Reservoir gas content prediction method and system based on feature extraction and fusion network

Abstract

The application provides a reservoir gas content prediction method and system based on feature extraction and fusion network. The method comprises the following steps: obtaining well-seismic data and geological information of a study area; preprocessing the obtained well-seismic data; performing fine well-seismic calibration by comprehensively using the geological information and the well-seismic data; converting time-domain seismic data to depth domain; performing feature extraction such as seismic texture, dispersion and formation quality factor on the depth-domain seismic data; obtaining target data by actual observation or forward simulation and constructing a data set; establishing a self-attention mechanism deep fusion network reservoir gas content prediction model and performing model training by using the constructed training set; testing the trained network model by using the constructed test set; and finally obtaining the reservoir gas content distribution of the target area. The reservoir gas content prediction method established by using the deep fusion network on the basis of extracting the features of the multi-parameter seismic data can effectively and accurately predict the gas content distribution of the deep reservoir.

Description

Technical Field

[0001] This invention relates to the field of oil and gas reservoir prediction, and in particular, to a method and system for predicting reservoir gas content based on feature extraction and fusion networks. Background Technology

[0002] Natural gas is a scarce strategic clean energy source in my country, and increasing the exploration and development of natural gas is an important part of the national energy strategy. Detecting the gas-bearing capacity of reservoirs is the core of natural gas exploration. Currently, most reservoirs are buried at great depths, exhibiting weak seismic response characteristics and pore fluid response, with minimal differences between reservoirs and non-reservoir areas, making the detection of gas-bearing capacity in natural gas reservoirs extremely difficult. Existing methods for detecting gas-bearing capacity, such as bright spot analysis and AVO analysis, are highly reliable for detecting gas-bearing capacity in shallow reservoirs, but their reliability is very low under deep reservoir conditions, rendering them largely unsuitable. The cepstral transform-based seismic ripple analysis method, due to its sensitivity to weak reflections, can detect dark spot-type gas reservoirs. Furthermore, its model independence allows us to overcome many drawbacks of modeling, enabling more reliable detection of gas-bearing capacity in deep and ultra-deep reservoirs.

[0003] For deep natural gas exploration, reservoir gas content detection is not just about determining whether a reservoir contains gas, but more importantly, determining the amount of gas, which is crucial in determining whether a reservoir has industrial exploitation value. To determine the gas content, it's necessary to analyze the reservoir's volumetric response, a possibility only achievable through low-frequency shadowing analysis. Low-frequency shadowing is a volumetric response, essentially a dispersion. If we combine it with methods such as seismic rhyme analysis, we can potentially determine the gas content of the reservoir and estimate its recoverable reserves from the reflection and dispersion of seismic waves.

[0004] Deep learning is a newly developed class of machine learning algorithms in the field of artificial intelligence. Machine learning uses computers to automatically analyze data to obtain patterns and then uses these patterns to predict unknown data. Given the current difficulty in establishing accurate reservoir gas content prediction models, reservoir gas content prediction models built using deep learning can achieve accurate predictions of reservoir gas content. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a method and system for predicting reservoir gas content based on feature extraction and fusion networks. This invention utilizes multi-parameter seismic data features extracted through depth-domain seismic texture analysis, low-frequency shadowing analysis, and formation quality factor inversion. It employs a self-attention mechanism deep fusion network to establish a complex nonlinear mapping relationship between the extracted multi-parameter seismic data features and reservoir gas content. This effectively and accurately predicts the gas content distribution of deep reservoirs, providing a reliable foundation for natural gas exploration and development.

[0006] To achieve the above objectives and technical effects, the present invention adopts the following technical solution:

[0007] In a first aspect, embodiments of this disclosure provide a method for predicting reservoir gas content based on feature extraction and fusion networks, the method comprising the following steps:

[0008] (1) Obtain well logging data, seismic data and geological information for the study area;

[0009] (2) Preprocess the acquired well logging data and seismic data;

[0010] (3) The geological information and pre-processed well logging data and seismic data are used to perform fine calibration of seismic geological comprehensive stratigraphy;

[0011] (4) Based on the fine stratigraphic calibration, perform high-precision time-depth conversion to convert time-domain seismic data to depth-domain seismic data;

[0012] (5) Perform seismic texture analysis, low-frequency shadow analysis and formation quality factor inversion on the depth domain seismic data to extract multi-parameter seismic data features such as seismic texture features, dispersion features and formation quality factors that characterize the reservoir.

