An attention-based graph convolutional neural network connectivity prediction method and system

Abstract

This invention discloses an attention-based graph convolutional neural network connectivity prediction method and system, comprising: collecting data covering geological features, geophysical properties, and production dynamics; employing data cleaning algorithms and standardization techniques to eliminate noise and outliers, achieving data normalization; dividing the reservoir into discrete nodes based on reservoir characteristics, constructing edges according to spatial relationships and physical connections, completing the data conversion to a graph structure; incorporating an attention mechanism into the graph convolutional neural network architecture, optimizing parameters through model training to obtain reservoir connectivity prediction results; and accurately identifying strongly connected, weakly connected, and disconnected regions based on the prediction results. The beneficial effects of this invention are: improving upon existing reservoir connectivity prediction methods that struggle to capture complex spatial relationships and nonlinear characteristics of reservoirs, enhancing the accuracy and reliability of reservoir connectivity prediction, assisting in the formulation of oil and gas field development plans, optimizing injection and production parameters, and improving recovery rates and resource extraction efficiency.

Description

Technical Field

[0001] This invention relates to the field of oil and gas development technology, and more specifically to an attention-based graph convolutional neural network connectivity prediction method and system. Background Technology

[0002] Reservoir connectivity refers to the unobstructed flow of fluids between different regions within a reservoir. Accurate prediction of reservoir connectivity is crucial for oil and gas field development planning, enhanced oil recovery, and rational resource extraction. Existing methods for predicting reservoir connectivity suffer from problems such as high subjectivity, high cost, insufficient accuracy, and difficulty in capturing complex spatial relationships. This invention combines attention mechanisms with graph convolutional neural networks to propose an innovative solution for this specific problem, aiming to provide the oil and gas industry with a more accurate and efficient method for predicting reservoir connectivity. Summary of the Invention

[0003] This invention overcomes the shortcomings of the prior art and provides an attention-based graph convolutional neural network connectivity prediction method and system.

[0004] A method for predicting reservoir connectivity using attention-based graph convolutional neural networks includes the following specific steps: S1. Collect geological attribute parameter data, geophysical parameter data and known reservoir connectivity label data of the reservoir, check the collected raw data, identify outliers and missing values, and standardize all processed feature data. S2. The collected reservoir data is converted into a graph structure, and the reservoir graph structure is constructed. The node features are selected based on geological attributes, and the edge features are selected based on distance and pressure difference. S3. Construct a graph convolutional neural network model based on the attention mechanism, input static data and production dynamic data from the reservoir connectivity label data after data processing, and obtain the prediction results of reservoir connectivity through the graph convolutional neural network model; S4. Train the graph convolutional neural network model constructed in S3 using the Adam optimizer. Input the node feature matrix, adjacency matrix, and labels of the training set into the model. Calculate the prediction results through forward propagation. Calculate the loss function value based on the prediction results and the true labels. Use the backpropagation algorithm to calculate the gradient of each parameter and update the parameters. After each training epoch, evaluate the model performance using the validation set and calculate the loss value of the validation set. If the validation set loss value does not decrease for 5 consecutive epochs, stop training and save the model parameters at this time as the optimal model parameters, then enter the prediction phase. Otherwise, adjust the model structure, hyperparameters, or increase training data for retraining and evaluation. S5. Input the node feature matrix and adjacency matrix to be predicted into the graph convolutional neural network model trained in S4. Calculate the connectivity probability corresponding to each node through forward propagation. Analyze the connectivity characteristics of each region of the reservoir based on the output connectivity probability to identify strongly connected regions, weakly connected regions, and disconnected regions. S6. Based on the prediction results of reservoir connectivity obtained in S5, determine the connected and non-connected regions of the reservoir and divide the injection and production units; set injection and production pressure constraints, injection and production volume constraints and reservoir connectivity constraints, and generate injection and production parameter adjustment schemes with the goal of maximizing recovery rate, and quantitatively evaluate the schemes.

[0005] Specific methods for outlier identification include: The Z-score method is used to calculate the Z-score value for each data point x in the original data. The formula for calculating the Z-score value is as follows. in, The mean of the data. Let x be the standard deviation, x be the original data points, and Z be the Z-score value; When |Z|>3, the data point is considered an outlier; When the number of outliers is less than the set threshold, the sample is deleted directly. When the number of outliers is more than the set threshold, the feature of the outliers is replaced with the median.

[0006] Specific methods for handling missing values ​​include: calculating the missing percentage of each feature; for features with a missing percentage of less than 5%, using the mean imputation method; for features with a missing percentage between 5% and 20%, using the median imputation method; and for features with a missing percentage of more than 20%, processing by re-collecting data or removing the feature.

[0007] All processed feature data are standardized using min-max standardization, which transforms each feature into a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensions. The calculation formula is as follows: Where: min(x) is the minimum value of feature x, max(x) is the maximum value of feature x, and the standardized data range is between [0,1].

[0008] The specific steps of S2 include: S21. Node determination: The entire reservoir is divided into N independent regions according to the geological unit division rules. Each region is defined as a node in the graph structure. S22. Node feature extraction: For each determined node in S21, collect the processed geological attributes, geophysical attributes, and production dynamic data from the known reservoir connectivity tag data in S1 corresponding to that region as data parameters, and integrate the collected parameters as the features of that node. S23. Construction of edges: For any two reservoir regions corresponding to nodes vi and vj, determine whether there is a potential connection relationship based on the spatial positional relationship and geological structural characteristics between the two nodes. If the distance between the two regions is within the set threshold range, or they belong to the same continuous geological structural unit, then construct an edge between the two nodes to represent the possible connection relationship between them. S24. Edge feature extraction: Extract parameters that reflect the strength of the association between the two corresponding regions for each constructed edge, and use the extracted parameters as the features of the edge. S25. Initialize the graph and construct an empty graph. Add the nodes partitioned in S21 to the set. Based on the edge construction conditions in S23, determine whether there are edges between nodes. If there are, add the edges to the edge set. S26. Construct an adjacency matrix, which is used to represent the connection relationships between nodes in the graph.

