are not limited to, USB flash drives, read-only memory, portable hard drives, magnetic disks, optical disks, and other media capable of storing computer programs.

[0181] Example 7: This embodiment of the invention discloses an electronic device, including a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement a method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network.

[0182] The processor described above can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. It can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The memory can include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, portable hard drives, magnetic disks, or optical disks.

[0183] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0184] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0185] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 Figure 1 The function specified in one or more boxes.

[0186] The above content is only a specific embodiment of the present invention, which has strong adaptability and implementation effect. However, the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention. Therefore, equivalent changes made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network, characterized in that, include: Three sets of two-dimensional low-resolution multi-attribute slices were obtained from the low-resolution three-dimensional reservoir attribute model. These three sets of two-dimensional low-resolution multi-attribute slices were obtained after processing the low-resolution three-dimensional reservoir attribute model using a three-way orthogonal slicing strategy and a multi-channel image encoding strategy. The slices include: A low-resolution three-dimensional reservoir property model was obtained, and a three-dimensional orthogonal slicing strategy was used to perform two-dimensional slicing operations in three orthogonal directions (XY, YZ, and XZ) to obtain three sets of two-dimensional low-resolution multi-attribute slices. After linear normalization of the three sets of low-resolution attribute slices, the corresponding three sets of two-dimensional low-resolution multi-attribute slices are obtained by using a multi-channel image coding strategy. Three sets of two-dimensional low-resolution multi-attribute slices are input into the resolution reconstruction model to obtain three sets of two-dimensional high-resolution multi-attribute slices. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model through several samples. Each sample includes three sets of two-dimensional low-resolution multi-attribute slices of the three-dimensional reservoir attribute model and the label information of the corresponding three sets of two-dimensional high-resolution multi-attribute slices. The Geo-RealESRGAN model adopts a generative adversarial network mechanism. The generator uses the generator to restore each set of input two-dimensional low-resolution multi-attribute slices into predicted two-dimensional high-resolution attribute slices. The discriminator uses patch-wise discrimination to determine the detailed differences between the real two-dimensional high-resolution multi-attribute slices and the predicted two-dimensional high-resolution attribute slices, and feeds back fine-grained spatial supervision signals to the generator. Three sets of two-dimensional high-resolution multi-attribute slices are decoded and fused to obtain a high-resolution three-dimensional reservoir attribute model, including: Each set of two-dimensional high-resolution multi-attribute slices is subjected to multi-channel image decoding and inverse normalization. Based on coordinate mapping lookup, each set of two-dimensional high-resolution multi-attribute slices is restored into a set of three-dimensional attribute point data, where each three-dimensional attribute point data includes three-dimensional physical coordinates (x, y, z) and its predicted attribute value; A multi-source prediction fusion mechanism is introduced to weight and integrate multiple predicted attribute values ​​for each 3D coordinate point to obtain the final predicted attribute value. A high-resolution three-dimensional reservoir attribute model is established based on the final predicted attribute values ​​of each three-dimensional coordinate point.

2. The method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network according to claim 1, characterized in that, The process of constructing the resolution reconstruction model includes: Obtain at least one original three-dimensional reservoir attribute model and obtain the corresponding sample, wherein the sample includes three sets of two-dimensional low-resolution multi-attribute slices and the label information of the corresponding three sets of real two-dimensional high-resolution multi-attribute slices; Data augmentation is performed on the samples to obtain multiple samples, which are then divided into training and test sets according to a certain ratio. The Geo-RealESRGAN model is trained using a training set. A loss function is introduced during training, and training ends when the value of the loss function is stable, resulting in a resolution reconstruction model. The Geo-RealESRGAN model uses a generative adversarial network mechanism. The generator restores each set of input two-dimensional low-resolution multi-attribute slices to predicted two-dimensional high-resolution attribute slices. The discriminator uses patch-wise discrimination to determine the detailed differences between the real two-dimensional high-resolution multi-attribute slices and the predicted two-dimensional high-resolution attribute slices, and feeds back fine-grained spatial supervision signals to the generator. The trained resolution reconstruction model is tested using a test set, the model parameters are optimized, and a resolution reconstruction model that meets the test evaluation requirements is output.

