Shale Gas Carbon Dioxide Displacement Well Selection Method and System Based on Interpretable Graph Neural Network

Abstract

This application focuses on the field of oil and gas development technology, proposing a method and system for shale gas carbon dioxide displacement well selection based on interpretable graph neural networks. This method innovatively employs dynamic topology modeling, leveraging the characteristics of graph neural networks to enhance node associations, mapping well group production time-series data into a topology graph containing supervariable attributes, and integrating Fourier graph neural networks and Bayesian optimization to accurately fit the dynamic laws of production. Simultaneously, it embeds a GNNExplainer interpretable module, generating a pressure crosstalk heatmap by quantifying inter-well weight coefficients, achieving visualized analysis of pressure crosstalk levels. Finally, a multi-dimensional decision model is constructed based on pressure crosstalk intensity and production indicators to complete well selection. This scheme solves the problem of quantifying nonlinear factors such as fracture closure through dynamic modeling of the well network structure, accurately identifying advantageous displacement channels without high-cost monitoring, providing an intelligent well selection solution for the large-scale application of CCUS technology, and contributing to the low-carbon and efficient development of shale gas reservoirs.

Description

Technical Field

[0001] This application relates to the field of petroleum development, and in particular to carbon dioxide displacement well selection methods, terminals and storage media for unconventional oil and gas reservoirs. Background Technology

[0002] Shale gas, as a clean energy source, is crucial for ensuring energy security. However, Chinese shale reservoirs are characterized by extremely low porosity and permeability (mostly Nadarcy-level) and strong adsorption properties, resulting in low natural gas desorption efficiency and limited recovery rates. Carbon dioxide displacement technology, by injecting CO2 to replace adsorbed methane, offers both increased reserves and storage benefits. However, this technology faces a core bottleneck: shale permeability is extremely low, requiring reliance on the natural fracture network formed by fracturing as a gas migration channel. If well selection is inappropriate, CO2 may not effectively reach the target reservoir, potentially leading to crossflow or storage failure.

[0003] Currently, shale gas CO2 displacement well selection mainly relies on wells with high "pressure channeling" to ensure effective CO2 flow, but the determination of pressure channeling has significant limitations. Field operations heavily depend on two types of methods. First, microseismic monitoring technology assesses the degree of pressure channeling between wells by capturing fracture extension signals, but the cost of monitoring a single well is as high as hundreds of thousands of yuan, and the resolution for deep reservoirs (>3500m) is insufficient, making large-scale promotion difficult. Second, empirical methods analyze changes in casing pressure in adjacent wells. A sudden increase in casing pressure of more than 1.0 MPa in adjacent wells during fracturing is considered a sign of pressure channeling, but this method is only applicable to the early stages of production. In the later stages of production, pore blockage caused by scaling (calcium and magnesium carbonate deposition), proppant embedding, and fracture closure effects caused by shale stress sensitivity significantly weaken the initial fracture conductivity, and empirical methods cannot dynamically capture such changes, leading to a disconnect between well selection models and actual geological conditions.

[0004] Patent document CN118296975A proposes a well connectivity prediction method based on a knowledge-interactive graph neural network, applicable to water-driven sandstone reservoirs. This method relies on the natural permeability of sandstone reservoirs to predict well connectivity through the seepage relationship after water injection. However, shale gas reservoirs require fracturing to create an artificial fracture network for development, and their connectivity prediction cannot be simplified to the water injection response of the natural reservoir. Therefore, the prediction method for sandstone reservoirs is not suitable for assessing the degree of channeling in shale gas reservoirs. Patent document CN118036477B introduces a well location and well control parameter optimization method based on a spatiotemporal graph neural network, primarily for new well development scenarios involving CO2 displacement in shale reservoirs. While this method can proactively design well network layouts and well control parameters, it relies on predicting future production dynamics data and is not suitable for the modification and evaluation of older wells in the later stages of development. For older wells, directly assessing the degree of channeling using historical production dynamics is more practical because this approach does not rely on predicting future production states but rather analyzes existing dynamic data, thereby improving the accuracy and efficiency of decision-making.

[0005] In summary, existing shale gas carbon dioxide displacement well selection technologies suffer from high costs, poor adaptability, strong subjectivity, and a lack of dynamic evaluation methods. They are particularly inadequate when dealing with the complex nonlinear changes in shale reservoirs during the later stages of production. Therefore, there is an urgent need for a cost-effective, highly accurate well selection method that can adapt to dynamic production changes to optimize the carbon dioxide displacement efficiency of shale gas reservoirs and overcome the limitations of existing technologies. Summary of the Invention

[0006] This application provides a shale gas carbon dioxide displacement well selection method, system, terminal, and storage medium based on interpretable graph neural networks, which is used to solve the problems of high cost and strong subjectivity in existing well selection technologies and achieve efficient and accurate well location decision-making.

[0007] The technical solution of this application is as follows:

[0008] This application provides a shale gas carbon dioxide displacement well selection method based on interpretable graph neural networks, including the following steps:

[0009] Step (1) Select candidate gas injection well groups based on the quantitative threshold of real-time production data and dynamic monitoring data of old wells in the well area. The production data includes pressure parameters, production parameters and efficiency indicators, and the dynamic monitoring data includes fluid characteristics and pressure test data.

[0010] Step (2) For the selected candidate well groups, obtain the full-cycle production dynamic time series data of the candidate well groups, and construct a standard dataset containing pressure time series curves and production time series curves for fitting the production dynamics of gas injection wells.

[0011] Step (3) Based on the standard dataset, construct a well network diagram structure model that can accurately fit the production dynamics of gas injection wells: map the production dynamics time series data into a hypervariable graph structure, where the production parameters of each time step are used as nodes, and the nodes are fully connected to form an adjacency matrix. Use a Fourier graph neural network to fit the production dynamics sequence of gas injection wells.

[0012] Step (4) Using the established well network diagram structure model, the weight coefficients between well groups are quantified through the graph neural network interpretable module (GNNExplainer), generating an edge weight matrix that reflects the pressure crosstalk intensity, obtaining the pressure crosstalk intensity between well groups, and outputting a heat map of the pressure crosstalk intensity between wells.

