Reservoir characterization method and device based on deep learning, equipment and storage medium
By using deep learning methods and variational autoencoders, reservoir structures are constructed using drilling data, which solves the problem of insufficient connectivity in multi-point geostatistical methods and achieves efficient and accurate reservoir characterization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2023-08-18
- Publication Date
- 2026-06-02
AI Technical Summary
Existing multi-point geostatistical methods for reservoir characterization suffer from poor connectivity of generated results and low accuracy of constructed reservoir structures.
A deep learning-based approach was adopted, utilizing selective attention mechanism and variational autoencoder to construct reservoir structure by training neural network, and drilling data was used as conditional data for representation.
This improves the efficiency and accuracy of reservoir structure simulation, ensuring that the generated reservoir structure has good connectivity and conforms to actual geological characteristics.
Smart Images
Figure CN117172954B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a reservoir characterization method, apparatus, device, and storage medium based on deep learning. Background Technology
[0002] Reservoir characterization and modeling are crucial technologies in oil and gas resource development. Accurate reservoir models can reflect the true structure of reservoirs, thus helping researchers correctly evaluate the oil and gas resources contained within. In actual production, our understanding of reservoirs is often obtained through drilling, but this data is often very sparse. Reservoir characterization utilizes sparse drilling data within a work area to model the overall distribution of reservoirs, and it has already played a role in oil and gas field exploration and development. Multipoint geostatistics is an important technique in reservoir modeling, addressing the problem of traditional variogram-based geostatistical methods' poor performance in characterizing anisotropic structures. Multipoint geostatistical methods directly extract spatial patterns of geological structures from training images instead of using variograms to describe the spatial distribution characteristics of geological structures. They use data templates to scan training images for pattern matching and store the resulting multipoint patterns in tree or linked list data structures, leading to a limited number of extracted spatial distribution feature patterns. Although applying multigrids can alleviate this problem to some extent, the extracted patterns remain limited. However, when the data template contains a large number of nodes, this can lead to enormous computational costs. Furthermore, in practice, the simulation structures generated by multi-point geostatistical methods often suffer from insufficient spatial connectivity. Obtaining good simulation results using multi-point geostatistical methods often requires adjusting a large number of parameters. Quickly selecting appropriate parameters requires users to have a sufficient understanding of multi-point statistical methods themselves, which reduces the efficiency of applying these methods.
[0003] Existing multi-point geological methods are insufficient in terms of efficiency and learning costs for practical applications in reservoir characterization. Furthermore, for reservoir structure, connectivity is a crucial indicator related to resource occurrence space, grade, and extraction difficulty. Therefore, the poor connectivity and low accuracy of reservoir structure generation results from traditional geostatistical methods are pressing technical problems that need to be addressed.
[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention
[0005] The main objective of this invention is to provide a reservoir characterization method, apparatus, device, and storage medium based on deep learning, aiming to solve the technical problem that the connectivity of the reservoir structure generated by traditional geostatistical methods in the prior art is not good enough, and the accuracy of the constructed reservoir structure is not high.
[0006] To achieve the above objectives, the present invention provides a reservoir characterization method based on deep learning, the deep learning-based reservoir characterization method comprising:
[0007] Training and test sets are generated based on the original reservoir structure;
[0008] The original reservoir structure in the training set is sampled to obtain the first conditional data of the borehole form;
[0009] The first conditional data is arranged according to its relative position. Blank positions in the arranged data are detected and filled to obtain second conditional data with the same size as the original reservoir structure.
[0010] Build the neural network to be trained and set the hyperparameters of the neural network to be trained;
[0011] Initialize the weight parameters of the neural network to be trained, and train the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network;
[0012] The test set and the corresponding third condition data are used to perform the test and obtain the test results.
[0013] If the test results reach the preset results, the reservoir structure is characterized using the corresponding target condition data;
[0014] If the test result does not meet the preset result, the process returns to the step of setting the hyperparameters of the neural network to be trained, and the neural network with the hyperparameters reset is trained again until the test result meets the preset result, and the reservoir structure is characterized by the corresponding current condition data.
[0015] Preferably, the neural network to be trained consists of three sub-networks, the sub-networks including a first encoder, a second encoder, and a decoder;
[0016] The step of training the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network specifically includes:
[0017] The number of iterations is determined based on the hyperparameters of the neural network to be trained;
[0018] In each loop, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure.
[0019] After completing all the aforementioned number of iterations, the target reservoir structure is generated based on the second part of the reservoir structure obtained in each iteration;
[0020] The first loss function is used to calculate the first difference between the target reservoir structure and the original reservoir structure, and the weight parameters in the first encoder and the decoder are updated by backpropagation.
