Reservoir geological modeling method based on progressive conditional generative adversarial network

Abstract

The invention provides a reservoir geologic modeling method based on a progressive conditional generative adversarial network, and relates to the field of reservoir geologic modeling, and the method comprises the steps: obtaining a geological image and logging data; constructing a training sample pair according to the geological image and the logging data; constructing a progressive conditional generative adversarial network based on logging data; training the progressive conditional generative adversarial network through the training sample pair in combination with a progressive training method of which the resolution is increased step by step; obtaining logging data to be predicted; and inputting to-be-predicted logging data into the progressive conditional generative adversarial network to generate a reservoir image conforming to both the geological mode and the logging data constraint. According to the technical scheme, the problems that in reservoir modeling, a geologic body space distribution mode is not enough to capture, and well logging data constraints are difficult to meet are solved.

Description

Technical Field

[0001] The present application relates to the field of reservoir geological modeling, and in particular to a reservoir geological modeling method based on a progressive conditional generative adversarial network. Background Art

[0002] Reservoir geological modeling is to construct a visual model of the spatial morphology, geological boundaries and attribute characteristics of a specific reservoir based on various geological images collected during regional geological surveys. Through reservoir geological modeling, researchers can more clearly understand the reservoir structure, predict the distribution of underground resources, and analyze the reservoir evolution process, thereby providing a scientific basis for practical applications such as mineral resource development, groundwater management, and geological disaster prediction.

[0003] Well logging data has high local accuracy and can accurately reflect the physical properties of the reservoir, such as porosity, permeability and lithology, and accurately describe the physical characteristics and spatial distribution of the reservoir at the logging location. These data can effectively identify the spatial consistency and local variation characteristics of the reservoir, providing a true and reliable basis for reservoir modeling. However, due to acquisition costs and technical limitations, well logging data are usually sparse in space, especially in large areas, and the distribution of logging points often cannot fully represent the geological characteristics of the entire reservoir. Therefore, in the process of reservoir modeling, it is necessary to combine sparse well logging data with expert knowledge to make up for the data gaps and improve the accuracy and reliability of the model.

[0004] Existing geostatistical methods have certain limitations when dealing with complex spatial distribution patterns of geological bodies. Traditional geostatistical methods mainly rely on empirical models and statistical inferences. Although they can capture the overall characteristics of underground geological bodies to a certain extent, the spatial distribution of geological bodies is usually affected by a variety of factors. Existing methods often find it difficult to accurately describe these complex spatial relationships and diverse geological change patterns. In addition, these methods often ignore spatial heterogeneity and local characteristics, resulting in a relatively rough geological model that cannot fully reflect the fine structure of the actual geological body.

[0005] As a deep learning method, generative adversarial networks (GANs) have powerful feature learning capabilities and can learn richer and more complex geological spatial distribution patterns by optimizing adversarial training of generators and discriminators. Compared with traditional geostatistical methods, GANs can automatically capture the complex patterns and diversity hidden in the data, especially in high-dimensional geological images, and can generate more realistic geological models. Although generative adversarial networks can improve the accuracy of geological modeling to some extent, in practical applications, it is still a challenge to build a geological model that satisfies both the constraints of well phase data and geological spatial patterns. To build a model that conforms to the geological spatial pattern, it is necessary to consider the local structural constraints of the geological body imposed by well phase data, while also maintaining the spatial consistency and rationality of the geological model. Summary of the invention

[0006] The purpose of the present invention is to provide a reservoir geological modeling method based on a progressive conditional generative adversarial network in order to solve the problem that it is difficult to establish a reservoir geological model that satisfies both the spatial distribution pattern of the geological body and the constraints of logging data in the existing technology.

