Method for predicting co2 flooding oil conversion in low permeability reservoirs with hypergraph multi-field space-time high-order coupling

Abstract

The application relates to the technical field of oil reservoir numerical simulation, and discloses a supergraph multi-field space-time high-order coupling low-permeability oil reservoir CO2 flooding oil conversion prediction method, which comprises the following steps: constructing a model input sample; constructing a node feature sequence of supergraph modeling; constructing a node-superedge super-association structure based on the node feature sequence; inputting the node-superedge super-association structure into a supergraph high-order coupling representation layer to obtain superedge features and spatially coordinated evolution features; inputting the model input sample into a time auxiliary layer to extract time evolution features of a physical field dynamic response process; taking dynamic physical field data at the last historical moment as a baseline item, inputting the spatially coordinated evolution features and the time evolution features into a feature fusion output layer for weighted fusion to obtain a target moment CO2 conversion related multi-field three-dimensional dynamic physical field prediction, accurate prediction of a physical field distribution state is realized, and conversion behavior of CO2 in an oil reservoir is characterized and analyzed based on the prediction result.

Description

Technical Field

[0001] This application relates to the field of reservoir numerical simulation technology, and to a method for predicting CO2-driven oil recovery conversion in low-permeability reservoirs using a high-order spatiotemporal coupling method with a supermap multi-field. Background Technology

[0002] Low-permeability reservoirs are an important type of oil and gas resource, but their efficient development has always faced significant challenges due to their low permeability, strong heterogeneity, and complex flow channels. CO2 flooding, as an important technology for enhancing oil recovery, can improve development results by reducing crude oil viscosity, improving crude oil flowability, and enhancing displacement capacity. It also possesses the potential for CO2 resource utilization and geological storage, thus holding significant theoretical and engineering value in the development of low-permeability reservoirs. However, the CO2 flooding process in low-permeability reservoirs is more complex. This process involves not only pressure propagation, multiphase flow, and gas channeling, but also significant CO2 component redistribution and phase transformation behaviors. Simultaneously, the interaction and mutual constraints between various physical fields, such as pressure, saturation, and component fields, exhibit significant multi-field coupling characteristics, making the dynamic process of CO2 flooding in low-permeability reservoirs highly complex.

[0003] The related technology discloses a numerical simulation method considering the time-varying seepage of CO2 gas drive in oil reservoirs. By obtaining two types of functional relationships—viscosity and residual liquid saturation, and residual gas phase relative permeability—a numerical simulation model of CO2 gas drive in oil reservoirs is established and meshed, with an initial formation oil viscosity assigned to the mesh. The crude oil viscosity at the next time step is obtained through numerical simulation, and the corresponding residual liquid saturation and residual gas phase relative permeability are determined and updated. This iteration is repeated until the simulation ends, and the cumulative oil production is output, thus achieving a numerical simulation of CO2 gas drive considering the time-varying seepage. However, this method has certain limitations: first, the mesh parameter homogenization process affects the simulation accuracy; second, it cannot fully capture the coupled time-varying seepage characteristics; and third, it relies on traditional numerical simulation iterations, making it difficult to meet the needs of rapid optimization of development schemes.

[0004] Furthermore, data-driven methods such as convolutional neural networks, recurrent neural networks, and graph neural networks provide new technical paths for rapid proxy modeling of complex reservoir processes, breaking through the efficiency bottleneck of traditional numerical simulation. However, their current applications still have several prominent shortcomings, specifically manifested in the following three aspects: First, at the prediction dimension level, most methods mainly focus on predicting single physical quantities or a few key variables, failing to fully capture the dynamic characteristics of the mutual coupling and influence of multiple physical fields. This results in a lack of comprehensiveness in the prediction results, making it difficult to support the scientific evaluation of multi-objective development schemes. Second, at the scene adaptation level, many methods remain at the shallow modeling level of simplified two-dimensional scenes or simple feature stacking, making it difficult to accurately adapt to the complex laws of multi-field dynamic evolution in three-dimensional space. The modeling accuracy is insufficiently aligned with the actual reservoir scene, limiting the method's field applicability. Third, at the relationship modeling level, most existing models focus on local neighborhood feature extraction or simple relationship modeling between pairs of grid cells or physical quantities, making it difficult to accurately characterize such complex relationships, thus affecting prediction accuracy and generalization ability.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0007] This disclosure provides a method for predicting CO2 flooding transformation in low-permeability reservoirs using a hypergraph-based multi-field spatiotemporal high-order coupling, in order to accurately predict the CO2 transformation behavior and the distribution of related multi-field three-dimensional dynamic physical fields during CO2 flooding in low-permeability reservoirs.

[0008] In some embodiments, the hypergraph multi-field spatiotemporal high-order coupling method for predicting CO2 flooding transformation in low-permeability reservoirs includes: S10, acquiring dynamic physical field data and static geological attribute data generated by numerical simulation of CO2 flooding in low-permeability reservoirs, and preprocessing the dynamic physical field data and the static geological attribute data to construct model input samples; S20, inputting the model input samples into a feature encoding layer, performing local three-dimensional convolutional encoding on the model input samples to obtain node feature sequences for hypergraph modeling; S30, inputting the node feature sequences into an adaptive hyperedge generation layer, generating hyperedge prototypes through an adaptive hyperedge generation mechanism based on the context information of the node feature sequences, and establishing hyper-association relationships between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequences and each hyperedge prototype, thereby constructing a node-hyperedge hyper-association structure; S40, inputting the node-hyperedge hyper-association structure into the hypergraph high-order coupling layer. The representation layer aggregates the feature vectors of multiple nodes associated with the same hyperedge to obtain hyperedge features, and backfeeds the hyperedge features to each node to obtain spatial co-evolution features; S50, the model input samples are input to the time auxiliary layer and rearranged according to the grid cell input. For each grid cell, a time series composed of continuous historical time steps is constructed for each dynamic physical field data. The time series is input to a gated recurrent unit (GRU) for temporal encoding to extract the temporal evolution features of each grid cell on different dynamic physical field data; S60, the dynamic physical field data of the last historical moment is used as the baseline term. The spatial co-evolution features and the temporal evolution features are input to the feature fusion output layer. The feature fusion output layer generates spatial branch prediction results and temporal branch prediction results. The baseline term, the spatial branch prediction results, and the temporal branch prediction results are weighted and fused to obtain the CO2 conversion related multi-field three-dimensional dynamic physical field prediction results at the target time.

[0009] In some embodiments, the hypergraph multi-field spatiotemporal high-order coupling low-permeability reservoir CO2 flooding transformation prediction system includes: a data preprocessing layer, configured to acquire dynamic physical field data and static geological attribute data generated by numerical simulation of CO2 flooding in low-permeability reservoirs, and preprocess the dynamic physical field data and the static geological attribute data to construct model input samples; a feature encoding layer, configured to input the model input samples into the feature encoding layer, perform local three-dimensional convolutional encoding on the model input samples, and obtain node feature sequences for hypergraph modeling; an adaptive hyperedge generation layer, configured to input the node feature sequences into the adaptive hyperedge generation layer, and based on the context information of the node feature sequences, the adaptive hyperedge generation layer generates hyperedge prototypes through an adaptive hyperedge generation mechanism, and establishes hyper-association relationships between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequences and each hyperedge prototype, thereby constructing a node-hyperedge hyper-association structure; and a hypergraph high-order coupling representation layer, configured to represent the node-hyperedge hyper-association structure. The input is fed into a high-order coupling representation layer of the hypergraph, where the feature vectors of multiple nodes associated with the same hyperedge are aggregated to obtain hyperedge features. These hyperedge features are then fed back to each node to obtain spatial co-evolution features. A temporal auxiliary layer is configured to input the model input samples and rearrange them according to grid cell input. For each grid cell's dynamic physics data, a time series consisting of consecutive historical time steps is constructed. This time series is then input into a gated recurrent unit (GRU) for temporal encoding to extract the temporal evolution features of each grid cell on different dynamic physics data. A feature fusion output layer is configured to use the dynamic physics data of the last historical moment as a baseline term. The spatial co-evolution features and temporal evolution features are input into the feature fusion output layer, which generates spatial branch prediction results and temporal branch prediction results. The baseline term, spatial branch prediction results, and temporal branch prediction results are then weighted and fused to obtain the CO2 conversion-related multi-field three-dimensional dynamic physics prediction results for the target time.