[0009] The S3 convolutional neural network model includes: The input layer is used to input the node feature matrix and adjacency matrix into the model and is responsible for data transmission. The graph convolutional layer assigns different attention weights to different neighbor nodes. In each graph convolutional operation, the attention weights between the current node and all neighbor nodes are calculated first, and then the features of the neighbor nodes are weighted and aggregated based on the attention weights. The attention mechanism layer is used to weight the contributions of different nodes and output the weighted features calculated by the attention weights. Fully connected layers are used to integrate the features output by the attention mechanism layer and map them to a low-dimensional space. The output layer uses the sigmoid activation function to map the output of the fully connected layer to the [0,1] interval to obtain the prediction probability.

[0010] The grading criteria for output connectivity probability in S5 are as follows: A region with a connectivity probability ≥ 0.8 is identified as a strongly connected region. When the connectivity probability is between 0.5 and 0.8, it is identified as a weakly connected region. When the connectivity probability is less than 0.5, it is identified as a disconnected region.

[0011] A system for predicting reservoir connectivity using attention-based graph convolutional neural networks, for performing a method for predicting reservoir connectivity using attention-based graph convolutional neural networks, includes: a data acquisition module, a graph structure construction module, an attention-enhanced graph convolutional neural network model, a model training module, a reservoir connectivity prediction module, and an injection-production optimization decision module; The data acquisition module is used to comprehensively collect multi-source data such as reservoir geological data, well logging data, production dynamic data, and 3D seismic data; The preprocessing module is used to check the raw data, identify outliers and missing values, and standardize all feature data after processing. The graph structure construction module uses well points and reservoir units in the reservoir as graph nodes, and uses the differences in reservoir physical properties and fluid flow capacity between nodes as edge weights to construct reservoir topology graph structure data. The graph convolutional neural network model includes an input layer, a graph convolutional feature extraction layer, an attention weight allocation layer, and an output layer. The graph convolutional feature extraction layer performs multi-scale feature fusion on graph structure data, and the attention weight allocation layer assigns dynamic weights to the features of different reservoir units through a self-attention mechanism. The model training module uses labeled reservoir connectivity sample data to train the attention-enhanced graph convolutional neural network model, and optimizes the model parameters through iterative optimization of the loss function to obtain a converged prediction model. The reservoir connectivity prediction module inputs the graph structure data of the reservoir to be predicted into the converged prediction model and outputs the connectivity probability values ​​between reservoir units. The injection-production optimization decision module constructs an injection-production well network connectivity matrix based on connectivity probability values, generates injection-production parameter adjustment schemes according to production capacity constraints, and performs quantitative evaluation of the schemes through the scheme evaluation unit.

[0012] The data acquisition module collects geological parameter data including porosity, permeability, oil saturation, and formation thickness. Geophysical parameters include seismic wave velocity, resistivity, acoustic transit time, and density; Known reservoir connectivity tag data includes production dynamics data such as daily oil production, water cut, daily fluid production, and bottom hole flowing pressure of oil production wells; and daily injection volume and injection pressure of water injection wells.

[0013] The goal of the injection-production optimization decision module is to maximize the recovery rate function, with the objective function being the recovery rate. The decision constraints include injection-production pressure constraints, injection-production volume constraints, and reservoir connectivity constraints. The injection-production pressure constraints are: the injection pressure must be less than the reservoir fracturing pressure, and the bottomhole flowing pressure must be greater than the minimum pressure for fluid flow. The injection-production volume constraints are: the daily water injection volume of the injection well must be less than the theoretical maximum water absorption, and the daily oil production of a single well must be less than the average daily oil production of the first year after production begins. The reservoir connectivity constraints are: based on the reservoir connectivity prediction results, for non-connected areas, reservoir stimulation measures are adopted; for weakly connected areas, acidizing, profile control, and displacement measures are adopted; and for strongly connected areas, optimized water injection and periodic water injection measures are implemented.

[0014] The beneficial effects of this invention are as follows: This solution deeply integrates the attention mechanism into the graph convolutional neural network architecture, constructing an intelligent model system suitable for reservoir connectivity prediction. Through adaptive learning algorithms, the model can quantify the relative importance of connections between reservoir units, assign differentiated weight parameters to key neighborhood nodes, thereby effectively capturing complex topological relationships and nonlinear feature mappings in geological space, significantly improving prediction accuracy. Methodologically, the model adopts a data-driven end-to-end prediction paradigm, constructing a training set based on historical production data and static geological data, avoiding the resource consumption of large-scale field testing in traditional methods, achieving a dual improvement in cost-effectiveness and prediction efficiency. Simultaneously, this model architecture effectively avoids subjective biases caused by expert experience, revealing reservoir connectivity patterns through objective data mining, ensuring the reliability and repeatability of prediction results. Under complex geological conditions, thanks to the powerful nonlinear fitting ability and feature learning mechanism of the graph convolutional neural network, the model exhibits good adaptability and generalization performance. Furthermore, the weight matrix generated by the attention mechanism provides interpretability support for the model's decision-making process, helping geological engineers to deeply understand the formation mechanism of reservoir connectivity. The model-innovative integrated injection-production optimization decision module can generate dynamic injection-production adjustment strategies based on connectivity prediction results and optimization algorithms. This provides data-driven scientific decision-making basis for solving injection-production contradictions in actual production and improving oil and gas recovery rates, demonstrating significant engineering application value. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention; Figure 2 This is a schematic diagram of the model prediction of the present invention; Figure 3 This is a schematic diagram of the tracer fitting process; Figure 4 This is a schematic diagram of the predicted injection-production connectivity results between actual well groups according to the present invention. Detailed Implementation

