CA3249948A1

CA3249948A1 - Reservoir property modeling method and apparatus

Info

Publication number: CA3249948A1
Application number: CA3249948A
Authority: CA
Inventors: Kexin Zhang; Chaoqian Zhang; Heping Chen; Fang Xu; Zheng Meng; Songwei GUO; Junchang WU; Yusheng Wang; Bin Han; Yuan Tian
Original assignee: Petrochina Co Ltd
Priority date: 2022-08-08
Filing date: 2023-08-08
Publication date: 2025-06-17
Also published as: CN117592350B; CN117592350A; WO2024032604A1

Abstract

The present disclosure discloses a method and apparatus for reservoir property modeling. The method includes: generating a reservoir lithology distribution model according to a sedimentary facies training template library, a lithology training template library, and a first generative adversarial network which is a deep convolutional generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; generating a reservoir property model according to work-area logging property data, the reservoir lithology distribution model, and a second generative adversarial network which is a geological property generation network based on lithology constraints; and correcting the reservoir property model according to work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a third generative adversarial network which is a property optimization network based on dynamic constraints. The present disclosure can effectively improve the conformity of dynamic and static data for stochastic modeling of reservoir properties.

Description

RESERVOIR PROPERTY MODELING METHOD AND APPARATUS TECHNICAL FIELD

[0001] The present disclosure relates to the field of reservoir development, and particularly, 5 to a method and apparatus for reservoir property modeling. BACKGROUND

[0002] This section is intended to provide background or context for the embodiments of the present disclosure set forth in the claims. The description here is not admitted to be the prior 10 art just because it is included in this section.

[0003] With the deepening of reservoir development, the requirements for the geological model are increasingly higher for fine reservoir development. Generally, the geological model includes a reservoir structure model, a reservoir property model, etc. The reservoir property model is particularly important for reservoir management and development plan optimization 15 because it can directly provide basic input parameters for a reservoir numerical simulation model.

[0004] At present, the reservoir property modeling methods usually include deterministic modeling and stochastic modeling. In this two modeling methods: ① The deterministic modeling method is mainly constructed from the work-area logging 20 data based on the methods such as Kriging interpolation, with only one deterministic output for each input. ② The stochastic modeling method, mainly including sequential Gaussian modeling, multi-point geostatistics modeling (MPS), etc., is mainly based on the input logging data to obtain statistical characteristics, including variance, variation, average, etc., and then generate a 25 stochastic model based on Monte Carlo sampling. In the stochastic modeling method, under the condition that the input is determined, the results of each Monte Carlo sampling modeling have certain randomness on the premise of meeting the statistical characteristics, and the modeling results are different each time.

[0005] Compared with the stochastic modeling, the results of the deterministic modeling, 30 adopting the spatial interpolation, lack response to the actual geological complexity. Therefore, at present, the reservoir property modeling mainly adopts the stochastic modeling method, but the existing stochastic modeling for the reservoir property has the problem of inconsistency between dynamic data and static data.

[0006] Therefore, there is an urgent need for a reservoir property modeling solution that2 can overcome the above problem. SUMMARY

[0007] The embodiments of the present disclosure provide a reservoir property modeling 5 method, so as to improve the conformity rate between dynamic data and static data of stochastic modeling for reservoir properties, the method including: [0008] obtaining work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, in which the work-area logging data includes work-area logging sedimentary facies data, work-area 10 logging sand body and interlayer data, and work-area logging property data; [0009] generating a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is 15 a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; [0010] generating a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial 20 network, in which the second generative adversarial network is a geological property generation network based on lithology constraints; and [0011] correcting the reservoir property model according to the work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third 25 generative adversarial network is a property optimization network based on dynamic constraints.

[0012] The embodiments of the present disclosure provide a reservoir property modeling apparatus, so as to improve the conformity rate between dynamic data and static data of stochastic modeling for reservoir properties, the apparatus including: [0013] a data obtaining module configured to obtain work-area logging data, work-area 30 production dynamic data, a sedimentary facies training template library and a lithology training template library, in which the work-area logging data includes work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; [0014] a reservoir lithology distribution model generation module configured to generate a3 reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is a deep convolution generative adversarial 5 network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; [0015] a reservoir property model generation module configured to generate a reservoir property model according to the work-area logging property data, the reservoir lithology 10 distribution model, and a pre-built second generative adversarial network, in which the second generative adversarial network is a geological property generation network based on lithology constraints; and [0016] a reservoir property model optimization module configured to correct the reservoir property model according to the work-area production dynamic data, the reservoir property 15 model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on dynamic constraints.

[0017] The embodiments of the present disclosure further provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable 20 on the processor, and when executing the computer program, the processor implements the aforementioned reservoir property modeling method.

[0018] The embodiments of the present disclosure further provide a computer-readable storage medium, storing a computer program, and when executed by a processor, the computer program implements the aforementioned reservoir property modeling method. 25 [0019] The embodiments of the present disclosure further provide a computer program product, including a computer program, and when executed by a processor, the computer program implements the aforementioned reservoir property modeling method.

[0020] Compared with the technical solution of stochastic modeling for reservoir properties in the prior art, the embodiments of the present disclosure obtain work-area logging 30 data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, in which the work-area logging data includes work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; generate a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data,4 the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary 5 facies training template library is taken as network initial input data; generate a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial network, in which the second generative adversarial network is a geological property generation network based on lithology constraints; and correct the reservoir property model according to the work-area production 10 dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on dynamic constraints. According to the embodiments of the present disclosure, dynamic production content is introduced during the reservoir property modeling, and a reservoir lithology distribution model 15 and a reservoir property model based on an artificial intelligence generative adversarial network are generated in sequence by combining the work-area logging data, the sedimentary facies training template library and the lithology training template library, and then the reservoir property model is corrected by combining the dynamic production content, so as to effectively build a reservoir property model with dynamic and static consistency, and improve the 20 conformity between the dynamic data and static data of stochastic modeling for reservoir properties. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the 25 present disclosure or in the prior art, the drawings to be used in the descriptions of the embodiments or the prior art will be briefly introduced as follows. Obviously, the drawings involved in the following description just illustrate some embodiments of the present disclosure, and those of ordinary skill in the art may obtain other drawings from these drawings without paying any inventive labor. In the drawings: 30 [0022] FIG. 1 illustrates a schematic diagram of a reservoir property modeling method according to an embodiment of the present disclosure;

[0023] FIG. 2 illustrates a schematic diagram of a template library building method according to an embodiment of the present disclosure;

[0024] FIG. 3 illustrates a flowchart of reservoir property modeling according to a specific5 embodiment of the present disclosure;