[0007] The above-mentioned purpose of the present application is achieved through the following technical solutions: S1: Acquire geological images and well logging data; S2: Construct training sample pairs based on geological images and well logging data; S3: Construct a progressive conditional generative adversarial network based on well logging data; S4: Train the progressive conditional generative adversarial network by training sample pairs and combining a progressive training method with a step-by-step resolution increase; S5: Obtain the logging data to be predicted; input the logging data to be predicted into the progressive conditional generative adversarial network to generate a reservoir image that conforms to both the geological model and the logging data constraints.

[0008] Optionally, step S2 includes: S21: The geological image is an original high-resolution geological image; the logging data includes: spatial distribution data and attribute data of the logging points; the attribute data is the physical property parameters or geological attribute values ​​of the logging points; S22: partitioning the geological image through a sliding window of a fixed size to obtain images of each sub-area; performing bilinear interpolation downsampling on each sub-area image to generate sub-area images of different resolution levels; S23: Match the sub-area images of different resolution levels with the logging data to construct training sample pairs.

[0009] Optionally, step S3 includes: Progressive conditional generative adversarial network, including: well logging data input channel module and total loss function based on well logging data; The logging data input channel module includes: spatial position channel and attribute channel; Spatial position channel, used to describe the spatial distribution of logging points; Attribute channel, used to record the physical parameters or geological attribute values ​​of logging points.

[0010] Optionally, step S4 includes: The progressive conditional generative adversarial network includes: a generator and a discriminator; S41: performing differential downsampling on the logging data in the training sample pair through the logging data input channel module of the generator to generate logging data with different resolution levels; S42: extracting the inherent spatial distribution and pore characteristics in the geological image, as well as the relationship between the logging data of different resolution levels and the geological images of different resolution levels in the training sample pairs through a progressive training method with gradually improved resolution, and obtaining a reservoir image generated by the generator; S43: The joint loss of the real reservoir image and the reservoir image generated by the generator is calculated through the total loss function based on the logging data, and the joint loss is back-propagated to the generator to train the progressive conditional generative adversarial network.

[0011] Optionally, step S41 includes: The maximum pooling method is used to downsample the input data of the spatial position channel of the well logging data input channel module; Average pooling is used to downsample the input data of the spatial position channel of the logging data input channel module.

[0012] Optionally, the total loss function based on the well logging data includes: a conditional loss function based on the well logging data and a priori loss function; The conditional loss function based on well logging data is calculated as follows:

[0013] In the formula, Represents a random variable and The joint probability distribution of and After sampling, the mathematical expectation of the objective function value is calculated; is the spatial location channel of the logging data, and US is the upsampling operation; Represents the samples generated by the generator; For the generator; represents random noise input; represents the well logging input; represents the real sample; represents element-wise product; The loss function of WGAN-GP is used as the prior loss function, which is defined as follows:

[0014] In the formula, represents the discriminator; is the score of the discriminator for the real sample; Represents the score of the discriminator on the samples generated by the generator; represents the gradient penalty weight; exist and Sampling between For real samples, The samples generated by the generator, The calculation formula is , is a variable, and its value is in the interval [0,1]; The total loss function based on well logging data is as follows:

[0015] In the formula, is the weight of the conditional loss function.

[0016] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes a reservoir geological modeling method based on a progressive conditional generative adversarial network.

[0017] A computer-readable storage medium stores instructions, which, when executed, perform a reservoir geological modeling method based on a progressive conditional generative adversarial network.

[0018] The beneficial effects of the technical solution provided by this application are: 1. Based on the progressive generative adversarial network model, the generator architecture is modified and the well logging data input channel is added to enable it to learn well logging data. The training starts with low-resolution geological images, first learning the basic spatial distribution and structural characteristics of the geological body; as the training progresses, the resolution of the generated image is gradually improved, thereby extracting more detailed geological features.

[0019] 2. A loss function based on well logging data is designed. By calculating the difference between the generated image and the reference image at the logging position, the generator is guided to learn the relationship between the well logging data and the spatial geological pattern, and finally a geological image that conforms to the geological spatial pattern and satisfies the constraints of the well logging data is generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present application will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a diagram of the structure of a progressive conditional generative adversarial network in an embodiment of the present application; Figure 2 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.

