A three-dimensional cus classification prediction method and system based on rfcn

Abstract

The application discloses a three-dimensional CUS classification prediction method and system based on RFCN, and the method comprises the following steps: S1, carbon utilization and storage CUS suitability classification model data analysis; S2, constructing a cyclic full convolutional neural network RFCN and optimizing parameters; S3, using the optimized cyclic full convolutional neural network RFCN to perform visual training and prediction on the carbon utilization and storage CUS suitability classification process, and completing prediction result analysis and reliability evaluation. The application can realize fast, quantitative and three-dimensional visual prediction of carbon utilization and storage CUS suitability classification.

Description

Technical Field

[0001] This invention relates to the field of carbon dioxide treatment technology, and in particular to a three-dimensional CUS classification and prediction method and system based on RFCN. Background Technology

[0002] CCUS (CO2 capture, utilization, and storage) is considered a powerful technology for achieving carbon neutrality. In the petroleum industry, its key application is CO2 flooding, which uses CO2 as a displacement agent to simultaneously displace crude oil from the formation into production wells and store CO2 within the formation, thus achieving CCUS. However, CO2 flooding technology requires comprehensive consideration of economic, social, and environmental factors. Therefore, before implementing this technology, a comprehensive characterization and modeling of the formation's CO2 storage and crude oil displacement capabilities is necessary to evaluate the economic, social, environmental, and safety benefits of using CO2 flooding technology in that formation. This allows for the development of an economical, environmentally friendly, efficient, and safe CO2 displacement scheme. Therefore, accurately and comprehensively evaluating the suitability of formation CO2 storage is the foundation and guarantee for the entire CO2 displacement technology.

[0003] Carbon utilization and storage (CUS) classification is essentially a classification and evaluation of CO2 flooding and storage suitability. Commonly used methods in this field include: 1) Utilizing reservoir and permeability parameters, rock mechanical parameters, physicochemical parameters, and structural characteristics of the target formation, employing fuzzy mathematics, empirical analysis, and cluster analysis to derive the parameter distribution under different suitability levels, forming a CUS suitability classification and evaluation table to model the CUS type of the entire target formation. 2) Using numerical simulation to simulate CO2 flooding, and directly obtaining the formation space CUS suitability classification and evaluation model based on the simulation results.

[0004] The current technologies mentioned above either use qualitative judgment or geological modeling or numerical simulation to evaluate the suitability of CUS. There is no evidence yet of using artificial intelligence to complete the three-dimensional quantitative prediction of CUS suitability classification. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a three-dimensional CUS classification and prediction method and system based on RFCN.

[0006] The objective of this invention is achieved through the following technical solution: a 3D CUS classification prediction method based on RFCN, comprising the following steps:

[0007] S1: Data analysis of the CUS suitability classification model for carbon utilization and storage;

[0008] S2: Construct a recurrent fully convolutional neural network (RFCN) and optimize its parameters;

[0009] S3: Utilize the optimized recurrent fully convolutional neural network RFCN to perform visualization training and prediction of the carbon utilization and storage (CUS) suitability classification process, and complete the prediction result analysis and reliability evaluation.

[0010] Furthermore, step S1 includes the following sub-steps:

[0011] S11: Continuity and adjacency analysis of the carbon utilization and storage CUS suitability classification model;

[0012] S12: Zoning analysis of the proportion of CUS suitability types for carbon utilization and storage.

[0013] Furthermore, step S11 specifically involves: analyzing the basic characteristics of the geometric shape, extension direction, extension length, extension width, thickness, and adjacency relationship of the carbon utilization and storage (CUS) suitability classification model, obtaining the overall distribution pattern of the CUS suitability classification type in the area to be predicted, and using this as the final evaluation of the reliability of the prediction results of the recurrent fully convolutional neural network (RFCN) for the CUS suitability classification model.

[0014] Furthermore, step S12 specifically involves: dividing the target carbon utilization and storage CUS suitability classification model into different sub-models according to different classification principles, statistically analyzing the proportion of carbon utilization and storage CUS suitability types in each sub-model, and forming a local proportion feature understanding of the carbon utilization and storage CUS suitability classification model; the classification principles include spatial dimension classification principles and geometric shape classification principles.