[0025] FIG. 4 illustrates a schematic diagram of generation of a reservoir property model according to an embodiment of the present disclosure;

[0026] FIG. 5 illustrates a structural diagram of a reservoir property modeling apparatus 5 according to the embodiment of the present disclosure; and

[0027] FIG. 6 illustrates a structural diagram of a computer device according to an embodiment of the present disclosure. DESCRIPTION OF THE EMBODIMENTS 10 [0028] In order to that the objectives, technical solutions and advantages of the embodiments of the present disclosure are clearer, the embodiments of the present disclosure will be further described in detail below with reference to the drawings. Here, the exemplary embodiments of the present disclosure and the descriptions thereof are used to explain the present disclosure, but are not used as limitations of the present disclosure. 15 [0029] Firstly, the technical terms in the embodiments of the present disclosure are introduced:

[0030] Artificial intelligence, which is implemented based on a computer to produce an intelligent model that responds in a similar way to human intelligence, mainly relating to robots, language recognition, image recognition, natural language processing and expert systems. 20 [0031] Dynamic data and static data: the dynamic data generally refers to the change of liquid production during the reservoir production. The dynamic data characteristics are directly related to the static parameter characteristics of the reservoir. The static data generally refers to an initial porosity, an initial permeability, an initial saturation, etc. in a reservoir analysis, and the static data also includes data which does not change with the development process, such as 25 seismic data.

[0032] Work-area production dynamic data, i.e., the dynamic data, which is the data such as oil production and water production changing over time during oilfield production. The data packet is accompanied with corresponding engineering parameter data.

[0033] Geological modeling: a three-dimensional (3D) static model of a reservoir is 30 generated using computer technologies based on logging geological data, logging data, seismic data, etc. in a work area of the reservoir. Generally, a reservoir porosity model, a reservoir permeability model and a reservoir saturation model are included, which are used for reservoir numerical simulation researches, prediction of the change of future production data, guidance of well location design, etc.6

[0034] Reservoir property modeling: a reservoir static property model is built based on work-area logging data, and a reservoir porosity model, a reservoir permeability model and a reservoir saturation model are included.

[0035] Sedimentary dynamics simulation: a sedimentary model conforming to sedimentary 5 facies in a research area is built based on sedimentary environment parameters of the research area, and carries out sedimentary simulation to obtain a 3D sedimentary data volume under specific sedimentary environment parameters.

[0036] Encoder, which is also called an auto-encoder, and is a widely used neural network for automatic encoding of two-dimensional image data and 3D image data in the field of 10 artificial intelligence. The auto-encoder includes two parts: an encoder and a decoder. The auto-encoder has a function of representing a learning algorithm in a general sense, and is applied to feature dimensionality reduction analysis.

[0037] Hidden variable, which is a dimensionality reduction feature array obtained by the encoder. 15 [0038] Generator, which is a neural network composed of combined of de-convolution, un-pooling and non-linear mapping, and can generate 3D data images according to an input feature array.

[0039] Discriminator, which is a neural network composed of combined of convolution, pooling and non-linear mapping, and can generate a 3D data image according to an input feature 20 array.

[0040] Multipoint geostatistics modeling (MPS modeling), which is a method that takes a 3D image as a training template, extracts statistical features between data from the template, and builds a model through Monte Carlo sampling based on the statistical features.

[0041] Template, which is a basic image used in MPS modeling, and the MPS method 25 statistics statistical characteristics between the property data based on the image. The template includes a 3D sedimentary facies template and a 3D lithology distribution template.

[0042] Lithology, which is a classification of a sand body and an interlayer in a reservoir, in which the sand body is 1, and the interlayer is 0.

[0043] In order to perform reservoir attribute modeling and improve the conformity rate 30 between dynamic data and static data of stochastic modeling for reservoir properties, an embodiment of the present disclosure provides a reservoir property modeling method as illustrated in FIG. 1, and the method include: [0044] step 101: obtaining work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, in which7 the work-area logging data includes work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; [0045] step 102: generating a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, 5 the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; 10 [0046] step 103: generating a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial network, in which the second generative adversarial network is a geological property generation network based on lithology constraints; and [0047] step 104: correcting the reservoir property model according to the work-area 15 production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on dynamic constraints.

[0048] As illustrated in FIG. 1, compared with the technical solution of stochastic 20 modeling for reservoir properties in the prior art, the embodiments of the present disclosure obtain work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, in which the work-area logging data includes work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; generate a reservoir lithology distribution model 25 according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected 30 from the sedimentary facies training template library is taken as network initial input data; generate a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial network, in which the second generative adversarial network is a geological property generation network based on lithology constraints; and correct the reservoir property model according to the8 work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on dynamic constraints. According to the embodiments of the present disclosure, dynamic 5 production content is introduced during the reservoir property modeling, and a reservoir lithology distribution model and a reservoir property model based on an artificial intelligence generative adversarial network are generated in sequence by combining the work-area logging data, the sedimentary facies training template library and the lithology training template library, and then the reservoir property model is corrected by combining the dynamic production content, 10 so as to effectively build a reservoir property model with dynamic and static consistency, and improve the conformity between the dynamic data and static data of stochastic modeling for reservoir properties.