[0021] The loop is repeated for the specified number of iterations. In each iteration, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure. At the same time, the second encoder encodes the second conditional data into the second latent space.
[0022] Repeat all the aforementioned number of iterations, and generate the target reservoir structure based on the second part of the reservoir structure obtained in each iteration;
[0023] The second loss function is used to calculate the second difference between the distributions of the first latent space and the second latent space. The weight parameters in the first encoder and the second encoder are updated by the backpropagation algorithm so that the second latent space approaches the first latent space, thus obtaining the trained neural network.
[0024] Preferably, the first encoder selectively reads a first portion of the reservoir structure based on local features of the original reservoir structure, encodes the portion of the reservoir structure into a first latent space, and the decoder decodes a first latent vector sampled from the first latent space to generate a second portion of the reservoir structure, including:
[0025] The first encoder uses a selective attention mechanism to selectively read the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder uses a selective attention mechanism to decode the first latent vector sampled from the first latent space to generate the second part of the reservoir structure.
[0026] Preferably, the first loss function is a reconstruction loss, the input target reservoir structure is represented by x, and the output target reservoir structure is represented by y. For a given x and y, the reconstruction loss of the network can be expressed by the following formula:
[0027] L recon = -xlogy-(1-x)log(1-y).
[0028] Preferably, the second loss function is a KL loss or a conditional loss, and the second loss function is used to measure the second difference between the distributions of the first latent space and the second latent space;
[0029] Suppose there are two normal distributions and The KL loss for the distributions of the first and second latent spaces is then:
[0030]
[0031] Preferably, the step of using the test set and the corresponding third condition data to perform the test and obtain the test results includes:
[0032] The reservoir structure in the test set and the corresponding sampled third condition data are used to conduct tests to obtain test results, and the test results are analyzed using variograms and connectivity functions.
[0033] Preferably, the step of generating training and testing sets based on the original reservoir structure includes:
[0034] The original reservoir structure is pruned using the sliding window method, and a complete dataset is constructed based on the pruned data.
[0035] A predetermined proportion of the data in the complete dataset is used as the training set, and the remaining data in the complete dataset is used as the test set.
[0036] Furthermore, to achieve the above objectives, the present invention also proposes a reservoir characterization device based on deep learning, the reservoir characterization device based on deep learning comprising:
[0037] The generation module is used to generate training and test sets based on the original reservoir structure.
[0038] The sampling module is used to sample the original reservoir structure in the training set to obtain the first conditional data of the borehole form;
[0039] The arrangement module is used to arrange the first condition data according to their relative positions, detect blank positions in the arranged data, and fill the blank positions to obtain second condition data with the same size as the original reservoir structure.
[0040] A module for building a neural network to be trained and setting the hyperparameters of the neural network to be trained;
[0041] The training module is used to initialize the weight parameters of the neural network to be trained, and to train the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network.
[0042] The testing module is used to perform tests using the test set and the corresponding third-condition data to obtain test results;
[0043] The characterization module is used to characterize the reservoir structure using corresponding target condition data if the test results reach the preset results.
[0044] The characterization module is also used to return to the step of setting the hyperparameters of the neural network to be trained if the test result does not meet the preset result, to retrain the neural network with the hyperparameters reset, until the test result meets the preset result, and to characterize the reservoir structure using the corresponding current condition data.
[0045] Furthermore, to achieve the above objectives, the present invention also proposes a reservoir characterization device based on deep learning, wherein the reservoir characterization device based on deep learning stores a reservoir characterization program based on deep learning, and when the reservoir characterization program based on deep learning is executed by a processor, it implements the steps of the reservoir characterization method based on deep learning as described above.
[0046] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a deep learning-based reservoir characterization program, which, when executed by a processor, implements the steps of the deep learning-based reservoir characterization method described above.
[0047] The beneficial effects of the above-described method of the present invention are as follows:
[0048] (1) The reservoir characterization method based on selective attention mechanism and variational autoencoder provided in this invention uses an attention mechanism to learn the connectivity of the reservoir structure and completes the construction of the reservoir structure step by step based on the connectivity, avoiding the problem of insufficient connectivity in the generated results of traditional geostatistical methods. The use of deep learning methods can also greatly shorten the simulation time of the reservoir structure and improve the simulation efficiency in practical applications. A variational autoencoder refers to a neural network structure that includes an encoder and a decoder. The neural network in this invention adopts the variational autoencoder structure and has been improved, containing two encoders and one decoder.