[0022] An embodiment of the present application provides a reservoir geological modeling method based on a progressive conditional generative adversarial network.

[0023] Please refer to Figure 1 , Figure 1 It is a progressive conditional generative adversarial network structure diagram of a reservoir geological modeling method based on a progressive conditional generative adversarial network in an embodiment of the present application, including: S1: Acquire geological images and well logging data; S2: Construct training sample pairs based on geological images and well logging data; S3: Construct a progressive conditional generative adversarial network based on well logging data; S4: Train the progressive conditional generative adversarial network by training sample pairs and combining a progressive training method with a step-by-step resolution increase; S5: Obtain the logging data to be predicted; input the logging data to be predicted into the progressive conditional generative adversarial network to generate a reservoir image that conforms to both the geological model and the logging data constraints.

[0024] As an embodiment, the present invention uses two different types of geological images to verify its effectiveness, namely, a steady-state classified river system and a non-steady-state continuous ice wedge system. The statistical characteristics of the steady-state system remain consistent in space, and there is no obvious trend in spatial distribution; while the statistical characteristics of the non-steady-state system change with spatial position, showing a certain trend or regularity. Through these two types of data, the applicability and robustness of the present invention under different spatial distribution conditions can be comprehensively evaluated.

[0025] Step S2 includes: S21: The geological image is an original high-resolution geological image; the logging data includes: spatial distribution data and attribute data of the logging points; the attribute data is the physical property parameters or geological attribute values ​​of the logging points; S22: partitioning the geological image through a sliding window of a fixed size to obtain images of each sub-area; performing bilinear interpolation downsampling on each sub-area image to generate sub-area images of different resolution levels; S23: Match the sub-area images of different resolution levels with the logging data to construct training sample pairs.

[0026] As an embodiment, in order to make the generative adversarial network suitable for reservoir geological modeling, the present invention first preprocesses the geological image and constructs a priori geological model of the study area as the original training image (i.e. Figure 1 The original high-resolution image (geological image) is partitioned through a sliding window of fixed size, and each sub-area image B is paired with randomly selected well logging data A to form a training sample pair (A, B) to better simulate the actual situation.

[0027] As an embodiment, since the present invention adopts a progressive training method, the resolution of the generated reservoir image is gradually improved, starting from 4×4 to 8×8, and finally to a higher resolution. Therefore, it is necessary to perform bilinear interpolation downsampling on each sub-area image B to generate a low-resolution image for training the CPGAN (Progressive Conditional Generative Adversarial Network) model.

[0028] Step S3 includes: Progressive conditional generative adversarial network, including: well logging data input channel module and total loss function based on well logging data; The logging data input channel module includes: spatial position channel and attribute channel; Spatial position channel, used to describe the spatial distribution of logging points; Attribute channel, used to record the physical parameters or geological attribute values ​​of logging points.

[0029] As an embodiment, in order to enable the CPGAN model to convert the input logging data into a reservoir image that conforms to the geological model and conditional data (well phase data constraints), a suitable input channel for logging data is specially designed. In the present invention, the logging data is processed into a dual-channel image form: the first channel is a spatial position channel, which is used to describe the distribution of logging points in space; the second channel is an attribute channel, which is used to record the physical parameters or other geological attribute values ​​of the logging points.

[0030] Step S4 includes: The progressive conditional generative adversarial network includes: a generator and a discriminator; S41: performing differential downsampling on the logging data in the training sample pair through the logging data input channel module of the generator to generate logging data with different resolution levels; S42: extracting the inherent spatial distribution and pore characteristics in the geological image, as well as the relationship between the logging data of different resolution levels and the geological images of different resolution levels in the training sample pairs through a progressive training method with gradually improved resolution, and obtaining a reservoir image generated by the generator; S43: The joint loss of the real reservoir image and the reservoir image generated by the generator is calculated through the total loss function based on the logging data, and the joint loss is back-propagated to the generator to train the progressive conditional generative adversarial network.