[0015] Furthermore, step S2 includes the following sub-steps:

[0016] S21: Construction of basic computational units for Recurrent Fully Convolutional Neural Networks (RFCN);

[0017] S22: Structure and parameter optimization of recurrent fully convolutional neural network (RFCN).

[0018] Furthermore, step S21 specifically involves: analyzing the computational logic of the fully convolutional neural network (FCN), establishing the relationship between network input and output parameters and features, analyzing the node structure of the recurrent neural network (RNN) learning data sequence features, combining the computational mathematical relationship of the FCN with the recurrent node structure of the RNN, and constructing the basic computational unit of the recurrent fully convolutional neural network (RFCN).

[0019] Furthermore, step S22 specifically involves: defining the adjustment direction of the RFCN structure and the adjustment range of each parameter based on the RFCN structure and important parameters, and optimizing to obtain a set of structure and parameter design schemes with better performance of the RFCN.

[0020] Furthermore, in step S3, the visualization training and prediction of the carbon utilization and storage (CUS) suitability classification process specifically involves: setting training-related parameters, visualizing the training process of the recurrent fully convolutional neural network (RFCN), analyzing the changing trend of parameters as training progresses, judging the generalization ability of the current parameters to predict unknown regions, and capturing the optimal combination of algorithm parameters.

[0021] Furthermore, in step S3, the prediction result analysis and reliability evaluation specifically involve: analyzing the distribution characteristics of each output image during the visualization training process of carbon utilization and storage CUS suitability classification, selecting the optimal distribution prediction result, and evaluating the reliability of the carbon utilization and storage CUS suitability classification prediction result.

[0022] A 3D CUS classification and prediction system based on RFCN includes a data analysis module, a recurrent fully convolutional neural network (RFCN) construction module, and a data prediction module. The data analysis module is used for data analysis of the carbon utilization and storage (CUS) suitability classification model. The RFCN construction module is used to construct a recurrent fully convolutional neural network (RFCN) and optimize its parameters. The data prediction module uses the optimized RFCN to perform visualization training and prediction of the CUS suitability classification process, and completes the prediction result analysis and reliability evaluation.

[0023] The beneficial effects of this invention are that it enables rapid, quantitative, and three-dimensional visualization prediction of carbon utilization and storage (CUS) suitability classification. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the present invention;

[0026] Figure 2 This is a system block diagram of the present invention;

[0027] Figure 3 An overview diagram of the CUS suitability classification model for case blocks;

[0028] Figure 4 A map showing the partitioning of the region using the model;

[0029] Figure 5 This is a partitioning diagram for Region 2 model;

[0030] Figure 6 This is a map showing the partitioning of the region using the three-model approach.

[0031] Figure 7 This is a map showing the partitioning of the region using the four-model approach.

[0032] Figures 8 to 17 This is a diagram illustrating the training process in the case study area.

[0033] Figure 18 This is the final prediction model diagram. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0035] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0036] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0037] Example 1:

[0038] See Figure 1 A 3D CUS classification prediction method based on RFCN includes the following steps:

[0039] S1: Data analysis of the CUS suitability classification model for carbon utilization and storage;

[0040] S2: Construct a recurrent fully convolutional neural network (RFCN) and optimize its parameters;

[0041] S3: Utilize the optimized recurrent fully convolutional neural network RFCN to perform visualization training and prediction of the carbon utilization and storage (CUS) suitability classification process, and complete the prediction result analysis and reliability evaluation.

[0042] The Recurrent Fully Convolutional Network (RFCN) combines the ability of recurrent neural networks to learn sequential features with the surface-to-surface computational mode of fully convolutional neural networks. It implements an algorithm trained on existing evaluation models for major oilfields, using known points in three-dimensional space (primarily CUS suitability classification models on individual wells) to derive the CUS suitability classification model for the entire space (primarily unknown regions between wells). Its high reusability and fast prediction computation speed mean that although the training process consumes a considerable amount of time, the overall computational efficiency of the algorithm is still higher than traditional modeling methods.