[0049] (2) The reservoir characterization method based on selective attention mechanism and variational autoencoder provided by this invention uses drilling data commonly used in resource exploration and development as conditional data to construct reservoir structures, which is closer to actual production. The addition of conditional data further ensures the accuracy of the reservoir structures constructed by the network.
[0050] (3) This invention can be widely applied in various three-dimensional geological information systems, geographic information systems, geological modeling and simulation systems and other software. Attached Figure Description
[0051] Figure 1 This is a schematic diagram of the structure of the reservoir characterization method based on deep learning in the hardware operating environment involved in the embodiments of the present invention;
[0052] Figure 2 This is a flowchart illustrating the first embodiment of the deep learning-based reservoir characterization method of the present invention.
[0053] Figure 3 This illustrates the data flow direction during neural network training and generation in the first embodiment of the deep learning-based reservoir characterization method of the present invention.
[0054] Figure 4 This is a schematic diagram of an experiment designed to verify the reconstructability of a two-dimensional reservoir in an embodiment of the present invention, wherein, Figure 4 (a) is a schematic diagram of the reservoir structure in the training set and the conditional data sampled from it. It can be seen that the conditional data used is borehole data. Figure 4 (b) is the reservoir structure generated by the network. Figure 4 (c) and Figure 4 (d) is a statistical evaluation of the network generation structure;
[0055] Figure 5 This is a schematic diagram of an experiment designed to verify the reconstruction capability of the present invention for three-dimensional structures in an embodiment of the present invention. Figure 5 (a) is a schematic diagram of the training concentrated fold structure. Figure 5 (b) is a diagram illustrating the conditions used. Figure 5 (c) is the reservoir structure generated by the network;
[0056] Figure 6 This is a schematic diagram showing the cross-sectional view of the three structures generated using conditional data in the X, Y, and Z directions in an embodiment of the present invention, illustrating the constraints of the conditional data on the generated structures;
[0057] Figure 7 This is a schematic diagram comparing statistical indicators of the three-dimensional structure generated by the neural network and the reference structure in an embodiment of the present invention, wherein... Figure 7 (a) is the variation function curve. Figure 7 (b) is the connected function curve in the X direction. Figure 7 (c) is the curve of the connected function in the Y direction.
[0058] Figure 8 This is a structural block diagram of the first embodiment of the reservoir characterization device based on deep learning of the present invention.
[0059] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0060] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0061] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a deep learning-based reservoir characterization device for the hardware operating environment involved in the embodiments of the present invention.
[0062] like Figure 1 As shown, the deep learning-based reservoir characterization device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, it may also include a standard wired interface or a wireless interface. In this invention, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.
[0063] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on deep learning-based reservoir characterization devices, which may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.
[0064] like Figure 1 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a reservoir representation program based on deep learning.
[0065] exist Figure 1 In the deep learning-based reservoir characterization device shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the user equipment; the deep learning-based reservoir characterization device calls the deep learning-based reservoir characterization program stored in the memory 1005 through the processor 1001 and executes the deep learning-based reservoir characterization method provided in the embodiment of the present invention.
[0066] Based on the above hardware structure, an embodiment of the reservoir characterization method based on deep learning of the present invention is proposed.
[0067] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the reservoir characterization method based on deep learning of the present invention, which presents the first embodiment of the reservoir characterization method based on deep learning of the present invention.
[0068] In the first embodiment, the deep learning-based reservoir characterization method includes the following steps:
[0069] Step S10: Generate training and test sets based on the original reservoir structure.
[0070] In this specific implementation, the execution entity of this embodiment is the deep learning-based reservoir characterization device. This device can be a personal computer or server, and this embodiment does not impose any limitations on it. The entire reservoir structure is cropped using a sliding window method to construct a complete dataset. The original large image of the reservoir structure is cropped into smaller images, shifting by a few pixels each time. The number of pixels the sliding window slides each time during the cropping process is determined based on the number of images to be cropped. A portion of the dataset is used as the training set, and the other portion as the test set. In this embodiment, step S10 specifically includes: cropping the original reservoir structure using a sliding window method; constructing a complete dataset based on the cropped data; using a preset proportion of data from the complete dataset as the training set; and using the remaining data from the complete dataset as the test set. The preset proportion can be set based on empirical values, such as 50% or 70%.
[0071] Step S20: Sample the original reservoir structure in the training set to obtain the first conditional data of the borehole form.
[0072] It should be noted that the training set and the test set are loaded, and the first conditional data of the borehole form are sampled from the reservoir structure of the training set.