[0031] The present application provides an embodiment as follows: the reservoir geological modeling framework based on progressive conditional generative adversarial network proposed in the present invention includes a generator network and a discriminator network such as Figure 1 As shown. In the process of adversarial learning, the discriminator network assists the generator network to produce higher quality reservoir images. The progressive training strategy adopted by the present invention gradually extracts feature information by gradually training the generator and the discriminator from low resolution to high resolution. It first starts training with low-resolution images (such as 4×4), so that the generator first learns the global structural features of the image (low-resolution image of the real geological image), while the discriminator focuses on the global consistency of low resolution. Then, as the training progresses, the resolution of the network is gradually increased, for example, from 8×8, 16×16 to higher resolution. This approach allows the network to first capture the rough features of the image, such as shape and layout, and then gradually refine finer textures and local details, reducing the complexity of high-resolution training, while improving the stability of training and the quality of generated reservoir images.

[0032] During the progressive training process, the network architecture of the generator and discriminator gradually becomes more complex. Specifically: Generator: The generator initially consists of fewer layers, including only one convolutional layer and one fully connected layer, which is used to generate low-resolution reservoir images. The design of these layers mainly focuses on the global structural features of the image.

[0033] As the resolution increases, new convolutional layers are gradually added to the generator to generate higher resolution details. The newly added convolutional layers gradually capture finer local geological patterns in the well logging data A and reservoir images B while retaining global consistency.

[0034] At each stage, new layers are trained simultaneously with existing layers, allowing the generator to progressively enhance its expressive power.

[0035] Discriminator: The initial architecture of the discriminator is simple, consisting of only one convolutional layer and two fully connected layers to distinguish between low-resolution real and generated reservoir images. These layers mainly focus on the global consistency of the image.

[0036] As the resolution increases, new layers are progressively added to the discriminator, enabling it to detect local details and texture differences at higher resolutions.

[0037] The newly added layers will work together with the existing layers to more accurately evaluate the authenticity and consistency of the generated images through multi-resolution information fusion.

[0038] The generator has two inputs, random noise z and well logging data A. Random noise z is a noise vector randomly sampled from a Gaussian distribution, which is used to provide generation diversity. Well logging data A is preprocessed through the well logging data input channel mentioned above. The processed dual-channel well logging data passes through a convolution layer to extract its feature map. These feature maps contain the spatial distribution and attribute information of the well logging data, which can effectively express the conditional constraint characteristics. Random noise z is processed by the main branch (multi-layer convolution layer) to extract its feature map, which is used to capture the randomness and diversity of the generated image. Then, these two parts of the feature map (the conditional feature map extracted from the well logging data A and the feature map extracted from the random noise z) are concatenated in the channel dimension to form a comprehensive feature representation. This comprehensive feature representation contains both the conditional information of the well logging data and the diversity brought by random noise, which serves as the input of the subsequent network generation module.

[0039] Through this design, the generator can not only learn the geological patterns in geological images, but also capture the relationship between geological patterns and logging data. Specifically, the generator combines the diversity provided by random noise z with the conditional constraints provided by logging data A, and learns two aspects at the same time during the generation process: one is the inherent spatial distribution and pore characteristics in geological images (i.e., geological patterns), and the other is how these characteristics are constrained or affected by logging data (i.e., the correlation between logging data and geological patterns). This correlation may be reflected in the control of logging data on reservoir distribution and attribute values. By making full use of the spatial location channel and attribute channel information provided by logging data during the training process, the generator can generate reservoir images that are consistent with both geological laws and input logging data, ensuring that the generated results are realistic and conditionally dependent, thereby more effectively meeting the needs of geological modeling.