[0049] The inventor finds that at present, the stochastic modeling for reservoir properties still has the problem of inconsistent between dynamic data and static data. Although the 15 work-area logging static data is fully considered in the current modeling, the work-area logging dynamic data is completely ignored. However, the dynamic data has a high reference value for some properties with a large uncertainty in logging interpretation. For example, the current logging method can accurately interpret the porosity based on an AC acoustic logging curve, but the uncertainties of the permeability and the saturation are far larger than that of the porosity in 20 practice although there are a plurality of models for the permeability and the saturation. At present, there is no direct logging interpretation solution for the permeability, and the interpretation mainly depends on a porosity-permeability relationship model. However, at present, it is difficult to express the porosity-permeability relationship model in a complex reservoir with a simple linear model, and a correlation of the porosity and the permeability is 25 also very poor for a tight reservoir. Thus, it is very difficult and uncertain to interpret the permeability model based on logging at present. The uncertainty of a permeability property model obtained based on the logging interpretation is much higher than the uncertainty of the porosity. Similarly, the logging interpretation of the reservoir saturation is also difficult, mainly because the argillaceous and coaly in the reservoir which interfere with the interpretation model, 30 resulting in a large deviation between the calculated water saturation and the actual value. The conventional stochastic modeling method only depends on the work-area logging data. When there is a large deviation in the work-area logging interpretation data, it inevitably leads to the failure of the model to accurately predict the dynamic data, resulting in the problem of the non-conformity between the dynamic data and the static data. Historical fitting is the main means9 to solve the problem of the geological model of the non-conformity between the dynamic data and the static data. The reservoir property model, as the basic input parameter, is input into the reservoir numerical simulation model to predict a work-area logging dynamic production curve. By comparing the work-area logging production curve obtained by numerical simulation and the 5 actual production curve of the oilfield, it can be determined whether there is the problem of the non-conformity between the dynamic data and the static data in the reservoir property model. If the difference between the predicted dynamic production area and the actual production curve is small, it is generally considered that there is substantially no the problem of the non-conformity between the dynamic data and the static data in the reservoir property model built based on the 10 conventional stochastic modeling method. Otherwise, it is necessary to make a reverse adjustment, so as to adjust the reservoir property model through historical fitting and a lot of artificial corrections until the dynamic data and static data are consistent. However, there are still many problems in solving the problem of the non-conformity between the dynamic data and the static data in the existing stochastic modeling for reservoir properties by the historical fitting 15 method. First of all, the simulation workload is very large, because in the stage of historical fitting, a reverse correction should be continuously made for the model, and each correction requires a reservoir numerical simulation, thereby consuming a lot of calculation data. Secondly, when the reservoir property model is reversely adjusted, the existing historical fitting method does not consider the geological rule between wells, and mainly dividing into zones and blocks, 20 and carrying out a rigid adjustment of the properties around a well with the problem of the non-conformity between the dynamic data and the static data, without considering the data of adjacent wells. Therefore, based on the artificial intelligence method, the embodiment of the present disclosure builds a 3D reservoir property model with the conformity between the dynamic data and the static data by combining the existing sedimentary dynamics simulation 25 data of the work area, the seismic data, the work-area logging sand body and interlayer data and the work-area dynamic production data. In this embodiment, the sedimentary dynamics data, the seismic data, the work-area logging sedimentary facies data, the sand body and interlayer data, the work-area logging porosity data, the work-area logging permeability data, the work-area logging saturation data and the work-area logging production data (including engineering data) 30 are the early-stage basic data needed by the present disclosure. The specific implementation includes three parts: sedimentary facies modeling based on artificial intelligence, lithology modeling based on artificial intelligence and property modeling based on artificial intelligence. Based on an encoder, a generator and a discriminator network, a property model conforming to static data and dynamic production data of a research area is generated using a hierarchical10 control method.

[0050] Each step is analyzed in detail below.

[0051] In step 101, work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library are obtained, 5 in which the work-area logging data includes work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data.

[0052] In an embodiment, as illustrated in FIG. 2, the sedimentary facies training template library and the lithology training template library are pre-built including following steps: [0053] step 201: obtaining sedimentary facies information and a sand body and interlayer 10 pattern result data volume, in which the sedimentary facies information includes any one or any combination of sediment dynamics data, an existing sedimentary facies map and a satellite image; [0054] step 202: building the sedimentary facies training template library according to the sedimentary facies information; and 15 [0055] step 203: building the lithology training template library according to the sand body and interlayer pattern result data volume.

[0056] During implementation, taking flume data as an example, it is possible to extract a first preset number of sedimentary facies training template data volumes from the sediment dynamics data, extract a second preset number of sedimentary facies training template data 20 volumes from a flume experimental data volume, and merge the extracted first preset number of sedimentary facies training template data volumes and the second preset number of sedimentary facies training template data volumes to establish a sedimentary facies training template library.

[0057] FIG. 3 illustrates a flowchart of reservoir property modeling according to an embodiment of the present disclosure. During implementation, p sedimentary facies training 25 templates, i.e., {SDi|i=1...p}, are obtained from a 3D data volume of sedimentary facies simulation result data in the sediment dynamics data, to build a sedimentary facies training template library. In the sedimentary facies training template library, each template data volume SDi has a length of SL, a width of SW, a height of SH and a channel number SC. The sedimentary flume simulation is completed using a flume scanning device, and a flume 30 experimental data volume is obtained, in which the flume scanning device is illustrated in FIG. 4. q sedimentary facies training templates, i.e., {SCi|i=1...q}, are obtained from a 3D data volume of flume experimental data, to build a sedimentary facies training template library. In the sedimentary facies training template library, each template data volume Sci has a length of SL, a width of SW, a height of SH and a channel number of SC. Two parts of sedimentary facies11 training templates are merged together to form {Si|i=1...n}, so as to build a sedimentary facies training template library. In the sedimentary facies training template library, each template data volume Si has a length of SL, a width of SW, a height of SH and a channel number of SC. h lithology training template data volumes, i.e., {Lis|i=1...h,s=1...n}, where s represents a 5 sedimentary facies serial number, for each sedimentary facies area are obtained from the 3D data volume of the sand body and interlayer pattern result data in the sediment dynamics data, so as to build a lithology training template library. Lis has a value range of {0,1}, where 0 represents sandstone and 1 represents mudstone. In the lithology training template library, each template data volume Lis has a length of LL, a width of LW, a height of LH, and a channel number of LC. 10 In this embodiment, the flume device is a device that simulates the transport and the sedimentation phenomenon of sedimentary particulates in liquid. The flume device may be used to reproduce the sedimentation process and analyze the internal configuration structure characteristics of a sedimentary body. The flume scanning device integrates a large-scale CT (Computed Tomography) scanning device and a flume device, and uses the large-scale CT 15 scanning device to perform 3D scanning of the flume sedimentation result to obtain a 3D model of the flume sedimentation result, which can clearly show the internal configuration structures of different sedimentation bodies, and the scanning results are stored in the form of 3D data volumes, to be applied to artificial intelligence for analysis.

[0058] In step 102, a reservoir lithology distribution model is generated according to the 20 work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary 25 facies training template library is taken as network initial input data.

[0059] In an embodiment, the first generative adversarial network includes a first generative adversarial sub-network based on a sedimentary facies pattern simulated by a sediment dynamics, a second generative adversarial sub-network based on a logging seismic sedimentary facies pattern, a third generative adversarial sub-network based on a lithology 30 pattern simulated by the sediment dynamics under sedimentary facies constraints, and a fourth generative adversarial sub-network based on a logging seismic lithology pattern under sedimentary facies constraints.