[0043] In this embodiment, step S1, CUS fitness classification model data analysis, involves analyzing the continuity of existing model data and the proportion of CUS fitness types, quantitatively determining the data characteristics of the CUS fitness model (geometric shape, extension direction, extension length, extension width, thickness, and adjacency relationship), providing a foundation for subsequent verification of the reliability of RFCN prediction results. Specifically, it includes the following sub-steps:

[0044] S11: Continuity and adjacency analysis of the carbon utilization and storage CUS suitability classification model;

[0045] S12: Zoning analysis of the proportion of CUS suitability types for carbon utilization and storage.

[0046] Furthermore, step S11 specifically involves: analyzing the basic characteristics of the geometric shape, extension direction, extension length, extension width, thickness, and adjacency relationship of the carbon utilization and storage (CUS) suitability classification model, obtaining the overall distribution pattern of the CUS suitability classification type in the area to be predicted, and using this as the final evaluation of the reliability of the prediction results of the recurrent fully convolutional neural network (RFCN) for the CUS suitability classification model.

[0047] See Figure 3Taking a block in eastern my country as an example, the entire model classifies CUS suitability into four categories: I (unsuitable for CUS operations), II (moderately suitable), III (relatively suitable), and IV (most suitable). Analysis of the model reveals that: Category II (moderately suitable) has the longest extension length, reaching approximately 2000m, with a minimum of 400m and an average of 1000m. The widest extension can exceed 1000m, with an average width of approximately 300m. The thickest extension can reach 100m. Most Category II areas extend in an east-west or north-south direction (positive direction), with a very small portion extending diagonally (southwest-northeast or southeast-northwest). The geometric shape is mainly sheet-like and irregularly shaped large areas, with most areas adjacent to Category I and a very small portion adjacent to Category III. Category III (relatively suitable) areas are smaller, with the main geometric shapes being slender strips and sheet-like structures, without large irregular areas. The elongated strip-shaped regions extend for over 4000m in length and are approximately tens of meters wide, with a thickness similar to their width. The extension direction is mostly circular, and their adjacency relationships are similar to Type II, meaning most areas are adjacent to Type I, with a very small portion adjacent to Type II. Type IV (the most suitable) regions have the smallest area, and their geometry in the model is mainly point-like, strip-like, and mat-like. Their length and width are within 800m, with most areas only about 200m in length and width. Their extension direction is similar to regions II and III. There are almost no continuous areas in terms of thickness, with most only a few meters thick.

[0048] Furthermore, step S12 specifically involves: dividing the target carbon utilization and storage CUS suitability classification model into different sub-models according to different classification principles, statistically analyzing the proportion of carbon utilization and storage CUS suitability types in each sub-model, and forming an understanding of the local proportion characteristics of the carbon utilization and storage CUS suitability classification model.

[0049] The classification methods include: 1) Classification based on spatial dimensions, such as dividing the model into sub-models with different thicknesses but equal planar areas according to the depth direction (one-dimensional), or dividing it into sub-models with different planar areas but equal thicknesses according to the plane (two-dimensional). 2) Classification based on geometric shapes, such as using geometric shapes like ellipsoids, spheres, and cylinders to deconstruct the CUS fitness classification model. Ellipsoids can examine the extension characteristics of different CUS fitness types in different directions, and cylinders can examine radial extension characteristics, etc.

[0050] In this embodiment, considering the characteristics of the RFCN network, a two-dimensional, planar partitioning method is used to divide the entire model, ensuring that the resulting sub-models have the same vertical thickness, i.e., the same depth range. The original model's geometric dimensions are: a 193×164 grid in the plane and 130 grids in the depth direction. The partitioning method involves dividing the plane into four approximately equal-sized rectangular regions using the midpoint of the width and height as the dividing line, thus dividing the original model into two 96×82×130 and two 97×82×130 sub-models (see [reference]). Figures 4 to 7 The statistical results of the type proportions of each sub-model and the overall model are shown in Table 1, which indicates that the type proportions of each sub-model are relatively close, and they have high similarity and representativeness.