[0073] Step S30: Arrange the first condition data according to their relative positions, detect the blank positions in the arranged data, and fill the blank positions to obtain the second condition data with the same size as the original reservoir structure.
[0074] Understandably, the condition data is arranged according to its relative position, and the blank spaces in the middle of the condition data are filled with invalid values. The invalid value can be 0, that is, 0 is used to fill the blank spaces, so as to obtain a second condition data with the same size as the original reservoir structure.
[0075] Step S40: Build the neural network to be trained and set the hyperparameters of the neural network to be trained.
[0076] It should be understood that, according to Figure 3 The network structure shown is used to build the neural network to be trained, and to set or adjust the hyperparameters of the neural network. The neural network to be trained consists of three sub-networks, including a first encoder, a second encoder, and a decoder; the second encoder is usually a conditional encoder, such as... Figure 3 As shown in the image.
[0077] Step S50: Initialize the weight parameters of the neural network to be trained, and train the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network.
[0078] In the specific implementation, the network training process is divided into two stages. First, an encoder and decoder are combined to generate the reservoir structure, and both the encoder and decoder use selective attention mechanisms. Then, a conditional encoder is used to encode the conditional data into the latent space distribution. The conditional encoder is a convolutional neural network. The neural network weight parameters are initialized, and training is performed using the training set and the corresponding conditional data. If suitable weight parameters are already available, they can be loaded and the process can proceed directly to step S70.
[0079] In this embodiment, the neural network to be trained consists of three sub-networks, including a first encoder, a second encoder, and a decoder.
[0080] The step of training the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network specifically includes:
[0081] The number of iterations is determined based on the hyperparameters of the neural network to be trained;
[0082] In each loop, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure.
[0083] After completing all the aforementioned number of iterations, the target reservoir structure is generated based on the second part of the reservoir structure obtained in each iteration;
[0084] The first loss function is used to calculate the first difference between the target reservoir structure and the original reservoir structure, and the weight parameters in the first encoder and the decoder are updated by backpropagation.
[0085] The loop is repeated for the specified number of iterations. In each iteration, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure. At the same time, the second encoder encodes the second conditional data into the second latent space.
[0086] Repeat all the aforementioned number of iterations, and generate the target reservoir structure based on the second part of the reservoir structure obtained in each iteration;
[0087] The second loss function is used to calculate the second difference between the distributions of the first latent space and the second latent space. The weight parameters in the first encoder and the second encoder are updated by the backpropagation algorithm so that the second latent space approaches the first latent space, thus obtaining the trained neural network.
[0088] It should be understood that the first encoder encodes the reservoir structure into a latent space distribution, employing a selective attention mechanism; the second encoder, called a conditional encoder, encodes the conditional data corresponding to the reservoir structure into another latent space distribution, which is a convolutional neural network; the decoder is used to decode the latent vectors sampled from the latent space distribution to generate the target reservoir structure, also applying a selective attention mechanism. The latent space is a Gaussian distribution, and a latent vector conforming to this Gaussian distribution is sampled from the latent space. The network training process mainly consists of the following two parts:
[0089] (1) The number of training iterations is determined based on the total number of simulation steps set in the network hyperparameters. In each iteration, the encoder selectively reads a portion of the reservoir structure based on local features such as connectivity and encodes it into the first latent space Latent1. The decoder decodes the first latent vector sampled in the first latent space to generate a portion of the reservoir structure. After all iterations are completed, the generated target reservoir structure is obtained. A loss function is used to calculate the first difference between the generated target reservoir structure and the original reservoir structure. Backpropagation is then used to update the weight parameters in the first encoder and the decoder. The above describes the first update process of the neural network.
[0090] (2) The number of loop steps in this process is the same as in (1). In each loop, the encoder and the decoder are used for encoding and decoding, and the second encoder, i.e., the conditional encoder, is used to encode the conditional data into another latent space, i.e., the second latent space Latent2. After the loop ends, the latent space distributions Latent1 and Latent2 in the generation process are obtained. The second difference between the first latent space and the second latent space distribution is calculated using the loss function. Then, the weight parameters in the first encoder and the second encoder are updated through the backpropagation algorithm to make Latent2 approach Latent1. The above is the second update process of the neural network.
[0091] Further, in this embodiment, the first encoder selectively reads a first portion of the reservoir structure based on the local features of the original reservoir structure, encodes the portion of the reservoir structure into a first latent space, and the decoder decodes the first latent vector sampled from the first latent space to generate a second portion of the reservoir structure, including:
[0092] The first encoder uses a selective attention mechanism to selectively read the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder uses a selective attention mechanism to decode the first latent vector sampled from the first latent space to generate the second part of the reservoir structure.