[0013] (6) Obtain target data through actual observation or forward simulation and construct a dataset by combining the extracted multi-parameter seismic data features;

[0014] (7) On this basis, a deep fusion network model for predicting reservoir gas content using a self-attention mechanism is established and the model is trained using the constructed training dataset to obtain a deep fusion network model for predicting reservoir gas content using a self-attention mechanism.

[0015] (8) Then, the trained reservoir gas content prediction deep fusion network model is tested using the constructed test dataset;

[0016] (9) Finally, the actual reservoir seismic response in the target area is used as input to predict the gas content distribution in the target area.

[0017] Preferably, the preprocessing of the acquired logging data and seismic data in step (2) includes: outlier removal, correction, and lateral standardization of the logging data, and trace editing, amplitude compensation, and pre-stack migration of the seismic data.

[0018] Preferably, the essence of step (3) in performing comprehensive seismic geological stratigraphic calibration by integrating the geological information and preprocessed well logging data with seismic data is to adjust the time-depth relationship, accurately and seamlessly corresponding the seismic information in the time domain with the well logging information in the depth domain, including:

[0019] Seismic wavelets are extracted from seismic data; reflection coefficients are calculated using well logging data and convolved with seismic wavelets to obtain synthetic seismic records; the correspondence between seismic reflections and geological strata is found by integrating geological information; the seismic data of the target layer is first aligned with the large set of strata in the synthetic record by using overall time shift, and then the smaller layers are fine-tuned by using local stretching or compression.

[0020] Preferably, step (4) involves performing high-precision time-depth conversion based on fine stratigraphic calibration to convert time-domain seismic data to depth-domain data, including: performing high-precision time-depth conversion on seismic data according to the time-depth relationship established in step (3); transforming seismic data to depth-domain data through scaling; resampling the depth-domain seismic data; and obtaining depth-domain seismic data.

[0021] Preferably, the seismic pattern analysis described in step (5) is a reservoir gas-bearing detection method developed by us based on the acoustic pattern analysis method and the introduction of the concept of seismic pattern. The reservoir gas-bearing detection method based on seismic pattern analysis is sensitive to weak reflections and can discover dark spot gas reservoirs. At the same time, because it is independent of the model, we can overcome many drawbacks brought about by modeling and realize reliable detection of gas-bearing properties in deep and ultra-deep reservoirs.

[0022] Preferably, the low-frequency shadow mentioned in step (5) refers to the phenomenon of relative enhancement of low-frequency component energy appearing below the gas-bearing layer on the seismic profile. In essence, the absorption or attenuation of seismic waves by the gas-bearing reservoir has dispersion, which reflects the volume response of the reservoir. Based on the dispersion analysis of low-frequency shadow, the gas content of the reservoir can be calculated, which requires obtaining the volume characteristic information of the reservoir.

[0023] Preferably, the formation quality factor described in step (5) is an important indicator of the hydrocarbon content of a formation and one of the important parameters for describing rock elasticity. It is closely related to the internal structural characteristics of the medium and the fluid properties such as saturation, porosity, and permeability of the medium. Accurate estimation of the formation quality factor can provide a lot of useful information for oil and gas detection.

[0024] Preferably, step (6) involves acquiring target data through actual observation or forward simulation and constructing a dataset by combining the extracted multi-parameter seismic data features, which includes: firstly, acquiring target data based on actual observation or forward simulation; then constructing a dataset using the acquired target data and the extracted multi-parameter seismic data features; and finally, dividing the constructed dataset into a training dataset and a test dataset according to specific needs.

[0025] Preferably, the training dataset described in step (6) includes actual reservoir seismic response data or low-frequency rock physics experimental observation data or a mixture of both, and should be selected and used reasonably according to the specific circumstances of the study area.