[0016] Example like Figure 1As shown, a method for predicting reservoir connectivity using attention-based graph convolutional neural networks includes the following steps: S1. Collect geological attribute parameter data, geophysical parameter data, and known reservoir connectivity label data of the reservoir. Check the collected raw data, identify outliers and missing values, and standardize all processed feature data.

[0017] Step S1 is implemented in an attention-based graph convolutional neural network system for predicting reservoir connectivity using a data acquisition module and a preprocessing module. The data acquisition module and the preprocessing module serve as the foundation of the entire system, responsible for collecting various types of raw data required by subsequent modules and performing checks and processing, providing data support for reservoir connectivity prediction and injection-production optimization decisions.

[0018] The geological attribute parameters in S1 include reservoir lithology, porosity, permeability, oil saturation, and sedimentary facies zones; The geophysical parameter data in S1 include seismic impedance, seismic amplitude, seismic frequency, resistivity, and acoustic transit time. The known reservoir connectivity tag data in S1 includes production dynamics data such as daily oil production, daily fluid production, water cut, daily water injection, and bottom hole flowing pressure.

[0019] The types of data collected include: Geological parameter data porosity Permeability K, Oil saturation S o Formation thickness (h) directly reflects the geological characteristics of the reservoir and is an important basis for judging reservoir connectivity. Geophysical parameters such as seismic wave velocity (V), resistivity (γ), acoustic transit time (T), and density (ρ) indirectly reveal the physical properties and structural characteristics of the reservoir. The marker information indicating whether different regions of the reservoir are connected is used for model training and validation. Production dynamics data includes the daily oil production (Q) of the oil wells. O Moisture content f w Daily Liquid Q L Bottom-hole flowing pressure P O Daily injection volume Q of injection well inj Injection pressure P inj .

[0020] Data collection methods and frequency: Geological parameter data acquisition methods: Porosity, permeability, etc., are obtained through core analysis; core samples retrieved during drilling are analyzed in the laboratory. Oil saturation is calculated using resistivity logging data combined with core analysis data. Formation thickness is determined through logging curve interpretation. Rock type and sedimentary facies are determined by combining geological surveys, core observations, and logging data. Data is collected annually to reflect changes in the basic geological characteristics of the reservoir.

[0021] Geophysical parameter data acquisition primarily relies on seismic exploration and well logging. Seismic wave velocity is obtained through seismic exploration, while resistivity, sonic transit time, and density are obtained through well logging. Data is collected every five years because geophysical parameters change relatively slowly compared to geological parameters.

[0022] The reservoir connectivity tag data is updated based on new well data, well test data, and daily production and injection data of oil and water wells. New relevant data is collected in a timely manner, with a collection frequency of real time.

[0023] Data inspection and processing: Outlier identification: The Z-score method is used to calculate the Z-score value for each data point x. When |Z| > 3, the data point is considered an outlier. Outliers are corrected or removed based on the actual situation.

[0024] in, The mean of the data. The standard deviation is denoted as .

[0025] Missing value handling: For features with a missing value ratio of less than 5%, the mean imputation method is used. That is, for missing values ​​of feature f, the mean of all non-missing values ​​of this feature is used. Imputation is performed; for features with a missing rate between 5% and 20%, the median imputation method is used, and the feature is imputed with the median(f) of all non-missing values; for features with a missing rate greater than 20%, data is re-collected or the feature is removed.

[0026] All processed feature data are standardized using min-max standardization, which transforms each feature into a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensions. The calculation formula is as follows: Where: min(x) is the minimum value of feature x, max(x) is the maximum value of feature x, and the standardized data range is between [0,1].

[0027] S2. The collected reservoir data is converted into a graph structure, and the reservoir graph structure is constructed. The node features are selected based on geological attributes, and the edge features are selected based on distance and pressure difference. In this step, the graph structure building module of the attention-based graph convolutional neural network system for predicting reservoir connectivity is used to transform the collected reservoir data into a graph structure to facilitate subsequent neural network model processing.

[0028] Node Definition and Partitioning: The reservoir is divided into grids, with each grid representing a node. Node features are selected from the processed geological attributes, geophysical attributes, and production dynamics data in S1. These features directly reflect the reservoir characteristics of the area and are the basis for determining connectivity. The feature vector of each node is represented as follows: in: Let be the feature vector of the i-th node; Porosity; K is permeability; So is oil saturation; h is formation thickness; V is seismic wave velocity; γ is resistivity; T is acoustic transit time; ρ is density; Q O Daily oil production; f w Q represents the moisture content. L Daily production liquid; P O Q is the bottom hole flowing pressure; inj Daily injection volume; P inj To inject pressure.

[0029] Edge definition and construction: Edges are used to represent potential associations between nodes. Edge features are selected from well distance, pressure difference change, and injection-production response correspondence rate. An edge is constructed when the values ​​between two nodes are within a certain threshold and the pressure difference change and injection-production response correspondence rate meet certain conditions.

