Reservoir property modeling method and apparatus

Abstract

This invention discloses a reservoir attribute modeling method and apparatus, comprising: generating a reservoir lithology distribution model based on a sedimentary facies training template library, a lithology training template library, and a first adversarial generative network (AGN), wherein the first AGN is a depth convolutional AGN based on multi-level well-seismic co-processing, and using preset-type sedimentary facies training template data selected from the sedimentary facies training template library as the initial input data for the network; generating a reservoir attribute model based on well logging attribute data of the work area, the reservoir lithology distribution model, and a second AGN, wherein the second AGN is a geological attribute generation network based on lithology constraints; and correcting the reservoir attribute model based on the work area production dynamic data, using a reservoir attribute model processed by a reservoir numerical simulation surrogate, and a third AGN, wherein the third AGN is an attribute optimization network based on dynamic constraints. This invention can effectively improve the dynamic and static data consistency rate of random reservoir attribute modeling.

Description

Technical Field

[0001] This invention relates to the field of reservoir development, and more particularly to a method and apparatus for reservoir property modeling. Background Technology

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] As reservoir development deepens, the requirements for geological models in refined reservoir development are becoming increasingly stringent. Geological models typically include reservoir structural models and reservoir attribute models. Among these, reservoir attribute models are particularly important for reservoir management and development scheme optimization because they directly provide fundamental input parameters for reservoir numerical simulation models.

[0004] Currently, reservoir property modeling methods typically include two types: deterministic modeling and stochastic modeling. Among them:

[0005] ① The deterministic modeling method is mainly based on well logging data in the work area and is constructed using methods such as Kriging interpolation. For each input, there is only one deterministic output.

[0006] ② Stochastic modeling methods, mainly including sequential Gaussian modeling and multi-point geostatistical modeling (MPS), are primarily based on input well logging data to obtain statistical characteristics, including variance, variation, and mean. Then, based on Monte Carlo sampling, a stochastic model is generated. With a fixed input, the results of each Monte Carlo sampling modeling session exhibit a degree of randomness, while still conforming to the statistical characteristics; therefore, each modeling result will be different.

[0007] Compared to stochastic modeling, deterministic modeling, which uses spatial interpolation, lacks a response to the actual geological complexity. Therefore, current reservoir attribute modeling mainly relies on stochastic modeling methods. However, existing stochastic reservoir attribute modeling suffers from inconsistencies between dynamic and static data.

[0008] Therefore, there is an urgent need for a reservoir property modeling scheme that can overcome the above problems. Summary of the Invention

[0009] This invention provides a reservoir attribute modeling method to improve the consistency rate of dynamic and static data in stochastic reservoir attribute modeling. The method includes:

[0010] Obtain well logging data for the work area, production dynamic data for the work area, sedimentary facies training template library and lithology training template library. The well logging data for the work area includes: sedimentary facies data for the work area, sand body interlayer data for the work area, and attribute data for the work area.

[0011] Based on the well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library and lithology training template library, and the pre-established first adversarial generative network, a reservoir lithology distribution model is generated. The first adversarial generative network is a deep convolutional adversarial generative network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network.

[0012] Based on the well logging attribute data of the work area, the reservoir lithology distribution model, and the pre-established second adversarial generation network, a reservoir attribute model is generated. The second adversarial generation network is a geological attribute generation network based on lithological constraints.

[0013] Based on the dynamic production data of the work area, the reservoir attribute model is modified through the reservoir attribute model of the reservoir numerical simulation agent and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints.

[0014] This invention provides a reservoir attribute modeling apparatus to improve the consistency rate of dynamic and static data in stochastic reservoir attribute modeling. The apparatus includes:

[0015] The data acquisition module is used to acquire well logging data of the work area, production dynamic data of the work area, sedimentary facies training template library and lithology training template library. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area and well logging attribute data of the work area.

[0016] The reservoir lithology distribution model generation module is used to generate a reservoir lithology distribution model based on well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library and lithology training template library, as well as a pre-established first adversarial generation network. The first adversarial generation network is a deep convolutional adversarial generation network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network.

[0017] The reservoir attribute model generation module is used to generate a reservoir attribute model based on well logging attribute data of the work area, reservoir lithology distribution model and pre-established second adversarial generation network. The second adversarial generation network is a geological attribute generation network based on lithological constraints.

[0018] The reservoir attribute model optimization module is used to modify the reservoir attribute model based on the dynamic production data of the work area, the reservoir attribute model of the reservoir numerical simulation agent, and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints.

[0019] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described reservoir property modeling method.

[0020] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described reservoir property modeling method.

[0021] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described reservoir attribute modeling method.

[0022] Compared with the existing technology of random modeling of reservoir attributes, the embodiments of the present invention obtain well logging data of the work area, production dynamic data of the work area, sedimentary facies training template library, and lithology training template library. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area, and well logging attribute data of the work area. Based on the well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area, sedimentary facies training template library, and lithology training template library, and a pre-established first adversarial generative network, a reservoir lithology distribution model is generated. The first adversarial generative network is a depth convolution adversarial network based on multi-level well-seismic joint operation. The network is generated by using sedimentary facies training template data of a preset type selected from a sedimentary facies training template library as the initial input data. Based on well logging attribute data of the work area, a reservoir lithology distribution model, and a pre-established second adversarial generation network, a reservoir attribute model is generated. The second adversarial generation network is a geological attribute generation network based on lithology constraints. Based on the work area's production dynamic data, the reservoir attribute model is modified using the reservoir numerical simulation proxy and a pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints. This embodiment of the invention introduces dynamic production content into the reservoir attribute modeling process. By combining the sedimentary facies training template library and the lithology training template library of the work area's well logging data, a reservoir lithology distribution model and a reservoir attribute model based on an artificial intelligence adversarial generation network are generated sequentially. Then, the reservoir attribute model is modified by combining the dynamic production content, effectively constructing a reservoir attribute model that matches both dynamic and static data, and improving the consistency rate of dynamic and static data in random reservoir attribute modeling. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0024] Figure 1 This is a schematic diagram of the reservoir property modeling method in an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the template library creation method in an embodiment of the present invention;

[0026] Figure 3 This is a flowchart of reservoir attribute modeling in a specific embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram illustrating the generation of the reservoir attribute model in an embodiment of the present invention;

[0028] Figure 5 This is a structural diagram of the reservoir property modeling device in an embodiment of the present invention;

[0029] Figure 6 This is a schematic diagram of the computer device structure according to an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0031] First, the technical terms used in the embodiments of this invention will be introduced:

[0032] Artificial intelligence: Artificial intelligence is a computer-based technology that produces models with a certain level of intelligence. These models can make responses similar to human intelligence and mainly involve robotics, speech recognition, image recognition, natural language processing, and expert systems.