[0093] Further, in this embodiment, the first loss function is the reconstruction loss, the input target reservoir structure is represented by x, and the output target reservoir structure of the neural network is represented by y. For a given x and y, the reconstruction loss of the network can be expressed by the following formula:
[0094] L recon = -xlogy-(1-x)log(1-y).
[0095] Understandably, the reconstruction loss function is used to guide the network in learning the spatial distribution pattern of reservoir structures, i.e., to endow the network with the ability to generate reservoir structures with corresponding spatial distribution patterns. The input reservoir structure is represented by x, and the target reservoir structure output by the neural network is represented by y. For a given x and y, the network's reconstruction loss can be expressed by the following formula: L recon = -xlogy-(1-)log(1-y).
[0096] Furthermore, in this embodiment, the second loss function is KL loss or conditional loss, and the second loss function is used to measure the second difference between the distributions of the first latent space and the second latent space;
[0097] Suppose there are two normal distributions and The KL loss for the distributions of the first and second latent spaces is then:
[0098]
[0099] It should be noted that the KL loss function is used to guide the conditional encoder to fit the encoder output using conditional data, that is, to give the network the ability to generate a reservoir model based on the conditional data. Assume the encoder output follows a normal distribution. The normal distribution of the conditional encoder output is μ represents the mean of the Gaussian distribution, and σ is the standard deviation of the Gaussian distribution. Through optimization of the KL loss, D p Used to fit D q The KL loss between two distributions can be written as:
[0100]
[0101] Training can be stopped once the loss function value stabilizes. The trained network can then be trained using the test set data. If the expected results are achieved, the next application step can be performed. If the results are unsatisfactory, the hyperparameters can be adjusted and the network retrained.
[0102] In practical applications, conditional encoders and decoders are used to generate reservoir structures. Using actual drilling data as conditional data, the network can output reservoir structures with the same spatial distribution pattern.
[0103] Step S60: Use the test set and the corresponding third condition data to perform the test and obtain the test results.
[0104] Understandably, after training, the reservoir structures in the test set and the corresponding sampled third-condition data are used for testing, and the results are analyzed using the variogram and connectivity function. If the results do not meet expectations, the process proceeds to step S40 to adjust the hyperparameters and retrain. A good hyperparameter range is first obtained through multiple training tests, and then fine-tuned within this range to obtain the optimal parameters. In this embodiment, step S60 specifically includes: using the reservoir structures in the test set and the corresponding sampled third-condition data for testing, obtaining test results, and analyzing the test results using the variogram and connectivity function.
[0105] In a specific implementation, the original reservoir structure in the test set is sampled to obtain fourth condition data in the form of boreholes; the fourth condition data is arranged according to relative positions, blank positions in the arranged data are detected, and the blank positions are filled to obtain third condition data with the same size as the original reservoir structure.
[0106] Step S70: If the test result reaches the preset result, the reservoir structure is characterized by the corresponding target condition data.
[0107] It should be noted that, after testing, this method can directly use borehole data from actual work as conditional data to characterize reservoir structure, and the weight parameters of the trained neural network can be saved.
[0108] Step S80: If the test result does not meet the preset result, return to the step of setting the hyperparameters of the neural network to be trained, retrain the neural network with the hyperparameters reset, until the test result meets the preset result, and characterize the reservoir structure using the corresponding current condition data.
[0109] In the specific implementation, the training process is completed, and the neural network stores the pattern information of the reservoir in the training set, which can complete the reservoir representation task under the constraints of borehole morphological condition data. The network weight parameters obtained after training can be used in the next simulation or transferred to other machines. If suitable weight parameters are available from previous storage, the process can start directly from step S40.
[0110] With the rapid development of artificial intelligence technologies, represented by deep learning, they have been widely applied in earth science fields such as mineral resource exploration and geological hazard identification. The essence of deep learning methods is to construct a neural network to fit the mapping relationship from a given input to a target output. This fitting process is called the training process of the neural network. During this process, we need to iteratively optimize the weight parameters of the neural network to achieve the best fitting effect. Before starting training, we need to design the neural network structure for the target task and collect a batch of training data to construct the training set, which includes the input and the corresponding target result. During training, the loss function is an important indicator for measuring the difference between the network's fitting result and the target, and it also guides the optimization direction of the network; therefore, the choice of loss function is also very important. After achieving the expected fitting effect, training ends. We can then use the trained network to map the known input to the unknown output; at this point, we can consider this output as the target result.