[0026] Preferably, the self-attention mechanism deep fusion network described in step (7) is a fusion network developed by combining a self-attention mechanism convolutional network and a self-attention mechanism long short-term memory network. The convolutional network is used to extract local morphological features of multi-parameter seismic data features, the long short-term memory network mines the trend information of data changes with depth, and the self-attention mechanism is used to improve the network's attention to data-related features, thereby improving the accuracy of reservoir gas content prediction.

[0027] Preferably, the self-attention mechanism described in step (7) is an improvement on the attention mechanism, which reduces reliance on external information and is better at capturing the internal correlations of data or features. The attention mechanism draws on the selective attention mechanism in the human visual system, which focuses attention on important information within the visual range and ignores non-essential information within the visual range. Thus, it filters out more valuable and important information for the current work goal from complex data information. It is a general idea and a mechanism for allocating resources according to their importance.

[0028] Preferably, the purpose of testing the trained reservoir gas content prediction model using the constructed test dataset in step (8) is to evaluate the performance of the self-attention mechanism deep fusion network model established in step (7).

[0029] Preferably, step (9) of using the actual reservoir seismic response of the target area as input to predict the gas content distribution of the target area includes: firstly, processing the seismic data of the target area through steps (1) to (5) to extract multi-parameter seismic data features in the depth domain of the target area; then inputting the extracted multi-parameter seismic data features into the trained reservoir gas content prediction depth fusion network model to obtain the reservoir gas content distribution of the target area.

[0030] Preferably, the reservoir gas content prediction in step (9) is performed in the depth domain, and the spatial location accuracy of depth domain seismic data is better than that of time domain seismic data.

[0031] As one specific implementation of this disclosure,

[0032] Secondly, this disclosure provides a reservoir gas content prediction system based on feature extraction and fusion networks. The system includes a processor and a memory, the memory storing a computer program executable on the processor. When the processor executes the computer program, it performs the following steps: acquiring well logging data, seismic data, and geological information for the study area; preprocessing the acquired well logging data and seismic data; comprehensively utilizing the geological information and the preprocessed well logging data and seismic data to perform fine seismic-geological stratigraphic calibration; performing high-precision time-depth conversion based on the fine stratigraphic calibration to convert the time-domain seismic data to the depth domain; and performing seismic texture analysis on the depth-domain seismic data. The seismic data features, including seismic texture characteristics, dispersion characteristics, and formation quality factors, are extracted from the reservoir through analysis, low-frequency shadowing analysis, and formation quality factor inversion. Target data is acquired through actual observation or forward simulation and combined with the extracted multi-parameter seismic data features to construct a dataset. Based on this, a deep fusion network model for predicting reservoir gas content using a self-attention mechanism is established and trained using the constructed training dataset to obtain the deep fusion network model for predicting reservoir gas content. Then, the trained reservoir gas content prediction model is tested using the constructed test dataset. Finally, the actual reservoir seismic response in the target area is used as input to predict the reservoir gas content distribution in the target area.

[0033] The beneficial effects of this invention are as follows: Based on high-precision time-depth conversion, this invention extracts multi-parameter seismic data features from depth-domain seismic data through seismic texture analysis, low-frequency shadow analysis, and formation quality factor inversion. The reservoir gas content prediction method established by the self-attention mechanism deep fusion network can effectively and accurately predict the gas content distribution of deep reservoirs, providing a reliable foundation for natural gas exploration and development. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments in this specification, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. In the drawings:

[0035] Figure 1 A flowchart illustrating a reservoir gas content prediction method based on feature extraction and fusion network provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram illustrating the data set construction and input of a reservoir gas content prediction method based on feature extraction and fusion network provided in an embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the specification, and not all embodiments.

[0038] This invention provides a method for predicting reservoir gas content based on feature extraction and network fusion. (Reference) Figure 1 As shown, in some embodiments, the method may include the following steps:

[0039] Step 101: Obtain well logging data, seismic data, and geological information for the study area;

[0040] Step 102: Preprocess the acquired well logging data and seismic data;

[0041] Step 103: Perform fine-scale seismic-geological comprehensive stratigraphic calibration by comprehensively utilizing the geological information, preprocessed well logging data, and seismic data;

[0042] Step 104: Based on the detailed stratigraphic calibration, perform high-precision time-depth conversion to convert the time-domain seismic data to the depth domain;

[0043] Step 105: Perform seismic texture analysis, low-frequency shadow analysis, and formation quality factor inversion on the depth domain seismic data to extract multi-parameter seismic data features that characterize the reservoir, such as seismic texture features, dispersion features, and formation quality factors.