[0033] Dynamic and static data: Dynamic data generally refers to the changes in produced fluids during reservoir production. The characteristics of dynamic data are directly related to the static parameters of the reservoir. Static data generally refers to initial porosity, permeability, and saturation in reservoir analysis; static data also includes seismic data and other data that do not change during development.

[0034] Production dynamic data of the work area: also known as dynamic data, is the data on changes in oil production, water production, etc. over time during the oilfield production process. This data package is accompanied by corresponding engineering parameter data.

[0035] Geological modeling: Based on well logging geological data, well logging data, seismic data, etc., from the reservoir work area, computer technology is used to generate a three-dimensional static model of the reservoir. This typically includes a reservoir porosity model, permeability model, and saturation model. This model will be used for reservoir numerical simulation studies, predicting future changes in production data, and guiding well location design, etc.

[0036] Reservoir attribute modeling: Based on well logging data of the work area, a static attribute model of the reservoir is constructed, including reservoir porosity model, permeability model, and saturation model.

[0037] Sedimentary dynamics simulation: Based on the sedimentary environment parameters of the study area, a sedimentary model that conforms to the sedimentary facies of the study area is established, sedimentary simulation is carried out, and three-dimensional sedimentary data volume under specific sedimentary environment parameters is obtained.

[0038] An encoder, also called an autoencoder, is a widely used neural network in the field of artificial intelligence for automatically encoding two-dimensional and three-dimensional image data. An autoencoder consists of two parts: an encoder and a decoder. Autoencoders possess the general functionality of representation learning algorithms and are applied to dimensionality reduction feature analysis.

[0039] Latent variables: the dimensionality-reduced feature array obtained by the encoder.

[0040] The generator is a network that combines a set of deconvolutions, unpooling, and nonlinear mappings to generate a 3D data image based on the input feature array.

[0041] Discriminator: It is a network application that combines a set of convolutions, pooling, and nonlinear mappings to generate a 3D data image based on the input feature array.

[0042] Multi-point geostatistical modeling (MPS modeling): A method that uses 3D images as training templates, extracts statistical features between data from the templates, and then constructs a model based on these statistical features through Monte Carlo sampling.

[0043] Templates: The base images used in MPS modeling. The MPS method is based on the statistical characteristics between the statistical attribute data of these images. Templates include: 3D sedimentary facies templates and 3D lithological distribution templates.

[0044] Lithology: Reservoir sand body and interlayer classification, sand body is 1, interlayer is 0.

[0045] To improve the consistency rate of dynamic and static data in random reservoir attribute modeling, this invention provides a reservoir attribute modeling method, such as... Figure 1 As shown, the method may include:

[0046] Step 101: Obtain well logging data for the work area, production dynamic data for the work area, sedimentary facies training template library and lithology training template library. The well logging data for the work area includes: sedimentary facies data for the work area, sand body interlayer data for the work area and attribute data for the work area.

[0047] Step 102: Based on the well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library and lithology training template library, and the pre-established first adversarial generation network, generate a reservoir lithology distribution model. The first adversarial generation network is a depth convolution adversarial generation network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library will be used as the initial input data of the network.

[0048] Step 103: Based on the well logging attribute data of the work area, the reservoir lithology distribution model and the pre-established second adversarial generation network, generate the reservoir attribute model. The second adversarial generation network is a geological attribute generation network based on lithology constraints.

[0049] Step 104: Based on the dynamic production data of the work area, the reservoir attribute model is modified using the reservoir attribute model of the reservoir numerical simulation agent and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints.

[0050] Depend on Figure 1As shown, compared with the prior art's random modeling of reservoir attributes, the embodiments of the present invention obtain well logging data of the work area, dynamic production data of the work area, sedimentary facies training template library, and lithology training template library. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area, and well logging attribute data of the work area. Based on the well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area, sedimentary facies training template library, and lithology training template library, as well as a pre-established first adversarial generative network, a reservoir lithology distribution model is generated. The first adversarial generative network is based on the depth of multi-level well-seismic combined model. The convolutional adversarial generative network (CGN) uses sedimentary facies training template data of a preset type selected from a sedimentary facies training template library as the initial input data. Based on well logging attribute data from the work area, a reservoir lithology distribution model, and a pre-established second adversarial generative network (GGN), a reservoir attribute model is generated. The second GGN is a geological attribute generation network based on lithology constraints. Based on dynamic production data from the work area, the reservoir attribute model is modified using a reservoir numerical simulation proxy and a pre-established third GGN, which is an attribute optimization network based on dynamic constraints. This embodiment of the invention introduces dynamic production content into the reservoir attribute modeling process. By combining well logging data from the work area with sedimentary facies and lithology training template libraries, a reservoir lithology distribution model and a reservoir attribute model based on an artificial intelligence adversarial generative network are generated sequentially. The reservoir attribute model is then modified using dynamic production content, effectively constructing a reservoir attribute model that matches both dynamic and static data, and improving the consistency rate of dynamic and static data in random reservoir attribute modeling.