3. The method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network according to claim 2, characterized in that, The process of obtaining at least one original three-dimensional reservoir property model and obtaining corresponding samples includes: Obtain the original three-dimensional reservoir property model; Using a three-dimensional orthogonal slicing strategy, two-dimensional slicing operations were performed on the original three-dimensional reservoir attribute model in three orthogonal directions (XY, YZ, and XZ) to obtain three sets of real two-dimensional high-resolution multi-attribute slices. Three sets of real two-dimensional high-resolution multi-attribute slices were downsampled to obtain corresponding two-dimensional low-resolution multi-attribute slices. Samples were constructed using the LR-HR image pairing strategy. The samples included the label information of the three sets of two-dimensional low-resolution multi-attribute slices and the corresponding three sets of real two-dimensional high-resolution multi-attribute slices. After linear normalization of the samples, a multi-channel image coding strategy is used to encode three sets of two-dimensional low-resolution multi-attribute slices in the samples.

4. The method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network according to claim 2 or 3, characterized in that, The enhancement process includes orientation transformation, cropping and scaling, blurring perturbation, and layer perturbation simulation.

5. The method for reconstructing the resolution of a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network according to claim 2 or 3, characterized in that, The loss function is as follows: in, Weights for each loss; These are pixel reconstruction loss, perceptual loss, and adversarial loss, respectively.

6. A resolution reconstruction device for a three-dimensional reservoir attribute model based on a perception-guided generative adversarial network, applying the method described in any one of claims 1 to 5, characterized in that, include: The low-resolution multi-attribute slice acquisition unit acquires three sets of two-dimensional low-resolution multi-attribute slices from a low-resolution three-dimensional reservoir attribute model. These three sets of two-dimensional low-resolution multi-attribute slices are obtained by processing the low-resolution three-dimensional reservoir attribute model using a three-way orthogonal slicing strategy and a multi-channel image encoding strategy. The slices include: A low-resolution three-dimensional reservoir property model was obtained, and a three-dimensional orthogonal slicing strategy was used to perform two-dimensional slicing operations in three orthogonal directions (XY, YZ, and XZ) to obtain three sets of two-dimensional low-resolution multi-attribute slices. After linear normalization of the three sets of low-resolution attribute slices, the corresponding three sets of two-dimensional low-resolution multi-attribute slices are obtained by using a multi-channel image coding strategy. The resolution reconstruction unit takes three sets of two-dimensional low-resolution multi-attribute slices as input to the resolution reconstruction model and obtains three sets of two-dimensional high-resolution multi-attribute slices. The resolution reconstruction model is obtained by deep learning the Geo-RealESRGAN model through several samples. Each sample includes three sets of two-dimensional low-resolution multi-attribute slices of the three-dimensional reservoir attribute model and the label information of the corresponding three sets of two-dimensional high-resolution multi-attribute slices. A high-resolution three-dimensional reservoir attribute model forming unit decodes and fuses three sets of two-dimensional high-resolution multi-attribute slices to obtain a high-resolution three-dimensional reservoir attribute model, including: Each set of two-dimensional high-resolution multi-attribute slices is subjected to multi-channel image decoding and inverse normalization. Based on coordinate mapping lookup, each set of two-dimensional high-resolution multi-attribute slices is restored into a set of three-dimensional attribute point data, where each three-dimensional attribute point data includes three-dimensional physical coordinates (x, y, z) and its predicted attribute value; A multi-source prediction fusion mechanism is introduced to weight and integrate multiple predicted attribute values ​​for each 3D coordinate point to obtain the final predicted attribute value. A high-resolution three-dimensional reservoir attribute model is established based on the final predicted attribute values ​​of each three-dimensional coordinate point.

7. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps of the method as claimed in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium stores a computer program that can be read by a computer, the computer program being configured to execute the steps of the method as described in any one of claims 1 to 5 when it is run.

Images