[0030] Distance calculation: The distance d between two nodes i and j is calculated using the Euclidean distance formula. ij : Where: (x i y i , z i ) and (x j y j , z j ) are the three-dimensional coordinates of node i and node j, respectively.

[0031] Pressure difference change: Calculate the pressure values ​​between node i and node j, and calculate the difference: Where: P i and P j These are the pressure values ​​for node i and node j, respectively.

[0032] Injection-production response rate: Calculate the difference between the daily oil production and daily liquid production changes between node i and node j. in: , and , Let $\mathbf$ be the daily oil production and daily liquid production of nodes $i$ and $j$, respectively.

[0033] The eigenvectors of an edge are represented as: Graph initialization method: The graph is initialized by constructing an empty graph: Where: V is the set of nodes; E is the set of edges.

[0034] The partitioned nodes are added to the set. Based on the edge construction conditions, it is determined whether there are edges between the nodes. If there are edges, they are added to the edge set.

[0035] Adjacency Matrix Construction: The adjacency matrix A is used to represent the connection relationships between nodes in the graph, and it is defined as follows: Meanwhile, to account for the influence of edge characteristics on connection strength, a weighted adjacency matrix W is constructed and calculated using the following formula: in: These are the weighting coefficients. .

[0036] S3. Construct a graph convolutional neural network model based on an attention mechanism. Input the processed static reservoir data and dynamic production data, and obtain the prediction results of reservoir connectivity through the graph convolutional neural network model.

[0037] The graph convolutional neural network model architecture includes an input layer, a graph convolutional layer, an attention mechanism layer, a fully connected layer, and an output layer. This model combines the ability of graph convolutional structures to express complex relationships between reservoir regions with the ability of attention mechanisms to focus on key connectivity information, thereby achieving accurate prediction of reservoir connectivity.

[0038] The input data consists of node feature matrices and adjacency matrices constructed from processed reservoir static data and production dynamic data. The node feature matrix is ​​derived from all vectors obtained in S2. Where: X is the node feature matrix; A is the adjacency matrix; N is the number of nodes; F is the feature dimension.

[0039] The output data is the reservoir connectivity prediction result. For each node pair, the probability of connectivity is between 0 and 1.

[0040] Calculation steps and parameter settings for each layer of the model: Input layer: Inputs the node feature matrix X and adjacency matrix A into the model. It does not perform any additional calculations and is only responsible for data transmission.

[0041] In S3, during the feature aggregation process of the graph convolutional neural network, different attention weights are assigned to different neighbor nodes. In each graph convolution operation, the attention weights between the current node and all neighbor nodes are calculated first. The weights are determined by the similarity or correlation strength of the node features. Then, the neighbor node features are aggregated in a weighted manner based on the attention weights.

[0042] Graph Convolutional Layers: The computation of graph convolutional layers is based on the Laplacian matrix of the graph, which is defined as: Where D is the degree matrix.

[0043] For ease of calculation, a normalized Laplace matrix is ​​used: The formula for calculating the convolutional layer is: in: This is the output of the l-th layer; Here is the weight matrix for the l-th layer; The ReLU function is used as the activation function.

[0044] Weight matrix in parameter settings The dimension is determined based on the input feature dimension. For example, if the input feature dimension is F and the output feature dimension is F', then: In S3, within the attention mechanism layer, the attention weights first map the node features to a new space through a linear transformation, calculate the correlation score between nodes i and j, normalize the score, and obtain the attention weights. After passing through multiple graph convolutional attention layers, the output layer outputs the connectivity probability through a fully connected layer and a sigmoid function.

[0045] Attention Mechanism Layer: The attention mechanism is used to weight the contributions of different nodes, focusing on key information in the static geological conditions and dynamic production changes. For the output of layer l+1... Calculate the attention weight α ij : Where: || denotes vector concatenation; For activation functions; Calculate the weighted features based on the attention weights: Parameter settings: The dimension of the attention weight vector α is the same as the feature dimension: Fully connected layer: Integrates the features output by the attention mechanism layer and maps them to a low-dimensional space. The calculation formula is: in: This is the weight matrix of the fully connected layer; This is a bias term.

[0046] Parameter settings: Output layer: The sigmoid activation function is used to map the output of the fully connected layer to the [0,1] interval to obtain the predicted probability. The calculation formula is as follows: in: This is the weight matrix of the output layer; This is a bias term.

[0047] Parameter settings: S4. Train the graph convolutional neural network model constructed in S3 using the Adam optimizer. Input the node feature matrix, adjacency matrix, and labels of the training set into the model. Obtain the prediction results through forward propagation, calculate the loss function value, calculate the gradient and update the parameters using backpropagation, evaluate the model with a validation set after each training round, adopt an early stopping strategy, save the optimal model parameters, evaluate the model with a test dataset, calculate the prediction accuracy, and if the model performance meets the requirements, enter the prediction phase; otherwise, adjust the model structure, hyperparameters, or increase training data, and retrain and evaluate. This step utilizes a model training module of an attention-based graph convolutional neural network system for predicting reservoir connectivity. The model training module employs a cross-entropy loss function as the objective function, corresponding to the classification problem of this invention, namely, whether reservoirs are connected. The loss function is calculated as follows: Where: M is the number of training samples; This is a real label; To predict the probability of connectivity.

[0048] The optimization algorithm in S4 uses the Adam optimizer, with the following parameters: learning rate α = 0.001, exponential decay rate β1 = 0.9 for the first moment estimate, exponential decay rate β2 = 0.999 for the second moment estimate, and a constant for numerical stability. .