[0051] The inventors discovered that current stochastic modeling of reservoir attributes still suffers from discrepancies between static and dynamic data. While current modeling fully considers static logging data from the work area, it completely neglects dynamic logging data. Dynamic data is highly valuable for attributes with significant uncertainty in well logging interpretation. For example, current logging methods based on AC sonic logging curves can interpret porosity characteristics relatively accurately, but for permeability and saturation characteristics, although multiple models exist, in practice, the uncertainty of these two attributes is far greater than that of porosity. This is because there is currently no direct well logging interpretation scheme for permeability; it mainly relies on porosity-permeability relationship models. However, current porosity-permeability relationship models in complex reservoirs are difficult to represent with simple linear models, and for tight reservoirs, the porosity-permeability correlation is also very poor. Therefore, interpreting permeability models based on well logging is currently difficult and uncertain. Furthermore, the uncertainty of permeability attribute models obtained based on well logging interpretation is far greater than that of porosity. Similarly, interpreting reservoir saturation logging data is challenging, primarily because the clay and coal components in the reservoir interfere with the interpretation model, leading to significant deviations in calculated water saturation from reality. Conventional stochastic modeling methods rely solely on well logging data from the work area. When significant deviations exist in the interpretation data, the model inevitably fails to accurately predict dynamic data, resulting in a discrepancy between dynamic and static data. History fitting is currently the primary method for resolving this discrepancy in geological models. The reservoir attribute model, as the fundamental input parameter, is input into the reservoir numerical simulation model to predict the dynamic production curve of the work area. By comparing the production curve obtained from the numerical simulation with the actual production curve of the oilfield, it can be determined whether a discrepancy exists in the reservoir attribute model. If the difference between the predicted dynamic production area and the actual production curve is small, it is generally considered that the reservoir attribute model constructed based on conventional stochastic modeling methods has virtually no discrepancy. Otherwise, reverse adjustments are necessary. Through history fitting and extensive empirical corrections, the reservoir attribute model is adjusted until the dynamic and static data are consistent. However, the history fitting method still has many problems in addressing the discrepancy between dynamic and static properties in existing stochastic reservoir property modeling. First, the simulation workload is enormous. During the history fitting phase, the model needs continuous back-correction, and each correction requires a reservoir numerical simulation, resulting in huge computational data consumption. Second, existing history fitting methods, when adjusting reservoir property models in reverse, lack consideration of inter-well geological patterns. They mainly focus on dividing areas and blocks, rigidly adjusting the properties around wells with discrepancies between dynamic and static properties, without considering data from adjacent wells. Therefore, this invention, based on artificial intelligence methods, combines existing sedimentary dynamics simulation data, seismic data, well logging data of sand bodies and interlayers in the work area, and dynamic production data to construct a three-dimensional reservoir property model that is consistent with both dynamic and static properties.The preliminary foundational data required for this invention includes sedimentary dynamics data, seismic data, well logging sedimentary facies data, sand body interlayer data, well logging porosity data, well logging permeability data, well logging saturation data, and well logging production data (including engineering data). The specific implementation comprises three parts: AI-based sedimentary facies modeling, AI-based lithological modeling, and AI-based attribute modeling. A hierarchical control method is employed, based on an encoder, generator, and discriminator network, to generate attribute models that conform to both static and dynamic production data of the study area.

[0052] The following is a detailed analysis of each step.

[0053] In step 101, well logging data of the work area, production dynamic data of the work area, sedimentary facies training template library and lithology training template library are obtained. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area and well logging attribute data of the work area.

[0054] In one embodiment, such as Figure 2 As shown, sedimentary facies training template libraries and lithology training template libraries are pre-established in the following manner:

[0055] Step 201: Obtain sedimentary facies information and sand body interlayer model result data. The sedimentary facies information includes: sedimentary dynamics data, existing sedimentary facies maps, satellite images, or any combination thereof.

[0056] Step 202: Based on the sedimentary facies information, establish a sedimentary facies training template library;

[0057] Step 203: Based on the data volume of the sand body interlayer pattern, establish a lithology training template library.

[0058] In specific implementation, taking flume data as an example, a first preset number of sedimentary facies training template data volumes can be extracted from sedimentary dynamics data; a second preset number of sedimentary facies training template data volumes can be extracted from flume experimental data volumes; the extracted first preset number of sedimentary facies training template data volumes and the second preset number of sedimentary facies training template data volumes are merged to establish a sedimentary facies training template library.

[0059] Figure 3 This is a flowchart illustrating the reservoir attribute modeling process in a specific embodiment of the present invention. In practice, p sedimentary facies training templates, i.e., {SDi|i=1...p}, are extracted from the three-dimensional data volume of the sedimentary facies simulation results in the sedimentary dynamics data to establish a sedimentary facies training template library. In the sedimentary facies training template library, the length, width, height, and number of channels of each template data volume SDi are SL, SW, SH, and SC. A scanning flume device is used to complete the sedimentary flume simulation and obtain the flume experimental data volume. The scanning flume device is as follows: Figure 4As shown. q sedimentary facies training templates, {SCi|i=1...q}, are extracted from the 3D data volume of the flume experiment data to establish a sedimentary facies training template library. In the sedimentary facies training template library, the length, width, height, and number of channels for each template data volume SCi are SL, SW, SH, and SC, respectively. The two sedimentary facies training templates are merged together to form {Si|i=1...n}, and a sedimentary facies training template library is established. In the sedimentary facies training template library, the length, width, height, and number of channels for each template data volume Si are SL, SW, SH, and SC, respectively. From the 3D data volume of the sand body interlayer model results data in the sedimentary dynamics data, h lithological training template data volumes, {Lis|i=1...h, s=1...n}, are extracted for each sedimentary facies region to establish a lithological training template library. The possible values ​​for Lis are {0,1}, where 0 represents sandstone and 1 represents mudstone. In the lithological training template library, each template data volume Lis has a length, width, height, and number of channels of LL, LW, LH, and LC, respectively. The flume apparatus is a device that simulates the transport and deposition of sedimentary particles in a liquid. It can be used to reproduce the depositional process and analyze the internal structural features of sedimentary bodies. The scanning flume apparatus combines a large-scale CT scanner with a flume apparatus. The large-scale CT scanner performs a three-dimensional scan of the flume depositional results, obtaining a three-dimensional model of the flume results. This model clearly displays the internal structural features of different sedimentary bodies, and the scan results are stored in the form of a three-dimensional data volume, which can be used for artificial intelligence analysis.

[0060] In step 102, a reservoir lithology distribution model is generated based on the well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library, lithology training template library, and a pre-established first adversarial generation network. The first adversarial generation network is a depth convolution adversarial generation network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network.

[0061] In one embodiment, the first adversarial generation network includes: a first adversarial generation subnetwork based on sedimentary facies model simulated by sedimentary dynamics, a second adversarial generation subnetwork based on sedimentary facies model simulated by well logging seismic logging, a third adversarial generation subnetwork based on lithology model simulated by sedimentary dynamics under sedimentary facies constraints, and a fourth adversarial generation subnetwork based on lithology model simulated by well logging seismic logging under sedimentary facies constraints.

[0062] Based on well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library, lithology training template library, and a pre-established first adversarial generative network, a reservoir lithology distribution model is generated, including:

[0063] Multiple sedimentary facies training template data volumes from the sedimentary facies training template library are input into the first adversarial generation sub-network for training.

[0064] The well logging sedimentary facies data of the work area are input into the second adversarial generation subnetwork for training. Based on the training results, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. The second adversarial generation subnetwork is established based on the trained first adversarial generation subnetwork.