[0010] The method for predicting CO2-induced oil displacement transformation in low-permeability reservoirs through hypergraph multi-field spatiotemporal high-order coupling provided in this disclosure can achieve the following technical effects: 1. This invention constructs an adaptive hyperedge generation layer and a hypergraph high-order coupling representation layer, and uses a node-hyperedge hyperassociation structure to uniformly model the node association relationships encoded by multiple grid cells. It can effectively characterize the multi-field high-order coupling relationships formed by reservoir heterogeneity, pressure propagation, phase saturation changes, fluid property changes and CO2 component distribution during CO2 flooding in low-permeability reservoirs. It solves the technical problem that existing methods mainly rely on local neighborhood feature extraction or pairwise relationship modeling, and are difficult to characterize complex nonlocal high-order coupling structures. 2. This invention incorporates the distribution characteristics of CO2 in different phase media into a unified prediction framework, and combines relevant multi-physics field information such as pressure field, saturation field, viscosity field, gas-oil interfacial tension field and CO2 component field to achieve joint prediction of dynamic responses related to CO2 component redistribution and phase transformation during CO2 flooding in low-permeability reservoirs, thereby improving the characterization ability of CO2 transformation process. 3. This invention integrates the high-order spatial coupling characteristics of the supergraph, the temporal evolution characteristics of GRU, and the dynamic physical field baseline term of the last historical moment. It can predict the three-dimensional dynamic physical field distribution state of CO2 transformation related to the target moment while maintaining the grid cell distribution structure. This helps to improve the accuracy and stability of dynamic prediction of CO2 flooding in low-permeability reservoirs and provides technical support for the optimization of CO2 flooding schemes and the evaluation of development effects.

[0011] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0012] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a schematic diagram of the CO2 flooding conversion prediction method for low-permeability reservoirs provided in the embodiments of this disclosure, which involves multi-field spatiotemporal high-order coupling of supergraphs. Figure 2 This is a schematic diagram of the framework of the supergraph multi-field spatiotemporal high-order coupling low-permeability reservoir CO2 flooding conversion prediction system provided in the embodiments of this disclosure; Figure 3 This is a schematic diagram of the interior of the adaptive hyperedge generation layer and the hypergraph higher-order coupling representation layer provided in the embodiments of this disclosure; Figure 4 This is a comparison chart of the actual and predicted values ​​of the gas phase saturation field in 3D provided by the embodiments of this disclosure; Figure 5 This is a comparison chart of the actual and predicted values ​​of the oil phase saturation field in 3D provided in the embodiments of this disclosure; Figure 6 This is a comparison chart of the actual and predicted values ​​of the CO2 gas phase component field in the embodiments of this disclosure; Figure 7 This is a comparison chart of the actual and predicted values ​​of CO2 oil phase components in the 3D field provided in the embodiments of this disclosure; Figure 8 This is a comparison chart of the actual and predicted values ​​of the physical field of CO2 aqueous phase components provided in the embodiments of this disclosure; Figure 9This is a comparison chart of the actual and predicted values ​​of the cross-section of the gas phase saturation field provided in this embodiment of the disclosure; Figure 10 This is a comparison chart of the actual and predicted values ​​of the cross-section of the oil phase saturation field provided in the embodiments of this disclosure; Figure 11 This is a comparison chart of the actual and predicted values ​​of the CO2 gas phase component field cross section provided in the embodiments of this disclosure; Figure 12 This is a comparison chart of the actual and predicted values ​​of the CO2 oil phase component field cross section provided in the embodiments of this disclosure; Figure 13 This is a comparison chart of the actual and predicted values ​​of the CO2 aqueous phase component field cross section provided in the embodiments of this disclosure; Figure 14 This is a comparison chart of the actual and predicted values ​​of the longitudinal section of the gas phase saturation field provided in this embodiment of the disclosure; Figure 15 This is a comparison chart of the actual and predicted values ​​of the longitudinal section of the oil phase saturation field provided in this embodiment of the disclosure; Figure 16 This is a comparison chart of the actual and predicted values ​​of the longitudinal section of the CO2 gas phase component field provided in this embodiment of the disclosure; Figure 17 This is a comparison chart of the actual and predicted values ​​of the CO2 oil phase component field longitudinal section provided in the embodiments of this disclosure; Figure 18 This is a comparison chart of the actual and predicted values ​​of the CO2 aqueous phase component field longitudinal section provided in the embodiments of this disclosure. Detailed Implementation

[0013] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0014] The terms "first," "second," etc., used in the specification and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0015] Unless otherwise stated, the term "multiple" means two or more.

[0016] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0017] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0018] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0019] Combination Figure 1 As shown, this disclosure provides a method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using a hypergraph-based multi-field spatiotemporal high-order coupling, comprising: S10: Obtain dynamic physical field data and static geological attribute data generated from the numerical simulation of CO2 flooding in low-permeability oil reservoirs, and preprocess the dynamic physical field data and static geological attribute data to construct model input samples. Specifically, this includes the following steps: S11, acquire dynamic physical field data and static geological attribute data generated by numerical simulation of CO2 flooding in low-permeability oil reservoirs.

[0020] The dynamic physical field data includes: pressure field, gas phase saturation field, oil phase saturation field, water phase saturation field, gas phase viscosity field, oil phase viscosity field, gas-oil interfacial tension field, CO2 component field in oil phase, CO2 component field in water phase, CO2 component field in gas phase, and global CO2 component field. Among these, the CO2 component fields in oil phase, water phase, gas phase, and global CO2 component field are used to characterize the spatial distribution, interphase allocation, and state of CO2 in different phase media. The gas phase saturation field, oil phase saturation field, water phase saturation field, and gas-oil interfacial tension field are used to characterize phase changes, interphase interactions, and fluid transport responses during CO2 displacement. Therefore, by using the aforementioned CO2 component field along with the pressure field, saturation field, viscosity field, and gas-oil interfacial tension field as input to the dynamic physical field data, the model can simultaneously learn the multi-field coupling characteristics related to CO2 component redistribution and phase transformation.

[0021] Static geological property data include porosity and permeability.

[0022] Assume a three-dimensional grid system for low-permeability reservoirs has a total of If there are 1, 2, 3 grid cells, then: ,in, These represent the three-dimensional mesh in... Maximum number of grid cells in three directions.

[0023] For the At each time step, the dynamic physical field data can be represented as: ,in, This represents the amount of dynamic physical field data, in this embodiment. ; , indicating the first The dynamic physical field data in the first Each time step corresponds to the value vector of all grid cells.

[0024] Static geological attribute data can be represented as: ,in, This represents the quantity of static geological attribute data, as shown in this embodiment. ; , indicating the first The value vector of a static geological attribute data across all grid cells.

[0025] S12, based on reservoir grid units Coordinates are used as the basis for positioning, mapping each physical field data to a unified grid cell. For any grid cell, its spatial index can be denoted as: .

[0026] The three-dimensional index is mapped through a one-to-one relationship. Convert to linear node index ,For example: , For any dynamic physical field data Its mapping result in the grid cell can be expressed as: , , in, Indicates the first The dynamic physical field data in the first The corresponding grid cell at each time step The value of .

[0027] Similarly, for any static geological attribute data The mapping result can be expressed as: , , in, Indicates the first Static geological attribute data in grid cells The value of .

[0028] Through the aforementioned spatial index mapping, dynamic physical field data and static geological attribute data at different time steps can be uniformly organized into the same grid cell, enabling subsequent models to carry out feature encoding, hypergraph modeling, and time modeling using the grid cell as a unified basic unit.

[0029] S13. Perform data preprocessing on the dynamic physical field data and static geological attribute data after spatial index mapping is completed. The preprocessing includes data cleaning, time step alignment and normalization.

[0030] Data cleaning is used to remove outliers, missing values, and invalid samples. If a certain physical field is in a grid cell... If outliers are found, neighborhood interpolation can be used for correction.

[0031] Time step alignment is used to ensure that different physical fields have consistent data representation under the same time index.

[0032] Normalization is used to reduce the impact of differences in dimensions and value ranges between different physical quantities on model training. For arbitrary dynamic physical field data and static geological attribute data, min-max normalization is used.

[0033] Through the above processing, different dynamic physical field data and static geological attribute data can be placed within a relatively uniform numerical range before entering the model.

[0034] S14, the first The dynamic physics data at each time step are constructed into dynamic physics feature tensors corresponding to each grid cell. The static geological attribute data is then constructed into static geological attribute tensors corresponding to each grid cell. .

[0035] In this embodiment, the first At each time step, the grid cell The corresponding dynamic physics data values ​​are concatenated along the channel dimension to form the dynamic physics characteristic tensor. Then we have: , ,in, The amount of dynamic physical field data; Indicates the first The dynamic physical field data in the first The corresponding grid cell at each time step The value of .