[0111] In practical production, the goal of reservoir characterization is to simulate and characterize the complete reservoir structure using known, sparse drilling data, thereby providing a reference for subsequent work. This goal aligns with the principle of deep learning technology in establishing a mapping from input to output; that is, deep learning technology can be used to establish a mapping from known drilling data to the unknown overall reservoir structure. Therefore, deep learning methods have great application prospects in reservoir characterization. A variational autoencoder (VAE) is a neural network architecture in deep learning methods, containing a pair of encoders and decoders. The encoder is used to encode the input into a latent space distribution, and the decoder is used to decode the latent vectors sampled from this distribution to obtain the target output. Note that during training, this latent space distribution is constrained to approximate a known distribution as closely as possible (generally a standard Gaussian distribution). In the field of Earth sciences, variational autoencoders have been used to complete tasks such as geophysical inversion and geochemical anomaly identification, achieving excellent performance.
[0112] This embodiment addresses the limitations of multi-point geostatistical reservoir characterization methods, such as high computational cost and insufficient connectivity of simulation results, by improving the variational autoencoder network structure and applying a selective attention mechanism to ensure the spatial connectivity of the generated structure, making it more suitable for reservoir characterization and modeling tasks. This method uses drilling data within the study area to be simulated as conditional data. The input conditional data is filled into blocks of the same size as the area to be simulated according to their relative positions. The result generated by this method is the characterization of the reservoir structure within the area. Before applying this method for reservoir characterization, it is necessary to train the neural network using a training set containing the corresponding distribution pattern to obtain the various weight parameters, or use other existing trained weight parameters. The network training set is derived from a large-sized reservoir structure containing the target spatial distribution pattern. By controlling the parameters during the trimming, the reservoir structure is trimmed into several structures of the same size. During training, the input conditional data is sampled from the training set structure, and the corresponding reservoir structure serves as the target output of the neural network. After obtaining the output during training, a loss function is used to evaluate the difference between the output and the target output and calculate the network loss value. This loss value is used for gradient calculation and backpropagation to optimize the parameters of the neural network. The loss value guides the direction of gradient descent. When the network output reaches the desired effect or the loss function value stabilizes, the neural network parameters are considered to have reached an optimal state, meaning the network can accurately construct the reservoir structure based on the input conditional data, and the training process ends. After training, drilling data is directly input into the neural network in the training format, and the neural network constructs a reservoir structure with the same pattern based on the geological pattern learned during training. Therefore, this method, by using a selective attention mechanism and a variational autoencoder architecture to design the neural network structure, achieves a fast and accurate method for constructing reservoir structure models, providing strong technical support for reservoir characterization.
[0113] To illustrate the feasibility of the method provided in this embodiment, two-dimensional and three-dimensional reservoir structure generation experiments were carried out according to the above steps.
[0114] Figure 4 The image shows a two-dimensional reservoir structure simulation experiment using this method. The experimental data used a typical fluvial reservoir structure. The white area represents the fluvial reservoir structure, and the black area represents mudstone. Figure 4 (a) is a schematic diagram of the reservoir structure in the training set and the conditional data sampled from it. It can be seen that the conditional data used is in the form of borehole data with a size of 64×64 pixels. Figure 4(b) shows the reservoir structures generated by the network using different conditional data as input. The red markings in the figure indicate the distribution of fluvial facies in the conditional data. It can be seen that the network can accurately construct reservoir structures with strong connectivity using conditional data, and the constructed structures can meet the constraints of the conditional data. Figure 4 (c) and Figure 4 (d) shows the variation function curves and connectivity function curves for 50 simulation results generated using the same conditional data, respectively, which further evaluate the simulation results using statistical methods. The curves show that the curves of the network simulation results are similar in variation and distribution to the curves of the reference structure. This indicates that the reservoir structure constructed using the method of this embodiment conforms to the spatial distribution pattern of the original structure and effectively recreates the original connectivity characteristics, enabling the reservoir to be characterized using conditional data in a two-dimensional context.