[0060] The generating the reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data,12 the sedimentary facies training template library, the lithology training template library, and the pre-built first generative adversarial network includes: [0061] inputting a plurality of sedimentary facies training template data volumes in the sedimentary facies training template library into the first generative adversarial sub-network for 5 training; [0062] inputting the work-area logging sedimentary facies data into the second generative adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to a training result to generate a reservoir sedimentary facies model, and the second generative adversarial sub-network is built according to the trained first generative adversarial 10 sub-network; [0063] inputting a plurality of lithology training template data volumes in the lithology training template library into the third generative adversarial sub-network for training; and [0064] inputting the work-area logging sand body and interlayer data into the fourth generative adversarial sub-network for training, and carrying out multipoint geostatistics 15 modeling according to a training result and the reservoir sedimentary facies model to generate a reservoir lithology distribution model, and the fourth generative adversarial sub-network is built according to the trained third generative adversarial sub-network.

[0065] In an embodiment, the inputting the plurality of sedimentary facies training template data volumes in the sedimentary facies training template library into the first generative 20 adversarial sub-network for training includes: [0066] inputting the sedimentary facies training template data with the preset type which is selected from the sedimentary facies training template library, as the network initial input data, into the first generative adversarial sub-network for training; [0067] making a conformity comparison between the selected sedimentary facies training 25 template data and an output result of the first generative adversarial sub-network to obtain a first error term; and [0068] correcting the first generative adversarial sub-network according to the first error term and the plurality of sedimentary facies training template data volumes until the first error term reaches a minimum value; and 30 [0069] the inputting the work-area logging sedimentary facies data into the second generative adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to the training result to generate the reservoir sedimentary facies model includes: building the second generative adversarial sub-network according to the trained first generative adversarial sub-network, inputting the work-area logging sedimentary facies data into13 the second generative adversarial sub-network for training to output a sedimentary facies template characterizing features of the work-area logging sedimentary facies data, and carrying out multipoint geostatistics modeling according to the sedimentary facies template characterizing the features of the work-area logging sedimentary facies data and actual well data to generate a 5 reservoir sedimentary facies model.

[0070] During implementation, corresponding encoder E1 and generator G1 are built according to the length, the width, the height and the channel number of each template data volume Si. An input layer of an encoder module EE1 of E1 has a size of SL×SW×SH×SC and a batch training number (Batch Size) of SM, and an output layer of a decoder module ED1 of E0 10 has a size of SL×SW×SH×SC and a batch training number of SM, for the sedimentary facies SC=1. G,1 is a copy of the decoder ED1 of E1. Each time, SM number of sedimentary facies training template data volumes, i.e., {Sj|j=k...k+SM-1}, are input into the encoder module EE1 of the encoder E1, and the decoder module ED1 of the encoder E1 outputs results {SOj|j=k...k+SM-1}; an input discriminator D1 is used to make a conformity comparison 15 between {Sj|j=k...k+SM-1} and {SOj|j=k...k+SM-1} to obtain an error of E1, i.e., a first error item; and errors of {Sj|j=k...k+SM-1} and {SOj|j=k...k+SM-1} are minimized by reversely correcting the encoder module of E1 and the decoder module of E1. At this time, the output of the encoder EE1 of E1 is a hidden variable SZ which effectively characterizes features of the 3D data volume of the sedimentary facies simulation result data, with a length of SZL, a width of 20 SZW, a height of SZH, a channel number of SZC and a designed size of SZL×SZW×SZH×SZC. The decoder module ED1 of E1 may generate a sedimentary facies template ED_S which characterizes features of a sedimentary facies simulation result data volume according to SZ and random noise.

[0071] Next, the coordinates and the work-area logging sedimentary facies data both of 25 which are of each point in all of work-area logging of K wells are sorted into whole area well data WS = [WSD1, WSS1, WSD2, WSS2, …, WSDK, WSSK], where WSD represents coordinate data, WSS represents sedimentary facies serial number data, WS represents a 2K×D image, and D represents the number of data points of the sedimentary facies in each work-area logging. An encoder E2 is built, in which an input layer of an encoder module EE2 of E2 has a 30 size of 2K×D and a batch training number of SW, and an output layer of a decoder module ED1 of E2 has a size of 2K×D and a batch training number of SW, where a hidden layer WZ output by EE2 has a length of SZL, a width of SZW, a height of SZH, a channel number of SZC, and a size of SZL×SZW×SZH×SZC which is the same as the size of SZ. The hidden layer WZ of E2 after training may be used as a hidden variable to characterize features of the work-area14 logging sedimentary facies data.

[0072] Furthermore, a generator G2 is built, of which an input layer has a size of SZL×SZW×SZH×SZC, and an output layer has a length of SL, a width of SW, a height of SH and a channel number of SC. The initially settings of G2 copy the parameters of the G1 to use 5 the learned sedimentation pattern. The hidden variable WZ output by E2 is input into G2 to generate a new 3D sedimentary facies template G2S, with a length of SL, a width of SW, a height of SH and a channel number of SC. By taking G2S as the template and the actual well data as the constraints, a reservoir sedimentary facies model FS is obtained by multipoint geostatistics modeling. 10 [0073] It should be noted that, firstly, the modeling is carried out using the well data and a sedimentary facies training network. Here, it may be artificially assumed that the research area is certain sedimentary facies at first, but if the artificial recognition is wrong, the sedimentary facies training network will adjust the output of sedimentary facies type by itself according to the error feedback, that is, correct the artificial recognition to optimize the sedimentary facies model. 15 Next, the output result of the interlayer pattern will be adjusted accordingly for the lithology data, and the principle is the same as above.

[0074] Specifically, the first step is to enable an artificial intelligence network to generate required different sedimentary facies, but the sedimentary facies actually to be generated is unknown to the artificial intelligence network. Therefore, according to the basic research 20 recognition, we tell the artificial intelligence network to generate a sedimentary facies type A to start modeling, and automatically correct the sedimentary facies type in the continuous modeling, and the final result may be a sedimentary facies B. However, the modeling result usually cannot pass the dynamic data verification, and in the last step dynamic data is added to correct the geological model, thereby unifying the static and dynamic data. Since the process is carried out 25 in stages and classifications, it is more intelligent and convenient than making the artificial intelligence network explore slowly by itself, the calculation is more stable and the efficiency is higher. In addition, it also makes the modeling method extensible, and if there is a completely new sedimentary facies type, it is only necessary to update the template library at the beginning. For example, if the type is artificially recognized as A at first but it is actually B, and the 30 calculation error is large when the network is trained with A, the model will automatically find B data from the template library. In addition, due to the characteristics of deep learning, a sedimentation type between A and B may be generated. The core idea is to use artificial intelligence network to help correct the recognition of sedimentation and increase the accuracy of modeling.15

[0075] In an embodiment, the reservoir property modeling method further includes: [0076] obtaining a work-area seismic data volume; [0077] making a conformity comparison between the work-area seismic data volume and the reservoir sedimentary facies model to obtain a second error term; and 5 [0078] correcting the second generative adversarial sub-network according to the second error term until the second error term reaches a minimum value; and [0079] the carrying out multipoint geostatistics modeling according to the sedimentary facies template characterizing the features of the work-area logging sedimentary facies data and actual well data to generate the reservoir sedimentary facies model includes: carrying out 10 multipoint geostatistical modeling according to the sedimentary facies template characterizing the features of the work-area logging sedimentary facies data output by the second generative adversarial sub-network when the second error term reaches the minimum value and actual well data when the fourth error term reaches the minimum value, so as to generate a reservoir sedimentary facies model. 15 [0080] During implementation, a discriminator D2 is used to calculate a comprehensive error between FS, and a seismic property data volume SEGY and the work-area logging data, i.e., 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 a minimum error with work-area logging and seism, and an optimal reservoir sedimentary facies 20 model OFS is output.