[0044] Step 106: Obtain target data through actual observation or forward modeling and construct a dataset by combining the extracted multi-parameter seismic data features;

[0045] Step 107: Based on this, establish a deep fusion network reservoir gas content prediction model with self-attention mechanism and train the model using the constructed training dataset to obtain the reservoir gas content prediction model.

[0046] Step 108: Next, the trained self-attention mechanism deep fusion network reservoir gas content prediction model is tested using the constructed test dataset.

[0047] Step 109: Finally, the actual seismic response of the target reservoir is used as input to predict the gas content distribution of the target reservoir.

[0048] In one example, the preprocessing of the acquired well logging data and seismic data described in step 102 includes: outlier removal, correction, and lateral standardization of the well logging data; and trace editing, amplitude compensation, and pre-stack migration of the seismic data.

[0049] In one example, step 103, which involves comprehensively utilizing the geological information and preprocessed well logging data and seismic data to perform fine-grained seismic geological stratigraphic mapping, essentially involves adjusting the time-depth relationship to accurately and seamlessly correlate seismic information in the time domain with well logging information in the depth domain. This includes:

[0050] Seismic wavelets are extracted from seismic data; reflection coefficients are calculated using well logging data and convolved with seismic wavelets to obtain synthetic seismic records; the correspondence between seismic reflections and geological strata is found by integrating geological information; the seismic data of the target layer is first aligned with the large set of strata in the synthetic record by using overall time shift, and then the smaller layers are fine-tuned by using local stretching or compression.

[0051] In one example, step 104, which describes performing high-precision time-depth conversion based on fine-grained stratigraphic calibration to convert time-domain seismic data to the depth domain, includes: performing high-precision time-depth conversion on the seismic data according to the time-depth relationship established in step 103; scaling the seismic data to the depth domain; resampling the depth-domain seismic data; and obtaining depth-domain seismic data.

[0052] In one example, the seismic pattern analysis described in step 105 is a reservoir gas-bearing detection method that we developed by drawing on acoustic pattern analysis and introducing the concept of seismic patterns. The reservoir gas-bearing detection method based on seismic pattern analysis is sensitive to weak reflections and can detect dark spot gas reservoirs. At the same time, because it is independent of the model, we can overcome many drawbacks brought about by modeling and achieve reliable detection of gas-bearing properties in deep and ultra-deep reservoirs.

[0053] In one example, the low-frequency shadow mentioned in step 105 refers to the relative enhancement of the low-frequency component energy appearing below the gas-bearing layer on the seismic profile. Essentially, the absorption or attenuation of seismic waves by the gas-bearing reservoir has dispersion, reflecting the volume response of the reservoir. Based on the dispersion analysis of the low-frequency shadow, the gas content of the reservoir can be calculated, which requires obtaining the volume characteristic information of the reservoir.

[0054] In one example, the formation quality factor mentioned in step 105 is an important indicator of the hydrocarbon content of a formation and one of the important parameters describing rock elasticity. It is closely related to the internal structural characteristics of the medium and fluid properties such as saturation, porosity, and permeability. Accurate estimation of the formation quality factor can provide a wealth of useful information for oil and gas detection.

[0055] In one example, step 106, which involves acquiring target data through actual observation or forward simulation and constructing a dataset by combining the extracted multi-parameter seismic data features, includes: first, acquiring target data based on actual observation or forward simulation; then, constructing a dataset using the acquired target data and the extracted multi-parameter seismic data features; and finally, dividing the constructed dataset into a training dataset and a test dataset according to specific needs.

[0056] In one example, the training dataset described in step 106 includes actual reservoir seismic response data, low-frequency rock physics experimental observation data, or a mixture of both, and should be selected and used appropriately according to the specific circumstances of the study area.

[0057] In one example, the self-attention mechanism deep fusion network described in step 107 is a fusion network developed by combining a self-attention mechanism convolutional network and a self-attention mechanism long short-term memory network. The convolutional network is used to extract local morphological features of multi-parameter seismic data features, the long short-term memory network mines the trend information of data changes with depth, and the self-attention mechanism is used to improve the network's attention to data-related features, thereby improving the accuracy of reservoir gas content prediction.