[0115] Figure 5 The image shows a simulation experiment of a three-dimensional folded structure using the method of this embodiment. The structure has strong connectivity features in three dimensions, which can illustrate the ability of this method in three-dimensional structure reconstruction. In this structure, the red markings represent the folded structure. Figure 5 (a) is an example reference structure with dimensions of 50×50×50. Figure 5 (b) is an instance of conditional data, which is in the form of three-dimensional logging data. Figure 5 (c) shows the generated results obtained by the network using data under different conditions, which demonstrates that it effectively represents the spatial distribution pattern of the original folded structure. Figure 6 The image shows a cross-sectional comparison of the three conditional data sets and the folded structures generated using them. It can be seen that the distribution of the folded structures follows the constraints of the conditional data sets. Figure 7 It is a statistical evaluation of 40 simulation results generated using data under the same conditions. Figure 7 (a) is the variation function curve. Figure 7 (b) is the connected function curve in the X direction. Figure 7 (c) shows the connectivity function curve in the Y direction. The consistent changes and distribution of the variation function curves indicate that the network simulation results can reproduce the spatial distribution pattern of the original structure. The two inverse connectivity function curves demonstrate that the network can generate a folded structure with consistent connectivity characteristics.
[0116] The two experimental cases above demonstrate that the simulation results of the reservoir structure characterization method provided by this invention are consistent with the reference structure in both visual and statistical indicators, proving that the method of this embodiment can reproduce the spatial characteristics of the structure in the training set. This also confirms that the method of this embodiment can quickly and accurately construct reservoir structures that conform to the original geological model and connectivity characteristics under the constraints of conditional data.
[0117] Furthermore, this embodiment of the invention also proposes a storage medium storing a deep learning-based reservoir characterization program, which, when executed by a processor, implements the steps of the deep learning-based reservoir characterization method described above.
[0118] In addition, refer to Figure 8 This invention also proposes a reservoir characterization device based on deep learning, the deep learning-based reservoir characterization device comprising:
[0119] Generation module 10 is used to generate training and test sets based on the original reservoir structure;
[0120] Sampling module 20 is used to sample the original reservoir structure in the training set to obtain the first conditional data of the borehole form;
[0121] The arrangement module 30 is used to arrange the first condition data according to the relative position, detect blank positions in the arranged data, and fill the blank positions to obtain second condition data with the same size as the original reservoir structure.
[0122] Module 40 is used to build the neural network to be trained and set the hyperparameters of the neural network to be trained.
[0123] The training module 50 is used to initialize the weight parameters of the neural network to be trained, and to train the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network.
[0124] Test module 60 is used to perform tests using the test set and the corresponding third condition data to obtain test results;
[0125] The characterization module 70 is used to characterize the reservoir structure using the corresponding target condition data if the test result reaches the preset result.
[0126] The characterization module 70 is further configured to, if the test result does not meet the preset result, return to the step of setting the hyperparameters of the neural network to be trained, retrain the neural network with the hyperparameters reset, until the test result meets the preset result, and characterize the reservoir structure using the corresponding current condition data.
[0127] Other embodiments or specific implementations of the deep learning-based reservoir characterization device described in this invention can be found in the above-described method embodiments, and will not be repeated here.
[0128] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.
[0129] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.
[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a magnetic disk or optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0131] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A reservoir characterization method based on deep learning, characterized in that, The deep learning-based reservoir characterization method includes: Training and test sets are generated based on the original reservoir structure; The original reservoir structure in the training set is sampled to obtain the first conditional data of the borehole form; The first conditional data is arranged according to its relative position. Blank positions in the arranged data are detected and filled to obtain second conditional data with the same size as the original reservoir structure. Build the neural network to be trained and set the hyperparameters of the neural network to be trained; Initialize the weight parameters of the neural network to be trained, and train the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network; The test set and the corresponding third condition data are used to perform the test and obtain the test results. If the test results reach the preset results, the reservoir structure is characterized using the corresponding target condition data; If the test result does not meet the preset result, return to the step of setting the hyperparameters of the neural network to be trained, retrain the neural network with the hyperparameters reset, until the test result meets the preset result, and characterize the reservoir structure using the corresponding current condition data. The neural network to be trained consists of three sub-networks, including a first encoder, a second encoder, and a decoder. The step of training the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network specifically includes: The number of iterations is determined based on the hyperparameters of the neural network to be trained; In each loop, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure. After completing all the aforementioned number of iterations, the target reservoir structure is generated based on the second part of the reservoir structure obtained in each iteration; The first loss function is used to calculate the first difference between the target reservoir structure and the original reservoir structure, and the weight parameters in the first encoder and the decoder are updated by backpropagation. The loop is repeated for the specified number of iterations. In each iteration, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure. At the same time, the second encoder encodes the second conditional data into the second latent space. Repeat all the aforementioned number of iterations, and generate the target reservoir structure based on the second part of the reservoir structure obtained in each iteration; The second loss function is used to calculate the second difference between the distributions of the first latent space and the second latent space. The weight parameters in the first encoder and the second encoder are updated by the backpropagation algorithm so that the second latent space approaches the first latent space, thus obtaining the trained neural network.