[0051] Table 1. Statistical Results of Block Partition Type Proportion in Case Studies

[0052]

[0053] In this embodiment, step S2, RFCN network construction and structural parameter optimization, involves: analyzing the computational logic of a fully convolutional neural network (FCN), introducing the concept of recurrent neural networks (RNNs) for learning data sequence features, and constructing the basic computational unit of the RFCN network. Based on factors such as network hierarchy, computational unit size, and layer thickness, a structural and parameter optimization scheme is formulated, and a better combination of structure and parameters is determined using a trial training sample set. Specifically, this includes the following sub-steps:

[0054] S21: Construction of basic computational units for Recurrent Fully Convolutional Neural Networks (RFCN);

[0055] S22: Structure and parameter optimization of recurrent fully convolutional neural network (RFCN).

[0056] Furthermore, step S21 specifically involves: analyzing the computational logic of the fully convolutional neural network (FCN), establishing the relationship between network input and output parameters and features, analyzing the node structure of the recurrent neural network (RNN) learning data sequence features, combining the computational mathematical relationship of the FCN with the recurrent node structure of the RNN, and constructing the basic computational unit of the recurrent fully convolutional neural network (RFCN).

[0057] The FCN network is characterized by replacing the final fully connected layer of a traditional convolutional neural network with a deconvolutional layer, and adding an unpooling layer. Therefore, the network mainly consists of convolutional layers, pooling layers, deconvolutional layers, and unpooling layers. A pooling layer can be considered a special type of convolutional layer, and similarly, an unpooling layer can be considered a special type of deconvolutional layer. Compared to a fully connected layer, a deconvolutional layer takes one image as input and outputs one image simultaneously. Thus, the input and output of the FCN network are both images. The image sizes of the input and output of the deconvolutional layer are exactly the opposite of those of the corresponding convolutional layer. If a convolutional layer receives an image of width W and height H and outputs an image of width W' and height H', then the corresponding deconvolutional layer receives an image of width W' and height W' and outputs an image of width W and height H. Using matrix X∈R... H×W and O∈R H′×W′ Let F represent the two images respectively, and let matrix F ∈ R. M×N Let X represent a convolutional layer with width N and height M, and a stride of S. Then, the parameters above satisfy H' = (HM) / S + 1 and W' = (WN) / S + 1. Generally, the stride is 1. In this case, the value of the h-th row and w-th column of the output image X' during the convolution operation is:

[0058]

[0059] Where m, m', and h satisfy the relation m' = h + m - 1; n, n', and w satisfy the relation n' = w + n - 1.

[0060] The specific operation of deconvolution is as follows: fill the outer perimeter of X' with a constant (usually 0) to expand its shape and form an image with a larger width and height. Then, perform the above convolution operation on this image to obtain an image with the same shape as X, thus realizing the calculation logic from the input image of size H'×W' to the output image of size H×W.

[0061] RNNs are classic deep learning algorithms characterized by their ability to associate previous and current samples using a weight matrix, thus effectively learning the latent sequential features of the data. In common neural network algorithms such as Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), and Deep Belief Networks (DBNs), each sample is trained independently, and their training processes do not interfere with each other. Taking DNNs as an example, the formula for calculating the output of each neuron is:

[0062] O=f(W·x+θ) (2)

[0063] Where f represents various activation functions, W represents the weight parameters between layers, x represents the input vector, and θ represents the bias. When introducing the RNN concept, it is necessary to consider features in the time dimension, therefore x is used. (t) O (t)

[0064] x represents the input of the current sample and the output of the neuron. (t-1) O (t-1） Given the input of the previous sample and the corresponding output of the neuron, one way to compute an RNN is as follows:

[0065] O (t) =f(W·x) (t) +W'·O (t-1) +θ) (3)

[0066] In the formula, W' is the weight matrix introduced by RNN, which is used to combine the data features between two samples with a certain sequence relationship. Therefore, the order of the input data needs to have an inherent pattern rather than be in an arbitrary order. In the CUS fitness classification model, the CUS fitness type in the depth direction has such a sequence feature. Therefore, the entire CUS fitness classification model can be decomposed into layers of sub-models in the depth direction.