[0040] Step S41 includes: The maximum pooling method is used to downsample the input data of the spatial position channel of the well logging data input channel module; Average pooling is used to downsample the input data of the spatial position channel of the logging data input channel module.

[0041] As an example, during the progressive training of the CPGAN model, the real reservoir image gradually transitions from low resolution to high resolution to gradually guide the network to learn multi-scale feature expression. Therefore, correspondingly, the well logging data also needs to be downsampled to different resolution levels. In order to retain the information of the well logging data to the greatest extent during the downsampling process, differentiated downsampling methods are designed according to the characteristics of different channels.

[0042] As an embodiment, the core function of the spatial position channel is to record the spatial distribution information of the logging points. When the spatial position channel of the logging data input channel module is downsampled, the maximum pooling method is used to represent the regional information by selecting the maximum value in the pooling window. Maximum pooling can effectively retain the significant features of the data, such as sharp changes or boundary information. For the spatial position channel, this method ensures that the significant spatial distribution characteristics of the logging points will not be blurred due to downsampling, thereby better retaining the information reflecting the spatial structure.

[0043] As an example, the attribute channel carries physical parameters or other geological attributes, which are usually presented as continuous and smoothly changing numerical distributions. When downsampling, the idea of ​​average pooling is adopted to represent regional information by taking the average value within the pooling window. Average pooling can smooth local changes and retain the overall trend and mean characteristics. This method is suitable for reflecting continuously changing geological attributes because it does not overemphasize local outliers, thereby better reflecting the overall distribution law of the attributes.

[0044] The total loss function based on the well logging data includes: a conditional loss function based on the well logging data and a priori loss function; The conditional loss function based on well logging data is calculated as follows:

[0045] In the formula, Represents a random variable and The joint probability distribution of and After sampling, the mathematical expectation of the objective function value is calculated; is the spatial location channel of the logging data, and US is the upsampling operation; Represents the samples generated by the generator; For the generator; represents random noise input; represents the well logging input; represents the real sample; represents element-wise product; The loss function of WGAN-GP is used as the prior loss function, which is defined as follows:

[0046] In the formula, represents the discriminator; is the score of the discriminator for the real sample; Represents the score of the discriminator on the samples generated by the generator; represents the gradient penalty weight; exist and Sampling between For real samples, The samples generated by the generator, The calculation formula is , is a variable, and its value is in the interval [0,1]; The total loss function based on well logging data is as follows:

[0047] In the formula, is the weight of the conditional loss function.

[0048] As an embodiment, in order to correctly guide the generator to learn the relationship between the geological model (i.e., the inherent spatial distribution and pore characteristics in the geological image) and the logging data, the present invention designs a conditional loss function based on the logging data. The conditional loss function based on the logging data quantifies the difference between the generated image and the actual conditional data (logging data) by calculating the L2 distance between the reservoir image generated by the generator and the logging data, and minimizes the conditional loss and back-propagates the ladder to the generator, thereby prompting the generator to better meet the constraints of the logging data during the optimization process. Since the resolution of the generated reservoir image is gradually improved during the progressive training process, while the resolution of the logging data is fixed, the present invention adopts an upsampling method to upsample the generated reservoir image and the reservoir image used for training to the same resolution as the conditional data image (logging data image).

[0049] As an embodiment, in order to guide the generator to correctly learn the geological model, the present invention adopts the loss function of WGAN-GP (Wasserstein GAN with Gradient Penalty) as the prior loss function. WGAN-GP is a generative adversarial network based on Wasserstein distance. It measures the difference between the real data distribution and the generated data distribution by optimizing the Wasserstein distance, thereby providing a more stable training process. WGAN-GP introduces a gradient penalty term to constrain the gradient of the discriminator to not exceed a certain value, ensuring that the output of the discriminator is a 1-Lipschitz continuous function, avoiding the problem of gradient explosion or disappearance during training.

[0050] The present application also discloses an electronic device. Figure 2 , Figure 2The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .

[0051] The communication bus 502 is used to realize the connection and communication between these components.