[0013] Step (5) Combine the pressure cross-connection intensity and production indicators between each well group to determine the final carbon dioxide displacement well group.

[0014] Preferably, the pressure parameters in step (1) include: current casing pressure and oil pressure; the production parameters include daily gas production, daily water production, cumulative gas production, and cumulative water production. The efficiency indicators include flowback rate, production efficiency, and commissioning cycle. The dynamic monitoring data also involve fluid characteristics, micro-fracturing test data, and static pressure gradient monitoring data.

[0015] Preferably, the production dynamic time series data mentioned in step (2) includes casing pressure sequence, oil pressure sequence, monthly gas production sequence and monthly water production sequence;

[0016] Preferably, the production indicators in step (5) refer to the current formation pressure, casing pressure, daily gas production, and backflow rate.

[0017] Preferably, the quantification threshold mentioned in step (1) is determined based on actual field development. Different thresholds are determined for different geological characteristics of normal pressure shale gas or overpressure shale gas. Applicable well groups usually have the following regional characteristics: casing pressure is lower than the critical pressure required to maintain fracture conductivity, daily gas production is in a continuous decreasing stage and the deceleration rate is greater than the regional average deceleration rate, flowback rate is higher than the regional average, and block pressure is lower than the pressure balance value of the reservoir structural unit.

[0018] More preferably, the critical pressure for the fracture conductivity is calculated using the following model:

[0019]

[0020] Among them, P c Here, E is the critical pressure for fracture conduction, and E is Young's modulus (GPa). Let S be the limiting strain, v be Poisson's ratio, and S be the limiting strain. h This refers to horizontal ground stress.

[0021] Preferably, step (2) of constructing the standard dataset further includes: performing data preprocessing to remove outliers, normalizing the data using min-max normalization, and dividing the dataset into a training set and a test set proportionally; wherein the min-max normalization formula is as follows:

[0022]

[0023] Where x represents the original feature data, min represents the minimum value among the same feature data, and max represents the maximum value among the same feature data. * These are the feature data after normalization.

[0024] Preferably, the specific steps for constructing the well network diagram structure model in step (3) include:

[0025] Node definition: The N-dimensional production variable at time step t is taken as a node, and its feature vector is X. t ∈R N This includes casing pressure, oil pressure, daily gas production, and daily water production; among which, X t Let N be the feature vector at time t, N be the dimension, i.e. the number of types of dynamic parameters, and R be the real vector space.

[0026] Edge connection mechanism: Initialize the fully connected adjacency matrix A∈R T×T Where A is the adjacency matrix and T is the length of the time series;

[0027] The sliding window method is used to process time series data, and hyperparameters such as learning rate, number of iterations, and dropout rate are automatically adjusted through Bayesian optimization algorithm.

[0028] Preferably, the specific steps for quantifying the weighting coefficients between well groups in step (4) include:

[0029] Optimize the loss function max using GNNExplainer Gs ,I(Y,G s |G), output edge weight matrix W∈[0,1] n×n Where G is the hypervariable graph, G s To interpret the subgraph, Y is the predicted output, W is the weight matrix, and n is the number of nodes;

[0030] The production contribution weights of the same well are summed to calculate the well group pressure cross-flow coefficient.

[0031] This application also provides a shale gas carbon dioxide displacement well selection device, which provides a method for efficiently and rapidly making decisions on shale gas carbon dioxide displacement well selection in a target block. The device includes:

[0032] Data acquisition module: used to integrate real-time production data and dynamic monitoring data of old wells in the well area, and to screen candidate gas injection well groups based on the quantitative thresholds of these data;

[0033] Standard dataset construction module: Constructs a standard dataset to fit the production dynamics of the candidate gas injection well groups, including processing and integrating the production dynamic parameters to form a time-series correlation dataset;

[0034] Graph structure modeling module: Based on Fourier graph neural network, a well network graph structure model is constructed, including mapping production dynamic time series data into a hypervariable graph structure, where the production parameters of each time step are used as nodes, and fully connected nodes form an adjacency matrix. Fourier graph neural network is used to fit the production dynamic sequence of gas injection wells and output the production dynamic fitting curve.

[0035] Pressure Crossing Analysis Module: The graph neural network interpretable module (GNNExplainer) quantifies the weight coefficients between well groups, generates an edge weight matrix that reflects the strength of pressure crossing connectivity, and outputs a heat map of the degree of pressure crossing between wells;

[0036] Decision output module: Based on the combined pressure cross-connection strength and production indicators, determine the final carbon dioxide displacement well group.

[0037] This application also provides a shale gas carbon dioxide displacement well selection terminal, including a memory, a processor, and a computer program for a shale gas carbon dioxide displacement well selection method stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the above-described shale gas carbon dioxide displacement well selection method.

[0038] This application also provides a computer-readable storage medium storing a program that can be loaded and executed by a processor to implement the above-described shale gas carbon dioxide displacement well selection method.

[0039] According to the specific embodiments provided in this application, this application has the following technical effects:

[0040] This application breaks through the bottlenecks of traditional well selection methods by pioneering a dynamic graph structure modeling technology. It constructs a data-driven intelligent decision-making model, fusing multi-dimensional dynamic production data through a Fourier graph neural network to output an inter-well weight coefficient matrix, achieving quantitative analysis of pressure crosstalk connectivity paths. The core value of graph neural networks (GNNs) lies in fusing well network topology with dynamic production data, emphasizing the connections between nodes. Specifically, this method maps the physical space of horizontal well deployment as a graph structure, where well locations are nodes and inter-well pressure crosstalk levels are edges. A node feature matrix is ​​constructed using dynamic production parameters (such as daily gas production, daily liquid production, casing pressure, and oil pressure changes in adjacent wells). Inter-well dynamic responses (such as sudden increases in casing pressure due to fracturing and the rate of production decline) serve as initial estimates of edge weights. The weight coefficients are iteratively optimized through a multi-layer message passing mechanism, ultimately forming a "well network graph structure model" reflecting the intensity of pressure crosstalk. Compared to traditional empirical methods, this approach improves identification accuracy and speed, and solves the quantitative characterization challenges of nonlinear factors such as model failure due to fracture closure in the later stages of production and particle migration. It provides an engineering solution that combines timeliness and robustness for the quantitative analysis of pressure channeling patterns. By outputting a pressure channeling degree heatmap through an interpretable module, high-potential well groups are accurately marked. This intelligent technology replaces traditional microseismic monitoring, significantly reducing the cost per well. This technical solution not only significantly improves the efficiency and accuracy of CO2 displacement well selection but also effectively enhances the CO2 replacement efficiency of well groups, providing an innovative path for cost reduction and efficiency improvement in the oil and gas development field. Attached Figure Description