[0067] By introducing the concepts of RNN networks into FCN networks, the basic computational units of RFCN networks can be constructed. One feasible construction method is as follows: during each convolution or deconvolution operation, the output value of the previous sample in the current convolutional or deconvolutional layer is added to the calculation of the current sample's output value using a weight matrix. The calculation formula is as follows:

[0068]

[0069] The process involves summing the Hadamard product of the current input value X, the output value of the previous sample in the convolutional or deconvolutional layer, and the weight matrix W', and then performing a convolution or deconvolution operation on the result. Both the weight matrix W' and the convolution kernel F are learnable parameters of the network and are continuously updated as the network is trained.

[0070] For the summation operation in equation (4) to be meaningful, X needs to be combined with O. (t-1) O (t) They have the same shape. For convolutional and deconvolutional layers, a layer of zeros can be padded around the image, and 3×3 computational units can be used, so that the image size does not change after the convolution operation; for pooling and depooling layers, the output result can be uniformly padded according to the size of the input image.

[0071] Furthermore, step S22 specifically involves: defining the adjustment direction of the RFCN structure and the adjustment range of each parameter based on the RFCN structure and important parameters, and optimizing to obtain a set of structure and parameter design schemes with better performance of the RFCN.

[0072] The structure of RFCN mainly includes the number of convolutional layers and deconvolutional layers before each pooling. The parameters of RFCN mainly refer to the number and size of convolutional kernels, the number of convolutional layers, and the number of pooling layers. Currently, there is no theoretical framework in the field of deep learning that can quantitatively design neural network structures and hyperparameters. Therefore, researchers usually rely on empirical judgment or actual training with data to examine the impact of various structures or parameters on network performance. Based on past experience and to reduce testing time, two schemes are used for the convolutional kernel size: 3×3 and 5×5. Two schemes are used for the number of pooling layers: 2 or 3. Two schemes are used for the number of convolutional layers before pooling: one or two. The number of deconvolutional layers corresponds to the number of pooling layers. In practical applications, as many schemes as possible should be designed to obtain a more optimal design. Here, the design scheme is simplified for the sake of illustration. A small block from the case block is used as a subset of test samples, and the minimum loss function value during training is used as the comparison metric.

[0073] The test sample subset needs to have similar data distribution characteristics to the overall dataset to ensure that the test set can approximately reflect the performance of different network models on the overall dataset. For 3D attribute distribution models, local regions can be selected as the training set, with the selection principle being that the local model is close to the overall model in terms of attribute distribution and type ratio. In this case, the type ratios of the four local models are relatively close to the overall model, but in terms of distribution characteristics, the model in region one is the most similar, so the model in region one is selected as the test set.

[0074] The following table shows the comparison results of trial training for different design schemes in this case study area. The table lists the minimum loss function values ​​for different structures in the first 100 training rounds, using cross-entropy as the loss function. Information gleaned from this table for this case study includes: the number of convolutions before pooling has a significant impact on model performance; when there are two convolutions before pooling, the kernel size and the total number of pooling layers have a smaller impact on model performance, and the loss function values ​​are relatively close, indicating that these four sets of structural parameters have similar performance. To improve computational efficiency, two pooling layers and a 3×3 convolution kernel are used for subsequent training and prediction.

[0075] Table 2 Performance comparison of one convolution before pooling

[0076]

[0077] In this embodiment, step S3, CUS fitness classification prediction and reliability evaluation, involves training all sample sets using an algorithm with optimized structural parameters. The training process is visualized to analyze real-time changes in algorithm model parameters, accurately assessing model performance and obtaining a superior algorithm model. Based on this, for the target region to be predicted, the distribution of CUS fitness types in the unknown region is inferred from known data points, and the model reliability is analyzed and evaluated.