[0036] Simultaneously, static geological attribute data are constructed into static geological attribute tensors. : , ,in, This indicates the quantity of static geological attribute data; Indicates the first Static geological attribute data in grid cells The value of .

[0037] Static geological property tensor This is used to characterize the distribution of static reservoir properties within grid cells. Since static geological property data does not change over time, the static geological property tensor can be used as input samples when constructing subsequent models. The data is replicated and expanded along the time dimension to serve as auxiliary input information for each historical time step, rather than as the final prediction object.

[0038] S15, select the dynamic physical field feature tensor of 5 consecutive historical time steps as the temporal dynamic input, that is: .

[0039] Subsequently, the dynamic physical field feature tensor of each historical time step is... With static geological property tensor The data is stitched together along the channel dimension to form the model input samples corresponding to the sliding time window. : , , , , ,in, for Historical time step multi-field three-dimensional feature tensor; This represents a splicing operation on the feature channel dimension; The amount of dynamic physical field data; This indicates the quantity of static geological attribute data.

[0040] In this way, the model receives dynamic physical field information and static geological attribute information at each historical time step, thereby enabling it to jointly learn the dynamic evolution laws and geological constraint characteristics of multiple fields.

[0041] S16, target time The dynamic physical field data is used to construct output labels with the dynamic physical field data as the prediction object, and forms supervised learning sample pairs. Let the target time be... The dynamic physical field characteristic tensor is The output label is then defined as: , , That is, the output labels only contain the distribution of dynamic physical field data in the grid cells at the target time. Static geological attribute data, since they do not change over time, are mainly used as model input constraint information in this invention, rather than as the final prediction target.

[0042] Therefore, supervised learning sample pairs can be represented as: ,in, This represents the model input samples consisting of five consecutive historical time steps. This represents the output label for the dynamic physics data at the next time step. In practical applications, the supervised learning sample set can be further divided into training, validation, and test sets for model training, parameter tuning, and performance evaluation. Through the above steps, the process of constructing a model from raw data of CO2 flooding numerical simulation in low-permeability reservoirs to usable input samples and dynamic physics prediction labels is realized.

[0043] S20: Input the model input samples into the feature encoding layer, perform local 3D convolutional encoding on the model input samples, and obtain the node feature sequence used for hypergraph modeling. Specifically, this includes the following steps: S21, Input samples to the model Decompose the data by time step to obtain the single-time multi-field three-dimensional feature tensor corresponding to each historical time step. .

[0044] For grid cells In its first The single-time multi-field input feature vector at each time step can be represented as: ,in, Indicates the first The time step Dynamic physics data in grid cells The value at; Indicates the first Static geological attribute data in this grid cell The value at that location. Therefore, the single-time multi-field three-dimensional feature tensor The single-time multi-field input feature vector corresponding to each grid cell The grid is arranged according to a three-dimensional spatial index, where each spatial index position corresponds to a grid cell, and the single-time multi-field input feature vector corresponding to that grid cell is stored at the spatial index position.

[0045] S22, the single-moment multi-field three-dimensional feature tensor corresponding to each historical time step. Local 3D convolutional encoding is performed separately, and local representation information of each encoded grid cell and its neighborhood local spatial region is extracted by downsampling to obtain the corresponding local spatial encoding features. : , ,in, Indicates the first Layer 3D convolutional encoding features; Represents a non-linear activation function; They represent the first Layer 3D convolution kernel parameters and bias terms.

[0046] In one embodiment, local spatial coding features are obtained through downsampling: , ,in, This indicates the dimensions of the encoded grid cells in three directions; Indicates the number of encoding channels.

[0047] Through the aforementioned local three-dimensional convolutional coding, the local spatial response characteristics of dynamic physical fields such as pressure field, saturation field, viscosity field, gas-oil interfacial tension field and CO2 component field within the grid cell and its neighborhood can be extracted. At the same time, the constraint effect of static geological attribute data such as porosity and permeability on local characterization is introduced.

[0048] S23, Encoding the local spatial features corresponding to each historical time step after convolutional encoding. Perform unified mapping and feature alignment to form multi-field fusion representation features corresponding to each historical time step. .

[0049] ,,in, Indicates the first Local spatial coding features corresponding to historical time steps; This is a mapping function.

[0050] In another embodiment, feature alignment can also be achieved using a linear mapping method: , ,in, These are learnable parameters; This indicates the node feature dimension required for subsequent hypergraph modeling.

[0051] S24 represents the multi-field fusion characterization features corresponding to each historical time step. The encoded grid cells are unfolded into a node feature sequence. The node feature sequence is then used as the input to the subsequent adaptive hyperedge generation layer.

[0052] Grid cell index With node number The following mapping relationship can be established between them: , , in, and These represent the position indices of the encoded grid cells in three directions; This indicates the dimensions of the encoded grid cells in three directions.

[0053] Then for the first Each historical time step can be represented by the node feature sequence as follows: , , , in, This represents the number of grid cells after encoding, with each grid cell corresponding to one hypergraph node. It also represents the number of nodes after encoding; Indicates the first This historical step is the first The node feature vector of each node.

[0054] Through the above processing, the multi-field fusion representation features in the form of a three-dimensional mesh can be transformed into a node feature sequence suitable for hypergraph modeling, so that each encoded mesh unit corresponds to a node, and this node feature sequence serves as the input to the subsequent adaptive hyperedge generation layer.

[0055] S30, the node feature sequence is input into the adaptive hyperedge generation layer. Based on the context information of the node feature sequence, the adaptive hyperedge generation layer generates hyperedge prototypes through the adaptive hyperedge generation mechanism, and establishes hyper-association relationships between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequence and each hyperedge prototype, thereby constructing a node-hyperedge hyper-association structure. Specifically, this includes the following steps: S31, input the node feature sequence into the adaptive hyperedge generation layer, and generate a context representation vector based on the global context information of the node feature sequence. .

[0056] In this embodiment, the node feature sequence obtained in step S20 is represented as follows: , ,in, Indicates the number of nodes after encoding; Indicates the first This historical step is the first The node feature vector of each node; Represents the dimension of node features.

[0057] Global context information consists of the statistical features of node feature sequences along the node dimension.

[0058] In one embodiment, the node feature sequence Perform mean statistics to obtain mean context features. That is, the mean information: ,in, Indicates the first This historical step is the first The node feature vector of each node.

[0059] In another embodiment, for node feature sequences Perform maximum value statistics to obtain the maximum value context features. That is, the maximum value information: ,in, Indicates the first This historical step is the first The node feature vectors of each node are statistically analyzed by taking the maximum value of each feature channel along the node dimension. This value is used to characterize the local region features that have a significant response at the current historical time step.

[0060] In a preferred embodiment, the mean context feature and the maximum context feature are concatenated, and a context representation vector is generated by a mapping function. That is, combined information: ,in, Indicates feature concatenation operation; This represents a linear or nonlinear mapping function.

[0061] S32, the context representation vector The input is fed into an adaptive hyperedge generation layer, which performs context modulation on a preset number of basic hyperedge prototypes to generate a set of hyperedge prototypes corresponding to the current input sample. .

[0062] In this embodiment, the preset basic hyperedge prototype set is set as follows: ,in, Indicates the number of superedges; Indicates the first A basic hyperedge prototype.

[0063] Generate the first The context modulation amount corresponding to each basic hyperedge prototype: ,in, It is a context representation vector; Indicates the relationship with the first Mapping functions corresponding to the basic hyperedge prototypes.

[0064] The basic hyperedge prototype With context modulation amount By merging, we obtain the first... This historical step is the first One hyperedge prototype: .

[0065] Thus, we obtain the set of hyperedge prototypes corresponding to the current historical time step: ,in, Indicates the first This historical step is the first A hyperedge prototype; This indicates the number of superedges.

[0066] S33. Extract the node feature vector of each node in the node feature sequence. Based on the association strength between the node feature vector corresponding to each node in the node feature sequence and each hyperedge prototype, calculate the association score from the node to the hyperedge, and normalize the association score to obtain the node-hyperedge assignment matrix. .

[0067] The correlation score can be represented as: ,in, A function representing the correlation between node feature vectors and hyperedge prototypes; Indicates the first The first step in history Each node's feature vector; Indicates the first The first step in history A hyper-edge prototype.

[0068] In one embodiment, the correlation calculation function can take the form of dot product similarity: .