2. The reservoir characterization method based on deep learning as described in claim 1, characterized in that, The first encoder selectively reads a first portion of the reservoir structure based on local features of the original reservoir structure, and encodes the portion of the reservoir structure into a first latent space. The decoder decodes a first latent vector sampled from the first latent space to generate a second portion of the reservoir structure, including: The first encoder uses a selective attention mechanism to selectively read the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder uses a selective attention mechanism to decode the first latent vector sampled from the first latent space to generate the second part of the reservoir structure.
3. The reservoir characterization method based on deep learning as described in claim 1, characterized in that, The first loss function is the reconstruction loss. The target reservoir structure input is represented by x, and the target reservoir structure output by the neural network is represented by y. For a given x and y, the reconstruction loss of the network can be expressed by the following formula: 。 4. The reservoir characterization method based on deep learning as described in claim 1, characterized in that, The second loss function is either KL loss or conditional loss, and it is used to measure the second difference between the distributions of the first latent space and the second latent space. Suppose there are two normal distributions and Then the KL loss for the distributions in the first and second hidden spaces is: 。 5. The reservoir characterization method based on deep learning as described in claim 1, characterized in that, The step of using the test set and the corresponding third condition data to perform tests and obtain test results includes: The reservoir structure in the test set and the corresponding sampled third condition data are used to conduct tests to obtain test results, and the test results are analyzed using variograms and connectivity functions.
6. The reservoir characterization method based on deep learning as described in claim 2, characterized in that, The generation of training and testing sets based on the original reservoir structure includes: The original reservoir structure is pruned using the sliding window method, and a complete dataset is constructed based on the pruned data. A predetermined proportion of the data in the complete dataset is used as the training set, and the remaining data in the complete dataset is used as the test set.
7. A reservoir characterization device based on deep learning, characterized in that, The deep learning-based reservoir characterization device includes: The generation module is used to generate training and test sets based on the original reservoir structure. The sampling module is used to sample the original reservoir structure in the training set to obtain the first conditional data of the borehole form; The arrangement module is used to arrange the first condition data according to their relative positions, detect blank positions in the arranged data, and fill the blank positions to obtain second condition data with the same size as the original reservoir structure. A module for building a neural network to be trained and setting the hyperparameters of the neural network to be trained; The training module is used to initialize the weight parameters of the neural network to be trained, and to train the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network. The testing module is used to perform tests using the test set and the corresponding third-condition data to obtain test results; The characterization module is used to characterize the reservoir structure using the corresponding target condition data if the test result reaches the preset result. The characterization module is also used to return to the step of setting the hyperparameters of the neural network to be trained if the test result does not meet the preset result, to retrain the neural network with the hyperparameters reset, until the test result meets the preset result, and to characterize the reservoir structure through the corresponding current condition data. The neural network to be trained consists of three sub-networks, including a first encoder, a second encoder, and a decoder. The step of training the neural network to be trained using the training set and the corresponding second conditional data to obtain the trained neural network specifically includes: The number of iterations is determined based on the hyperparameters of the neural network to be trained; In each loop, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure. After completing all the aforementioned number of iterations, the target reservoir structure is generated based on the second part of the reservoir structure obtained in each iteration; The first loss function is used to calculate the first difference between the target reservoir structure and the original reservoir structure, and the weight parameters in the first encoder and the decoder are updated by backpropagation. The loop is repeated for the specified number of iterations. In each iteration, the first encoder selectively reads the first part of the reservoir structure based on the local features of the original reservoir structure and encodes the part of the reservoir structure into the first latent space. The decoder decodes the first latent vector sampled from the first latent space to generate the second part of the reservoir structure. At the same time, the second encoder encodes the second conditional data into the second latent space. Repeat all the aforementioned number of iterations, and generate the target reservoir structure based on the second part of the reservoir structure obtained in each iteration; The second loss function is used to calculate the second difference between the distributions of the first latent space and the second latent space. The weight parameters in the first encoder and the second encoder are updated by the backpropagation algorithm so that the second latent space approaches the first latent space, thus obtaining the trained neural network.
8. A reservoir characterization device based on deep learning, characterized in that, The deep learning-based reservoir characterization device stores a deep learning-based reservoir characterization program, which, when executed by a processor, implements the steps of the deep learning-based reservoir characterization method as described in any one of claims 1 to 6.
9. A storage medium, characterized in that, The storage medium stores a reservoir characterization program based on deep learning, which, when executed by a processor, implements the steps of the reservoir characterization method based on deep learning as described in any one of claims 1 to 6.