[0078] Furthermore, step S3 includes sub-step S31: Visualization training and prediction of the RFCN network model process, specifically: setting training-related parameters (training rounds, batch size, learning rate, etc.), visualizing the training process of the recurrent fully convolutional neural network RFCN, analyzing the changing trend of parameters as training progresses, judging the generalization ability of the current parameters to predict unknown regions, and capturing the optimal combination of algorithm parameters.

[0079] For practical CUS fitness classification problems, the model should be trained using as much data as possible to improve its generalization ability. In this embodiment, three of the four sub-models are used to train the network, and the distribution of data for the last model is predicted. The training runs for 200 epochs, and the lowest layer network surface of the predicted model is output every 20 epochs (see the training process diagram in the case study area). Figures 8 to 17 Each graph represents the network prediction results after every 20 rounds of training, allowing observation of changes in the network's generalization ability. The final prediction model can be found in [reference needed]. Figure 18 .

[0080] Furthermore, step S3 also includes sub-step S32: CUS fitness classification prediction result analysis and reliability evaluation. Specifically, this involves analyzing the distribution characteristics of each output image during the visualization training process of carbon utilization and storage CUS fitness classification, selecting the optimal distribution prediction result, and evaluating the reliability of the CUS fitness classification prediction results for carbon utilization and storage. The results show that the model initially exhibits strong randomness. As the training rounds increase, the CUS fitness types in the prediction results begin to show a certain trend. However, with further increases in training rounds, the algorithm model begins to show an "overfitting" trend, meaning that the CUS fitness types begin to become more homogenized. This indicates that the parameters and structure of the algorithm model have not yet reached the optimal solution. The algorithm has already demonstrated preliminary learning ability for the morphological features of various types. If the algorithm model's level is increased and more rounds of training are conducted, the algorithm can better learn the characteristics of different CUS fitness classification levels. Therefore, the RFCN network already possesses certain application value and the ability to quickly, quantitatively, and with three-dimensional visualization prediction capabilities for CUS fitness classification models.

[0081] See Figure 2Based on the same inventive concept, this invention also provides a 3D CUS classification and prediction system based on RFCN to implement the above-mentioned 3D CUS classification and prediction method based on RFCN. The system includes a data analysis module, a recurrent fully convolutional neural network (RFCN) construction module, and a data prediction module. The data analysis module is used for data analysis of the carbon utilization and storage CUS suitability classification model; the recurrent fully convolutional neural network (RFCN) construction module is used to construct a recurrent fully convolutional neural network (RFCN) and optimize its parameters; the data prediction module is used to use the optimized recurrent fully convolutional neural network (RFCN) to perform visualization training and prediction of the carbon utilization and storage CUS suitability classification process, and to complete the prediction result analysis and reliability evaluation.

[0082] This invention introduces deep learning algorithms into the three-dimensional quantitative prediction of CUS suitability classification, proposing a Recurrent Fully Convolutional Network (RFCN). This network model combines the ability of recurrent neural networks to learn sequence features with the surface-to-surface computational mode of fully convolutional neural networks. It achieves the derivation of the entire CUS suitability classification model (mainly the unknown region between wells) using known points in three-dimensional space (primarily the CUS suitability classification model on a single well) trained on existing evaluation models of major oilfields. Its high reusability and fast prediction calculation speed mean that although the training process consumes a considerable amount of time, the overall computational efficiency of the algorithm is still higher than that of traditional modeling methods. In summary, introducing deep learning algorithms into CUS suitability classification evaluation can improve the reliability and efficiency of the evaluation.

[0083] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to this application.

[0084] The above embodiments describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Modifications and variations made by those skilled in the art without departing from the spirit and scope of the invention should be within the protection scope of the appended claims.