[0069] In another embodiment, the correlation calculation function can take the form of a bilinear function with a learnable mapping: ,in, It is a learnable parameter matrix used to map node feature vectors and hyperedge prototypes to a unified correlation calculation space.

[0070] Furthermore, regarding the first The association scores from each node to all hyperedges are normalized to obtain the node-hyperedge assignment matrix: , , where matrix elements Indicates the first The node is assigned to the first The association weights of each hyperedge; Indicates the number of superedges; This indicates the number of nodes after encoding.

[0071] Therefore, for any node The association weights between it and all hyperedges satisfy: , .

[0072] S34, Based on node-hyperedge allocation matrix Establish hyper-association relationships between nodes and hyperedges, and organize them in a time dimension to construct a node-hyperedge hyper-association structure corresponding to each historical time step. .

[0073] In this embodiment, the node-hyperedge hyperassociation structure is represented as follows: ,in, Indicates the first Node feature sequence at each historical time step; Indicates the first A set of hyperedge prototypes adaptively generated at each historical time step; Represents the node-hyperedge assignment matrix.

[0074] S35. Organize the node-hyperedge-hyperassociation structure corresponding to continuous historical time steps in the time dimension, and use the result of this time dimension organization as the input of the subsequent hypergraph higher-order coupling representation layer.

[0075] In this embodiment, since each historical time step corresponds to a different dynamic physical field state, the context representation vector corresponding to each time step is... Hyperedge prototype set Node-hyperedge assignment matrix and node-hyperedge-hyperassociation structure All of them can be adaptively updated as the time step changes.

[0076] S40: The node-hyperedge hyperassociation structure is input into the higher-order coupling representation layer of the hypergraph. Feature vectors of multiple nodes associated with the same hyperedge are aggregated to obtain hyperedge features. These hyperedge features are then backpropagated to each node to obtain spatial co-evolutionary features. Specifically, the steps include: S41, the node-hyperedge hyperassociation structure The input is fed into the higher-order coupling representation layer of the hypergraph. Based on the node-hyperedge assignment matrix that represents the relationship between nodes and hyperedges in the node-hyperedge association structure, the contribution weight of each node's feature vector to each hyperedge is determined.

[0077] For the The node and the first Each hyperedge is assigned using a node-hyperedge distribution matrix. Determine the first The node is the first Contribution weight of each superedge: , in, Indicates the first The node is the first The contribution weight of each superedge; The node-hyperedge assignment matrix represents the node-hyperedge assignment matrix. The node is assigned to the first The association weights of each hyperedge; Indicates the number of nodes after encoding; This represents a constant used to prevent the denominator from being zero.

[0078] S42, for any hyperedge Based on the contribution weights of each node, the feature vectors of multiple nodes associated with the hyperedge are weighted and aggregated to obtain the first... The initial hyperedge features corresponding to each hyperedge : , in, Indicates the first The first step in history Each node's feature vector.

[0079] S43, normalize and / or nonlinearly map the initial hyperedge features to obtain the first... The hyperedge features corresponding to each hyperedge : , in, Indicates the first The initial hyperedge features corresponding to each hyperedge; This indicates normalization processing; and These represent the learnable weight matrix and bias term, respectively; This represents a non-linear activation function.

[0080] Furthermore, for all hyperedges, we can obtain the hyperedge feature set: , This indicates the number of superedges.

[0081] S44, the hyperedge feature obtained from S43 A mapping transformation is performed to obtain an updated hyperedge representation, thereby enhancing the ability of the hyperedge to express higher-order relationships of multiple mesh cells and various physical quantities.

[0082] In this embodiment, for the first The updated hyperedge representation is obtained by performing a nonlinear mapping on each hyperedge feature. : ,in, These are learnable parameters; It is a non-linear activation function; Indicates the first The hyperedge in the th... Each historical time step yields the corresponding hyperedge features.

[0083] In a preferred embodiment, a feedforward mapping structure can be used to enhance the expressive power of hyperedge features: ,in, All of these are learnable parameters.

[0084] Furthermore, the hyperedge prototype can also be... Hyperedge features obtained by aggregation To preserve the prior structural information obtained from the adaptive hyperedge generation layer, the following steps are performed: ,in, Indicates feature concatenation operation; These are learnable parameters.

[0085] For all hyperedges, the updated set of hyperedge representations can be expressed as: .

[0086] S45, Based on Node-Hyperedge Allocation Matrix The updated hyperedge representation The data is transmitted back to each node to obtain the higher-order coupling propagation characteristics received by the nodes. .

[0087] For the The higher-order coupling propagation characteristics received by a node from each hyperedge can be expressed as follows: , where matrix elements Indicates the first The node is assigned to the first The association weights of each hyperedge; This represents the updated hyperedge representation.

[0088] S46, node feature vector Propagation characteristics with higher-order coupling Perform residual update to obtain the updated node feature vector. .

[0089] Specifically, for the first The residual update process for each node is represented as follows: .

[0090] In a preferred embodiment, to control the influence of higher-order coupling propagation features on node updates, learnable gating coefficients can be introduced. The calculation formula is as follows: ,in, These are learnable parameters.

[0091] Accordingly, the updated node features can be represented as: ,in, This indicates element-wise multiplication.

[0092] S47, update the node feature vector Feedforward mapping and feature normalization are performed, and temporal organization is applied to obtain the spatial co-evolutionary features corresponding to each historical time step. .

[0093] In this embodiment, for the first The updated features of each node are then subjected to feedforward mapping and normalization: , ,in, These are learnable parameters; This indicates a feature normalization operation.

[0094] By organizing the spatial co-evolutionary features of all nodes, we obtain the spatial co-evolutionary features of the nodes: .

[0095] S48, organizing the spatial co-evolutionary features corresponding to consecutive historical time steps in a temporal dimension, yields: This will be used as a multi-field spatial collaborative feature sequence required for subsequent spatial branching and temporal fusion.

[0096] Since the CO2 phase transformation process is usually not determined independently by a single grid cell or a single physical field, but is jointly affected by pressure propagation, phase saturation changes, interfacial tension changes, viscosity changes, and the component distribution of CO2 in different phases, high-order aggregation of nodes encoded by multiple grid cells through a node-hyperedge hyperassociation structure can characterize the nonlocal interphase component redistribution relationship of CO2 in three-dimensional reservoir space and the multi-field cooperative response caused by phase changes.

[0097] S50: The model input samples are input to the time-assisted layer and rearranged according to the grid cell input. For each grid cell's dynamic physics data, a time series consisting of continuous historical time steps is constructed. The time series is then input into a gated recurrent unit (GRU) for time-series encoding to extract the temporal evolution characteristics of each grid cell on different dynamic physics data. Specifically, the steps include: S51, Input the model into the sample Input to the time-aided layer, which inputs samples into the model. The data is rearranged according to grid cells and dynamic physics fields. For each dynamic physics field in each grid cell, a time series consisting of dynamic physics field data from multiple consecutive historical time steps is constructed. For the first... The first grid cell and the first A dynamic physical field data, its time series It can be represented as: ,in, Indicates the first The dynamic physical field data in the first The corresponding grid cell at each time step The value of .

[0098] S52, which sets the time series of each dynamic physics data corresponding to each grid cell. The data are input into gated recurrent units (GRUs) for time-series encoding, and the evolution of each dynamic physics data in the time dimension is independently encoded to obtain the corresponding hidden state representation. .

[0099] For the The first grid cell of the first grid cell A dynamic physical field data, its hidden state representation It can be represented as: ,in, This represents the encoding process of a GRU unit, used to extract the dynamic change characteristics of the corresponding dynamic physical field data from a univariate time series of continuous historical time steps.

[0100] Using GRU encoding, the set of hidden state representations corresponding to all grid cells and all dynamic physics data can be represented as: ,in, The set of all hidden state features; The number of grid cells; This indicates the amount of dynamic physical field data.

[0101] S53, Representation of Hidden States Perform a linear mapping and reorganize to obtain the time evolution features. This is used for further feature fusion and target prediction.

[0102] The characteristics of temporal evolution can be expressed as: ,in, For the first Learnable weight matrices and bias terms corresponding to dynamic physical field data; For activation functions; This represents the hidden state.

[0103] S54 organizes the temporal evolution features of all grid cells in each dynamic physical field according to the grid cell dimension and the dynamic physical field dimension to obtain a set of temporal evolution features in the form of nodes, and rearranges them according to the three-dimensional grid space index corresponding to each grid cell to obtain the temporal evolution features corresponding to the target time.