[0049] The dataset was divided into training, validation, and test sets in a 7:1:2 ratio. That is, 70% of the data was used for model training, 10% for validating model performance and tuning parameters, and 20% for final model performance evaluation.

[0050] Training epochs and batch size: The initial maximum number of training epochs is set to 200, with the specific number determined based on the early stopping strategy. The batch size is set to 32 based on the dataset size and hardware performance.

[0051] The entire model training process is as follows: The node feature matrix X, adjacency matrix A, and labels y from the training set are input into the model, and the prediction result P is calculated through forward propagation. The loss function value L is calculated based on the prediction result and the true labels. The gradients of each parameter are calculated using the backpropagation algorithm, and the parameters are updated using the Adam optimizer. After each training epoch, the model performance is evaluated using a validation set, and the loss value of the validation set is calculated. An early stopping strategy is adopted: when the validation set loss does not decrease for five consecutive epochs, training is stopped, and the model parameters at this point are saved as the optimal model parameters.

[0052] Preferably, based on the above method, a connectivity prediction accuracy assessment method can be introduced to evaluate and assess the model. Specifically, the model is evaluated using a test dataset, and the prediction accuracy is calculated using the following formula: Where: TP represents predicted connectivity and actual connectivity; TN represents predicted non-connectivity and actual non-connectivity; FP represents predicted connectivity but actual non-connectivity; and FN represents predicted non-connectivity but actual connectivity.

[0053] If the model's performance accuracy reaches 0.85 or higher, it is considered to meet the requirements and proceeds to the prediction stage; otherwise, the model structure, hyperparameters, or training data are adjusted, and retraining and evaluation are carried out.

[0054] S5: Reservoir connectivity prediction module: This module constructs the node feature matrix and adjacency matrix from the trained model and the input reservoir data to be predicted. in: The number of nodes to be predicted.

[0055] S51: Input the feature matrix and adjacency matrix of the nodes to be predicted into the trained model, and calculate the connectivity probability of each node pair through forward propagation. According to the set threshold , : S52: The output is a matrix. .

[0056] S6: The objective of the injection-production optimization decision module is to maximize the recovery rate function, which is: in: For recovery rate; N represents cumulative oil production; N represents geological reserves.

[0057] Decision-making constraints: Injection and production pressure constraints: The injection pressure must be lower than the reservoir fracture pressure to prevent formation damage; the bottom hole flowing pressure must be greater than the minimum pressure for fluid flow to reduce the constraint of the start-up pressure on fluid flow.

[0058] Injection-production constraints: The daily water injection rate of the injection well is less than the theoretical maximum water absorption rate, which is determined by the water injection thickness, water absorption index per meter, and maximum injection pressure. The daily oil production of a single well is less than the average daily oil production in the first year of production. Considering the decline in oil well production, the initial daily oil production is generally the highest historical production capacity of the well.

[0059] Reservoir connectivity constraints: Based on the reservoir connectivity prediction results, reservoir stimulation measures such as pressure flooding are carried out in non-connected areas to improve reservoir connectivity; acidizing, profile control and flooding measures are adopted in weakly connected areas to improve reservoir connectivity; and optimized water injection and periodic water injection measures are carried out in strongly connected areas to reduce ineffective water circulation and improve water drive sweep efficiency.

[0060] The injection and extraction optimization decision module used in this solution is an existing module.

[0061] The specific application methods of the injection and extraction optimization decision-making module are as follows: First, the decision-making results are delivered to on-site production personnel in the form of a visual report and a list of core parameters. This report covers key information such as optimized injection-production rates for well groups, pressure control ranges, and recommended production rhythms for individual wells, providing a clear and actionable basis for adjusting injection-production operations. Simultaneously, based on the geological characteristics of the oilfield block and current production dynamics, targeted injection-production parameter adjustment plans are generated. These plans must clearly define specific operational indicators such as the injection-production intensity per well, the injection-production ratio of the well group, and the frequency of periodic adjustments, ensuring that each adjustment action has a quantifiable standard. Subsequently, the plans are quantitatively evaluated from both technical and economic dimensions: technically, the recovery rate improvement, water cut control effect, and formation energy maintenance level are calculated; economically, the operating costs after parameter adjustments, changes in cost per ton of oil, and expected benefits are calculated to select the optimal plan.

[0062] like Figures 2-4 As shown, in order to make the technical solution of this embodiment clearer, the specific implementation process of this embodiment will be described in detail below by way of example, including the following implementation steps.

[0063] S1: The example is a complex fault-block reservoir. This reservoir is characterized by thin interbedded layers, multiple faults, high-permeability channels, and injection-production mismatch. Targeted data collection and processing are required to provide a reference for subsequent injection-production optimization decisions.

[0064] Data Acquisition: Geological parameter data acquisition includes porosity, permeability, oil saturation, and formation thickness; targeted coverage is given to areas on both sides of faults and high-permeability channels to reflect reservoir heterogeneity. Geophysical parameter acquisition includes seismic wave velocity, resistivity, sonic transit time, and density, with emphasis on seismic response characteristics near faults; Production dynamic data acquisition includes daily oil production, daily fluid production, water cut, and bottom hole flowing pressure of production wells; and daily injection volume and injection pressure of injection wells. This data needs to be updated dynamically in real time to ensure timeliness.