[0065] Multiple lithology training template data volumes from the lithology training template library are input into the third adversarial generation subnetwork for training;

[0066] The well logging data of the sand body interlayer in the work area is input into the fourth adversarial generation subnetwork for training. Based on the training results and the reservoir sedimentary facies model, multi-point geological statistical modeling is performed to generate a reservoir lithology distribution model. The fourth adversarial generation subnetwork is established based on the trained third adversarial generation subnetwork.

[0067] In one embodiment, inputting multiple sedimentary facies training template data volumes from the sedimentary facies training template library into a first adversarial generation sub-network for training includes:

[0068] The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network and input into the first adversarial generation sub-network for training;

[0069] The first error term is obtained by comparing the consistency between the selected sedimentary facies training template data and the output of the first adversarial generation sub-network.

[0070] The first adversarial generation subnetwork is corrected based on the first error term and multiple sedimentary phase training template data volumes until the first error term reaches its minimum value.

[0071] The well logging sedimentary facies data of the work area are input into the second adversarial generation subnetwork for training. Based on the training results, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. This includes: establishing a second adversarial generation subnetwork based on the trained first adversarial generation subnetwork; inputting the well logging sedimentary facies data of the work area into the second adversarial generation subnetwork for training; outputting a sedimentary facies template characterizing the characteristics of the well logging sedimentary facies data of the work area; and performing multi-point geological statistical modeling based on the sedimentary facies template characterizing the characteristics of the well logging sedimentary facies data of the work area and actual well data to generate a reservoir sedimentary facies model.

[0072] In practice, based on the length, width, height, and number of channels of each template data volume Si, a corresponding encoder E1 and generator G1 are established. The input layer size of the encoder module EE1 of E1 is SL×SW×SH×SC, and the batch training number is SM. The output layer size of the decoder module ED1 of E0 is SL×SW×SH×SC, and the batch training number is SM, with SC=1 for the deposition phase. G,1 is a copy of the decoder ED1 of E1. Each time, SM sedimentary facies training template data volumes {Sj|j=k...k+SM-1} are input to the encoder module EE1 of encoder E1. The output of the decoder module ED1 of E1 is {SOj|j=k...k+SM-1}. The discriminator D1 compares the conformity between {Sj|j=k...k+SM-1} and {SOj|j=k...k+SM-1} to obtain the error of E1, which is the first error term. By reverse-correcting the encoder and decoder modules of E1, the errors of {Sj|j=k...k+SM-1} and {SOj|j=k...k+SM-1} are minimized. At this time, the output of the encoder EE1 of E1 is a latent variable SZ that can effectively characterize the three-dimensional data volume features of the sedimentary facies simulation results. Its length, width, height, and number of channels are SZL, SZW, SZH, and SZC, and its designed size is SZL×SZW×SZH×SZC. The E1 decoder module ED1 can generate a sedimentary facies template ED_S that characterizes the volumetric features of the sedimentary facies simulation results data based on SZ and random noise.

[0073] Then, the coordinates of each point from all K well logging points and the sedimentary facies data for the entire area are compiled into a comprehensive well data set WS = [WSD1, WSS1, WSD2, WSS2, ..., WSDK, WSSK], where WSD represents coordinate data, WSS represents sedimentary facies numbering data, and WS is a 2K×D image, where D is the number of data points for each sedimentary facies in each well logging point. An encoder E2 is established, with an input layer size of 2K×D and a batch training size of SW. The output layer size of the decoder ED1 is also 2K×D, with a batch training size of SW. The hidden layer WZ output by EE2 has dimensions SZL, SZW, SZH, and SZC, with a size of SZL×SZW×SZH×SZC, consistent with the size of SZ. The trained hidden layer WZ of E2 can serve as a latent variable characterizing the sedimentary facies data features of the well logging data for the entire area.

[0074] Next, a generator G2 is established, with an input layer size of SZL×SZW×SZH×SZC and output layer length, width, height, and number of channels SL, SW, SH, and SC. G2 is initially set to copy the parameters of G1 to utilize the learned sedimentary pattern. The latent variable WZ from E2 is input into G2 to generate a new three-dimensional sedimentary facies template G2S, with length, width, height, and number of channels SL, SW, SH, and SC. Using G2S as the template and actual well data as constraints, multi-point geostatistical modeling is employed to obtain the reservoir sedimentary facies model FS.

[0075] It should be noted that, firstly, modeling is carried out based on well data and sedimentary facies training network. Here, we can first assume that the study area is a certain sedimentary facies. However, if the human understanding is incorrect, the sedimentary facies training network will adjust the sedimentary facies type output according to the error feedback, that is, correct the human understanding and achieve the optimal sedimentary facies model. Then, for lithological data, it will also adjust the interlayer mode output results accordingly, on the same principle.

[0076] Specifically, the first step is to enable the AI ​​network to generate different sedimentary facies as needed, but the AI ​​network doesn't actually know which sedimentary facies it is. Therefore, based on our understanding from basic research, we tell the AI ​​network to generate sedimentary facies type A to start modeling, and automatically correct the sedimentary facies type during the continuous modeling process, possibly resulting in sedimentary facies B at the end. However, our modeling results often cannot withstand dynamic data verification, so the final step incorporates dynamic data to correct the geological model, thus unifying static and dynamic aspects. Because this is done in a hierarchical and categorized manner, this process is more intelligent and convenient than letting the AI ​​network figure it out on its own, resulting in more stable calculations and higher efficiency. It also increases the scalability of the entire modeling method; if a completely new sedimentary facies type is discovered, the initial template library can be updated. For example, if someone initially believes it to be A, but it is actually B, and then uses A to train the network, the calculated error will be large, and the model will automatically look for B data in the template library. Furthermore, due to the characteristics of deep learning, it may even generate a sedimentary type that falls between A and B. The core idea is to use the AI ​​network to help correct sedimentary understanding and increase the accuracy of modeling.

[0077] In one embodiment, the reservoir property modeling method further includes:

[0078] Obtain the seismic data volume of the work area;

[0079] The consistency between the seismic data volume of the work area and the reservoir sedimentary facies model was compared to obtain the second error term;

[0080] The second adversarial generative subnetwork is corrected based on the second error term until the second error term reaches its minimum value;

[0081] Based on the sedimentary facies template characterizing the characteristics of well logging sedimentary facies data in the work area and actual well data, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. This includes: based on the sedimentary facies template characterizing the characteristics of well logging sedimentary facies data in the work area output by the second adversarial generation sub-network when the second error term reaches its minimum value and actual well data, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model.