[0104] The set of temporal evolution features in node form can be represented as: ,in, This represents the dimension of the time evolution characteristics corresponding to each dynamic physical field.

[0105] Furthermore, the set of temporal evolution features in node form is reorganized to obtain the temporal evolution features corresponding to the target time: ,in, , This indicates that the temporal evolution characteristics are rearranged according to the correspondence between the grid cell number and the three-dimensional grid spatial index.

[0106] Through the aforementioned time encoding process, the time-assisted layer can extract the variation trend of CO2 component fields in the oil, water, and gas phases over historical time steps. Combined with the dynamic responses of pressure, saturation, viscosity, and gas-oil interfacial tension fields, it can characterize the time evolution features related to CO2 component redistribution and phase transformation.

[0107] Through this process, the time evolution features not only include the time series features of each grid cell, but also capture the dynamic changes of each physical field, further enhancing the model's ability to model time evolution characteristics.

[0108] S60 uses the dynamic physics data of the last historical moment as the baseline term. Spatial co-evolutionary features and temporal evolutionary features are input to the feature fusion output layer. The feature fusion output layer then separately calculates the spatial branch prediction results and the temporal branch prediction results, and performs a weighted fusion of the baseline term, spatial branch prediction results, and temporal branch prediction results to obtain the CO2 conversion-related multi-field three-dimensional dynamic physics prediction results for the target moment. Specifically, this includes the following steps: S61, Input samples from the model Extract the last historical moment The dynamic physical field data is used as the baseline term for the prediction at the target time. .

[0109] In this embodiment, the last historical time step is selected. The corresponding dynamic physical field characteristic tensor As a baseline term, that is: , ,in, These represent the three-dimensional mesh in... Maximum number of grid cells in three directions; This indicates the amount of dynamic physical field data.

[0110] For any grid cell Its baseline term can be expressed as: ,in, Indicates the first This historical step is the first Dynamic physics data in the grid cell The value at that location.

[0111] It should be noted that the baseline item Only the dynamic physical field data of the last historical moment is extracted, excluding static geological attribute data. The final prediction target in this step is the target time. Dynamic physical field data.

[0112] S62, perform time-weighted fusion and spatial decoding on the spatial co-evolutionary features corresponding to the continuous historical time steps obtained in step S40 to obtain the spatial branch prediction results. .

[0113] To integrate spatial co-evolutionary features from consecutive historical time steps, temporal fusion weights can be set for each historical time step. And the spatial co-evolutionary feature sequence obtained from S40 After weighted fusion, we get: ,in, Indicates the first The time fusion weights corresponding to each historical time step; Indicates the first The characteristics of spatial co-evolution at each historical time step.

[0114] In one embodiment, the temporal fusion weights can be learnable parameters, obtained through a normalization function: ,in, In order to be with the first Learnable weight parameters corresponding to each historical time step.

[0115] Furthermore, the spatial co-evolutionary features of individual grid nodes are extracted, and the spatial co-evolutionary feature vector corresponding to all nodes is denoted as: , .

[0116] Furthermore, the characteristics of spatial co-evolution after fusion. Spatial decoding is performed to restore the original grid cells from the encoded node feature space, and the spatial branch prediction results are obtained. The process can be represented as follows: ,in, This indicates a spatial decoding function.

[0117] In one embodiment, the spatial decoding function includes node feature rearrangement, upsampling, and 3D convolutional mapping. First, the fused node features... Rearranged into a 3D feature tensor according to the mapping relationship between node numbers and encoded grid indices: , .

[0118] Then, to Upsample to restore it to the original 3D mesh size: , .

[0119] Furthermore, spatial branch prediction results are obtained using 3D convolutional mapping: , ,in, For spatial branch decoding parameters, This indicates the amount of dynamic physical field data.

[0120] Through this step, the spatial branch prediction results... It can characterize the higher-order spatial coupling changes reflected by spatial co-evolutionary features.

[0121] S63, to make the time branch prediction results consistent with the baseline term and spatial branch prediction results Maintaining consistency across the channel dimension allows for the extraction of temporal evolution characteristics from S50. Perform linear mapping or 3D convolution mapping to obtain the temporal branch prediction results: , .in, These are the parameters for time branch mapping.

[0122] S64, representing the baseline terms. Time branch prediction results Set learnable fusion weights and and the baseline item Spatial branch prediction results and time branch prediction results Weighted fusion is performed to obtain the prediction results of multi-field three-dimensional dynamic physical field data at the target time. .

[0123] In this embodiment, the fusion process can be represented as: ,in, Indicates baseline term Learnable fusion weights; Indicates the time branch prediction result Learnable fusion weights.

[0124] At this point, for the first The prediction result for a given dynamic physical field data can be expressed as: ,in, and They represent the first The baseline fusion weights and time branch fusion weights corresponding to each dynamic physical field data.

[0125] Furthermore, for any lattice cell Passing the exam For a given dynamic physical field data point, the predicted value can be expressed as: ,in, Indicates the first A dynamic physical field in the grid cell Spatial branch prediction results at the location; Indicates the last historical moment. The baseline value of a dynamic physical field at this grid cell; This indicates the time branch prediction result of the dynamic physical field at this grid cell.

[0126] S65, Output the prediction results of the multi-field three-dimensional dynamic physics field data at the target time. This is used to characterize the distribution of dynamic physical fields in grid cells at the target time during CO2 flooding in low-permeability reservoirs.

[0127] In this embodiment, the prediction result of the multi-field three-dimensional dynamic physics field at the target time is expressed as follows: ,in, This indicates the number of predicted dynamic physical fields.

[0128] In one embodiment, the dynamic physical field includes a pressure field, a gas phase saturation field, an oil phase saturation field, a water phase saturation field, a gas phase viscosity field, an oil phase viscosity field, a gas-oil interfacial tension field, a CO2 composition field in the oil phase, a CO2 composition field in the water phase, a CO2 composition field in the gas phase, and a global CO2 composition field. Further, for any grid cell... The target time prediction result can be expressed as: ,in, Indicates the target time. Dynamic physics data in grid cells The predicted value at that location.

[0129] S60 output The results represent the predictions of the multi-field three-dimensional dynamic physical fields at the target time, with static geological properties not included as the final prediction target output. Specifically, the predictions of the CO2 component fields in the oil, water, and gas phases at the target time, as well as the global CO2 component field, can be used to reflect the spatial distribution and component redistribution of CO2 in different phase media. Combined with the predictions of the saturation field, viscosity field, and gas-oil interfacial tension field at the target time, the phase transformation during CO2 oil displacement and its impact on the multi-field dynamic response can be further characterized.

[0130] Through the above steps, the feature fusion output layer can simultaneously utilize the dynamic physical field baseline information of the last historical moment, the spatial co-evolution features obtained based on the high-order coupling relationship of the hypergraph, and the temporal evolution features obtained based on GRU to achieve prediction of the multi-field three-dimensional dynamic physical field of CO2 flooding in low-permeability reservoirs at the target moment.

[0131] The CO2 flooding transformation prediction method for low-permeability reservoirs using the hypergraph multi-field spatiotemporal high-order coupling method provided in this disclosure can achieve the following technical effects: 1. This invention constructs an adaptive hyperedge generation layer and a hypergraph high-order coupling representation layer, and uses a node-hyperedge hyperassociation structure to uniformly model the node association relationships encoded by multiple grid cells. It can effectively characterize the multi-field high-order coupling relationships formed by reservoir heterogeneity, pressure propagation, phase saturation changes, fluid property changes and CO2 component distribution during CO2 flooding in low-permeability reservoirs. It solves the technical problem that existing methods mainly rely on local neighborhood feature extraction or pairwise relationship modeling, and are difficult to characterize complex nonlocal high-order coupling structures. 2. This invention incorporates the distribution characteristics of CO2 in different phase media into a unified prediction framework, and combines relevant multi-physics field information such as pressure field, saturation field, viscosity field, gas-oil interfacial tension field and CO2 component field to achieve joint prediction of dynamic responses related to CO2 component redistribution and phase transformation during CO2 flooding in low-permeability reservoirs, thereby improving the characterization ability of CO2 transformation process. 3. This invention integrates the high-order spatial coupling characteristics of the supergraph, the temporal evolution characteristics of GRU, and the dynamic physical field baseline term of the last historical moment. It can predict the three-dimensional dynamic physical field distribution state of CO2 transformation related to the target moment while maintaining the grid cell distribution structure. This helps to improve the accuracy and stability of dynamic prediction of CO2 flooding in low-permeability reservoirs and provides technical support for the optimization of CO2 flooding schemes and the evaluation of development effects.