[0065] Data processing and inspection: Due to fault cutting and drilling conditions, data from complex fault-block reservoirs often exhibit concentrated missing values ​​and frequent outliers. The processing methods are as follows: Outliers were identified using the Z-score method for anomalies in permeability and pressure near faults caused by fracture development. If the sample size was small, outliers were directly removed. If outliers accounted for more than 30% of a feature dimension, median replacement was used to preserve the spatial distribution trend. For features with significant missing logging data near faults, mean imputation was used for features with a missing percentage less than 5%, median imputation was used for features with a missing percentage between 5% and 20%, and re-acquiring or removing the feature for features with a missing percentage greater than 20%. Finally, the processed data were standardized, transforming all features to the [0,1] interval to eliminate the influence of different parameters' dimensions.

[0066] S2: Based on the data imported in S1, the heterogeneity and fault characteristics of complex fault-block reservoirs are transformed into a graphical structure to express non-Euclidean features.

[0067] Node partitioning and feature definition: Using fault block units as basic nodes, each node contains all well data and average geological parameters within that fault block. This approach avoids the problem of grid cell fragmentation caused by complex fault cutting, making the nodes more closely resemble the geological entity. Node characteristics include both static and dynamic features. In complex fault block reservoirs, the dynamic data within the same fault block exhibits high consistency, and connectivity can be reflected through differences in dynamic features.

[0068] Edge construction and feature extraction: Edge existence determination: When the distance between two nodes is less than or equal to the set threshold, the distance threshold is set to D=500m to consider the influence of the fault, and the pressure difference threshold is set to ΔP=0.5MPa based on production history experience, and an edge is constructed; even if the nodes on both sides of the fault are close, if the pressure difference ΔP>1MPa, no edge is constructed.

[0069] Edge characteristics include well spacing D, pressure difference ΔP, and injection-production response rate ΔQ. For edge characteristics in high-permeability channel areas, the weight of the injection-production response rate needs to be strengthened to increase the proportion of ineffective water circulation.

[0070] Weighted adjacency matrix: ,in The influence of faults on seepage resistance is taken as 0.6.

[0071] S3: To address the complexity of connectivity in complex fault-block reservoirs, an attention mechanism is used to focus on the characteristics of key areas near faults and high-permeability channels.

[0072] Input layer: Receives node feature matrix with dimensions N×F, where N is the number of fragmented nodes and F is a 12-dimensional feature and weighted adjacency matrix.

[0073] Graph convolutional layer: Consists of two layers. The first layer uses 64 convolutional kernels to perform initial aggregation of node features, as shown in the formula: Where: D is the degree matrix; This is the weight matrix, with dimensions 12×64.

[0074] The second layer uses 32 convolutional kernels to deepen feature extraction and adapt to the multi-scale connectivity relationships of complex blocks.

[0075] Attention mechanism layer: Calculates the attention weights between node i and its neighbor j. The formula is: In complex fault-block reservoirs, predicting connectivity across faults is a significant challenge. An attention mechanism dynamically adjusts weights to focus the model on key influencing factors, assigning higher weights to nodes near the fault and enhancing the model's attention to structurally fractured zones. Node pairs with a pressure difference ΔP < 0.5 MPa are considered to have potential connectivity. Automatic lifting; for closed faults, Reduce to 10% of the base value to reduce interference from invalid associations.

[0076] Output layer: Outputs the connectivity probability of node pairs [0, 1] using the sigmoid function.

[0077] S4: Model Training: The dataset is divided into training, validation, and test sets in a 7:1:2 ratio. The division is based on the following: the training set contains historically stable fault block data, which needs to cover 80% of the known connectivity relationships obtained from field tracer testing; the validation set focuses on fault-dense area data to monitor the model's ability to fit complex structures; and the test set contains newly developed fault blocks to simulate actual prediction scenarios.

[0078] Loss Function and Optimizer: The loss function used is cross-entropy loss, and the formula is as follows: The Adam optimizer was selected, with a learning rate of 0.001, an exponential decay rate β1=0.9 for first moment estimation, and an exponential decay rate β2=0.999 for second moment estimation, to meet the non-convex optimization requirements of complex data.

[0079] Overfitting control: L2 regularization with a coefficient of 0.0001 is added to limit excessive weights; an early stopping strategy is adopted because complex block data fluctuates greatly and requires a longer observation period. Therefore, training is stopped when the validation set loss does not decrease for 8 consecutive rounds.

[0080] S5: Model Evaluation and Prediction. To ensure overall prediction reliability and to focus on the predictive ability for complex structures and the balance between precision and recall, the output connectivity probability matrix is ​​graded according to the following criteria: Among them, a strongly connected zone can be considered as a reservoir within the same fault block without fault barriers; a weakly connected zone can be considered as a region divided by a non-closed fault; and a non-connected zone can be considered as a region on both sides of a closed fault or where the well distance exceeds the effective distance.

[0081] S6: Generate decision-making recommendations and guide oilfield production. First, extract features from the prediction results: For strongly connected areas, which are mostly continuous reservoirs within the same fault block, the production dynamic data are highly consistent, and the injection-production adjustment effect is significant; For weakly connected areas, affected by non-closed faults, there are local connected channels, which manifest as pressure transmission lag; For unconnected areas, closed faults block or reservoir properties change abruptly, resulting in large differences in production data and basically no response to water injection.

[0082] Impact assessment on production and decision-making on measures: The reason why highly interconnected zones are prone to water channeling and rapid water cut increases is essentially due to the existence of highly efficient seepage channels in the reservoir, causing injected water to propagate in a "short-circuit" manner. According to Darcy's law, seepage velocity is positively correlated with permeability. When reservoir connectivity is extremely high, injected water will preferentially flow rapidly along high-permeability channels, bypassing most low-permeability oil-bearing areas, forming "water channeling channels." In this situation, controlling the injection volume is primarily achieved by reducing the driving pressure gradient, slowing down the flow velocity of injected water in high-permeability channels, and forcing some water flow to divert to low-permeability oil-bearing areas, thereby extending the waterless production period. Field practice shows that reducing the injection volume in highly interconnected zones by 30%-40% can slow the rate of water cut increase by more than 50%, and extend the average stable production period of a single well by 1.5 years.