Claims

1. A 3D CUS classification prediction method based on RFCN, characterized in that, Includes the following steps: S1: Data analysis of the CUS suitability classification model for carbon utilization and storage; S2: Construct a recurrent fully convolutional neural network (RFCN) and optimize its parameters; S3: Using the optimized recurrent fully convolutional neural network RFCN, perform visualization training and prediction of the carbon utilization and storage (CUS) suitability classification process, and complete the prediction result analysis and reliability evaluation. Step S1 includes the following sub-steps: S11: Continuity and adjacency analysis of the carbon utilization and storage CUS suitability classification model; S12: Zoning analysis of the proportion of CUS suitability types for carbon utilization and storage; The specific steps of step S11 are as follows: Analyze the basic characteristics of the geometric shape, extension direction, extension length, extension width, thickness and adjacency relationship of the carbon utilization and storage (CUS) suitability classification model, obtain the overall distribution pattern of the carbon utilization and storage (CUS) suitability classification type in the area to be predicted, and use this as the basis for the final evaluation of the reliability of the prediction results of the recurrent fully convolutional neural network (RFCN) for the carbon utilization and storage (CUS) suitability classification model. The specific steps of step S12 are as follows: the target carbon utilization and storage CUS suitability classification model is divided into different sub-models according to different classification principles, the proportion of carbon utilization and storage CUS suitability types in each sub-model is statistically analyzed, and the local proportion characteristics of the carbon utilization and storage CUS suitability classification model are understood. The division principles include spatial dimension division principles and geometric shape division principles. The spatial dimension division principle includes dividing the model into sub-models with different thicknesses but equal planar areas according to the depth direction, or dividing the model into sub-models with different planar areas but equal thicknesses according to the plane. The geometric shape division principle includes using geometric shapes to deconstruct the CUS fitness classification model. The geometric shapes include ellipsoids, spheres, and cylinders. Ellipsoids can examine the extension characteristics of different CUS fitness types in different directions, and cylinders can examine radial extension characteristics.

2. The 3D CUS classification prediction method based on RFCN as described in claim 1, characterized in that, Step S2 includes the following sub-steps: S21: Construction of basic computational units for Recurrent Fully Convolutional Neural Networks (RFCN); S22: Structure and parameter optimization of recurrent fully convolutional neural network (RFCN).

3. The 3D CUS classification prediction method based on RFCN as described in claim 2, characterized in that, Step S21 specifically involves: analyzing the computational logic of the fully convolutional neural network (FCN), establishing the relationship between network input and output parameters and features, analyzing the node structure of the recurrent neural network (RNN) learning data sequence features, combining the computational mathematical relationship of the fully convolutional neural network (FCN) with the recurrent node structure of the recurrent neural network (RNN), and constructing the basic computational unit of the recurrent fully convolutional neural network (RFCN).

4. The 3D CUS classification prediction method based on RFCN as described in claim 2, characterized in that, Specifically, step S22 involves: defining the adjustment direction of the RFCN structure and the adjustment range of each parameter based on the RFCN structure and important parameters, and optimizing to obtain a set of structure and parameter design schemes with better performance of the RFCN.

5. The 3D CUS classification prediction method based on RFCN as described in claim 1, characterized in that, In step S3, the visualization training and prediction of the carbon utilization and storage (CUS) suitability classification process specifically involves: setting training-related parameters, visualizing the training process of the recurrent fully convolutional neural network (RFCN), analyzing the changing trend of parameters as training progresses, judging the generalization ability of the current parameters to predict unknown regions, and capturing the optimal combination of algorithm parameters.

6. The 3D CUS classification prediction method based on RFCN as described in claim 1, characterized in that, In step S3, the prediction result analysis and reliability evaluation specifically involve: analyzing the distribution characteristics of each output image during the visualization training process of carbon utilization and storage CUS suitability classification, selecting the optimal distribution prediction result, and evaluating the reliability of the carbon utilization and storage CUS suitability classification prediction result.

7. A 3D CUS classification and prediction system based on RFCN, used to implement the 3D CUS classification and prediction method based on RFCN as described in any one of claims 1 to 6, characterized in that, The system includes a data analysis module, a recurrent fully convolutional neural network (RFCN) construction module, and a data prediction module. The data analysis module is used for data analysis of the carbon utilization and storage (CUS) suitability classification model. The RFCN construction module is used to construct a recurrent fully convolutional neural network (RFCN) and optimize its parameters. The data prediction module is used to perform visualization training and prediction of the CUS suitability classification process using the optimized RFCN, and to complete the prediction result analysis and reliability evaluation.

Images