[0041] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0042] Figure 1 The steps of the shale gas carbon dioxide displacement well selection method based on interpretable graph neural networks described in the embodiments of this application are as follows;

[0043] Figure 2 This is an example of the dynamic production sequence of the well group described in the embodiments of this application;

[0044] Figure 3 This is the fitting curve for the monthly gas production sequence using the Fourier graph neural network described in the embodiments of this application;

[0045] Figure 4 This is the Fourier transform neural network overlay sequence fitting curve described in the embodiments of this application;

[0046] Figure 5 Heatmap showing the contribution weights of pressure cross-contamination production as described in the embodiments of this application;

[0047] Figure 6This is a structural block diagram of the shale gas carbon dioxide displacement well device described in the embodiments of this application;

[0048] Figure 7 This is a schematic diagram of the electronic equipment used for shale gas carbon dioxide displacement wells as described in the embodiments of this application. Detailed Implementation

[0049] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0050] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product or device.

[0051] Please see Figure 1 This embodiment provides a shale gas carbon dioxide displacement well selection method based on interpretable graph neural networks, including the following steps:

[0052] Step 1: Collect current production data and dynamic monitoring data of old wells in the well area. Based on screening criteria such as low casing pressure, low production, low daily gas production, high flowback rate, and low block pressure, preliminarily determine the candidate gas injection well groups.

[0053] Step 2: For the selected candidate well groups, obtain their full-cycle production dynamic data, construct a standard dataset for fitting the production dynamics of gas injection wells, and provide data support for subsequent modeling.

[0054] Step 3: Based on the above standard dataset, build a well network structure model that can accurately fit the production dynamics of gas injection wells, laying the foundation for analyzing well group relationships.

[0055] Step 4: Using the established well network diagram structure model, and employing the interpretable module of a graph neural network, the pressure cross-connection strength between well groups can be intuitively reflected by quantifying the weight coefficients between each node.

[0056] Step 5: Taking into account the real-time production data, dynamic monitoring data and pressure cross-connection strength of each well group, the final carbon dioxide displacement well group is determined after system evaluation.

[0057] In a further embodiment, step 1 specifically includes:

[0058] Step 1.1: Collect dynamic production parameters and obtain real-time production data of old wells in the well area, including but not limited to: pressure parameters: current casing pressure, oil pressure; production parameters: daily gas production, daily water production, cumulative gas production, cumulative water production; efficiency indicators: flowback rate, production efficiency, and commissioning cycle. The casing pressure, daily gas production, and flowback rate are used as quantitative indicators for screening candidate gas injection well groups. Oil pressure and daily water production are used for graph neural network fitting of production dynamics. Cumulative gas production, cumulative water production, production efficiency, and commissioning cycle are used for optimization.

[0059] Step 1.2: Integrate dynamic monitoring data and collect engineering monitoring data for the well area, including but not limited to: fluid characteristics: high-pressure physical properties (PVT) data from the benchmark well; pressure testing: pressure recovery test (PTA), micro-fracture test, and static pressure gradient monitoring data. The pressure recovery test (PTA) is used to obtain data such as static pressure and pressure gradient of the current producing well, and then calculates the formation pressure. This data mainly serves well groups where optimal formation pressure is desired. Micro-fracture tests are commonly used in the early stages of development to obtain formation pressure. After obtaining the current formation pressure, this application selects well groups with lower formation pressure as the preferred targets for CO displacement, thereby fully utilizing the energy-enhancing effect of CO, optimizing the seepage field, and improving displacement efficiency.

[0060] Step 1.3: Screening candidate gas injection well groups. Target well groups are typically selected based on the following regional characteristics as quantitative criteria: casing pressure is lower than the critical pressure required to maintain fracture conductivity; daily gas production is in a continuously declining phase with a deceleration rate greater than the regional average deceleration rate; flowback rate is higher than the regional average; and block pressure is lower than the pressure balance value of the reservoir structural unit. The critical pressure for fracture conductivity is calculated using the following model:

[0061]

[0062] Among them, P c Here, E is the critical pressure for fracture conduction, and E is Young's modulus (GPa). Let S be the limiting strain, v be Poisson's ratio, and S be the limiting strain. h This refers to horizontal ground stress.

[0063] For example, in this embodiment, based on the geological characteristics of atmospheric pressure shale gas, the current formation pressure field is inverted based on benchmark well monitoring data, and combined with real-time production data, target well groups can be screened according to the following quantitative criteria: casing pressure threshold: <2MPa (reflecting reservoir energy depletion); production capacity decline: daily gas production <5000m³ / h 3 / d and shows a continuous decreasing trend (monthly decrease rate > 5%); abnormal flowback: flowback rate > 30% (characterizing deterioration of fracture conductivity); reservoir pressure status: low pressure in well group blocks (enhancing CO2 energy).

[0064] It should be noted that in practical use, different standards can be formulated for different blocks based on different geological characteristics and engineering conditions.

[0065] Step 1 above clarifies the quantitative criteria for screening candidate gas injection well groups. These criteria are based on the geological characteristics and production patterns of shale gas reservoirs, providing a clear reference for well selection. By setting specific quantitative thresholds, such as casing pressure, daily gas production, flowback rate, and block pressure, well groups with high potential for pressure cross-connection can be effectively identified, providing preliminary screening of target wells for subsequent carbon dioxide displacement.