[0083] The root cause of inefficient water injection circulation in weakly connected zones lies in the mismatch between the injection-production well network and the reservoir's heterogeneity. Although these areas lack obvious high-permeability channels, the reservoir permeability distribution is uneven, and the well spacing and well row direction are inconsistent with the main seepage direction of the reservoir, resulting in "excessive" water injection in some areas and "insufficient" water injection in others. From a development perspective, the area sweep efficiency of inefficient circulation zones is typically below 0.3, far below the 0.6-0.7 required for reasonable development. The key to adjusting the injection-production well network lies in optimizing the well density and direction, ensuring the well row direction is parallel to the main fracture direction of the reservoir. This increases the water injection sweep efficiency to 0.55, increases the average daily oil production per well by 4.2 tons, and reduces the area of ​​inefficient circulation zones by 60%. The core challenge to low recovery rates in unconnected reservoirs lies in poor reservoir permeability and high fluid flow resistance. These areas are typically tight reservoirs with pore throat radii less than 0.1 micrometers, making it difficult for injected water to drive crude oil through conventional injection methods. The core of reservoir artificial stimulation is to construct effective seepage channels through physical or chemical means; currently, mainstream technologies include volumetric fracturing and acidizing. For unconnected reservoirs, artificial stimulation often combines a "fracture network + water injection" synergistic model, utilizing the artificial fracture network created by fracturing to expand the water-bearing volume and further address the problem of "injectable but unrecoverable."

[0084] From an overall perspective, the development contradictions in strongly connected areas, weakly connected areas, and unconnected areas present a progressive relationship—from "water flow too fast" to "uneven water flow" and then to "water flow not moving." The corresponding measures also change from "speed control" to "direction adjustment" and then to "channel creation." Together, they constitute a differentiated water injection development system based on reservoir connectivity. The core objective of all of them is to improve the volumetric sweep efficiency and oil displacement efficiency of the injected water, and ultimately achieve efficient recovery of oil and gas resources.

[0085] Evaluation and adjustment of policy recommendations: Based on feedback from the actual implementation on-site, post-implementation evaluations were conducted for the effectiveness of measures in different areas. If the profile control effect in strongly connected areas was not good, chemical flooding could be used instead. If the production increase after the infill wells in weakly connected areas did not meet expectations, sidetracking to develop horizontal wells could be considered to increase the drainage area and improve the sweep efficiency. If there was no response after fracturing in unconnected areas, it could be considered not to develop the area for the time being.

[0086] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A method for predicting reservoir connectivity using attention-based graph convolutional neural networks, characterized in that, The specific steps are as follows: S1. Collect geological attribute parameter data, geophysical parameter data and known reservoir connectivity label data of the reservoir, check the collected raw data, identify outliers and missing values, and standardize all processed feature data. S2. The collected reservoir data is converted into a graph structure, and the reservoir graph structure is constructed. The node features are selected based on geological attributes, and the edge features are selected based on distance and pressure difference. S3. Construct a graph convolutional neural network model based on the attention mechanism, input static data and production dynamic data from the reservoir connectivity label data after data processing, and obtain the prediction results of reservoir connectivity through the graph convolutional neural network model; S4. Train the graph convolutional neural network model constructed in S3 using the Adam optimizer. Input the node feature matrix, adjacency matrix, and labels of the training set into the model. Calculate the prediction results through forward propagation. Calculate the loss function value based on the prediction results and the true labels. Use the backpropagation algorithm to calculate the gradient of each parameter and update the parameters. After each training epoch, evaluate the model performance using the validation set and calculate the loss value of the validation set. If the validation set loss value does not decrease for 5 consecutive epochs, stop training and save the model parameters at this time as the optimal model parameters, then enter the prediction phase. Otherwise, adjust the model structure, hyperparameters, or increase training data for retraining and evaluation. S5. Input the node feature matrix and adjacency matrix to be predicted into the graph convolutional neural network model trained in S4. Calculate the connectivity probability corresponding to each node through forward propagation. Analyze the connectivity characteristics of each region of the reservoir based on the output connectivity probability to identify strongly connected regions, weakly connected regions, and disconnected regions. S6. Based on the prediction results of reservoir connectivity obtained in S5, determine the connected and non-connected regions of the reservoir and divide the injection and production units; set injection and production pressure constraints, injection and production volume constraints and reservoir connectivity constraints, and generate injection and production parameter adjustment schemes with the goal of maximizing recovery rate, and quantitatively evaluate the schemes.

2. The method for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 1, characterized in that, Specific methods for outlier identification include: The Z-score method is used to calculate the Z-score value for each data point x in the original data. The formula for calculating the Z-score value is as follows. in, The mean of the data. Let x be the standard deviation, x be the original data points, and Z be the Z-score value; When |Z|>3, the data point is considered an outlier; When an outlier is less than the set threshold, the sample is deleted directly. When an outlier exceeds the set threshold, the feature of the outlier is replaced with the median.

3. The method for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 1, characterized in that, Specific methods for handling missing values ​​include: calculating the missing percentage of each feature; for features with a missing percentage of less than 5%, using the mean imputation method; for features with a missing percentage between 5% and 20%, using the median imputation method; and for features with a missing percentage of more than 20%, processing by re-collecting data or removing the feature.