[0082] In practice, discriminator D2 is used to calculate the comprehensive error between FS, the seismic attribute data volume SEGY, and the well logging data of the work area, which is also the second error term. The error calculated by D2 is used to correct G2 until the sedimentary facies model generated based on the sedimentary facies template generated by G2 has the smallest error with the well logging and seismic data of the work area, and the optimal reservoir sedimentary facies model OFS is output.

[0083] In one embodiment, multiple lithology training template datasets from the lithology training template library are input into a third adversarial generation subnetwork for training, including:

[0084] Multiple lithology training template data volumes are input into the third adversarial generative subnetwork for training;

[0085] The consistency between the lithology training template data volume and the output of the third adversarial generation subnetwork is compared to obtain the third error term.

[0086] The third adversarial generative subnetwork is modified based on the third error term until the third error term reaches its minimum value;

[0087] The well logging data of sand bodies and interlayers in the work area are input into the fourth adversarial generation subnetwork for training. Based on the training results and the reservoir sedimentary facies model, multi-point geological statistical modeling is performed to generate a reservoir lithology distribution model. This includes: establishing a fourth adversarial generation subnetwork based on the trained third adversarial generation subnetwork; inputting the well logging data of sand bodies and interlayers in the work area into the fourth adversarial generation subnetwork for training; outputting a lithology distribution template that characterizes the lithology data of the work area; and performing multi-point geological statistical modeling based on the lithology distribution template that characterizes the lithology data of the work area and actual well data to generate a reservoir lithology distribution model.

[0088] In practice, an encoder E3 and a generator G3 are established based on the length, width, height, and number of channels of each template data volume Lis. The input layer size of the encoder module EE3 of E3 is LL×LW×LH×LC, and the batch training number is LM. The output layer size of the decoder module ED3 of E3 is LL×LW×LH×LC, and the batch training number is LM. For the deposition phase, LC=1. G3 is a copy of the decoder module ED3 of E3. Each time, LM lithological training template data volumes {Lj|j=k...k+LM-1} are input to the encoder module EE3 of encoder E3. The output of the decoder module ED3 of E3 is {LOj|j=k...k+LM-1}. The consistency module compares {Lj|j=k...k+LM-1} and {LOj|j=k...k+LM-1} to obtain the error of E3. By reverse correction of the encoder and decoder modules of E3, the errors of {Lj|j=k...k+LM-1} and {LOj|j=k...k+LM-1} are minimized. At this time, the output of the encoder module EE3 of E3 is the latent variable LZ, which can effectively characterize the three-dimensional data volume of the sand body interlayer model result data. Its length, width, height and number of channels are LZL, LZW, LZH and LZC, and the designed size is LZL×LZW×LZH×LZC. The E3 decoder module ED3 can generate a lithological distribution template ED_L that characterizes the data volume of the sand body interlayer model results based on LZ and random noise.

[0089] Then, the coordinates of each point from all K well logging sites and the sand body interlayer data from the well logging sites are organized into a full-area well data WL = [WLD1, WLL1, WLD2, WLL2, ..., WLDK, WLLK], where WLD is the coordinate data, WLL is the sand body interlayer number data, and WL is a 2K×D image, where D is the number of data points for the sand body interlayer in each well logging site. An encoder E4 is established. The input layer size of the E4 encoder module EE4 is 2K×D, and the batch training size is LW. The output layer size of the E4 decoder module ED4 is 2K×D, and the batch training size is LW. The hidden layer LZ output by EE4 has a length, width, height, and number of channels LZL, LZW, LZH, and LZC, with a size of LZL×LZW×LZH×LZC, consistent with the size of LZ. The hidden layer WZ of E4 after training can be used as a parameter to characterize the logging core characteristics of the entire upper working area.

[0090] Furthermore, a generator G4 is established, with an input layer size of LZL×LZW×LZH×LZC and output layer length, width, height, and number of channels of LL, LW, LH, and LC, respectively. G4 is initially set to copy the parameters of G3 to utilize the learned sedimentary patterns. The latent variable LZ from E4 is input into G4 to generate a new lithological distribution template G1L, representing the characteristics of the well logging lithological data in the work area, with length, width, height, and number of channels of LL, LW, LH, and LC, respectively. Using G1L as the template and actual well data as constraints, multi-point geostatistical modeling is employed to obtain the reservoir lithological distribution model FLs for the s-th sedimentary facies region.

[0091] In one embodiment, the reservoir property modeling method further includes:

[0092] Obtain the seismic data volume of the work area;

[0093] A comparison of the consistency between the seismic data volume of the work area and the reservoir lithology distribution model yields the fourth error term.

[0094] The fourth adversarial generative subnetwork is modified based on the fourth error term until the fourth error term reaches its minimum value;

[0095] Based on the lithological distribution template characterizing the lithological characteristics of the well logging data in the work area and the actual well data, multi-point geological statistical modeling is performed to generate a reservoir lithological distribution model. This includes: based on the lithological distribution template characterizing the lithological characteristics of the well logging data in the work area output by the fourth adversarial generation sub-network when the fourth error term reaches its minimum value and the actual well data, multi-point geological statistical modeling is performed to generate a reservoir lithological distribution model.

[0096] In practice, discriminator D4 is used to calculate the comprehensive error between FLs and the seismic attribute data body SEGY and the well logging data of the work area, i.e., the fourth error term. D4 calculates the error and is used to correct G4 until the sedimentary facies model generated based on the 3D sand body interlayer template generated by G4 has the smallest error with the well logging and seismic data of the work area. The optimal reservoir lithology distribution model OFLs for the s-th sedimentary facies region is then output. This modeling process is repeated for each of the n sedimentary regions to complete the intelligent lithology modeling of all n sedimentary facies regions, thus obtaining the optimal reservoir lithology distribution model OFLs.

[0097] In step 103, a reservoir attribute model is generated based on the well logging attribute data of the work area, the reservoir lithology distribution model, and the pre-established second adversarial generation network. The second adversarial generation network is a geological attribute generation network based on lithology constraints.

[0098] In one embodiment, the well logging attribute data for the work area includes: well logging porosity attribute data for the work area, well logging permeability attribute data for the work area, and well logging saturation attribute data for the work area.

[0099] Based on the well logging attribute data of the work area, the reservoir lithology distribution model, and the pre-established second adversarial generation network, a reservoir attribute model is generated, including:

[0100] The well logging porosity attribute data and reservoir lithology distribution model of the work area are input into the second adversarial generative network, and the corresponding reservoir porosity attribute model is output.