[0081] In an embodiment, the inputting the plurality of lithology training template data volumes in the lithology training template library into the third generative adversarial sub-network for training includes: [0082] inputting the plurality of lithology training template data volumes into the third 25 generative adversarial sub-network for training; [0083] making a conformity comparison between the lithology training template data volume and an output result of the third generative adversarial sub-network to obtain a third error term; and [0084] correcting the third generative adversarial sub-network according to the third error 30 term until the third error term reaches a minimum value; and [0085] the inputting the work-area logging sand body and interlayer data into the fourth generative adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to the training result and the reservoir sedimentary facies model to generate the reservoir lithology distribution model includes: building the fourth generative adversarial16 sub-network according to the trained third generative adversarial sub-network, inputting the work-area logging sand body and interlayer data into the fourth generative adversarial sub-network for training to output a lithology distribution template characterizing features of work-area logging lithology data, and carrying out multipoint geostatistics modeling according 5 to the lithology distribution template characterizing the features of work-area logging lithology data and actual well data to generate the reservoir lithology distribution model.

[0086] During implementation, corresponding encoder E3 and generator G3 are built according to the length, the width, the height and the channel number of each template data volume Lis. An input layer of an encoder module EE3 of E3 has a size of LL×LW×LH×LC and a 10 batch training number of LM, and an output layer of a decoder module ED3 of E3 has a size of LL×LW×LH×LC and a batch training number of LM, for the sedimentary facies LC=1. G3 is a copy of the decoder module ED3 of E3. Each time, LM number of lithology training template data volumes, i.e., {Lj|j=k...k+LM-1}, are input into the encoder module EE3 of the encoder E3, and the decoder module ED3 of the encoder E3 outputs results {LOj|j=k...k+LM-1}; a 15 conformity module is used to compare {Lj|j=k..k+LM-1} and {LOj|j=k...k+LM-1} to obtain an error of E3; and errors of {Lj|j=k...k+LM-1} and {LOj|j=k...k+LM-1} are minimized by reversely correcting the encoder module of E3 and the decoder module of E3. At this time, the output of the encoder module EE3 of E3 is a hidden variable LZ which effectively characterizes features of the 3D data volume of the sand body and interlayer mode result data, with a length of 20 LZL, a width of LZW, a height of LZH, a channel number of LZC and a designed size of LZL× LZW×LZH×LZC. The decoder module ED3 of E3 may generate a lithology distribution template ED_L which characterizes features of a sand body and interlayer pattern result data volume according to LZ and random noise.

[0087] Next, the coordinates and the work-area logging sand body and interlayer data both 25 of which are of each point in all of work-area logging of K wells are sorted into whole area well data WL = [WLD1, WLL1, WLD2, WLL2, …, WLDK, WLLK], where WLD represents coordinate data, WLL represents sand body and interlayer serial number data, WL represents a 2K×D image, and D represents the number of data points of the sand body and interlayer in each work-area logging. An encoder E4 is built, in which an input layer of an encoder module 30 EE4 of E4 has a size of 2K×D and a batch training number of LW, and an output layer of a decoder module ED4 of E4 has a size of 2K×D and a batch training number of LW, where a hidden layer LZ output by EE4 has a length of LZL, a width of LZW, a height of LZH, a channel number of LZC, and a size of LZL×LZW×LZH×LZC which is the same as the size of LZ. The hidden layer WZ of E4 after training may be used as a parameter to characterize features of17 the lithology of the whole area work-area logging.

[0088] Furthermore, a generator G4 is built, of which an input layer has a size of LZL× LZW×LZH×LZC, and an output layer has a length of LL, a width of LW, a height of LH and a channel number of LC. The initially settings of G4 copy the parameters of the G3 to use the 5 learned sedimentation pattern. The hidden variable LZ output by E4 is input into G4 to generate a new lithology distribution template G1L characterizing features of the work-area logging lithology data, with a length of LL, a width of LW, a height of LH and a channel number of LC. By taking G1L as the template and the actual well data as the constraints, a reservoir lithology distribution model FLs of an s-th sedimentary facies area is obtained by multipoint geostatistics 10 modeling.

[0089] In an embodiment, the reservoir property modeling method further includes: [0090] obtaining a work-area seismic data volume; [0091] making a conformity comparison between the work-area seismic data volume and the reservoir lithology distribution model to obtain a fourth error term; and 15 [0092] correcting the fourth generative adversarial sub-network according to the fourth error term until the fourth error term reaches a minimum value; and [0093] the carrying out multipoint geostatistics modeling according to the lithology distribution template characterizing the features of work-area logging lithology data and actual well data to generate the reservoir lithology distribution model includes: carrying out multipoint 20 geostatistics modeling according to the lithology distribution template characterizing the features of work-area logging lithology data output by the fourth generative adversarial sub-network when the fourth error term reaches the minimum value and actual well data when the fourth error term reaches the minimum value, so as to generate the reservoir lithology distribution model.

[0094] During implementation, a discriminator D4 is used to calculate a comprehensive 25 error between FLs, and a seismic property data volume SEGY and the work-area logging data, i.e., the fourth error term. The error calculated by D4 is used to correct G4 until the sedimentary facies model generated based on a 3D sand body and interlayer template generated by G4 has a minimum error with work-area logging and seism, and an optimal reservoir lithology distribution model OFLs of an s-th sedimentary facies area is output. All of the n sedimentary facies areas are 30 traversed and the above modeling process is performed for each sedimentary facies area, thereby completing the lithology intelligent modeling of all of the n sedimentary facies areas to obtain an optimal reservoir lithology distribution model OFL.

[0095] In step 103, a reservoir property model is generated according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second18 generative adversarial network, in which the second generative adversarial network is a geological property generation network based on lithology constraints.