[0052] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.

[0053] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0054] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for a processor to load to execute the above-mentioned reservoir geological modeling method based on a progressive conditional generative adversarial network.

[0055] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure.

[0056] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A reservoir geological modeling method based on a progressive conditional generative adversarial network, characterized in that: The method comprises the following steps: S1: Acquire geological images and well logging data; S2: Construct training sample pairs based on geological images and well logging data; S3: Construct a progressive conditional generative adversarial network based on well logging data; S4: Train the progressive conditional generative adversarial network by training sample pairs and combining a progressive training method with a step-by-step resolution increase; S5: Obtain the logging data to be predicted; input the logging data to be predicted into the progressive conditional generative adversarial network to generate a reservoir image that conforms to both the geological model and the logging data constraints.

2. A reservoir geological modeling method based on a progressive conditional generative adversarial network according to claim 1, characterized in that: Step S2 includes: S21: The geological image is an original high-resolution geological image; the logging data includes: spatial distribution data and attribute data of the logging points; the attribute data is the physical property parameters or geological attribute values ​​of the logging points; S22: partitioning the geological image through a sliding window of a fixed size to obtain images of each sub-area; performing bilinear interpolation downsampling on each sub-area image to generate sub-area images of different resolution levels; S23: Match the sub-area images of different resolution levels with the logging data to construct training sample pairs.

3. A reservoir geological modeling method based on a progressive conditional generative adversarial network according to claim 1, characterized in that: Step S3 includes: Progressive conditional generative adversarial network, including: well logging data input channel module and total loss function based on well logging data; The logging data input channel module includes: spatial position channel and attribute channel; Spatial position channel, used to describe the spatial distribution of logging points; Attribute channel, used to record the physical parameters or geological attribute values ​​of logging points.

4. A reservoir geological modeling method based on a progressive conditional generative adversarial network as claimed in claim 3, characterized in that: Step S4 includes: The progressive conditional generative adversarial network includes: a generator and a discriminator; S41: performing differential downsampling on the logging data in the training sample pair through the logging data input channel module of the generator to generate logging data with different resolution levels; S42: extracting the inherent spatial distribution and pore characteristics in the geological image, as well as the relationship between the logging data of different resolution levels and the geological images of different resolution levels in the training sample pairs through a progressive training method with gradually improved resolution, and obtaining a reservoir image generated by the generator; S43: The joint loss of the real reservoir image and the reservoir image generated by the generator is calculated through the total loss function based on the logging data, and the joint loss is back-propagated to the generator to train the progressive conditional generative adversarial network.

5. A reservoir geological modeling method based on a progressive conditional generative adversarial network as claimed in claim 4, characterized in that: Step S41 includes: The maximum pooling method is used to downsample the input data of the spatial position channel of the well logging data input channel module; Average pooling is used to downsample the input data of the spatial position channel of the logging data input channel module.

6. A reservoir geological modeling method based on a progressive conditional generative adversarial network as claimed in claim 3, characterized in that: The total loss function based on the well logging data includes: a conditional loss function based on the well logging data and a priori loss function; The conditional loss function based on well logging data is calculated as follows: In the formula, Represents a random variable and The joint probability distribution of and After sampling, the mathematical expectation of the objective function value is calculated; is the spatial location channel of the logging data, and US is the upsampling operation; Represents the samples generated by the generator; For the generator; represents random noise input; represents the well logging input; represents the real sample; represents element-wise product; The loss function of WGAN-GP is used as the prior loss function, which is defined as follows: In the formula, represents the discriminator; is the score of the discriminator for the real sample; Represents the score of the discriminator on the samples generated by the generator; represents the gradient penalty weight; exist and Sampling between For real samples, The samples generated by the generator, The calculation formula is , is a variable, and its value is in the interval [0,1]; The total loss function based on well logging data is as follows: In the formula, is the weight of the conditional loss function.

7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images