[0101] The well logging permeability attribute data and reservoir lithology distribution model of the work area are input into the second adversarial generative network, and the corresponding reservoir permeability attribute model is output.

[0102] The well logging saturation attribute data and reservoir lithology distribution model of the work area are input into the second adversarial generative network, and the corresponding reservoir saturation attribute model is output.

[0103] In practice, the coordinates of each point from all K well logging points and the logging attribute data for the entire area are compiled into a comprehensive well data set WPq = [WPD1q, WPP1q, WPD2q, WPP2q, ..., WPDKq, WPPKq]. WPD represents the coordinate data, WPP represents the porosity data, and WP represents a 2K×D image, where D is the number of data points for each well logging attribute in the area, and q is the attribute number to be analyzed. The attribute number for porosity is q = 1, for permeability it is q = 2, and for saturation it is q = 3. Encoder E5 is established. The input layer size of E5's encoder module EE5 is 2K×D, and the batch training size is PW. The output layer size of E5's decoder module ED5 is 2K×D, and the batch training size is PW. The hidden layer PZ output by EE5 has a length, width, height, and number of channels PZL, PZW, PZH, and PZC, with a size of PZL×PZW×PZH×PZC, consistent with the size of PZ. The hidden layer PZq of E5 after training can be used as a parameter to characterize the q-th attribute data feature of the well logging in the entire work area, that is, the latent variable characterizing the saturation attribute data feature of the well logging in the work area. Generator G5 is established, with an input layer size of PZL×PZW×PZH×(2×PZC) and an output layer size of PZL×PZW×PZH. The first PZC channels in the input layer contain PZ data, and the remaining PZC channels extract lithological data blocks of size PZL×PZW×PZH from the OFL using a translational traversal. After G5 completes the OFL traversal, as shown... Figure 5As shown, the reservoir attribute model OFPp for the p-th attribute is generated. After G4 traverses the OFL, a three-dimensional reservoir attribute model OFPp is generated. Compared to inputting the entire OFL into G4 at once, moving the traversal reduces the size of the G4 input layer, thus reducing hardware requirements. Following the above process, reservoir attribute models OPFS = {OPF1, OFP2, OPF3} for porosity, permeability, and saturation attributes are generated sequentially, where OPF1 is the optimal reservoir porosity attribute model, OPF2 is the optimal reservoir permeability attribute model, and OPF3 is the optimal reservoir saturation attribute model.

[0104] In step 104, based on the dynamic production data of the work area, the reservoir attribute model is modified using the reservoir attribute model of the reservoir numerical simulation agent and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints.

[0105] In one embodiment, based on the dynamic production data of the work area, the reservoir attribute model is modified using the reservoir attribute model of the reservoir numerical simulation agent and a pre-established third adversarial generative network, including:

[0106] The fifth error term is obtained by comparing the consistency between the dynamic production data of the work area and the reservoir attribute model after the reservoir numerical simulation agent.

[0107] The third adversarial generative network is modified based on the fifth error term until the fifth error term reaches its minimum value;

[0108] The reservoir property model output by the third adversarial generative network when the fifth error term reaches its minimum value is taken as the corrected reservoir property model.

[0109] In practical implementation, based on OPFS, the reservoir numerical simulation surrogate RS is used to predict the production parameters of the work area. The discriminator D5 is then used to calculate the error Err1 between the predicted data and the actual production parameters FO of the work area, as well as the error Err2 between the generated porosity, permeability, and saturation attributes and the well logging attribute data of the work area. This yields the total error Err = Err1 + Err2 for both dynamic and static data. Based on this error Err, the generator G5 is corrected in reverse until the error obtained by D5 is minimized, at which point the optimal reservoir attribute model OFP is output.

[0110] The following is a specific embodiment illustrating the application of reservoir attribute modeling in this invention. This specific embodiment uses the X block in South America as an example. First, a sedimentary dynamics model conforming to the sedimentary environment of the X block was constructed, resulting in 2000 three-dimensional sedimentary facies simulation data volumes and 4000 three-dimensional sand body-interlayer model data volumes, which were used as training data volumes. Simultaneously, well logging sedimentary facies data from 400 wells in the study area, sand body-interlayer data from 400 wells, and porosity, permeability, and saturation data from 400 wells were collected. Dynamic production data from the 400 wells in the study area were compiled. Seismic data volumes from the study area were collected, achieving full coverage of the study area. Based on the AMAX deep learning environment, using an environment with eight 48GB RTX8000 graphics cards, encoders E0, G0, G1, E1, G2, E2, G3, E3, and G4, as well as discriminators D1, D2, D3, and D4, were built. Through 21,000 reverse training iterations, the generated 3D reservoir attribute model achieved a 98.2% agreement rate with the well logging porosity data, a 97.3% agreement rate with the well logging permeability data, a 98.5% agreement rate with the well logging saturation data, and a 94.1% agreement rate with the well logging production data. Using existing MPS non-AI methods, the resulting 3D reservoir attribute model achieved an 82.4% agreement rate with the well logging porosity data, an 83.2% agreement rate with the well logging permeability data, an 81.1% agreement rate with the well logging saturation data, and a 79.2% agreement rate with the well logging production data. The dynamic and static data agreement rate of the reservoir attribute model obtained by this invention is higher than that of existing stochastic modeling methods.

[0111] This invention presents a multi-point geostatistical 3D template generation method based on artificial intelligence and combining dynamic and static methods. It effectively addresses the issue of existing MPS (Multi-Point Geostatistical System) templates failing to consider dynamic production data. Furthermore, it utilizes multi-level control based on artificial intelligence for reservoir attribute modeling. First, a sedimentary facies model is established using artificial intelligence methods. Then, under the control of the sedimentary facies model, a lithology model based on artificial intelligence is established. Finally, under lithology control, an attribute model is established based on artificial intelligence. This multi-level control not only reduces training difficulty but also improves model accuracy.

[0112] Based on the same inventive concept, embodiments of the present invention also provide a reservoir property modeling apparatus, as described in the following embodiments. Since the principles of solving these problems are similar to those of the reservoir property modeling method, the implementation of the reservoir property modeling apparatus can be referred to the implementation of the method, and repeated details will not be elaborated further.