[0066] In a further embodiment, step 2 specifically includes:

[0067] Step 2.1: Integrate the full-cycle production dynamic time series data and extract the full-cycle production dynamic data of the candidate well groups, including: pressure time series curves: casing pressure and oil pressure change sequences over time; production time series curves: historical trends of daily gas production, monthly gas production, daily water production, and monthly water production.

[0068] Step 2.2: Construct an initial dataset, using the production dynamic data as feature data (input variables), and the casing pressure sequence and daily gas production sequence of the target injection well as response data (prediction target), to establish a multidimensional time-series correlation dataset.

[0069] Step 2.3, Data Standardization and Cleaning: Perform the following preprocessing on the dataset:

[0070] Outlier removal: Deletes outliers and missing values ​​caused by well shut-in, equipment failure, or data acquisition errors.

[0071] Normalization: Min-MaxScaling is used to linearly map the feature data to the [0,1] interval, outputting a standard dataset; the Min-MaxScaling formula is as follows:

[0072]

[0073] Where x represents the original feature data, min represents the minimum value among the same feature data, and max represents the maximum value among the same feature data. * These are the feature data after normalization.

[0074] Dataset partitioning: The dataset is divided into training and test sets in a 7:3 ratio.

[0075] Here, the min-max normalization method is used to normalize the data, which can eliminate the impact of dimensional differences on model training and enable the model to better learn the intrinsic features of the data. In other embodiments, other preprocessing techniques and normalization methods, such as z-score normalization and wavelet transform, can also be used to adapt to the processing needs of different types of data and further improve the model's generalization ability and prediction accuracy.

[0076] In a further embodiment, step 3 specifically includes:

[0077] Initial graph construction: Based on the production dynamic time series data of the gas injection well and its adjacent wells, a hypervariable spatiotemporal fully connected graph is constructed, where nodes represent the set of production dynamic parameters at each time step, and edges represent the potential interaction between data at any two time steps;

[0078] Step 3.1: Construct a hypervariate graph structure to map the production dynamic time-series data of the gas injection well and adjacent wells into a spatiotemporal fully connected graph. Node definition: Each sampling point (an N-dimensional variable at time step t) is an independent node, and the node feature vector is X. t ∈R N (Including dynamic parameters such as casing pressure, oil pressure, daily gas production, and daily water production, where X) t Let N be the feature vector at time t, N be the dimension (i.e., the number of dynamic parameter types), and R be the real vector space. Edge connection mechanism: Initialize the fully connected adjacency matrix A∈R. T×T (Where A is the adjacency matrix and T is the length of the time series), representing the potential interaction between any two nodes; Graph size: For time series data of length T and dimension N, generate T*N nodes.

[0079] Step 3.2, Fourier Graph Neural Network Training: Historical Production Dynamic Nodes {X1, X2, X3, ..., X} of Gas Injection Wells and Adjacent Wells t As the input layer of the model, the next time step node X of the gas injection well is... t+1 The output layer data is used to train and determine the weights and biases within the model, resulting in a Fourier graph neural network model that can accurately fit the production dynamics of gas injection wells.

[0080] Step 3.3, Hyperparameter Optimization: Bayesian optimization is used to iteratively select hyperparameter combinations. For the training set, a surrogate model is constructed based on historical experimental results. The configuration with the highest expected improvement is evaluated first. Optimization stops when the maximum number of trials is reached or when there is no significant improvement in the evaluation metric after 10 consecutive iterations. Hyperparameters to be optimized include, but are not limited to: sliding input window size, learning rate, number of iterations, batch size, dropout parameter, weight decay, Fourier hidden layers, and neighbor sampling number. Here, our model is trained using the RMSProp optimizer, with Mean Squared Error (MSE) as the loss function.

[0081] Step 3.4, Multi-index Model Validation: Calculate the model predictions on the independent test set. The deviation from the true value y is used to evaluate performance through performance evaluation metrics. These metrics can be, but are not limited to, any of the following: coefficient of determination (R²), mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute deviation (MAE), mean square error (RSM), etc.

[0082] Step 3 above, by using the production variables of each time step as nodes to construct a hypervariable graph structure, can fully reflect the complexity and connectivity of the well group's production dynamics. The automatic adjustment of hyperparameters through Bayesian optimization improves the efficiency of model training and the accuracy of predictions, ensuring that the model can effectively fit the production dynamic sequence of gas injection wells, providing strong support for the quantitative analysis of pressure crosstalk intensity.

[0083] In a further embodiment, step 4 specifically includes:

[0084] Step 4.1: Integrate the interpretable module and use GNNExplainer to generate subgraph masks. That is, the trained graph neural network is input into GNNExplainer, and the importance weight of each node is output by GNNExplainer.

[0085] Optimize the loss function max Gs ,I(Y,G s |G), where G is a hypervariable graph, G s To interpret the subgraph, Y is the predicted output, and the final output is the edge weight matrix W∈[0,1]. n×n The edge weights of the production data of the same well are summed to obtain the correlation weights between the production changes of the gas injection well and the production changes of the gas injection well itself and the adjacent wells. Finally, the pressure cross-flow coefficient of the well group is obtained.

[0086] Specifically, GNNExplainer is implemented by learning the key substructure and feature mask of the input graph, and maximizing the mutual information between the explanatory subgraph and the prediction of the original model. The specific steps are as follows:

[0087] 1. Input and Output

[0088] Input: trained graph neural network GNN model, prediction result of target node / graph

[0089] Output: explanatory subgraph structure Gs (via edge mask Edge_Mask) and key node feature subset Xs (via feature mask Feat_Mask)

[0090] Objective: optimize mask parameters to maximize the mutual information between the subgraph prediction and the original model prediction, and the formula is:

[0091] max Gs,Xs , I(Y; (G s , X s )) = H(Y) - H(Y丨G s , X s ), where H is information entropy, representing the uncertainty of a random variable, and Y is the prediction output

[0092] 2. Mask Mechanism

[0093] Edge Mask: an n×n matrix (n is the number of nodes) with values in [0,1], representing the importance of edges, activated by Sigmoid or ReLU

[0094] Feature Mask: a d-dimensional vector (d is the feature dimension) that identifies key feature dimensions