[0096] In an embodiment, the work-area logging property data includes work-area logging porosity property data, work-area logging permeability property data and work-area logging 5 saturation property data; and [0097] the generating the reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and the pre-built second generative adversarial network includes: [0098] inputting the work-area logging porosity property data and the reservoir lithology 10 distribution model into the second generative adversarial network to output a corresponding reservoir porosity property model; [0099] inputting the work-area logging permeability property data and the reservoir lithology distribution model into the second generative adversarial network to output a corresponding reservoir permeability property model; and 15 [00100] inputting the work-area logging saturation property data and the reservoir lithology distribution model into the second generative adversarial network to output a corresponding reservoir saturation property model.

[00101] During implementation, the coordinates and the work-area logging property data both of which are of each point in all of work-area logging of K wells are sorted into whole 20 area well data WPq = [WPD1q, WPP1q, WPD2q, WPP2q, …, WPDKq, WPPKq], where WPD represents coordinate data, WPP represents porosity data, WP represents a 2K×D image, D represents the number of data points of the property data in each work-area logging, and q represents a serial number of the property to be analyzed. The porosity has a property serial number q=1, the permeability has a property serial number q=2, and the saturation has a property 25 serial number q=3. An encoder E5 is built, and an input layer of an encoder module EE5 of E5 has a size of 2K×D and a batch training number of PW, and an output layer of a decoder module ED5 of E5 has a size of 2K×D and a batch training number of PW, where a hidden layer PZ output by EE5 has a length of PZL, a width of PZW, a height of PZH, a channel number of PZC, and a size of PZL×PZW×PZH×PZC which is the same as the size of PZ. The hidden layer 30 PZq of E5 after training may be used as a parameter to characterize features of q-th property data of the whole work-area logging, i.e., a hidden variable characterizing features of the work-area logging saturation property data. A generator G5 is built, where an input layer 1 has a size of PZL×PZW×PZH×(2×PZC), and an output layer has a size of PZL×PZW×PZH. The first PZC channels in the input layer are used to place the PZ data, and the following PZC channels19 are used to extract lithology data blocks with the size of PZL×PZW×PZH from the OFL by means of translational traversal. After traversing the OFL, G5 generates a reservoir property model OFPp of a p-th property, as illustrated in FIG. 5. G4 generates a 3D reservoir property model OFPp after traversing the OFL, which may reduce, by means of translational traversal, the 5 size of the input layer of G4 and the requirement on hardware compared with inputting all the OFLs into G4 at one time. The reservoir property model OPFS= {OPF1, OFP2, OPF3} of the properties of porosity, permeability and saturation are respectively generated orderly according to the above procedure, where OPF1 represents a generated optimal reservoir porosity property model, OPF2 represents a generated optimal reservoir permeability property model, and OPF3 10 represents a generated optimal reservoir saturation property model.

[00102] In step 104, the reservoir property model is corrected according to the work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on 15 dynamic constraints.

[00103] In an embodiment, the correcting the reservoir property model according to the work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network includes: 20 [00104] making a conformity comparison between the work-area production dynamic data and the reservoir property model which has been subjected to a reservoir numerical simulation agent to obtain a fifth error term; [00105] correcting the third generative adversarial network according to the fifth error term until the fifth error term reaches a minimum value; and 25 [00106] taking a reservoir property model output by the third generative adversarial network when the fifth error term reaches the minimum value as a corrected reservoir property model.

[00107] During implementation, the reservoir numerical simulation agent RS is used to predict a work-area production parameter based on the OPFS, and a discriminator D5 is used to 30 calculate an error Err1 between the predicted data and the actual work-area production parameter FO, and an error Err2 between the generated properties of porosity, the generated permeability, and the generated saturation, and the work-area logging property data, thereby obtaining a total error of dynamic data and static data, i.e., Err=Err1+Err2. According to the error Err, the generator G5 is corrected reversely until the error obtained by D5 is the minimum to output an20 optimal reservoir property model OFP.

[00108] A specific embodiment is given below to illustrate the specific application of reservoir property modeling according to the embodiments of the present disclosure. In this specific embodiment, Block X in South America is taken as an application example. Firstly, a 5 sediment dynamic model conforming to the sediment environment of Block X is built, and 2,000 3D sedimentary facies simulation result data volumes and 4,000 3D sand body-interlayer pattern result data volumes are obtained as training data volumes. Meanwhile, the work-area logging sedimentary facies data of 400 wells in the research area, the sand body-interlayer data of 400 wells in the research area, and the data of porosity, permeability and saturation of 400 wells in 10 the research area are collected. The dynamic production data of 400 wells in the research area are sorted. The seismic data volumes of the research area are collected to realize the full coverage of the work area. Based on an AMAX deep learning environment, an RTX 8000 graphics card having eight 48GB memories is used, an encoder E0, a generator G0, a generator G1 and encoder E1, a generator G2 and an encoder E2, a generator G3 and an encoder E3, a generator 15 G4, and discriminators D1, D2, D3 and D4 are built. After 21,000 times of reverse training, the generated 3D reservoir property model has a conformity of 98.2% with the work-area logging porosity data, a conformity of 97.3% with the work-area logging permeability data, a conformity of 98.5% with the work-area logging saturation data, and a conformity of 94.1% with the work-area logging liquid production data. The 3D reservoir property model obtained using 20 existing MPS non-artificial intelligence method has a conformity of 82.4% with the work-area logging porosity data, a conformity of 83.2% with the work-area logging permeability data, a conformity of 81.1% with the work-area logging saturation data, and a conformity of 79.2% with the work-area logging liquid production data. The conformity between the dynamic data and static data of the reservoir property model obtained by the present disclosure is higher than that 25 of the existing stochastic modeling method.

[00109] The embodiments of the present disclosure provide a 3D template generation method according to artificial intelligence and a multipoint geostatistical with dynamic and static integration, which effectively solves the problem that the existing MPS method template does not consider the dynamic production data, and is the reservoir property modeling under the 30 multilevel control of artificial intelligence. Firstly, a sedimentary facies model is built based on an artificial intelligence method; a lithology model based on artificial intelligence is built under the control of the sedimentary facies model; and finally, a property model is built based on artificial intelligence under the lithology control. Through the multilevel control, not only the training difficulty is reduced, but also the model accuracy is improved.21

[00110] Based on the same invention concept, the embodiments of the present disclosure further provide a reservoir property modeling apparatus as described below. Since the principles for the reservoir property modeling apparatus to solve the problems are similar to those for the reservoir property modeling method, the implementation of the reservoir property 5 modeling apparatus can refer to the implementation of the method, and the repeated content will be omitted here.