[0058] In one example, the self-attention mechanism described in step 107 is an improvement on the attention mechanism, which reduces the reliance on external information and is better at capturing the internal correlations of data or features. The attention mechanism draws on the selective attention mechanism in the human visual system, which focuses attention on important information within the visual range and ignores non-essential information within the visual range, thereby filtering out more valuable and important information for the current work goal from complex data information.

[0059] In one example, the purpose of testing the trained reservoir gas content prediction model using the constructed test dataset in step 108 is to evaluate the performance of the self-attention mechanism deep fusion network model established in step 107.

[0060] In one example, step 109, which uses the actual reservoir seismic response of the target area as input to predict the gas content distribution of the target area, includes: firstly, processing the seismic data of the target area through steps 101 to 105 to extract multi-parameter seismic data features in the depth domain of the target area; then, inputting the extracted multi-parameter seismic data features into the trained reservoir gas content prediction depth fusion network model to obtain the reservoir gas content distribution of the target area.

[0061] In one example, the reservoir gas content prediction described in step 109 is performed in the depth domain, where the spatial location accuracy of depth domain seismic data is superior to that of time domain seismic data.

[0062] Tests were conducted using actual data from a certain exploration area. Figure 2 This diagram illustrates how a dataset is constructed and input into a model. The diagram shows five inline and five xline lines, with a depth of 300 meters. (Example:) Figure 2First, seismic data feature volumes, including well-side seismic texture features, dispersion features, and formation quality factors, are extracted. Then, seismic data feature profiles are extracted along the four directions of these feature volumes to increase the amount of training data and improve the model's generalization ability. Next, well gas-bearing data is matched one track at a time along the extracted seismic data features and input into a self-attention mechanism deep fusion network model in the form of point-window pairs for iterative learning and training. Finally, a multi-parameter seismic data feature reservoir gas-bearing prediction self-attention mechanism deep fusion network model is obtained. As needed, the extracted multi-parameter seismic data features in the depth domain of the target area are input into the trained reservoir gas-bearing prediction model to obtain the reservoir gas-bearing distribution in the target area.

[0063] Corresponding to the aforementioned reservoir gas content prediction method based on feature extraction and fusion networks, this invention provides a reservoir gas content prediction system based on feature extraction and fusion networks. In some embodiments, the prediction system includes a processor and a memory. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it performs the following steps: acquiring well logging data, seismic data, and geological information of the study area; preprocessing the acquired well logging data and seismic data; comprehensively utilizing the geological information and the preprocessed well logging data and seismic data to perform fine seismic-geological stratigraphic calibration; performing high-precision time-depth conversion based on the fine stratigraphic calibration to convert the time-domain seismic data to the depth domain; and performing seismic texture analysis, low-frequency shadowing analysis, and formation quality factor inversion on the depth-domain seismic data to improve... Multi-parameter seismic data features, such as seismic texture characteristics, dispersion characteristics, and formation quality factors, are extracted to characterize the reservoir. Target data is acquired through actual observation or forward simulation, and a dataset is constructed by combining the extracted multi-parameter seismic data features. Based on this, a deep fusion network model for predicting reservoir gas content using a self-attention mechanism is established and trained using the constructed training dataset to obtain the deep fusion network model for predicting reservoir gas content. Subsequently, the trained deep fusion network model for predicting reservoir gas content is tested using the constructed test dataset. Finally, the actual seismic response of the target area reservoir is used as input to predict the reservoir gas content distribution in the target area.

[0064] The reservoir gas content prediction system based on feature extraction and fusion network provided in this embodiment of the invention can perform the above-mentioned tasks. Figure 1 The specific implementation process of the method embodiment shown is detailed in the method embodiment and will not be repeated here.

[0065] It should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art should understand that various changes and modifications can be made to the present invention without departing from the spirit and scope of the present invention, and all such changes and modifications fall within the scope of the claimed present invention.