[0095] Step 4.2, Visual Output: draw a heat map of inter-well weight coefficients to provide intuitive visualization results

[0096] In the above step 4, by embedding an interpretable module into the GNN, quantitative analysis of the weight coefficients between well groups is realized. Through optimizing the loss function, an edge weight matrix reflecting the fracturing channeling connectivity between wells can be generated, and then the fracturing channeling coefficient of the well group can be calculated, which provides intuitive quantitative indicators for well selection decision-making

[0097] In a further embodiment, step 5 determines the final standard for the carbon dioxide displacement well group as follows: the historical production data contribution of the target injection well is ≤0.5, and the pressure crosstalk weight coefficient with at least two adjacent wells is ≥0.3. This standard is derived with the optimization of CO2 displacement efficiency as the core: the historical production data contribution of the target injection well should be as low as possible, because the lower the value, the greater the pressure crosstalk with adjacent wells; for the pressure crosstalk weight coefficient of adjacent wells, under the normalization constraint that the sum is 1, in order to expand the CO2 sweepability, more adjacent wells need to be included and the weight coefficient distribution should tend to be even, so as to avoid the formation of dominant channels. This logic highlights the pressure crosstalk interaction of adjacent wells by reducing the influence of single-well historical data, and expands the fluid sweep volume by using the uniform weight configuration of multiple wells, ultimately achieving the balanced development goal of suppressing the "single-well dominance effect" and the "dominant seepage channel".

[0098] Step 5, by setting thresholds for the historical production data contribution of the target injection well and the pressure channeling weight coefficient of adjacent wells, effectively identifies well groups with good connectivity to neighboring wells and whose own production potential has not been fully explored. These well groups are preferred targets for carbon dioxide displacement, ensuring that they possess good pressure channeling connectivity and displacement potential, thus improving the efficiency and effectiveness of carbon dioxide displacement. In other embodiments, the thresholds of these standards can be adjusted according to actual conditions to adapt to the well selection needs of different geological conditions and production stages, further optimizing the well selection strategy.

[0099] To enable those skilled in the art to better understand the solutions of this application, more detailed embodiments are now provided for illustration.

[0100] S1: Collect current production data and dynamic monitoring data of old wells in the well area. Based on screening criteria such as low casing pressure, low production, low average daily gas production, high flowback rate, and low block pressure, preliminarily determine the candidate gas injection well groups.

[0101] First, based on the actual production needs on site, and taking into account factors such as geological conditions, mining progress, and equipment compatibility, the target operating block was selected. This block covers 33 production wells. Subsequently, the system carried out production parameter acquisition work, comprehensively collecting pressure and production data of each well during the current period. Specific parameters are detailed in Table 1.

[0102] Table 1 Current production parameters of each production well in the block

[0103]

[0104] After completing the collection of basic production parameters, dynamic monitoring data integration was immediately initiated. To accurately grasp the engineering status of the well area, this study focused on collecting pressure recovery monitoring data from benchmark wells on each platform in the block. Strictly adhering to industry standards, high-precision instruments were used to collect measured shut-in pressures, and specialized software was employed to perform fitting analysis on the data to obtain the fitted extrapolated pressure. Simultaneously, core data such as formation pressure and pressure coefficients were collected to provide a scientific basis for subsequent reservoir dynamic analysis and production plan optimization, as shown in Table 2.

[0105] Table 2. Pressure Recovery Monitoring Parameters for Benchmark Wells on Each Platform

[0106]

[0107] Based on pre-established quantitative screening criteria, a systematic evaluation of the well group was conducted. Specific criteria included: casing pressure threshold below 2 MPa and daily gas production less than 5000 m³. 3 The pressure of the well group is continuously decreasing, with a flowback rate exceeding 30%, while the overall pressure in the block where the well group is located must be at a low level. Data verification and trend analysis revealed that the formation pressure of well groups X and Y is significantly lower than the regional average. Wells X-2 and Y-3, in particular, basically meet the established standards in key indicators such as casing pressure, daily gas production, and flowback rate, and are therefore identified as target well groups. Subsequent evaluation of the pressure channeling degree will be conducted on these two well groups.

[0108] S2: For the selected candidate well groups, obtain their full-cycle production dynamic data, construct a standard dataset for fitting the production dynamics of gas injection wells, and provide data support for subsequent modeling.

[0109] Wells X-2 and Y-3, intended for gas injection, were selected as typical cases. Their production dynamic data were systematically collected, including casing pressure curves, oil pressure curves, and monthly gas and water production variation curves. For detailed data distribution characteristics, please refer to [link to relevant documentation]. Figure 2 .

[0110] To construct a dataset suitable for model training, the dynamic production sequence of well groups was used as feature data (input variables), and the casing pressure sequence and daily gas production sequence of the target injection well were set as response data (prediction targets). A multidimensional time-series correlation dataset containing 2656 nodes was successfully constructed. Based on this, the raw data underwent systematic preprocessing: outliers and missing values ​​caused by factors such as well shut-in operations, equipment failures, and data acquisition errors were strictly removed, and finally 2344 valid data nodes were selected to ensure data quality.

[0111] To eliminate the interference of data dimension differences on model training, the Min-MaxScaling method is further used to normalize the data. A linear transformation maps the feature data to […]. 0 ,1 The specific conversion formula for the interval is as follows:

[0112]

[0113] Where x represents the original feature data, min represents the minimum value among the same feature data, and max represents the maximum value among the same feature data. * These are the feature data after normalization.

[0114] Finally, the dataset was divided into training and test sets in a 7:3 ratio.

[0115] S3: Establish a well network diagram structure model and optimize the hyperparameters within the model.

[0116] Based on the preprocessed standard dataset, the well network diagram structure model was trained to construct a model architecture that can effectively represent the spatial distribution of well networks and the correlation with production data. During this process, the hyperparameters of the Fourier graph neural network model were deeply optimized. Hyperparameters in machine learning, such as model architecture and learning rate, need to be manually set and play a decisive role in the model's training accuracy, convergence speed, and generalization ability.