[00111] FIG. 5 illustrates a structural diagram of a reservoir property modeling apparatus according to an embodiment of the present disclosure. As illustrated in FIG. 5, the reservoir property modeling apparatus includes: 10 [00112] a data obtaining module 501 configured to obtain work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, the work-area logging data including work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; 15 [00113] a reservoir lithology distribution model generation module 502 configured to generate a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is a deep convolution 20 generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; [00114] a reservoir property model generation module 503 configured to generate a reservoir property model according to the work-area logging property data, the reservoir 25 lithology distribution model, and a pre-built second generative adversarial network, in which the second generative adversarial network is a geological property generation network based on lithology constraints; and [00115] a reservoir property model optimization module 504 configured to correct the reservoir property model according to the work-area production dynamic data, the reservoir 30 property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on dynamic constraints.

[00116] Based on the above invention concept, as illustrated in FIG. 6, the embodiments of the present disclosure further provide a computer device 600, which includes a22 memory 610, a processor 620, and a computer program 630 stored in the memory 610 and executable on the processor 620, and when executing the computer program 630, the processor 620 implements the aforementioned reservoir property modeling method.

[00117] Based on the above invention concept, the embodiments of the present 5 disclosure further provide a computer-readable storage medium which stores a computer program, and when executed by a processor, the computer program implements the aforementioned reservoir property modeling method.

[00118] The embodiment of the present disclosure also provides a computer program product, including a computer program, and when executed by a processor, the computer 10 program implements the aforementioned reservoir property modeling method.

[00119] Compared with the technical solution of stochastic modeling for reservoir properties in the prior art, the embodiments of the present disclosure obtain work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, in which the work-area logging data includes work-area 15 logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; generate a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, in which the first generative adversarial network is 20 a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; generate a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial network, in which the second 25 generative adversarial network is a geological property generation network based on lithology constraints; and correct the reservoir property model according to the work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, in which the third generative adversarial network is a property optimization network based on dynamic constraints. 30 According to the embodiments of the present disclosure, dynamic production content is introduced during the reservoir property modeling, and a reservoir lithology distribution model and a reservoir property model based on an artificial intelligence generative adversarial network are generated in sequence by combining the work-area logging data, the sedimentary facies training template library and the lithology training template library, and then the reservoir23 property model is corrected by combining the dynamic production content, so as to effectively build a reservoir property model with dynamic and static consistency, and improve the conformity between the dynamic data and static data of stochastic modeling for reservoir properties. 5 [00120] Those skilled in the art should appreciate that any embodiment of the present disclosure can be provided as a method, a system or a computer program product. Therefore, the present disclosure can take the form of a full hardware embodiment, a full software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure can take the form of a computer program product implemented on one or more computer usable storage 10 mediums (including, but not limited to, a magnetic disc memory, CD-ROM, optical storage, etc.) including therein computer usable program codes.

[00121] The present disclosure is described with reference to the flowcharts and/or block diagrams of the method, apparatus (system) and computer program product according to the embodiments of the present disclosure. It should be appreciated that each flow and/or block 15 in the flowchart and/or block diagram, and the combinations of the flows and/or blocks in the flowchart and/or block diagram can be implemented through computer program instructions. The computer program instructions may be provided to a general computer, a dedicated computer, an embedded processor or a processor of other programmable data processing device, to form a machine so that the instructions, which are executed through the computer or the processor of 20 other programmable data processing device, generate devices for realizing the functions specified in one or more flows in the flowchart and one or more blocks in the block diagram.

[00122] The computer program instructions may also be stored in a computer readable memory which is capable of guiding the computer or other programmable data processing device to work in a specific mode, so that the instructions stored in the computer readable memory 25 generate a product including instructing device for realizing the functions specified in one or more flows in the flowchart and one or more blocks in the block diagram.

[00123] The computer program instructions may also be loaded to a computer or any other programmable data processing device, so that a series of operation steps can be performed in the computer or other programmable device to produce a processing realized by the computer, 30 thus the instructions executed in the computer or other programmable device provide the steps for realizing the functions specified in one or more flows in the flowchart and one or more blocks in the block diagram.

[00124] The specific embodiments described above further illustrate the objectives, technical solutions and advantageous effects of the present disclosure in detail. It should be24 understood that those described above are only specific embodiments of the present disclosure and are not intended to limit the protection scope of the present disclosure. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present disclosure should fall within the protection scope of the present disclosure.