4. The method for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 1, characterized in that, All processed feature data are standardized using min-max standardization, which transforms each feature into a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensions. The calculation formula is as follows: Where: min(x) is the minimum value of feature x, max(x) is the maximum value of feature x, and the standardized data range is between [0,1].

5. The method for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 1, characterized in that: The specific steps of S2 include: S21. Node determination: The entire reservoir is divided into N independent regions according to the geological unit division rules. Each region is defined as a node in the graph structure. S22. Node feature extraction: For each determined node in S21, collect the processed geological attributes, geophysical attributes, and production dynamic data from the known reservoir connectivity tag data in S1 corresponding to that region as data parameters, and integrate the collected parameters as the features of that node. S23. Construction of edges: For any two reservoir regions corresponding to nodes vi and vj, determine whether there is a potential connection relationship based on the spatial positional relationship and geological structural characteristics between the two nodes. If the distance between the two regions is within the set threshold range, or they belong to the same continuous geological structural unit, then construct an edge between the two nodes to represent the possible connection relationship between them. S24. Edge feature extraction: Extract parameters that reflect the strength of the association between the two corresponding regions for each constructed edge, and use the extracted parameters as the features of the edge. S25. Initialize the graph and construct an empty graph. Add the nodes partitioned in S21 to the set. Based on the edge construction conditions in S23, determine whether there are edges between nodes. If there are, add the edges to the edge set. S26. Construct an adjacency matrix, which is used to represent the connection relationships between nodes in the graph.

6. The method for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 5, characterized in that, The S3 convolutional neural network model includes: The input layer is used to input the node feature matrix and adjacency matrix into the model and is responsible for data transmission. The graph convolutional layer assigns different attention weights to different neighbor nodes. In each graph convolutional operation, the attention weights between the current node and all neighbor nodes are calculated first, and then the features of the neighbor nodes are weighted and aggregated based on the attention weights. The attention mechanism layer is used to weight the contributions of different nodes and output the weighted features calculated by the attention weights. Fully connected layers are used to integrate the features output by the attention mechanism layer and map them to a low-dimensional space. The output layer uses the sigmoid activation function to map the output of the fully connected layer to the [0,1] interval to obtain the prediction probability.

7. The method for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 1, characterized in that, The grading criteria for output connectivity probability in S5 are as follows: A region with a connectivity probability ≥ 0.8 is identified as a strongly connected region. When the connectivity probability is between 0.5 and 0.8, it is identified as a weakly connected region. When the connectivity probability is less than 0.5, it is identified as a disconnected region.

8. A system for predicting reservoir connectivity using an attention-based graph convolutional neural network, for performing the method for predicting reservoir connectivity using an attention-based graph convolutional neural network as described in any one of claims 1-7, characterized in that, include: The system includes a data acquisition module, a graph structure construction module, an attention-enhanced graph convolutional neural network model, a model training module, a reservoir connectivity prediction module, and an injection-production optimization decision module. The data acquisition module is used to comprehensively collect multi-source data such as reservoir geological data, well logging data, production dynamic data, and 3D seismic data; The preprocessing module is used to check the raw data, identify outliers and missing values, and standardize all feature data after processing. The graph structure construction module uses well points and reservoir units in the reservoir as graph nodes, and uses the differences in reservoir physical properties and fluid flow capacity between nodes as edge weights to construct reservoir topology graph structure data. The graph convolutional neural network model includes an input layer, a graph convolutional feature extraction layer, an attention weight allocation layer, and an output layer. The graph convolutional feature extraction layer performs multi-scale feature fusion on graph structure data, and the attention weight allocation layer assigns dynamic weights to the features of different reservoir units through a self-attention mechanism. The model training module uses labeled reservoir connectivity sample data to train the attention-enhanced graph convolutional neural network model, and optimizes the model parameters through iterative optimization of the loss function to obtain a converged prediction model. The reservoir connectivity prediction module inputs the graph structure data of the reservoir to be predicted into the converged prediction model and outputs the connectivity probability values ​​between reservoir units. The injection-production optimization decision module constructs an injection-production well network connectivity matrix based on connectivity probability values, generates injection-production parameter adjustment schemes according to production capacity constraints, and performs quantitative evaluation of the schemes through the scheme evaluation unit.

9. The system for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 8, characterized in that: The data acquisition module collects geological parameter data including porosity, permeability, oil saturation, and formation thickness. Geophysical parameters include seismic wave velocity, resistivity, acoustic transit time, and density; Known reservoir connectivity tag data includes production dynamics data such as daily oil production, water cut, daily fluid production, and bottom hole flowing pressure of oil production wells; and daily injection volume and injection pressure of water injection wells.

10. A system for predicting reservoir connectivity using an attention-based graph convolutional neural network according to claim 8, characterized in that: The goal of the injection-production optimization decision module is to maximize the recovery rate function, with the objective function being the recovery rate. The decision constraints include injection-production pressure constraints, injection-production volume constraints, and reservoir connectivity constraints. The injection-production pressure constraints are: the injection pressure must be less than the reservoir fracturing pressure, and the bottomhole flowing pressure must be greater than the minimum pressure for fluid flow. The injection-production volume constraints are: the daily water injection volume of the injection well must be less than the theoretical maximum water absorption, and the daily oil production of a single well must be less than the average daily oil production of the first year after production begins. The reservoir connectivity constraints are: based on the reservoir connectivity prediction results, for non-connected areas, reservoir stimulation measures are adopted; for weakly connected areas, acidizing, profile control, and displacement measures are adopted; and for strongly connected areas, optimized water injection and periodic water injection measures are implemented.

Images