[0117] In the hyperparameter optimization stage of the Fourier graph neural network model, a hybrid optimization strategy combining manual adjustment and automatic adjustment using the Bayesian optimization algorithm is adopted. Manual adjustment can initially determine the parameter range based on experience, quickly narrowing the optimization space; the Bayesian optimization algorithm, by constructing a probabilistic model, adaptively explores the optimal parameter combination, significantly improving optimization efficiency. Simultaneously, to objectively measure the hyperparameter optimization effect, a comprehensive evaluation index is set as an optimization guide; when this index reaches its maximum value, the corresponding model's predictive performance is optimal. In this example, the coefficient of determination R² is selected as the comprehensive evaluation index, and its calculation formula is as follows:

[0118]

[0119] In the above formula, R² is the coefficient of determination. To predict the response data value using a Fourier graphical neural network model, y i In response to the true value of the data, The average value of the response data is n, where n is the total number of samples.

[0120] After data standardization, based on data characteristics, the response data in the standard dataset was deconstructed into two major sequence modules: casing pressure curves and monthly gas production curves. To explore the optimal activation function of the Fourier transform neural network model, a systematic parameter tuning experiment was conducted, using the casing pressure curve data as a starting point.

[0121] In this example, the model is trained using the RMSProp optimizer, with Mean Squared Error (MSE) as the loss function. In the activation function selection phase, a manual adjustment strategy was employed, choosing sigmoid, tanh, and ReLU as candidate activation functions. The experimental process strictly followed the machine learning model training and validation paradigm: taking the traverse pressure curve dataset as an example, the model was trained iteratively multiple times using the training set, and the model parameters were dynamically adjusted using the backpropagation algorithm to allow the model to gradually learn the data features. After training, the model performance was independently validated using the test set, and the coefficient of determination (R²) was calculated to quantitatively evaluate the model's goodness of fit and prediction accuracy to the traverse pressure curve data. A comparative analysis of the coefficients of determination was conducted for the model training results under different activation functions. Experimental results showed that when the ReLU activation function was used, the coefficient of determination (R²) reached its peak, exhibiting optimal performance in the traverse pressure curve data prediction task, as shown in Table 3. Therefore, the ReLU function was established as the suitable activation function for Fourier transform neural network models processing traverse pressure curve data.

[0122] Table 3 Comparison of the effects of each activation function

[0123]

[0124] Based on the preliminary optimization results of the hyperparameters obtained through manual adjustment, the optimal hyperparameter interval is accurately located, and the initial values ​​for Bayesian optimization are selected accordingly. For the Fourier graph neural network model, the parameters requiring automatic adjustment, the corresponding optimization space, and the initial settings are detailed in Table 3. Subsequently, strictly following the optimization space shown in Table 3, the search space domain of the Bayesian optimization algorithm is finely configured, and the number of optimization iterations is set to 150 to ensure that the algorithm efficiently explores the optimal hyperparameter combination under limited computational resources.

[0125] Table 4 Hyperparameter optimization space and initial values

[0126]

[0127] After optimization, the corresponding hyperparameter combination is the optimal hyperparameter of the Fourier plotted neural network model under the current data. The same optimization process was performed on the monthly gas production curve, and the optimal hyperparameter combination of the final determined casing pressure sequence and daily gas production sequence is shown in Table 5.

[0128] Table 5 Optimal Hyperparameter Combinations

[0129]

[0130] Furthermore, the trained model is used to fit the production dynamics of the gas injection well, and the model performance is validated using test set data, taking well Y-3 as an example. The test set here includes entirely new cases not covered during training, and the predicted monthly gas production curves and casing pressure curves are compared with actual production data. Figure 3 , Figure 4 As shown in Table 6, the model's prediction indicators are detailed in the table. The results clearly demonstrate that the model exhibits excellent fitting performance to the production dynamics of gas injection wells, with a high degree of agreement between the predicted and actual values.

[0131] Table 6. Test Set Fitting Performance Indicators

[0132]

[0133] S4: Based on the well network diagram structure model, the GNNExplainer module is used to calculate the well network pressure cross-linking coefficient.

[0134] First, the pre-trained graph neural network (GNN) model is initialized, and well network data is loaded. A Fourier graph neural network is used here, where nodes represent the production dynamics data at each time step, and edges represent the dynamic relationships between production data. The contribution of each time step of the same well to the production dynamics of the well to be injected is summed, including the contributions of neighboring wells and the historical data of the injection well itself. The results for well group X are shown in Table 7, and the results for well group Y are shown in Table 8. Then, the pressure crosstalk is visualized, and a heatmap of inter-well weighting coefficients is plotted. Figure 5 As shown.

[0135] Table 7 Production Contribution of Each Well in Well Group X

[0136]

[0137] Table 8 Production Contribution of Each Well in Well Group Y

[0138]

[0139] 5: System evaluation to determine the final carbon dioxide displacement well group.

[0140] In step S1, based on formation pressure distribution characteristics and real-time production dynamic monitoring data, well groups X and Y were selected as research objects for a special analysis of pressure crosstalk. Through systematic analysis of the relevant data in Tables 7 and 8, a screening standard for pressure crosstalk correlation was first established: wells with a historical production data contribution greater than 0.5 were defined as wells with dominant self-source influence. The gas production changes and casing pressure fluctuations of these wells are mainly controlled by their own production dynamics, and their pressure crosstalk connectivity with adjacent wells is weak. Typical examples include well X-1 (contribution 0.64), well X-3 (0.61), well Y-1 (0.49), and well Y-2 (0.54), whose data characteristics directly reflect their low degree of pressure crosstalk connectivity with adjacent wells.

[0141] Further analysis revealed that the pressure channeling weighting coefficients of wells X-2 and X-4, along with their adjacent wells, exceeded the 0.5 threshold, indicating that they had formed highly interconnected pressure channeling pathways with specific neighboring wells. Such pathways tend to become dominant seepage paths during CO2 displacement, leading to inefficient circulation of the displacement fluid and significantly reducing reservoir sweep efficiency.