Claims

CLAIMS 1. A reservoir property modeling method, comprising: obtaining work-area logging data, work-area production dynamic data, a sedimentary facies 5 training template library and a lithology training template library, wherein the work-area logging data comprises work-area logging sedimentary facies data, work-area logging sand body and interlayer data, and work-area logging property data; generating a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary 10 facies training template library, the lithology training template library, and a pre-built first generative adversarial network, wherein the first generative adversarial network is a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data; 15 generating a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial network, wherein the second generative adversarial network is a geological property generation network based on lithology constraints; and correcting the reservoir property model according to the work-area production dynamic data, 20 the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, wherein the third generative adversarial network is a property optimization network based on dynamic constraints.
2. The reservoir property modeling method according to claim 1, wherein the sedimentary facies training template library and the lithology training template library are pre-built 25 comprising following steps: obtaining sedimentary facies information and a sand body and interlayer pattern result data volume, wherein the sedimentary facies information comprises any one or any combination of sediment dynamics data, an existing sedimentary facies map and a satellite image; building the sedimentary facies training template library according to the sedimentary facies 30 information; and building the lithology training template library according to the sand body and interlayer pattern result data volume.
3. The reservoir property modeling method according to claim 1, wherein the first26 generative adversarial network comprises: a first generative adversarial sub-network according to a sedimentary facies pattern simulated by a sediment dynamics, a second generative adversarial sub-network based on a logging seismic sedimentary facies pattern, a third generative adversarial sub-network based on a lithology pattern simulated by the sediment dynamics under 5 sedimentary facies constraints, and a fourth generative adversarial sub-network based on a logging seismic lithology pattern under sedimentary facies constraints; and the generating the reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and the pre-built first 10 generative adversarial network comprises: inputting a plurality of sedimentary facies training template data volumes in the sedimentary facies training template library into the first generative adversarial sub-network for training; inputting the work-area logging sedimentary facies data into the second generative 15 adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to a training result to generate a reservoir sedimentary facies model, wherein the second generative adversarial sub-network is built according to the trained first generative adversarial sub-network; inputting a plurality of lithology training template data volumes in the lithology training 20 template library into the third generative adversarial sub-network for training; and inputting the work-area logging sand body and interlayer data into the fourth generative adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to a training result and the reservoir sedimentary facies model to generate a reservoir lithology distribution model, wherein the fourth generative adversarial sub-network is built 25 according to the trained third generative adversarial sub-network.
4. The reservoir property modeling method according to claim 3, wherein the inputting the plurality of sedimentary facies training template data volumes in the sedimentary facies training template library into the first generative adversarial sub-network for training comprises: inputting the sedimentary facies training template data with the preset type which is selected 30 from the sedimentary facies training template library, as the network initial input data, into the first generative adversarial sub-network for training; making a conformity comparison between the selected sedimentary facies training template data and an output result of the first generative adversarial sub-network to obtain a first error27 term; and correcting the first generative adversarial sub-network according to the first error term and the plurality of sedimentary facies training template data volumes until the first error term reaches a minimum value; and 5 the inputting the work-area logging sedimentary facies data into the second generative adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to the training result to generate the reservoir sedimentary facies model comprises: building the second generative adversarial sub-network according to the trained first generative adversarial sub-network, inputting the work-area logging sedimentary facies data into the second 10 generative adversarial sub-network for training to output a sedimentary facies template characterizing features of the work-area logging sedimentary facies data, and carrying out multipoint geostatistics modeling according to the sedimentary facies template characterizing the features of the work-area logging sedimentary facies data and actual well data to generate a reservoir sedimentary facies model. 15
5. The reservoir property modeling method according to claim 4, further comprising: obtaining a work-area seismic data volume; making a conformity comparison between the work-area seismic data volume and the reservoir sedimentary facies model to obtain a second error term; and correcting the second generative adversarial sub-network according to the second error term 20 until the second error term reaches a minimum value; and the carrying out multipoint geostatistics modeling according to the sedimentary facies template characterizing the features of the work-area logging sedimentary facies data and actual well data to generate the reservoir sedimentary facies model comprises: carrying out multipoint geostatistics modeling according to the sedimentary facies template characterizing the features of 25 the work-area logging sedimentary facies data output by the second generative adversarial sub-network when the second error term reaches the minimum value and actual well data when the second error term reaches the minimum value, so as to generate a reservoir sedimentary facies model.
6. The reservoir property modeling method according to claim 3, wherein the inputting the 30 plurality of lithology training template data volumes in the lithology training template library into the third generative adversarial sub-network for training comprises: inputting the plurality of lithology training template data volumes into the third generative adversarial sub-network for training;28 making a conformity comparison between the lithology training template data volume and an output result of the third generative adversarial sub-network to obtain a third error term; and correcting the third generative adversarial sub-network according to the third error term until the third error term reaches a minimum value; and 5 the inputting the work-area logging sand body and interlayer data into the fourth generative adversarial sub-network for training, and carrying out multipoint geostatistics modeling according to the training result and the reservoir sedimentary facies model to generate the reservoir lithology distribution model comprises: building the fourth generative adversarial sub-network according to the trained third generative adversarial sub-network, inputting the 10 work-area logging sand body and interlayer data into the fourth generative adversarial sub-network for training to output a lithology distribution template characterizing features of work-area logging lithology data, and carrying out multipoint geostatistics modeling according to the lithology distribution template characterizing the features of work-area logging lithology data and actual well data to generate the reservoir lithology distribution model. 15
7. The reservoir property modeling method according to claim 6, further comprising: obtaining a work-area seismic data volume; making a conformity comparison between the work-area seismic data volume and the reservoir lithology distribution model to obtain a fourth error term; and correcting the fourth generative adversarial sub-network according to the fourth error term 20 until the fourth error term reaches a minimum value; and the carrying out multipoint geostatistics modeling according to the lithology distribution template characterizing the features of work-area logging lithology data and actual well data to generate the reservoir lithology distribution model comprises: carrying out multipoint geostatistics modeling according to the lithology distribution template characterizing the features 25 of work-area logging lithology data output by the fourth generative adversarial sub-network when the fourth error term reaches the minimum value and actual well data when the fourth error term reaches the minimum value, so as to generate the reservoir lithology distribution model.
8. The reservoir property modeling method according to claim 1, wherein the work-area logging property data comprises: work-area logging porosity property data, work-area logging 30 permeability property data and work-area logging saturation property data; and the generating the reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and the pre-built second generative adversarial network comprises:29 inputting the work-area logging porosity property data and the reservoir lithology distribution model into the second generative adversarial network to output a corresponding reservoir porosity property model; inputting the work-area logging permeability property data and the reservoir lithology 5 distribution model into the second generative adversarial network to output a corresponding reservoir permeability property model; and inputting the work-area logging saturation property data and the reservoir lithology distribution model into the second generative adversarial network to output a corresponding reservoir saturation property model. 10
9. The reservoir property modeling method according to claim 8, wherein the correcting the reservoir property model according to the work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and the pre-built third generative adversarial network comprises: making a conformity comparison between the work-area production dynamic data and the 15 reservoir property model which has been subjected to a reservoir numerical simulation agent to obtain a fifth error term; correcting the third generative adversarial network according to the fifth error term until the fifth error term reaches a minimum value; and taking a reservoir property model output by the third generative adversarial network when 20 the fifth error term reaches the minimum value as a corrected reservoir property model.
10. A reservoir property modeling apparatus, comprising: a data obtaining module configured to obtain work-area logging data, work-area production dynamic data, a sedimentary facies training template library and a lithology training template library, wherein the work-area logging data comprises work-area logging sedimentary facies data, 25 work-area logging sand body and interlayer data, and work-area logging property data; a reservoir lithology distribution model generation module configured to generate a reservoir lithology distribution model according to the work-area logging sedimentary facies data, the work-area logging sand body and interlayer data, the sedimentary facies training template library, the lithology training template library, and a pre-built first generative adversarial network, 30 wherein the first generative adversarial network is a deep convolution generative adversarial network based on multilevel well-seismic integration, and a preset type of sedimentary facies training template data selected from the sedimentary facies training template library is taken as network initial input data;30 a reservoir property model generation module configured to generate a reservoir property model according to the work-area logging property data, the reservoir lithology distribution model, and a pre-built second generative adversarial network, wherein the second generative adversarial network is a geological property generation network based on lithology constraints; 5 and a reservoir property model optimization module configured to correct the reservoir property model according to the work-area production dynamic data, the reservoir property model which has been subjected to a reservoir numerical simulation agent, and a pre-built third generative adversarial network, wherein the third generative adversarial network is a property optimization 10 network based on dynamic constraints.
11. A computer program product, comprising a computer program, wherein when executed by a processor, the computer program implements a method according to claim 1.