[0132] Comprehensive performance analysis of this method In this embodiment, the implementation environment of the supergraph multi-field spatiotemporal high-order coupling prediction method for CO2 flooding transformation in low-permeability reservoirs is based on the Ubuntu 22.04 operating system, the programming language is Python 3.12, the deep learning framework is PyTorch 2.8.0, and CUDA 12.8 is used for GPU acceleration computing.

[0133] like Figure 2As shown, the prediction system provided in this embodiment includes a data preprocessing layer, a feature encoding layer, an adaptive hyperedge generation layer, a hypergraph high-order coupling representation layer, a temporal auxiliary layer, and a feature fusion output layer. The model extracts the spatial co-evolution features and temporal evolution features of the multi-physics field, and then weights and fuses them with the dynamic physics field data of the last historical moment to obtain the multi-field three-dimensional dynamic physics field prediction result at the target moment.

[0134] like Figure 3 As shown, the adaptive hyperedge generation layer generates hyperedge prototypes and node-hyperedge assignment matrices based on the global context information of node features; the hypergraph high-order coupling representation layer sequentially performs node-to-hyperedge aggregation, hyperedge representation update, and hyperedge-to-node backpropagation to obtain spatial co-evolution features representing the high-order nonlocal coupling relationship between grid cells.

[0135] In this embodiment, the reservoir's 3D mesh size is set to 10×64×64, i.e., K=10, I=64, J=64, and the sliding time window length is set to 5. The input data includes 11 dynamic physical field data points and 2 static geological attribute data points. The dynamic physical field data points include pressure field, gas phase saturation field, oil phase saturation field, water phase saturation field, gas phase viscosity field, oil phase viscosity field, gas-oil interfacial tension field, CO2 composition field in the oil phase, CO2 composition field in the water phase, CO2 composition field in the gas phase, and global CO2 composition field. The static geological attribute data points include porosity and permeability. The model uses multi-field 3D physical field data from 5 consecutive historical time steps as input and 11 dynamic physical field data points at the target time as the prediction output.

[0136] To verify the effectiveness of the adaptive hypergraph generation layer and hypergraph higher-order coupled representation layer structure in this invention, an ablation experiment was conducted in this embodiment. The proposed model was compared with a model that removed the hypergraph module and a model that replaced the hypergraph module with a multilayer perceptron module. The proposed model is denoted as Ours; the model that removed the hypergraph module is denoted as w / ohypergraph (i.e., removing the adaptive hypergraph generation layer and the hypergraph higher-order coupled representation layer); and the model that replaced the hypergraph module with a multilayer perceptron module is denoted as w / MLP (i.e., using ordinary feature mapping to replace the adaptive hypergraph generation layer and the hypergraph higher-order coupled representation layer).

[0137] The experiment used normalized mean absolute error (MAE) and coefficient of determination (R²) as evaluation metrics. MAE measures the average deviation between predicted and actual values; a smaller MAE value indicates a lower prediction error. R² measures the model's ability to explain changes in the real dynamic physical field; a larger R² value indicates a better model fit.

[0138] The experiments statistically analyzed the normalized mean absolute error (MAE) of the proposed model, the model without the hypergraph module, and the model with the hypergraph module replaced by a multilayer perceptron module on 11 dynamic physics datasets. The experimental results are shown in Table 1, and the coefficient of determination (R²) results for the 11 dynamic physics datasets are shown in Table 2.

[0139] Table 1. Comparison of MAE of different methods on dynamic physics data

[0140] Table 2. Comparison of R² values ​​for different methods on dynamic physics data.

[0141] As can be seen from the experimental results in Tables 1 and 2, the model of this invention (Ours) achieved optimal performance in all 11 prediction tasks of dynamic physics fields.

[0142] Regarding the MAE index, compared to models with the hypergraph module removed (w / o hypergraph) and models with the hypergraph module replaced by a multilayer perceptron (w / MLP), the model of this invention exhibits significantly lower errors across all physical fields, indicating a smaller deviation between the predicted results and the actual values. Regarding the R² index, the model of this invention approaches 1 across most physical fields, significantly outperforming the comparative models, demonstrating its stronger fitting and explanatory power for the dynamic evolution of low-permeability reservoirs across multiple physical fields.

[0143] The above results show that the adaptive hypergraph generation layer and the hypergraph higher-order coupling representation layer can effectively model the higher-order nonlocal spatial dependencies between grid cells, thereby significantly enhancing the ability of the model of this invention to characterize the complex coupling relationships between multiple physics fields.

[0144] Furthermore, compared to the traditional multilayer perceptron structure, the hypergraph structure can more fully explore the high-order correlation information between spatial nodes, enabling the model of this invention to achieve stable and significant performance improvements in various physical field prediction tasks such as pressure field, saturation field, and CO2 component field, thereby achieving high-precision prediction of the dynamic evolution process of multi-physical fields in low-permeability reservoirs.

[0145] Figures 4 to 8 This paper presents a comparison between the actual and predicted three-dimensional spatial distributions of the gas phase saturation field, oil phase saturation field, and CO2 component fields in the gas, oil, and water phases at the target time in the embodiments of this disclosure.

[0146] from Figure 4It can be seen that the gas phase saturation exhibits obvious high-value and low-value distribution characteristics in space. The high-value areas are mainly concentrated in the upper part of the model and near the flow channel of this invention. The prediction results can accurately reproduce the location, shape and expansion range of the gas phase accumulation area in the real value, indicating that the model of this invention can accurately capture the gas phase transport process.

[0147] from Figure 5 It can be seen that the oil phase saturation exhibits a clear inverse relationship with the gas phase distribution in space. The gas phase enrichment area corresponds to the oil phase saturation reduction area. The prediction results can effectively reflect the spatial decay trend during the oil phase displacement process, indicating that the model of this invention can characterize the gas-oil mutual replacement process.

[0148] from Figure 6 It can be seen that the components of CO2 in the gas phase exhibit a spatial distribution characteristic of diffusion and gradual enrichment from the injection end outwards. The high-value region forms a continuous migration channel over time, and the prediction results can accurately reproduce the migration path and enrichment region changes of CO2 in the gas phase.

[0149] from Figure 7 It can be seen that the component distribution of CO2 in the oil phase exhibits a significant lag diffusion characteristic compared to the gas phase. Its high value region is mainly distributed in the gas phase front and transition region. The prediction results can better characterize the process characteristics of CO2 dissolving and diffusing from the gas phase into the oil phase.

[0150] from Figure 8 It can be seen that CO2 components in the aqueous phase are generally distributed in a low-value dispersed state, but there is still obvious enrichment in local seepage channels and areas with large pressure gradients. The prediction results can effectively capture the spatial distribution characteristics of CO2 diffusion into the aqueous phase and cross-phase migration.

[0151] Further comparison Figures 4 to 8 It can be seen that the various physical fields have significant coupling correlations in spatial distribution. The compositional changes of CO2 in different phases correspond to the spatial migration trends of gas phase saturation and oil phase saturation, exhibiting a consistent displacement-dissolution-redistribution evolution law. The predicted results are not only highly consistent with the actual values ​​in terms of overall distribution, but also can accurately reconstruct the cross-phase migration and component redistribution process of CO2, thus reflecting the CO2 transformation behavior from the perspective of physical field evolution.

[0152] Therefore, this method can accurately characterize the multiphase migration and transformation process of CO2 in low-permeability reservoirs and make high-precision predictions of the three-dimensional dynamic distribution under strong coupling conditions of multiple physics fields.

[0153] Figures 9 to 13The figures show a comparison of the actual values, predicted values, and error distributions of the gas phase saturation field, oil phase saturation field, and CO2 component fields in the gas, oil, and water phases in the embodiments of this disclosure on typical cross sections (K=1 and K=2) at the target time.

[0154] Depend on Figure 9 and Figure 10 It can be seen that under different cross-sectional conditions, the gas phase saturation exhibits significant local enrichment and diffusion regions within the two-dimensional cross-section, forming a continuous banded structure along the main seepage direction. The prediction results can well maintain the morphological characteristics and distribution boundaries of this two-dimensional spatial structure, showing high consistency across different vertical layers (K=1 and K=2).

[0155] Depend on Figure 11 It can be seen that the components of CO2 in the gas phase exhibit obvious non-uniform diffusion characteristics in the cross-section. The high-value regions form a continuous migration channel structure in the two-dimensional plane, reflecting the dominant migration path of CO2 in the gas phase. The prediction results can accurately reproduce the spatial location and expansion trend of this migration channel, indicating that the model of this invention has a good spatial characterization ability for CO2 migration behavior at the two-dimensional cross-sectional scale.