[0113] Figure 5 This is a structural diagram of the reservoir property modeling device in an embodiment of the present invention, such as... Figure 5 As shown, the reservoir property modeling apparatus includes:

[0114] The data acquisition module 501 is used to acquire well logging data of the work area, production dynamic data of the work area, sedimentary facies training template library and lithology training template library. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area and well logging attribute data of the work area.

[0115] The reservoir lithology distribution model generation module 502 is used to generate a reservoir lithology distribution model based on well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library and lithology training template library, and a pre-established first adversarial generation network. The first adversarial generation network is a deep convolutional adversarial generation network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network.

[0116] The reservoir attribute model generation module 503 is used to generate a reservoir attribute model based on the well logging attribute data of the work area, the reservoir lithology distribution model and the pre-established second adversarial generation network. The second adversarial generation network is a geological attribute generation network based on lithological constraints.

[0117] The reservoir attribute model optimization module 504 is used to modify the reservoir attribute model based on the dynamic production data of the work area, the reservoir attribute model of the reservoir numerical simulation agent, and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints.

[0118] Based on the aforementioned inventive concept, such as Figure 6 As shown, this embodiment of the invention also provides a computer device 600, including a memory 610, a processor 620, and a computer program 630 stored in the memory 610 and executable on the processor 620. When the processor 620 executes the computer program 630, it implements the above-mentioned reservoir attribute modeling method.

[0119] Based on the foregoing inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described reservoir property modeling method.

[0120] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described reservoir attribute modeling method.

[0121] Compared with the existing technology of random modeling of reservoir attributes, the embodiments of the present invention obtain well logging data of the work area, production dynamic data of the work area, sedimentary facies training template library, and lithology training template library. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area, and well logging attribute data of the work area. Based on the well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area, sedimentary facies training template library, and lithology training template library, and a pre-established first adversarial generative network, a reservoir lithology distribution model is generated. The first adversarial generative network is a depth convolution adversarial network based on multi-level well-seismic joint operation. The network is generated by using sedimentary facies training template data of a preset type selected from a sedimentary facies training template library as the initial input data. Based on well logging attribute data of the work area, a reservoir lithology distribution model, and a pre-established second adversarial generation network, a reservoir attribute model is generated. The second adversarial generation network is a geological attribute generation network based on lithology constraints. Based on the work area's production dynamic data, the reservoir attribute model is modified using the reservoir numerical simulation proxy and a pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints. This embodiment of the invention introduces dynamic production content into the reservoir attribute modeling process. By combining the sedimentary facies training template library and the lithology training template library of the work area's well logging data, a reservoir lithology distribution model and a reservoir attribute model based on an artificial intelligence adversarial generation network are generated sequentially. Then, the reservoir attribute model is modified by combining the dynamic production content, effectively constructing a reservoir attribute model that matches both dynamic and static data, and improving the consistency rate of dynamic and static data in random reservoir attribute modeling.

[0122] 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 embodied 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.

[0123] 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 1A device that provides the functions specified in one or more boxes.

[0124] 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 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A reservoir property modeling method, characterized in that, include: Obtain well logging data for the work area, dynamic production data for the work area, sedimentary facies training template library and lithology training template library. The well logging data for the work area includes: sedimentary facies data for the work area, sand body interlayer data for the work area, and attribute data for the work area. Based on the well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library and lithology training template library, and the pre-established first adversarial generative network, a reservoir lithology distribution model is generated. The first adversarial generative network is a deep convolutional adversarial generative network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network. Based on the well logging attribute data of the work area, the reservoir lithology distribution model, and the pre-established second adversarial generation network, a reservoir attribute model is generated. The second adversarial generation network is a geological attribute generation network based on lithological constraints. Based on the dynamic production data of the work area, the reservoir attribute model is corrected through the reservoir attribute model of the reservoir numerical simulation agent and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints. The first adversarial generation network includes: a first adversarial generation subnetwork based on sedimentary facies model simulated by sedimentary dynamics, a second adversarial generation subnetwork based on sedimentary facies model simulated by well logging seismic logging, a third adversarial generation subnetwork based on lithology model simulated by sedimentary dynamics under sedimentary facies constraints, and a fourth adversarial generation subnetwork based on lithology model simulated by well logging seismic logging under sedimentary facies constraints. Based on well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library, lithology training template library, and a pre-established first adversarial generative network, a reservoir lithology distribution model is generated, including: Multiple sedimentary facies training template data volumes from the sedimentary facies training template library are input into the first adversarial generation sub-network for training. The well logging sedimentary facies data of the work area are input into the second adversarial generation subnetwork for training. Based on the training results, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. The second adversarial generation subnetwork is established based on the trained first adversarial generation subnetwork. Multiple lithology training template data volumes from the lithology training template library are input into the third adversarial generation subnetwork for training; The well logging data of the sand body interlayer in the work area is input into the fourth adversarial generation subnetwork for training. Based on the training results and the reservoir sedimentary facies model, multi-point geological statistical modeling is performed to generate a reservoir lithology distribution model. The fourth adversarial generation subnetwork is established based on the trained third adversarial generation subnetwork.

2. The reservoir property modeling method as described in claim 1, characterized in that, Pre-establish sedimentary facies training template libraries and lithology training template libraries as follows: The data body contains sedimentary facies information and sand body interlayer model results. The sedimentary facies information includes: sedimentary dynamics data, existing sedimentary facies maps, and satellite images, or any combination thereof. Based on the sedimentary facies information, a sedimentary facies training template library shall be established; Based on the data volume of the sand body interlayer pattern, a lithology training template library is established.

3. The reservoir property modeling method as described in claim 1, characterized in that, Multiple sedimentary facies training template data volumes from the sedimentary facies training template library are input into the first adversarial generation sub-network for training, including: The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network and input into the first adversarial generation sub-network for training; The first error term is obtained by comparing the consistency between the selected sedimentary facies training template data and the output of the first adversarial generation sub-network. The first adversarial generation subnetwork is corrected based on the first error term and multiple sedimentary phase training template data volumes until the first error term reaches its minimum value. The well logging sedimentary facies data of the work area are input into the second adversarial generation subnetwork for training. Based on the training results, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. This includes: establishing a second adversarial generation subnetwork based on the trained first adversarial generation subnetwork; inputting the well logging sedimentary facies data of the work area into the second adversarial generation subnetwork for training; outputting a sedimentary facies template characterizing the characteristics of the well logging sedimentary facies data of the work area; and performing multi-point geological statistical modeling based on the sedimentary facies template characterizing the characteristics of the well logging sedimentary facies data of the work area and actual well data to generate a reservoir sedimentary facies model.