[0142] The focus was on the dynamic characteristics of the Y-well group: Well Y-3's historical production data contribution was only 0.08, exhibiting typical multi-well pressure channeling response characteristics. Its pressure channeling coefficients with three adjacent wells within the group reached 0.38, 0.34, and 0.20, respectively, demonstrating extensive reservoir connectivity. Well Y-4 only showed limited pressure channeling correlation with wells Y-2 and Y-3, and its CO2 sweep efficiency was significantly lower than that of well Y-3. Based on a comprehensive evaluation of reservoir connectivity, displacement efficiency, and well group adaptability, the Y-well group was ultimately selected as the target well group for the CO2 displacement test, and the highly connected well Y-3 was chosen as the main injection well, laying a data foundation for the subsequent implementation of enhanced oil recovery technology.

[0143] In a further embodiment, this application also provides a shale gas carbon dioxide displacement well system, see [link to relevant documentation]. Figure 6 The system includes multiple functional modules such as data acquisition, dataset construction, graph structure modeling, crosstalk analysis, and decision output.

[0144] Data acquisition module: Used to integrate production data and dynamic monitoring data from old wells in the well area, and to screen candidate gas injection well groups based on the quantitative standards of these data. Production data includes, but is not limited to: pressure parameters: current casing pressure, oil pressure; production parameters: daily gas production, daily water production, cumulative gas production, cumulative water production; efficiency indicators: flowback rate, production efficiency, production cycle. Dynamic monitoring data includes, but is not limited to: fluid characteristics: high pressure physical properties (PVT) data from benchmark wells; pressure tests: pressure recovery test (PTA), micro-fracture test, static pressure gradient monitoring data. Specifically, target well groups are screened based on the following regional characteristics of the applicable well groups as quantitative standards: casing pressure is lower than the critical pressure required to maintain fracture conductivity, daily gas production is in a continuous decreasing phase and the rate of decrease is greater than the regional average rate of decrease, flowback rate is higher than the regional average, and block pressure is lower than the pressure balance value of the reservoir structural unit.

[0145] Standard Dataset Construction Module: This module constructs a standard dataset to fit the production dynamics of the candidate gas injection well groups. This includes processing and integrating the production dynamic parameters to form a time-series correlation dataset. The production dynamic time-series data includes casing pressure sequences, oil pressure sequences, monthly gas production sequences, and monthly water production sequences. These data form the basis for analyzing the production dynamics of the well groups. By extracting and integrating this time-series data, a multidimensional dataset reflecting the production history and dynamic changes of the well groups can be constructed, providing a rich source of information for subsequent model training and validation.

[0146] Graph structure modeling module: Based on Fourier graph neural network, a well network graph structure model is constructed, including mapping production dynamic time series data into a hypervariable graph structure, where the production parameters of each time step are used as nodes, and fully connected nodes form an adjacency matrix. Fourier graph neural network is used to fit the production dynamic sequence of gas injection wells and output the production dynamic fitting curve.

[0147] Pressure Crossing Analysis Module: The graph neural network interpretable module (GNNExplainer) quantifies the weight coefficients between well groups, generates an edge weight matrix that reflects the strength of pressure crossing connectivity, and outputs a heat map of the degree of pressure crossing between wells;

[0148] Decision output module: Based on the combined pressure cross-connection strength and production indicators, determine the final carbon dioxide displacement well group.

[0149] In a further embodiment, a shale gas carbon dioxide displacement well selection terminal is also provided, including a memory, a processor, and a computer program for a shale gas carbon dioxide displacement well selection method stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps of the shale gas carbon dioxide displacement well selection method described in the above embodiments. Please refer to... Figure 7The computer equipment specifically includes input devices, a processor, and a memory. The input devices are used to input current production data and dynamic monitoring data of existing wells in the well area, as well as the production dynamic curves of the well group, including casing pressure curves, oil pressure curves, and monthly gas production and water production change curves. The processor is used to perform preliminary screening of gas injection well groups using the collected well group production data and dynamic monitoring data; to establish a dataset using the production dynamic curves; to establish a Fourier graph neural network model by training the training dataset; to verify the reliability of the established Fourier graph neural network using a test dataset; to calculate the pressure crosstalk degree of the well network using the graph neural network interpretation module; and to comprehensively evaluate the system and determine the final carbon dioxide displacement well group. The memory is used to store current production data and dynamic monitoring data of the target well area, as well as the production dynamic curves of the well group and the well selection results.

[0150] In this embodiment, the input device can specifically be one of the main devices for information exchange between the user and the computer system. Input devices may include keyboards, mice, cameras, scanners, light pens, handwriting input tablets, voice input devices, etc.; input devices are used to input raw data and programs that process this data into the computer. Input devices can also receive data transmitted from other modules, units, and devices. The processor can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) that can be executed by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. The memory can specifically be a memory device used to store information in modern information technology. Memory can include multiple layers; in digital systems, anything that can store binary data can be memory; in integrated circuits, a circuit without physical form but with storage function is also called memory, such as RAM, FIFO, etc.; in a system, a storage device with physical form is also called memory, such as a memory module, TF card, etc.

[0151] In this embodiment, the specific functions and effects implemented by the electronic device can be explained by comparison with other embodiments, and will not be repeated here.

[0152] In a further embodiment, a computer-readable storage medium is provided, on which instructions are stored for causing a machine to perform the shale gas carbon dioxide displacement well selection method described in the embodiments of this application.

[0153] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0154] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0155] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0156] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0157] The optional embodiments of this application have been described in detail above with reference to the accompanying drawings. However, the embodiments of this application are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of this application, various simple modifications can be made to the technical solutions of the embodiments of this application, and these simple modifications all fall within the protection scope of the embodiments of this application. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of this application will not describe the various possible combinations separately.

[0158] Furthermore, various different implementations of this application can be combined arbitrarily, as long as they do not violate the spirit of the implementation of this application, they should also be regarded as the content disclosed in the implementation of this application.