[0156] Depend on Figure 12 It can be seen that the component distribution of CO2 in the oil phase exhibits a significant hysteresis diffusion characteristic compared to the gas phase. Its high-value regions are mainly distributed at the phase interface and the transition region at the gas phase front, and it shows an asymmetric expansion characteristic in cross-section. The prediction results can well reconstruct the two-dimensional spatial evolution process of CO2 penetration and dissolution diffusion from the gas phase to the oil phase.

[0157] Depend on Figure 13 It can be seen that CO2 exhibits a generally low-concentration, dispersed distribution in the aqueous phase, but local enrichment occurs in areas with large seepage channels and pressure gradients, reflecting the local diffusion behavior of CO2 migrating into the aqueous phase across phases. The prediction results can effectively capture this spatial distribution pattern of low-concentration diffusion and local enrichment.

[0158] Further comparison of the cross sections with K=1 and K=2 reveals significant differences in the distribution of physical fields in different vertical layers. CO2 exhibits non-uniform migration and stratified diffusion characteristics in different depth layers, indicating that the model of this invention can effectively characterize the influence of formation heterogeneity on the flow path in a two-dimensional cross section.

[0159] As can be seen from the error distribution diagram, the prediction errors of each physics field are mainly concentrated in the areas near the phase interface and the CO2 migration front where the gradient changes are more intense, while the errors are smaller in the large-scale uniformly distributed areas. This indicates that the model of the present invention can stably reconstruct the main spatial structure features of the multiphysics field at the cross-sectional scale.

[0160] Furthermore, it can be seen that this method can achieve high-precision prediction of the spatial distribution of multi-physics fields in low-permeability reservoirs at the cross-sectional scale, and can accurately characterize the two-dimensional migration and distribution characteristics of CO2 between different phases, thereby verifying the predictive ability and stability of the model of this invention at the local spatial scale.

[0161] Figures 14 to 18 The figures show a comparison of the true values, predicted values, and error distributions of the gas phase saturation field, oil phase saturation field, and CO2 component fields in the gas, oil, and water phases at a typical longitudinal section at a target time, as provided in the embodiments of this disclosure.

[0162] Depend on Figures 14 to 18 It can be seen that, along the longitudinal section, the physical fields exhibit a clear stratified distribution along the vertical layers, with significant differences in the saturation field and CO2 component field at different depths. Gas phase saturation shows higher values ​​in the upper layers and in the local dominant channel regions, while oil phase saturation shows a relatively decreasing trend in the corresponding regions, indicating a clear correspondence between gas phase intrusion and oil phase transfer during CO2 displacement in the longitudinal profile.

[0163] Further observation of the CO2 composition fields in the gas, oil, and aqueous phases reveals that the CO2 composition is not uniformly distributed in the longitudinal section, but rather exhibits zonal expansion characteristics along specific strata and local channels. The gas phase CO2 composition mainly reflects the rapid migration path of CO2 in the gas phase, the oil phase CO2 composition mainly reflects the transition region of CO2 dissolution and diffusion into the oil phase, and the aqueous phase CO2 composition reflects the low concentration distribution characteristics of CO2 after migrating into the aqueous phase in local areas.

[0164] The comparison between the predicted results and the actual values ​​shows that the model of the present invention can better reproduce the interlayer distribution differences, longitudinal extension boundaries and local channel characteristics of various physical fields on the longitudinal section. In particular, it can maintain the spatial correspondence between the high gas saturation area and the CO2 enrichment area, indicating that the model of the present invention can effectively capture the migration, displacement and cross-phase distribution process of CO2 in the longitudinal strata of low-permeability reservoirs.

[0165] As can be seen from the error distribution, the prediction error is mainly concentrated in the drastically changing regions such as the phase interface, interlayer transition zone, and CO2 migration front in the longitudinal section, while the error is smaller in the relatively stable vertical distribution region. This indicates that the model of the present invention has a good reconstruction capability for the multiphysics-dominated structure in the longitudinal section.

[0166] Furthermore, it can be concluded that this method can accurately predict the vertical distribution characteristics of multi-physics fields in low-permeability reservoirs at the longitudinal section scale, and can reflect the migration, displacement and cross-phase distribution processes of CO2 between different phases from the longitudinal profile level, thereby verifying the accuracy of the model of this invention in predicting the physical processes related to CO2 transformation under vertical heterogeneity conditions.

[0167] In summary, this invention addresses the limitations of existing multi-field three-dimensional physical field prediction methods for CO2 flooding in low-permeability reservoirs, including limited local spatial modeling capabilities, insufficient expression of high-order coupling relationships, inadequate extraction of multi-physics temporal evolution features, and insufficient synergistic utilization of static geological attributes and dynamic physical fields. This invention, through local three-dimensional convolutional coding, adaptive hyperedge generation, hypergraph high-order coupling propagation, and GRU time-assisted coding, achieves joint modeling of the spatial synergistic relationships and temporal evolution laws of dynamic physical fields such as pressure field, saturation field, viscosity field, gas-oil interfacial tension field, and CO2 component field in continuous historical time steps under static geological attribute constraints. Furthermore, it characterizes the distribution, component redistribution, and phase transformation-related responses of CO2 in the oil, water, and gas phases. Based on this, the invention uses the dynamic physical field of the last historical moment as a baseline term and weightedly fuses it with spatial synergistic evolution features and temporal evolution features to obtain the multi-field three-dimensional dynamic physical field prediction results for the target moment, thereby improving the accuracy and stability of dynamic physical field prediction for CO2 flooding in low-permeability reservoirs.

[0168] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0169] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

[0170] This embodiment also provides a hypergraph multi-field spatiotemporal high-order coupled low-permeability reservoir CO2 flooding transformation prediction system, including: The data preprocessing layer is configured to acquire dynamic physical field data and static geological attribute data generated by the numerical simulation of CO2 flooding in low-permeability reservoirs, and to preprocess the dynamic physical field data and static geological attribute data to construct model input samples. The feature encoding layer is configured to input the model input samples into the feature encoding layer, perform local 3D convolutional encoding on the model input samples, and obtain the node feature sequence for hypergraph modeling; The adaptive hyperedge generation layer is configured to input the node feature sequence into the adaptive hyperedge generation layer. Based on the context information of the node feature sequence, the adaptive hyperedge generation layer generates hyperedge prototypes through the adaptive hyperedge generation mechanism, and establishes hyper-association relationships between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequence and each hyperedge prototype, thereby constructing a node-hyperedge hyper-association structure. The hypergraph higher-order coupling representation layer is configured to input the node-hyperedge hyperassociation structure into the hypergraph higher-order coupling representation layer, aggregate the feature vectors of multiple nodes associated with the same hyperedge to obtain the hyperedge features, and backfeed the hyperedge features to each node to obtain the spatial co-evolution features. The time-assisted layer is configured to input the model input samples into the time-assisted layer and rearrange them according to the grid cell input. For each grid cell, a time series consisting of continuous historical time steps is constructed for each dynamic physical field data. The time series is then input into a gated recurrent unit (GRU) for time-series encoding to extract the temporal evolution characteristics of each grid cell on different dynamic physical field data. The feature fusion output layer is configured to take the dynamic physical field data of the last historical moment as the baseline term, input the spatial co-evolution features and temporal evolution features into the feature fusion output layer, generate spatial branch prediction results and temporal branch prediction results by the feature fusion output layer, and perform weighted fusion of the baseline term, spatial branch prediction results and temporal branch prediction results to obtain the CO2 conversion related multi-field three-dimensional dynamic physical field prediction results at the target time.

[0171] The specific implementation process of this system can be found in the description of the above method embodiments, and will not be repeated here.