4. The reservoir property modeling method as described in claim 3, characterized in that, Also includes: Obtain the seismic data volume of the work area; The consistency between the seismic data volume of the work area and the reservoir sedimentary facies model was compared to obtain the second error term; The second adversarial generative subnetwork is corrected based on the second error term until the second error term reaches its minimum value; Based on the sedimentary facies template characterizing the characteristics of well logging sedimentary facies data in the work area and actual well data, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. This includes: based on the sedimentary facies template characterizing the characteristics of well logging sedimentary facies data in the work area output by the second adversarial generation sub-network when the second error term reaches its minimum value and actual well data, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model.

5. The reservoir property modeling method as described in claim 1, characterized in that, Multiple lithology training template datasets from the lithology training template library are input into a third adversarial generative subnetwork for training, including: Multiple lithology training template data volumes are input into the third adversarial generative subnetwork for training; The consistency between the lithology training template data volume and the output of the third adversarial generation subnetwork is compared to obtain the third error term. The third adversarial generative subnetwork is modified based on the third error term until the third error term reaches its minimum value; The well logging data of sand bodies and interlayers in the work area are input into the fourth adversarial generation subnetwork for training. Based on the training results and the reservoir sedimentary facies model, multi-point geological statistical modeling is performed to generate a reservoir lithology distribution model. This includes: establishing a fourth adversarial generation subnetwork based on the trained third adversarial generation subnetwork; inputting the well logging data of sand bodies and interlayers in the work area into the fourth adversarial generation subnetwork for training; outputting a lithology distribution template that characterizes the lithology data of the work area; and performing multi-point geological statistical modeling based on the lithology distribution template that characterizes the lithology data of the work area and actual well data to generate a reservoir lithology distribution model.

6. The reservoir property modeling method as described in claim 5, characterized in that, Also includes: Obtain the seismic data volume of the work area; A comparison of the consistency between the seismic data volume of the work area and the reservoir lithology distribution model yields the fourth error term. The fourth adversarial generative subnetwork is modified based on the fourth error term until the fourth error term reaches its minimum value; Based on the lithological distribution template characterizing the lithological characteristics of the well logging data in the work area and the actual well data, multi-point geological statistical modeling is performed to generate a reservoir lithological distribution model. This includes: based on the lithological distribution template characterizing the lithological characteristics of the well logging data in the work area output by the fourth adversarial generation sub-network when the fourth error term reaches its minimum value and the actual well data, multi-point geological statistical modeling is performed to generate a reservoir lithological distribution model.

7. The reservoir property modeling method as described in claim 1, characterized in that, The well logging attribute data for the work area includes: well logging porosity attribute data, well logging permeability attribute data, and well logging saturation attribute data. Based on the well logging attribute data of the work area, the reservoir lithology distribution model, and the pre-established second adversarial generation network, a reservoir attribute model is generated, including: The well logging porosity attribute data and reservoir lithology distribution model of the work area are input into the second adversarial generative network, and the corresponding reservoir porosity attribute model is output. The well logging permeability attribute data and reservoir lithology distribution model of the work area are input into the second adversarial generative network, and the corresponding reservoir permeability attribute model is output. The well logging saturation attribute data and reservoir lithology distribution model of the work area are input into the second adversarial generative network, which outputs the corresponding reservoir saturation attribute model.

8. The reservoir property modeling method as described in claim 7, characterized in that, Based on the dynamic production data of the work area, the reservoir attribute model is modified using the reservoir numerical simulation surrogate model and a pre-established third adversarial generative network, including: The fifth error term is obtained by comparing the consistency between the dynamic production data of the work area and the reservoir attribute model after the reservoir numerical simulation agent. The third adversarial generative network is modified based on the fifth error term until the fifth error term reaches its minimum value; The reservoir property model output by the third adversarial generative network when the fifth error term reaches its minimum value is taken as the corrected reservoir property model.

9. A reservoir property modeling device, characterized in that, include: The data acquisition module is used to acquire well logging data of the work area, production dynamic data of the work area, sedimentary facies training template library and lithology training template library. The well logging data of the work area includes: well logging sedimentary facies data of the work area, well logging sand body interlayer data of the work area and well logging attribute data of the work area. The reservoir lithology distribution model generation module is used to generate a reservoir lithology distribution model based on well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library and lithology training template library, as well as a pre-established first adversarial generation network. The first adversarial generation network is a deep convolutional adversarial generation network based on multi-level well-seismic joint operation. The preset type of sedimentary facies training template data selected from the sedimentary facies training template library is used as the initial input data of the network. The reservoir attribute model generation module is used to generate a reservoir attribute model based on well logging attribute data of the work area, reservoir lithology distribution model and pre-established second adversarial generation network. The second adversarial generation network is a geological attribute generation network based on lithological constraints. The reservoir attribute model optimization module is used to modify the reservoir attribute model based on the dynamic production data of the work area, the reservoir attribute model of the reservoir numerical simulation agent, and the pre-established third adversarial generation network. The third adversarial generation network is an attribute optimization network based on dynamic constraints. The first adversarial generation network includes: a first adversarial generation subnetwork based on sedimentary facies model simulated by sedimentary dynamics, a second adversarial generation subnetwork based on sedimentary facies model simulated by well logging seismic logging, a third adversarial generation subnetwork based on lithology model simulated by sedimentary dynamics under sedimentary facies constraints, and a fourth adversarial generation subnetwork based on lithology model simulated by well logging seismic logging under sedimentary facies constraints. Based on well logging sedimentary facies data, well logging sand body interlayer data, sedimentary facies training template library, lithology training template library, and a pre-established first adversarial generative network, a reservoir lithology distribution model is generated, including: Multiple sedimentary facies training template data volumes from the sedimentary facies training template library are input into the first adversarial generation sub-network for training. The well logging sedimentary facies data of the work area are input into the second adversarial generation subnetwork for training. Based on the training results, multi-point geological statistical modeling is performed to generate a reservoir sedimentary facies model. The second adversarial generation subnetwork is established based on the trained first adversarial generation subnetwork. Multiple lithology training template data volumes from the lithology training template library are input into the third adversarial generation subnetwork for training; The well logging data of the sand body interlayer in the work area is input into the fourth adversarial generation subnetwork for training. Based on the training results and the reservoir sedimentary facies model, multi-point geological statistical modeling is performed to generate a reservoir lithology distribution model. The fourth adversarial generation subnetwork is established based on the trained third adversarial generation subnetwork.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

12. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

Images