Claims

1. A shale gas carbon dioxide displacement well selection method based on interpretable graph neural networks, characterized in that, Includes the following steps: Step (1) Select candidate gas injection well groups based on the quantitative threshold of real-time production data and dynamic monitoring data of old wells in the well area. The production data includes pressure parameters, production parameters and efficiency indicators, and the dynamic monitoring data includes pressure recovery test data. Step (2) Obtain the full-cycle production dynamic time series data of the candidate gas injection well group and construct a standard dataset containing pressure time series curves and production time series curves; Step (3) Based on the standard dataset, construct a well network diagram structure model: map the production dynamic time series data of the gas injection well and its adjacent wells into a hypervariable graph structure, where the N-dimensional production variables at time step t are used as nodes, and the feature vector is X. t ∈R N This includes casing pressure, oil pressure, daily gas production, and daily water production. X t Let N be the feature vector at time t, N be the dimension (i.e., the number of types of dynamic parameters), and R be the real vector space. Fully connected nodes form an adjacency matrix. A Fourier graph neural network is used to train the hypervariable graph structure, with the historical production dynamic nodes of the gas injection well and its neighboring wells as inputs and the next time step node of the gas injection well as the output. The model is trained to obtain a Fourier graph neural network model that can accurately fit the production dynamic sequence of the gas injection well. Step (4) Input the trained Fourier graph neural network model into the graph neural network interpretable module, output the importance weight of each node through the graph neural network interpretable module, optimize the loss function, output the edge weight matrix reflecting the pressure crosstalk connection strength between well groups, sum the edge weights of the same well group to obtain the correlation weight between the production change of the gas injection well and the production change of the gas injection well itself and the adjacent wells, finally obtain the pressure crosstalk coefficient of the well group, and output the heat map of the pressure crosstalk degree between wells; Step (5) Combine the pressure cross-connection strength and production indicators to determine the final carbon dioxide displacement well group.

2. The method according to claim 1, characterized in that, The pressure parameters mentioned in step (1) include: current casing pressure and oil pressure; the production parameters include daily gas production, daily water production, cumulative gas production, and cumulative water production; the efficiency indicators include flowback rate, production efficiency, and commissioning cycle; the dynamic monitoring data also involves fluid characteristics, micro-fracturing test, and static pressure gradient monitoring data; the production dynamic time series data mentioned in step (2) includes casing pressure sequence, oil pressure sequence, monthly gas production sequence, and monthly water production sequence; the production indicators mentioned in step (5) refer to the current formation pressure, casing pressure, daily gas production, and flowback rate.

3. The method according to claim 2, characterized in that, The quantification threshold mentioned in step (1) is determined based on actual field development. Different thresholds are set for different geological characteristics of normal-pressure shale gas or overpressure shale gas. The applicable well groups have the following regional characteristics: The casing pressure is lower than the critical pressure required to maintain the fracture conductivity, the daily gas production is in a continuous decreasing phase and the rate of decrease is greater than the regional average rate of decrease, the flowback rate is higher than the regional average, and the block pressure is lower than the pressure balance value of the reservoir structural unit.

4. The method according to claim 3, characterized in that, The critical pressure required for the fracture conductivity is calculated using the following model: ， Among them, P c The critical pressure required for fracture conductivity, where E is Young's modulus in gigabytes per second (GPa). Let S be the limiting strain, v be Poisson's ratio, and S be the limiting strain. h This refers to horizontal ground stress.

5. The method according to any one of claims 1 to 4, characterized in that, Step (2) in constructing the standard dataset also includes: performing data preprocessing to remove outliers, normalizing the data using min-max normalization, and dividing the dataset into training and testing sets proportionally; wherein the min-max normalization formula is as follows: ， Where x represents the original feature data, min represents the minimum value among the same feature data, and max represents the maximum value among the same feature data. * These are the feature data after normalization.

6. The method according to any one of claims 1-4, characterized in that, The specific steps for constructing the well network diagram structure model described in step (3) also include: Edge connection mechanism: Initialize the fully connected adjacency matrix A∈R T×T Where A is the adjacency matrix and T is the length of the time series; The sliding window method is used to process time series data, and the learning rate, number of iterations, and dropout rate are automatically adjusted through a Bayesian optimization algorithm.

7. The method according to claim 6, characterized in that, The specific calculation process of the well group pressure channeling coefficient in step (4) includes: The graph neural network can be used to interpret the module optimization loss function max. Gs ,I(Y,G s |G), output edge weight matrix W∈[0,1] n×n Where G is the hypervariable graph, G s To interpret the subgraph, Y is the predicted output, W is the weight matrix, and n is the number of nodes; The production contribution weights of the same well are summed to calculate the well group pressure cross-flow coefficient.

8. The method according to any one of claims 1-4, characterized in that: The criteria for determining the final displacement well group in step (5) are: the contribution of the historical production data of the target gas injection well is ≤0.5, and the pressure cross-flow weight coefficient with at least two adjacent wells is ≥0.

3.

9. A shale gas carbon dioxide displacement well system for implementing the method of any one of claims 1-8, characterized in that, include: Data acquisition module: used to integrate production data and dynamic monitoring data of old wells in the well area, and to screen candidate gas injection well groups based on the quantitative standards of these data; Standard dataset construction module: Constructs a standard dataset to fit the production dynamics of the candidate gas injection well groups, including processing and integrating the production dynamic parameters to form a time-series correlation dataset; Graph structure modeling module: Based on Fourier graph neural network, a well network graph structure model is constructed, including mapping production dynamic time series data into a hypervariable graph structure, where the production parameters of each time step are used as nodes, and fully connected nodes form an adjacency matrix. Fourier graph neural network is used to fit the production dynamic sequence of gas injection wells and output the production dynamic fitting curve. Pressure Crossing Analysis Module: The graph neural network interpretable module quantifies the weight coefficients between well groups, generates an edge weight matrix that reflects the strength of pressure crossing connectivity, and outputs a heat map of the degree of pressure crossing between wells; Decision output module: Based on the combined pressure cross-connection strength and production indicators, determine the final carbon dioxide displacement well group.

10. A shale gas carbon dioxide displacement well terminal, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: When the processor executes the program, it implements the steps of the method according to any one of claims 1-8.

11. A computer-readable storage medium storing computer instructions, characterized in that: When the instructions are loaded by the processor, they execute the steps of the method according to any one of claims 1-8.

Images