[0172] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0173] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using a hypergraph-based multi-field spatiotemporal high-order coupling, characterized in that... include: S10, acquire dynamic physical field data and static geological attribute data generated by the numerical simulation of CO2 flooding in low-permeability oil reservoirs, and preprocess the dynamic physical field data and static geological attribute data to construct model input samples; S20, the model input sample is input into the feature encoding layer, and local three-dimensional convolutional encoding is performed on the model input sample to obtain the node feature sequence used for hypergraph modeling; S30, the node feature sequence is input to the adaptive hyperedge generation layer. Based on the context information of the node feature sequence, the adaptive hyperedge generation layer generates hyperedge prototypes through the adaptive hyperedge generation mechanism, and establishes a hyper-association relationship between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequence and each hyperedge prototype, thereby constructing a node-hyperedge hyper-association structure. S40, the node-hyperedge hyperassociation structure is input to the hypergraph higher-order coupling representation layer, the feature vectors of multiple nodes associated with the same hyperedge are aggregated to obtain hyperedge features, and the hyperedge features are back-transmitted to each node to obtain spatial co-evolution features; S50, the model input samples are input to the time auxiliary layer and rearranged according to the grid cell input. For each grid cell, a time series consisting of continuous historical time steps is constructed for each dynamic physical field data. The time series is input to the gated recurrent unit (GRU) for time-series encoding to extract the time evolution characteristics of each grid cell on different dynamic physical field data. S60, the dynamic physical field data of the last historical moment is used as the baseline term, the spatial co-evolution feature and the temporal evolution feature are input to the feature fusion output layer, the feature fusion output layer generates the spatial branch prediction result and the temporal branch prediction result, and the baseline term, the spatial branch prediction result and the temporal branch prediction result are weighted and fused to obtain the CO2 conversion related multi-field three-dimensional dynamic physical field prediction result at the target time.

2. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using hypergraph multi-field spatiotemporal high-order coupling as described in claim 1, characterized in that, In step S10, the preprocessing of the dynamic physical field data and the static geological attribute data to construct model input samples includes: Spatial index mapping is performed on the dynamic physical field data and the static geological attribute data, using the coordinates of the reservoir grid unit as the positioning basis, to map the dynamic physical field data and the static geological attribute data to a unified grid unit; The dynamic physical field data and the static geological attribute data that have been mapped are preprocessed by data cleaning, time step alignment and normalization; The first The dynamic physics data at each time step are constructed into a dynamic physics feature tensor corresponding to each of the grid cells. The static geological attribute data is constructed into static geological attribute tensors corresponding to each of the grid cells. ; The dynamic physical field feature tensor of multiple consecutive historical time steps is concatenated with the static geological attribute tensor in the feature channel dimension to construct the model input sample.

3. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using hypergraph multi-field spatiotemporal high-order coupling according to claim 1, characterized in that, In step S20, performing local 3D convolutional encoding on the model input samples to obtain a node feature sequence for hypergraph modeling includes: The input samples of the model are split according to time steps to obtain the single-time multi-field three-dimensional feature tensor corresponding to each historical time step; Local three-dimensional convolutional coding is performed on the single-time multi-field three-dimensional feature tensors corresponding to each historical time step, and the local representation information of each encoded grid cell and its neighborhood local spatial region is extracted by downsampling to obtain the corresponding local spatial coding features; The local spatial coding features are uniformly mapped and aligned to form multi-field fusion representation features corresponding to each historical time step; The multi-field fusion representation features corresponding to each historical time step are expanded into a node feature sequence according to the encoded grid cell order, and the node feature sequence is used as the input of the subsequent adaptive hyperedge generation layer.

4. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using hypergraph multi-field spatiotemporal high-order coupling as described in claim 1, characterized in that, In step S30, based on the context information of the node feature sequence, the adaptive hyperedge generation layer generates a hyperedge prototype through an adaptive hyperedge generation mechanism, including: A context representation vector is generated based on the global context information of the node feature sequence; wherein, the global context information is composed of statistical features of the node feature sequence, including the mean information, maximum information, or a combination of the mean information and the maximum information of the node feature sequence in the node dimension; The context representation vector is input into the adaptive hyperedge generation layer, and the adaptive hyperedge generation layer is used to perform context modulation on a preset number of basic hyperedge prototypes to generate a set of hyperedge prototypes corresponding to the current input sample.

5. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using hypergraph multi-field spatiotemporal high-order coupling according to claim 4, characterized in that, In step S30, establishing the hyper-association relationship between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequence and each hyperedge prototype includes: Extract the node feature vector of each node in the node feature sequence, calculate the association score from the node to the hyperedge based on the association strength between each node feature vector and each hyperedge prototype, and normalize the association score to obtain the node-hyperedge assignment matrix. Based on the node-hyperedge allocation matrix, the hyper-association relationship is established and organized in the time dimension to construct the node-hyperedge hyper-association structure.

6. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using hypergraph multi-field spatiotemporal high-order coupling according to claim 5, is characterized in that... In step S40, aggregating the feature vectors of multiple nodes associated with the same hyperedge to obtain hyperedge features includes: Based on the node-hyperedge assignment matrix that represents the relationship between nodes and hyperedges in the node-hyperedge hyperassociation structure, determine the contribution weight of each node feature vector to each hyperedge. For any hyperedge, the feature vectors of multiple nodes associated with the hyperedge are weighted and aggregated according to the corresponding contribution weight to obtain the corresponding initial hyperedge features; The initial hyperedge features are normalized and / or nonlinearly mapped to obtain the hyperedge features.

7. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using hypergraph multi-field spatiotemporal high-order coupling according to claim 6, characterized in that, In step S40, the step of backpropagating the hyperedge features to each node to obtain spatial co-evolution features includes: The hyperedge features are mapped and transformed to obtain the updated hyperedge representation; Based on the node-hyperedge allocation matrix, the hyperedge representation is back propagated to each node to obtain the high-order coupling propagation characteristics received by the node. The node feature vector is updated by performing residual updates on the higher-order coupled propagation features to obtain the updated node feature vector. The updated node feature vectors are subjected to feedforward mapping and feature normalization, and then organized in the time dimension to obtain the spatial co-evolutionary features.

8. The method for predicting CO2-driven oil recovery transformation in low-permeability reservoirs using super-map multi-field spatiotemporal high-order coupling according to any one of claims 1 to 7, characterized in that, In step S50, inputting the time series into a gated recurrent unit (GRU) for time-series encoding to extract the temporal evolution characteristics of each grid cell on different dynamic physical field data includes: The time series corresponding to each dynamic physical field data of each grid cell is input into the gated recurrent unit (GRU) to independently encode the evolution process of each dynamic physical field data in the time dimension, thereby obtaining the corresponding hidden state representation. The hidden state representations are linearly mapped and reorganized to obtain the various temporal evolution features. The temporal evolution characteristics of all grid cells in each dynamic physical field are organized according to the grid cell dimension and the dynamic physical field dimension to obtain a set of temporal evolution characteristics in the form of nodes. Then, the set is rearranged according to the three-dimensional grid space index corresponding to each grid cell to obtain the temporal evolution characteristics corresponding to the target time.

9. A high-order spatiotemporal coupling supergraph multi-field CO2 flooding transformation prediction system for low-permeability reservoirs, characterized in that, include: The data preprocessing layer is configured to acquire dynamic physical field data and static geological attribute data generated by the numerical simulation of CO2 flooding in low-permeability reservoirs, and to preprocess the dynamic physical field data and the static geological attribute data to construct model input samples. The feature encoding layer is configured to input the model input samples into the feature encoding layer, perform local three-dimensional convolutional encoding on the model input samples, and obtain a node feature sequence for hypergraph modeling; An adaptive hyperedge generation layer is configured to input the node feature sequence into the adaptive hyperedge generation layer. Based on the context information of the node feature sequence, the adaptive hyperedge generation layer generates hyperedge prototypes through an adaptive hyperedge generation mechanism, and establishes a hyper-association relationship between nodes and hyperedges based on the correlation strength between the node feature vectors corresponding to each node in the node feature sequence and each hyperedge prototype, thereby constructing a node-hyperedge hyper-association structure. The hypergraph higher-order coupling representation layer is configured to input the node-hyperedge hyperassociation structure into the hypergraph higher-order coupling representation layer, aggregate the feature vectors of multiple nodes associated with the same hyperedge to obtain hyperedge features, and backfeed the hyperedge features to each node to obtain spatial co-evolution features; The time-assisted layer is configured to input the model input samples into the time-assisted layer and rearrange them according to the grid cell input. For each grid cell, a time series consisting of continuous historical time steps is constructed for each dynamic physical field data. The time series is then input into a gated recurrent unit (GRU) for time-series encoding to extract the time evolution characteristics of each grid cell on different dynamic physical field data. The feature fusion output layer is configured to take the dynamic physical field data of the last historical moment as the baseline term, input the spatial co-evolution feature and the temporal evolution feature into the feature fusion output layer, generate spatial branch prediction results and temporal branch prediction results from the feature fusion output layer, and perform weighted fusion of the baseline term, the spatial branch prediction results and the temporal branch prediction results to obtain the CO2 conversion related multi-field three-dimensional dynamic physical field prediction results at the target moment.

Images