Claims

1. A method for predicting reservoir gas content based on feature extraction and fusion networks, characterized in that, The method may include: Acquire well logging data, seismic data, and geological information for the study area; The acquired well logging data and seismic data are preprocessed; By comprehensively utilizing geological information and pre-processed well logging data and seismic data, a fine-scale seismic-geological integrated stratigraphic determination is performed. Based on the fine stratigraphic calibration, high-precision time-depth conversion is performed to convert time-domain seismic data to depth-domain seismic data; Seismic texture analysis, low-frequency shadowing analysis, and formation quality factor inversion are performed on the depth domain seismic data to extract multi-parameter seismic data features characterizing the reservoir, including seismic texture features, dispersion features, and formation quality factor. The target data is obtained through actual observation or forward simulation and combined with the extracted multi-parameter seismic data features to construct a dataset, which is then divided into a training dataset and a test dataset. Based on this, a deep fusion network model for predicting reservoir gas content using a self-attention mechanism is established, and the model is trained using the constructed training dataset to obtain a deep fusion network model for predicting reservoir gas content using a self-attention mechanism. Then, the trained reservoir gas content prediction self-attention mechanism deep fusion network model was tested using the constructed test dataset. Finally, the actual reservoir seismic response in the target area is used as input to predict the gas content distribution in the target area.

2. The method for predicting reservoir gas content based on feature extraction and fusion network according to claim 1, characterized in that: The seismic pattern analysis described above is a reservoir gas-bearing detection method developed by drawing on acoustic pattern analysis and introducing the concept of seismic patterns. Seismic texture analysis-based reservoir gas-bearing detection methods are sensitive to weak reflections and can detect dark spot-type gas reservoirs.

3. The reservoir gas content prediction method based on feature extraction and fusion network according to claim 1, characterized in that: The low-frequency shadow refers to the relative enhancement of low-frequency component energy appearing below the gas-bearing layer on the seismic profile. Essentially, it is the dispersion of the absorption or attenuation of seismic waves by the gas-bearing reservoir, reflecting the volume response of the reservoir. Based on the dispersion analysis of the low-frequency shadow, it is necessary to obtain the volume characteristic information of the reservoir to calculate the gas content of the reservoir.

4. The method for predicting reservoir gas content based on feature extraction and fusion network according to claim 1, characterized in that: The training dataset includes actual reservoir seismic response data, low-frequency rock physics experimental observation data, or a mixture of both, and should be selected and used appropriately according to the specific conditions of the study area.

5. A method for predicting reservoir gas content based on feature extraction and fusion networks according to claim 1, characterized in that: The self-attention mechanism deep fusion network is a fusion network developed by combining a self-attention mechanism convolutional network and a self-attention mechanism long short-term memory network. The convolutional network is used to extract local morphological features of multi-parameter seismic data, the long short-term memory network mines the trend information of data changes with depth, and the self-attention mechanism is used to improve the network's attention to data-related features, thereby improving the accuracy of reservoir gas content prediction.

6. A reservoir gas content prediction system based on feature extraction and fusion network, characterized in that, The system includes a processor and a memory, the memory storing a computer program that can run on the processor, the processor executing the computer program to perform the following steps: Acquire well logging data, seismic data, and geological information for the study area; The acquired well logging data and seismic data are preprocessed; By comprehensively utilizing geological information and pre-processed well logging data and seismic data, a fine-scale seismic-geological integrated stratigraphic determination is performed. Based on the fine stratigraphic calibration, high-precision time-depth conversion is performed to convert time-domain seismic data to depth-domain seismic data; Seismic texture analysis, low-frequency shadowing analysis, and formation quality factor inversion are performed on the depth domain seismic data to extract multi-parameter seismic data features characterizing the reservoir, including seismic texture features, dispersion features, and formation quality factor. The target data is obtained through actual observation or forward simulation and combined with the extracted multi-parameter seismic data features to construct a dataset, which is then divided into a training dataset and a test dataset. Based on this, a deep fusion network model for predicting reservoir gas content using a self-attention mechanism is established, and the model is trained using the constructed training dataset to obtain a deep fusion network model for predicting reservoir gas content using a self-attention mechanism. Then, the trained reservoir gas content prediction self-attention mechanism deep fusion network model was tested using the constructed test dataset. Finally, the actual reservoir seismic response in the target area is used as input to predict the gas content distribution in the target area.

Images